Joint Research Program

XXII Meeting of the Central Bank

Researchers Network

Inflation Expectations,
Their Measurement and the
Estimate of Their Degree of
Anchoring

Editors:
Alexander Guarín, Luis Fernando Melo and Eliana González
Inflation Expectations,
Their Measurement
and the Estimate of
Their Degree of
Anchoring
Inflation Expectations,
Their Measurement and
the Estimate of Their
Degree of Anchoring

JOINT RESEARCH PROGRAM 2017

CENTRAL BANK RESEARCHERS NETWORK

CENTER FOR LATIN AMERICAN MONETARY STUDIES

Editors
Alexander Guarín
Senior Researcher associated to Macroeconomic
Modeling Department, Banco de la República
<[email protected]>

Luis Fernando Melo

Senior Econometrician, Banco de la República
<[email protected]>

Eliana González
Chief of Statistics division, Banco de la República
<[email protected]>

First edition, 2020

© Center for Latin American Monetary Studies, 2020

Durango núm. 54, Colonia Roma Norte,
Delegación Cuauhtémoc, 06700
Ciudad de México, México.
All rights reserved)
ISBN: In process
Made in Mexico

vi
EDITORS

Alexander Guarín
Currently, Alexander is senior researcher at Macroeco-
nomics Modeling Department of Banco de la República. He
studied Economics and did a master in Economic Sciences
at Universidad Nacional de Colombia. He also obtained a
master degree in Industrial Engineering from the Univer-
sidad de los Andes (Colombia) and earned a PhD in Com-
putational Finance from the University of Essex (United
Kingdom). His current research interests include the de-
velopment and application of both numerical and econo-
metric methods for the solution of problems in finance and
macroeconomics. In particular, he is interested in meshfree
computational methods and Bayesian econometrics. Some
of his research works have been published in specialized
scientific journals (for example, Journal of Economic, Dy-
namics and Control, European Journal of Operational Re-
search, and Emerging Markets Review). < http://investiga.
banrep.gov.co/es/profile/81>.

Luis Fernando Melo

Statistician from Universidad Nacional of Colombia with
a master’s degree in statistics from University of Michigan-
Ann Arbor. He currently works as a Senior Econometrician in
the Econometric Unit of Banco de la República. His research
interests include Econometrics and time series applied to
macroeconomic and finance. He is author of several papers
published in the Journal of Development Economics, Applied
Economics, Empirical Economics, Contemporary Economic
Policy, Research in International Business and Finance, Eco-
nomic Systems, International Finance, Finance Research Let-
ters and Studies in Economics, and Finance, among others. <
http://investiga.banrep.gov.co/es/profile/354>.

vii
Eliana González
Statistician from Universidad Nacional of Colombia with a
master’s degree in Finance and Econometrics from Univer-
sity of London, Queen Mary College. She currently works as
a Chief of Statistics division of Banco de la República. Her
research areas of interests include Econometrics and time se-
ries applied to macroeconomic and finance. She is author of
several working papers published in Borradores de Economía
series of Banco de la República, besides she has published in
some journals as the American Journal of Agricultural Eco-
nomics and Money affairs, among others.

viii
PREFACE
In 2005, CEMLA’s Board of Governors agreed to bols-
ter economic research and collaboration among its
membership through the establishment of research
activities on topics of common interest. After a careful
analysis of the best way to implement such a program,
the heads of economic studies of the central banks on
the Steering Committee of CEMLA’s Central Bank Re-
search Network identified topics of interest and agreed
that papers on these topics should be presented at the
Network’s Annual Meetings and subsequently publis-
hed. The terms of reference for the first joint research
project were established in 2006, and the first Joint Re-
search Program book was published in 2008, entitled
Estimating and Using Unobservable Variables in the Region.
Since then, research topics have been selected
annually by the heads of economic studies at central
banks within the Research Network Steering Commit-
tee, while representatives from the participating cen-
tral banks have acted voluntarily as coordinators for
each of these projects. Additional volumes have been
published on topics such as inflationary dynamics,
persistence, and prices and wages formation; domes-
tic assets prices, global fundamentals, and financial
stability; monetary policy and financial stability in
Latin America and the Caribbean; international spi-
llovers of monetary policy, and financial decisions of
households and financial inclusion in Latin America
and the Caribbean, among others.
All of the aforementioned subjects are of particular
importance for the design and conduct of monetary
policy and the preservation of financial stability. In its
2017 Meeting, the Research Network focused on a topic
of particular interest for central banking, and whose
importance has increased over recent years: that of the

ix
measurement of inflation expectations and their an-
choring to an inflation target. One particular motive
for this renewed interest was the shocks that affected
inflation trends in the global economy in recent years,
such as commodities price fluctuations and those asso-
ciated with climate change phenomena, among others.
As argued in the literature (an overview of it is offe-
red in the Introduction to the present volume), infla-
tion expectations and, in particular, their degree of
anchoring, are fundamental for determining price
evolution and volatility developments. Therefore, an
accurate measurement of inflation expectations and
a better understanding of their determinants are fun-
damental for the design of an effective monetary poli-
cy. Nevertheless, such a measurement is a challenging
task, which has been approached through survey-based
or model-based methods, including their inference
from market prices of financial instruments. Moreo-
ver, there has been a lively debate among authorities
and researchers about the potential links between po-
licy decisions and agents’ expectations, and whether
long-term expectations may be well-anchored.
The papers included in the present volume address
these and other closely related topics (e.g. forecasts of
inflation using novel techniques). They represent an
effort by researchers of the central banks of Argenti-
na, Bolivia, Brazil, Colombia, Costa Rica, Guatemala,
Mexico, Paraguay, Peru, Spain, as well as researchers
from CEMLA and the Bank for International Settle-
ments (BIS), all of them coordinated by the Banco de
la República (Colombia) with support provided by the
Financial Stability Group of the Inter-American De-
velopment Bank.
We at CEMLA would like to thank the collaborators
in this project, and hope that these documents serve
as a showcase of the analysis carried out in the region
and contribute towards the improvement of policy de-
sign related to the core activities of central banking in
Latin America and the Caribbean.

x
 Table of Contents

Editors......................................................................................... vii
Preface.......................................................................................... ix

Introduction..................................................................................1
Alexander Guarín
Luis Fernando Melo
Eliana González
1. The Formation and Measurement
of Inflation Expectations......................................................... 9
2. The Degree of Anchoring of Inflation Expectations............ 10
3. Inflation Forecasting and Its Performance Evaluation......... 11
4. Inflation Expectations and Its Relation
with Economic Policy............................................................ 12
References............................................................................. 13

FORMATION AND MEASUREMENT OF

INFLATION EXPECTATIONS

Extraction of Inflation Expectations

from Financial Instruments......................................................21
Alberto Fuertes
Ricardo Gimeno
José Manuel Marqués
Abstract........................................................................................ 21
1. Introduction............................................................................ 22
2. Financial Instruments with Information
about Inflation Expectations................................................ 24

xi
 2.1. Inflation-linked Bonds.................................................. 24
2.2. Inflation-linked Swaps................................................... 26
2.3. Inflation-linked Options............................................... 27
2.4. Inflation Expectations from Financial
Instruments in Latin America..................................... 28
3. Modeling Interest Rates from Public Debt Markets............. 29
3.1. The Model...................................................................... 29
4. Results of Inflation Expectations from Public
Debt Markets.......................................................................... 31
4.1 Yield Curve Estimation.................................................. 31
4.2. Empirical Results........................................................... 34
5. Conclusions............................................................................. 45
References............................................................................. 46

The Information Rigidities and Rationality

of Costa Rican Inflation Expectations......................................49
Alonso Alfaro Ureña
Aarón Mora Meléndez
Abstract........................................................................................ 49
1. Introduction............................................................................ 50
2. Inflation Expectations Survey................................................. 52
2.1 Disagreement Among Expectations.............................. 55
2.2 Realized Bias................................................................... 60
3. Sticky Information Model....................................................... 61
3.1 Evidence for Information Rigidities............................. 64
3.2 Simulating a Sticky Information model........................ 66
4. Conclusions............................................................................. 71
Annex........................................................................................... 73
Annex A. Monthly Inflation and Exchange Rates
Expectations Survey..................................................... 73

xii
 Annex B. Expected Inflation, Responses
and Dispersion by Group............................................. 75
References................................................................................... 77

Formation and Evolution of Inflation

Expectations in Paraguay.........................................................79
Pablo Agustín Alonso Méndez
Abstract........................................................................................ 79
1. Introduction............................................................................ 80
2. Monetary Policy in Paraguay.................................................. 81
3. Influence of expectations on inflation.................................. 84
4. Empirical Model for Paraguay................................................ 88
4.1 Data Features.................................................................. 88
4.2 Estimation of the empirical model................................ 91
5. Conclusion.............................................................................. 95
Annex........................................................................................... 97
Annex A. Credibility Index.................................................. 97
References............................................................................. 97

THE DEGREE OF INFLATION EXPECTATION ANCHORING

Anchoring of Inflation Expectations in Latin America...............101

Rocío Gondo
James Yetman
Abstract............................................................................... 101
1. Introduction........................................................................ 102
2. Methodology........................................................................ 105
3. Data........................................................................................ 107
4. Results.................................................................................. 109
4.1 Decay Function............................................................. 109
4.2 Estimated Inflation Anchors........................................ 114
4.3 Effect of it...............................................................................120

xiii
 5. Conclusions.......................................................................... 124
Annex ........................................................................................ 127
References................................................................................. 134

The Time-Varying Degree of Inflation

Expectation Anchoring in Bolivia...........................................137
Mauricio Mora Barrenechea
Juan Carlos Heredia Gómez
David Esteban Zeballos Coria
Abstract...................................................................................... 137
1. Introduction........................................................................ 138
2. Inflation Expectations in Bolivia........................................ 140
3. Empirical Analysis of the Short Term ................................ 145
3.1 bcb Projection against Headline Inflation................. 148
3.2. bcb Projection against Other Variables..................... 151
4. Empirical Analysis in the Medium Term............................. 156
5. Some Considerations Regarding the Results..................... 166
6. Conclusions........................................................................... 168
Annexes..................................................................................... 169
Annex 1. bcb Projections................................................... 169
Annex 2.............................................................................. 171
Annex 3............................................................................... 172
References........................................................................... 173

Expectations Anchoring Indexes for Brazil

Using Kalman Filter: Exploring Signals of Inflation
Anchoring in the Long Term..................................................177
Fernando Nascimento de Oliveira
Wagner Piazza Gaglianone
Abstract...................................................................................... 177
1. Introduction ......................................................................... 178
2. Data........................................................................................ 184

xiv
 3. Empirical Analysis ................................................................ 192
3.1 Signals of Long-term Inflation Anchoring.................. 194
3.1.1 Signals from Survey Data................................... 194
3.1.2 Signals from Market Data.................................. 196
3.1.3 Selection of Signals Based on Correlation
Analysis .............................................................. 196
3.2 Factor Analysis ............................................................. 196
3.3 State-space Model ........................................................ 198
3.4 Baseline eais................................................................. 209
3.5 Robustness Analyses..................................................... 210
4. Conclusion............................................................................ 212
References........................................................................... 216

INFLATION FORECASTING AND ITS

PERFORMANCE EVALUATION

Forecasting Inflation in Argentina.............................................223

Lorena Garegnani
Mauricio Gómez Aguirre
Abstract...................................................................................... 223
1. Introduction.......................................................................... 224
2. Bayesian var Methodology................................................... 225
2.1 Level or Growth Rate.................................................... 228
2.2 Choice of Hyperparameters
and Lag Length Strategy............................................ 229
2.3 Comparison Strategy.................................................... 230
2.4 Model Specification...................................................... 232
3. Data........................................................................................ 233
4. Results.................................................................................... 233
4.1 Estimation of the bvar Model..................................... 233
4.1.1 The Optimal Hyperparameters........................ 233

xv
 4.1.2 Forecasting Exercise........................................... 235
4.2 Forecast Evaluation....................................................... 236
5. Conclusions........................................................................... 237
Annex A. Data Characteristics........................................... 238
Annex B. Results characteristics........................................ 239
References........................................................................... 242

MIDAS Modeling for Core Inflation Forecasting.......................245

Luis Libonatti
Abstract...................................................................................... 245
1. Introduction.......................................................................... 246
2. midas Regression Models...................................................... 246
3. Literature Review.................................................................. 251
4. Data, Exercise, and Results................................................... 255
4.1 Data............................................................................... 255
4.2 Forecasting Exercise..................................................... 257
4.3 Empirical Results.......................................................... 259
5. Conclusion............................................................................ 270
Annex......................................................................................... 270
Appendix A: midas Specifications..................................... 270
midas-dl:..........................................................................................270

midas-adl:........................................................................................271

midas-adl-cf:...................................................................................272

Other:.................................................................................. 272
Appendix B: Additional Tables.......................................... 273
References........................................................................... 281

xvi
Evaluation of Inflation Forecasting Models in Guatemala..........285
Juan Carlos Castañeda-Fuentes
Carlos Eduardo Castillo-Maldonado
Héctor Augusto Valle-Samayoa
Douglas Napoleón Galindo-Gonzáles
Juan Carlos Catalán-Herrera
Guisela Hurtarte-Aguilar
Juan Carlos Arriaza-Herrera
Edson Roger Ortiz-Cardona
Mariano José Gutiérrez-Morales
Abstract...................................................................................... 285
1. Introduction.......................................................................... 286
2. Forecasting Evaluation at the Central
Bank of Guatemala.............................................................. 286
2.1 Unconditional-forecasts Models.................................. 287
2.2. Conditional-forecasts Models..................................... 288
3. Data and Forecast Evaluation Methodology........................ 289
3.1 Data............................................................................... 290
3.2. Forecast Evaluation Methodology.............................. 290
4. Results.................................................................................... 293
4.1. Unconditional Forecast Evaluation............................ 293
4.1.1. Skewness and Normality................................... 293
4.1.2. rmse and mpe..................................................... 294
4.1.3. Diebold-Mariano Test....................................... 294
4.1.4. Pesaran-Timmerman Test................................. 297
4.1.5. Giacomini-Rossi Fluctuation Test.................... 298
4.1.6. Weak Efficiency Test......................................... 299
4.1.7. Strong Efficiency Test....................................... 300
4.2 Conditional Forecast Evaluation................................. 301
5. Conclusions........................................................................... 303
Annex......................................................................................... 303

xvii
 Annex 1. Tables of the Unconditional
Forecast Evaluation.................................................... 303
Tables of the Conditional Forecast Evaluation................. 305
References........................................................................... 307

Forecasting Inflation Expectations from the cesifo

World Economic Survey: An Empirical Application
in Inflation Targeting Countries.............................................309
Héctor M. Zárate Solano
Daniel R. Zapata Sanabria
Abstract...................................................................................... 309
1. Introduction.......................................................................... 310
2. World Economic Survey Data and Their Suitability for
Forecasting Inflation............................................................ 312
2.1 Quantitative Forecasting Inflation Expectations........ 318
2.1.1 Statistical Analysis of the Forecasting Error...... 318
2.1.2 Quantitative Inflation Expectation Results...... 319
3. Methodology......................................................................... 330
3.1 Artificial Neural Networks........................................... 330
3.2 Self-Organizing Maps................................................... 331
3.3 Nonlinear Auto-Regressive Neural Network............... 332
3.4 arima.............................................................................. 336
4. Results.................................................................................... 338
4.1 Self-Organizing Maps of Agents’ Expectations........... 338
4.2 Overlapping Agents’ Inflation Expectations
by Country.................................................................. 340
4.3 Non-Linear Auto-Regressive Neural
Network Results.......................................................... 343
4.4 Forecast Accuracy.......................................................... 348
5. Conclusions........................................................................... 349
Annex A. Data........................................................................... 351
A.1 Qualitative Series......................................................... 351

xviii
 A.2 wes Survey Questionnaire........................................... 357
A.3 Quantitative Forecasting Inflation Expectations....... 365
A.3.1 Equations of the Statistical Analysis
Forecasting Error............................................... 365
Annex B. Self-Organizing Map Validation............................... 380
B.1 Choice of Topology...................................................... 380
B.2 Post-Training Analysis.................................................. 380
B.3 Non-Linear Auto-Regressive Neural
Networks Validation and Other Results.................... 386
B.3.1 Lag Selection..................................................... 386
B.3.2 Post-Training Analysis........................................ 390
B.3.3 MSE Evaluation................................................. 392
B.3.4 Results, Other Countries.................................. 394
B.4 arima............................................................................. 398
References........................................................................... 400

INFLATION EXPECTATIONS
AND ITS RELATION WITH ECONOMIC POLICY

Did the Introduction of Inflation Targeting

Represent a Regime Switch of Monetary Policy
in Latin America?...................................................................405
Sebastián Cadavid Sánchez
Alberto Ortiz Bolaños
Abstract...................................................................................... 405
1. Introduction.......................................................................... 406
2. Model..................................................................................... 409
3. Solution and estimation of the Markov-Switching
dsge model........................................................................... 411
3.1 Counterfactuals............................................................ 415
4. Results.................................................................................... 415
4.1 Regime probabilities.................................................... 415

xix
 4.1.1 High interest rate response regimes................. 416
4.1.2 High volatility shock regimes............................ 420
4.2 Estimation results......................................................... 420
4.3 Impulse response functions......................................... 425
4.4 Counterfactuals............................................................ 442
5. Conclusions........................................................................... 450
Annex......................................................................................... 452
A. Estimated Parameters.................................................... 452
References........................................................................... 462

Fiscal Policy and Inflation: Understanding

the Role of Expectations in Mexico.......................................465
Bernabe López Martin
Alberto Ramírez de Aguilar
Daniel Sámano
Abstract...................................................................................... 465
1.Introduction........................................................................... 466
2. The Baseline Model.............................................................. 469
2.1 The Money Demand and the Government
Budget Constraint...................................................... 469
2.2 Inflation Expectations.................................................. 471
2.3 The Process for Fiscal Deficits...................................... 474
2.4 Model Restrictions on Expectations............................ 477
3. Baseline Model Results......................................................... 478
3.1 Baseline Model Estimation.......................................... 478
3.2 Fiscal Deficits, Inflation, and Expectations................ 481
3.3 Fiscal Deficits: Data and Model Simulation................ 485
4. Beyond the Baseline Model.................................................. 489
4.1 The Role of the Exchange Rate................................... 493
4.2 The Role of the embi Spread........................................ 497
4.3 The Exchange Rate: An Alternative Channel............. 502

xx
5. Concluding Remarks............................................................ 505
6. Appendix............................................................................... 505
6.1 Parameter Estimation................................................... 505
6.2 Adaptive vs. Rational Expectations.............................. 508
6.3 Exchange Rate Regimes............................................... 510
References........................................................................... 511

xxi
Introduction

Alexander Guarín
Luis Fernando Melo
Eliana González

T
he control of inflation and its volatility are fun-
damental issues for any country. Economies with
a high level of inflation or uncertainty on its fu-
ture value can lead, for example, to high costs for eco-
nomic agents, distortions on future investment plans
and welfare implications for society. On the contrary,
economies with low levels of inflation and volatility,
for instance, can enhance their population living
conditions, access to credit sources, and confidence
indicators for international investors (e.g., Madeira
and Zafar, 2015; and Strohsal and Winkelmann, 2015).
Accordingly, keeping inflation under control be-
comes a crucial task for the monetary authority. In this
regard, a strand of the economic literature has estab-
lished an explicit relation between inflation, its long
term expectations and their anchoring to a target lev-
el. In particular, the literature has underlined the re-
lation of this anchoring to the ability of central banks
to control inflation, set up an effective monetary policy
strategy, and improve the transmission mechanisms
(e.g., Haubrich et al., 2012; Autrup and Grothe, 2014;
and Strohsal et al., 2016).

1
 In this context, the appropriate measurement of inflation expec-
tations and their degree of anchoring are essential elements for mak-
ing monetary policy decisions by central banks. Nevertheless, these
variables are unobservable and, hence, their monitoring and assess-
ment are not straightforward.
In practice, inflation expectations are measured through surveys
of specific population groups (e.g., financial market agents, firms,
and consumers), or inferred from financial instruments’ market
prices (e.g., break-even inflation rates, inflation-linked bonds, swaps,
and options). However, the analyses of such expectations from these
two sources of information do not necessarily lead to the same conclu-
sions (e.g., Pierdzioch and Rülke, 2013; and Nautz and Strohsal, 2015).
These measures have different features associated with their
empirical counterparts. Survey-based expectations are a direct es-
timate of the probability distribution of inflation rates from differ-
ent economic sectors. Nonetheless, these expectations are usually
only available at low-frequencies (e.g., monthly or quarterly) and for
a limited number of short-term horizons (typically, one or two years)
(e.g., Autrup and Grothe, 2014; and Pierdzioch and Rülke, 2013).
By contrast, financial market-based expectations can be accessi-
ble in real time, at a higher-frequency (e.g., daily), and with multiple
time horizons, including the long-term ones (e.g., five or ten-year).
Nonetheless, these data are indirect measures of inflation expecta-
tions, whose measurement can be contaminated by several factors.
For instance, break-even inflation rate1 is considered a measure
of inflation compensation that, in addition to inflation expectations,
includes the inflation risk and liquidity premiums. The latter is as-
sociated with market conditions and the availability of liquid nom-
inal and inflation-linked bonds (e.g., Antunes, 2015; and Strohsal
and Winkelmann, 2015).
For authorities, another fundamental aspect is the formation pro-
cess of inflation expectations. This process is essential to understand-
ing how monetary policy decisions are transmitted to expectations
(e.g., economic channels and their speed) and, in turn, to inflation.
This enables central banks to design an effective policy strategy (e.g.,
Evans and Honkapohja, 2001; and Maertens and Rodríguez, 2013).

1
These rates are derived from the spread between nominal and inflation-
linked government bond yields.

2
 The academic literature has directed its attention to two main
schemes of expectations, namely, adaptive and rational. The former
considers that inflation dynamics are based only on their own past
values, and hence agents form their expectations using the observed
price information (that is, a backward-looking rule). Under the lat-
ter scheme, each time expectations are formed, individuals consider
all available information including, for example, the learning from
previous prediction errors, the probable future actions of the central
bank as well as the agents’ beliefs (that is, a forward-looking rule)
(e.g., Taylor, 1985; Kiley, 2007; and Golden and Monks, 2009). There
have been other expectations formation mechanisms proposed
in literature. For example, Gerberding (2001) has studied a combi-
nation of both adaptive and rational schemes, while Ekeblom (2012)
has proposed some degree of learning in the formation of expecta-
tions. Other examples within this literature are Carlson and Valev
(2002), Heinemann and Ullrich (2006), and Oral et al. (2011).
As mentioned, the anchoring of inflation expectations is funda-
mental for monetary policy. In fact, the literature points out that well-
anchored expectations reduce the inflation risk premium, improve
investment decisions, enhance the valuation of long-term assets, low-
er the volatility on long-term interest rates and make them less sen-
sitive to shocks (e.g., Gürkaynak et al., 2010; Mehrotra and Yetman,
2014; and Berument and Froyen, 2015).
Inflation expectations are well-anchored (that is, central bank’s
credibility is strong) if shocks affecting current inflation and its
short-term expectation do not lead to long-run deviations from
the target level. If they are well-anchored, then long-term expecta-
tions should be insensitive to macroeconomic shocks or other sur-
prises, so that once shocks have dissipated, inflation should return
to its long-run target. On the contrary, if central bank’s credibility
is weak, economic shocks could deviate long-term inflation expecta-
tions away from its inflation target (e.g., Demertzis et al., 2009; Galati
et al., 2011; and Pagenhardt et al., 2015).
Recent literature has carried out the measurement of the degree
of anchoring through several methodologies, which capture theo-
retical aspects from two main lines of research. The first one evalu-
ates if long-term inflation expectations are moving close to a target
level, so the degree of anchoring depends on the deviation of these
expectations regarding a specific inflation target (e.g., Mehrotra
and Yetman, 2014; and Strohsal and Winkelmann, 2015).

3
 The second line studies the dependence relation between short-
and long-term inf lation expectations. This literature assesses
if shocks that affect short-term inflation expectations have effects
on those of the long-term, so that the degree of anchoring depends
on how statistically significant is the joint movement between short-
and long-term inflation expectations in response to shocks (e.g.,
Gürkaynak et al., 2010; and Antunes, 2015).
Figure 1 illustrates recent research works on the anchoring of in-
flation expectations. Each of these studies is characterized accord-
ing to both the methodology considered and the source of data used
in its empirical exercises.
A broad segment of this literature has investigated mainly two is-
sues. The first one is the assessment and characterization of differ-
ences in the degree of anchoring between countries with and without
an inflation-targeting regime. For instance, Gürkaynak et al. (2005),
Gürkaynak et al. (2007), Demertzis et al. (2009), Gürkaynak et al.
(2010), and Beechey et al. (2011) examined this matter for the United
States (US) and the euro area, for sample periods between the 1990s
and the end of 2000s. These studies find that a credible inflation-
targeting strategy improves the anchoring of long-term inflation
expectations, reduces their volatility and makes them less sensitive
to inflation shocks.
The second issue is the evolution of the degree of anchoring
over time. For example, the dynamics of anchoring in the pre-
and post-Global Financial Crisis periods in the US between 2004
and 2014 is studied by Galati et al. (2011), Autrup and Grothe (2014),
Nautz and Strohsal (2015), and Strohsal et al. (2016). The first three
works state that inflation expectations have been deanchored since
the Global Financial Crisis, while the latter work points out that
the deanchoring lasted a short period in 2008, after which expecta-
tions were anchored again .
The works by Lemke and Strohsal (2013), Antunes (2015), Pagen-
hardt et al. (2015), and Scharnagl and Stapf (2015) carry out similar
research for the euro area for sample periods between 2000 and 2015.
Lemke and Strohsal (2013) and Scharnagl and Stapf (2015) stated
that although the European Sovereign Debt Crisis increased the vol-
atility of inflation expectations in 2011, these were not deanchored.
On the other hand, Antunes (2015) and Pagenhardt et al. (2015)
found that the same crisis’ events increased the joint movement

4
of short- and long-term inflation expectations, and since then the lat-
ter have been responding to economic shocks.
The variation in the degree of anchoring has also been studied
in other countries for diverse samples between 1996 and 2013.
De Pooter et al. (2014) find that inflation expectations in Brazil, Chile,
and Mexico are anchored, and that these react to us news’ surprises.
Kabundi and Schaling (2013), and Çiçek and Akar (2014) provide
evidence on the unsuccessful anchoring of inflation expectations
in South Africa and Turkey. These are due to low credibility in each
country. Mehrotra and Yetman (2014), and Berument and Froyen
(2015) show that inflation expectations are more firmly anchored
after the adoption of credible inflation-targeting regimes. Other
recent examples are the studies about the degree of anchoring
in Singapore by Ee and Supaat (2008); the US, European Monetary
Union, United Kingdom and Sweden by Strohsal and Winkelmann
(2015), and Colombia by Gamba et al. (2016).
Another topic associated directly with inflation expectations is the
continuous monitoring and forecasting of inflation. This is highly
relevant for central banks and their monetary policy strategies, par-
ticularly in economies with inflation-targeting regimes. Inflation
forecasts are computed using various types of macroeconometric
and time series methodologies. Recently, forecasting models based
on large data sets, high numbers of predictors and direct combina-
tion of different forecast models have attracted the attention of mod-
elers and practitioners. These techniques are useful considering that
central banks have inflation forecasts coming from different models.
Recent examples of these forecasting tools are the Bayesian
model averaging (BMA) (e.g., Wright, 2009), factor-augmented
vector autoregressive (FAVAR) models (e.g., Bernanke et al., 2005)
and schemes for combining forecasts, proposed by Reid (1968),
and Bates and Granger (1969). Hall and Mitchell (2007), and Geweke
and Amisano (2011) consider combinations of forecasting densities
instead of punctual predictions. Tian and Anderson (2014) pro-
posed new schemes for combining forecasts with possible structural
changes, and Kapetanios et al. (2015) extended the previous litera-
ture with weighting schemes.
A fundamental topic in forecasting is the performance evalua-
tion of prediction models and their comparison with respect to a
benchmark or other forecasts. The works by Giacomini and White
(2006), and Giacomini and Rossi (2010) are recent examples of static

5
 Figure 1
RECENT LITERATURE ON THE ANCHORING
OF INFLATION EXPECTATIONS

Copulas Antunes (2015)

Option implied prob.

Scharnagl & Stapf (2015)
density function

Pagenhardt et al. (2015)

Multiple endogenous
break point tests
Nautz & Strohsal (2015)

Gürkaynak et al. (2007)

Gürkaynak et al. (2010)

News regressions Beechey et al. (2011)

Galati et al. (2011)

De Pooter et al. (2013)

FBEI

News regressions
+ models Strohsal & Winkelmann (2015)

Demertzis et al. (2009)

 Lemke & Strohsal (2013)

Berument & Froyen (2015)

Non-linear term Potter & Rosenberg (2007)

Time varying
Strohsal et al. (2016)
parameter models

Regression+
Autrup & Grothe (2014)
model

Kalman learning
Davis & Mack (2013)
process

6
 Figure 1 (cont.)
RECENT LITERATURE ON THE ANCHORING
OF INFLATION EXPECTATIONS

Exponential decay functions Mehrotra & Yetman (2014)

 Demertzis et al. (2009)

Bayesian  Clark & Nakata (2008)

Surveys

Quantile regression Çiçek & Akar (2014)

Castelnuovo et al. (2003)

Linear regression
Kabundi & Schaling (2013)

Tsenova (2011)

Panel regression model

Pierdzioch & Rulke (2013)

and dynamic predictive ability tests, while Rossi and Sekhposyan

(2010) is an application of these tests in inflation forecasting.
Currently, the literature on inflation expectations has gotten
the attention of academics and policy makers. Their renewed interest
in these issues is the result of recent shocks that have affected infla-
tion. In particular, between the end of 2014 and the beginning of 2017,
the global economy suffered a sudden and abrupt fall in oil prices
with diverse effects on other prices and macroeconomic variables.
Likewise, between 2015 and the first-half of 2016, some economies

7
were affected by a climate phenomenon known as El Niño 2 with di-
rect effects on the food supply and its prices, as well as indirect ef-
fects on core inflation through indexation mechanisms. The impact
of these shocks on current inflation has underlined the relevance
of bringing up old and new questions about the formation and mea-
surement of inflation expectations, the estimation of their degree
of anchoring, as well as the development of more accurate forecasts
of future inflation and their relation with expectations.
This is inconsistent. S ometimes they use, for exa mple,
the empirical identification of inf lation expectation formation
processes (e.g., adaptive, rational, hybrid, or adaptive learning),
their changes over time, the statistical validation of these schemes
and the characterization of their main determinants. Likewise,
these queries relate to the measurement of an informative signal
of expectations, the choice of a suitable source of data and time
horizons as well as the theoretical and empirical implications of such
an election for monetary policy decision making.
Other questions are addressed, for instance, the measurement
of the degree of inflation expectations anchoring over time and un-
der different policy regimes, the implementation of existing meth-
odologies, the design of new methods and the comparison of their
results. A recent challenge is the prediction of variations in the de-
gree of anchoring in response to diverse shocks (e.g., climate related
shocks and commodity price’s shocks). Other discussions arise on the
evaluation of the measures of expectations as forecasts of future infla-
tion, structural changes in these predictions and how to model them.
With the aim of providing empirical and theoretical support
to the economic research and the policy decisions of central banks,
the Center for Latin American Monetary Studies (cemla) in coor-
dination with the Banco de la República (that is, the Central Bank
of Colombia) organized the 2017 Joint Research Annual Program
to study inflation expectations and other relevant topics associat-
ed with them. In the development of this program, the Financial
Stability Group of the Inter-American Development Bank and the
cemla provided academic support to the research groups through

2
This is a season of high temperatures, shortage of rains and droughts.

8
academic feedback given by professors Olivier Coibion, 3 Massimil-
iano Marcellino,4 and Andrea Tambalotti. 5
This joint program was an opportunity to deal with some of the
previous questions, learn about the current research on inflation
expectations in central banks and contribute to the burgeoning eco-
nomic literature on these issues. The results of this research agenda
are compiled in this book, which includes 13 chapters. The first one is
this Introduction. The remaining 12 chapters correspond to works
from 10 central banks (Argentina, Bolivia, Brazil, Colombia, Costa
Rica, Guatemala, Mexico, Paraguay, Peru, and Spain) and two in-
ternational institutions (Bank for International Settlements – BIS,
and cemla). These works address topics on the formation of infla-
tion expectations, their measurement through surveys and finan-
cial market data, the estimation of the degree of anchoring adopting
several methodological approaches, and the forecasts of inflation
using novel techniques. The works were divided into four main sec-
tions, as follows.

1. THE FORMATION AND MEASUREMENT

OF INFLATION EXPECTATIONS
In chapter 2, Alberto Fuertes, Ricardo Gimeno and José Manuel
Marqués of the Banco de España use the affine model proposed by Gi-
meno and Marqués (2009) to decompose the nominal interest rates
from Chile, Mexico, Colombia, and Brazil into real risk-free rates,
inflation expectations and risk premium. For each country, the em-
pirical exercises consider different sample periods between 2001
and 2016, depending on the availability of data on nominal govern-
ment bonds. Results suggest that expectations in Mexico and Chile
were anchored during the periods of study. On the contrary, in Co-
lombia and Brazil, during the sample period analyzed, the inflation
expectations deanchored and fluctuated over time.

3
Associate Professor at University of Texas at Austin.
4
Full Professor of Econometrics at Bocconi University.
5
Assistant Vice President and Function Head, Macroeconomic and
Monetary Studies Function, Research and Statistics Group, Federal
Reserve Bank of New York.

9
 Chapter 3 presents the work by Alonso Alfaro and Aarón Mora
from the Banco Central de Costa Rica. The authors use the model
by Mankiw and Reis (2002) to examine information rigidities in in-
flation expectations of agents from several economic sectors between
2006 and 2017. Although previous studies suggest the existence
of these rigidities in the expectations formation process in Costa
Rica, the results of this research do not support these claims. Esti-
mates show that the magnitude of the rigidities captured from data
is not large enough to validate that statement.
The work by Pablo Alonso of the Banco Central del Paraguay is pre-
sented in chapter 4. Alonso estimates a model of determinants of the
formation of inflation expectations in Paraguay since the adoption
of the inflation-targeting scheme in 2011. His results show that ex-
pectations are a function of past inflation and the credibility in the
central bank. Other variables such as the foreign exchange rate de-
preciation and the changes in oil prices do not seem to play a key role
in their determination.

2. THE DEGREE OF ANCHORING

OF INFLATION EXPECTATIONS
In chapter 5, Rocío Gondo and James Yetman of the Banco Cen-
tral de Reserva del Perú and the bis, respectively, use the work
by Mehrotra and Yetman (2014) to infer from inflation expecta-
tions, for several Latin American countries between 1993 and 2016,
an implicit anchor in the data. They also assess how it has evolved
over time and compare it with the central bank’s target level. Results
show that most countries have an anchor whose importance has in-
creased over time as a result of improvements in the credibility after
the adoption of an inflation-targeting regime.
The research work by Mauricio Mora, Juan Carlos Heredia
and David Zeballos of the Banco Central de Bolivia (bcb) is pre-
sented in chapter 6. Authors assess whether inflation expectations
in Bolivia between 2005 and 2017 were anchored, in the sense that
they were coherent with the inflation future path and the target lev-
el announced by the central bank. 6 Results indicate that long-term

6
Bolivia is under a monetary-targeting scheme, such that the main refer-
ence for future inflation are the central bank’s projections.

10
expectations were strongly anchored since 2014 due to a greater
credibility of the bcb linked with a larger intervention in the money
market, a more active communication strategy and a stable macro-
economic environment.
Chapter 7 presents the research by Fernando Nascimento
de Oliveira and Wagner Gaglianone of Banco Central do Brasil
(bcbr). They build several time-varying expectation anchoring index-
es of the bcbr from 2002 to 2017, which are based on the monetary au-
thority’s capability to anchor long-term inflation expectations. Those
indexes consider variables of fiscal and monetary policy in their es-
timation. Authors state that estimated indexes are consistent with
the central bank’s credibility perceived by economic agents in Bra-
zil over the sample period.

3. INFLATION FORECASTING
AND ITS PERFORMANCE EVALUATION
Chapters 8 and 9 present the research works developed by Lorena
Garegnani and Maximiliano Gómez, and Luis Libonatti of the
Banco Central de la República Argentina, respectively. Garegnani
and Gómez estimate Bayesian var models with Argentinian data
from 2004 to 2017, and forecast the headline inflation for several
time horizons under a rolling window scheme. In the same line
of research, Libonatti uses a mixed data sampling regression mod-
el to forecast the monthly core inflation of Argentina between 2015
and 2017 using a daily online price index captured by web scraping.
In both works, authors compare their results to forecasts from tradi-
tional benchmark models and show, in general, a good performance
of their predictions.
In chapter 10, the Economic Research Department of the Banco
de Guatemala (Banguat) presents its work. This research assesses
the performance of both unconditional and conditional inflation
forecasts for several time horizons between 2011 and 2017. These
predictions are built using time series tools and structural macro-
economic models used by the Banguat. In line with the traditional
literature, their results show that forecasts computed with time series
tools provide more accuracy in the shortest terms while structural
macroeconometric models provide better predictions for medium-
and long-term horizons.

11
 In chapter 11, Héctor Zárate and Daniel Zapata from Banco de la
República (Colombia) use artificial neural networks to forecast in-
flation expectations in a set of 16 countries with inflation-targeting
regimes and a sample period between 1991 and 2016. Their predic-
tions consider different expectations patterns depending on percep-
tions about the oil shock in 2014. Authors show that their exercises
provide more accurate forecasts than the benchmark model and,
anticipate turning points of inflation in most of the cases.

4. INFLATION EXPECTATIONS AND ITS

RELATION WITH ECONOMIC POLICY
In chapter 12, Sebastián Cadavid and Alberto Ortiz from cemla
examine empirically if the economic reforms implemented in Bra-
zil, Chile, Colombia, Mexico, and Peru between 1999 and 2002 –par-
ticularly the adoption of an inflation-targeting regime and a flexible
exchange rate– led to the observed reduction of inflation in these
countries. Their empirical exercises consider counterfactual sce-
narios in an open economy with monetary factors. The authors show
that if these reforms had not been adopted in these Latin Ameri-
can countries, they would have experienced higher inflation rates,
variations in gross domestic product with small gains in economic
growth and a large volatility in nominal variables.
Finally, chapter 13 presents the work by Bernabé López-Martin,
Alberto Ramírez de Aguilar and Daniel Sámano from Banco de Mex-
ico. They analyze the interaction between inflation, its expectations
and fiscal deficits in Mexico between 1969 and 2016. The authors ex-
tend the model developed by Sargent et al. (2009) to study how fiscal
policy can affect inflation expectations in a context of central bank
independence. Their results suggest that the fiscal deficits financed
through money creation are central to explain the behavior of Mexi-
can inflation and its expectations during the sample period.
The editors trust that this brief introduction will motivate readers
to carry on with each of the research works in this book. This publi-
cation is an opportunity to learn about relevant issues on inflation
expectations for central banks in the region, as well as the current
state of their research. We also think that this book will encourage
relevant policy discussions on inflation, its expectations and other
related issues, contributing to literature on monetary policy.

12
References

Antunes, António Armando (2015), “Co-movement of Revisions in

Short- and Long-Term Inflation Expectations,” Banco de Portugal
Economic Studies, Vol. 1, No. 1, pp. 1-19.
Autrup, Søren, and Magdalena Grothe (2014), Economic Surprises
and Inflation Expectations: Has Anchoring of Expectations Sur-
vived the Crisis? Working Paper Series, European Central
Bank, No. 1671.
Bates, J. M. and C. W. J. Granger (1969), “The Combination of
Forecasts,” Operations Research Quarterly, Vol. 20, No. 4, De-
cember, pp. 451-468, <doi: 10.2307/3008764>.
Beechey, Meredith J., Benjamin K. Johannsen, and Andrew T.
Levin (2011), “Are Long-run Inflation Expectations Anchored
More Firmly in the Euro Area than in the United States?,”
American Economic Journal: Macroeconomics, Vol. 3, No. 2, pp.
104-129, <doi: 10.1257/mac.3.2.104>.
Bernanke, Ben S., Jean Boivin, and Piotr Eliasz (2005), “Mea-
suring Monetary Policy: A Factor Augmented Vector Au-
toregressive ( favar ) Approach,” Quarterly Journal of Eco-
nomics, Vol. 120, issue 1, pp. 387-422, <https://doi.
org/10.1162/0033553053327452>.
Berument, Hakan, and Richard T. Froyen (2015), “Monetary Policy
and Interest Rates Under Inflation Targeting in Australia
and New Zealand,” New Zealand Economic Papers, Vol. 49,
issue 2, pp. 171-188, <https://doi.org/10.1080/0077995
4.2014.929608>.
Carlson, John A., and Neven T. Valev (2002), “A Disinflation
Trade‐off: Speed Versus Final Destination,” Economic Inquiry,
Vol. 40, issue 3, pp. 450-456, <https://doi.org/10.1093/
ei/40.3.450>.
Çiçek, Serkan, and Cüneyt Akar (2014), “Do Inflation Expecta-
tions Converge Toward Inflation Target or Actual Inflation?
Evidence from Expectation Gap Persistence,” Central Bank
Review, Vol. 14, No. 1, pp. 15-21.
Davis, Scott, and Adrienne Mack (2013), Cross-country Variation in
the Anchoring of Inflation Expectations, Staff Papers, Federal
Reserve Bank of Dallas, No. 21.

13
De Pooter, Michiel, Patrice Robitaille, Ian Walker, and Michael
Zdinak (2014), Are Long-term Inflation Expectations Well-anchored
in Brazil, Chile, and Mexico?, International Finance Discussion
Papers, Board of Governors of the Federal Reserve System,
No. 1098, March.
Demertzis, Maria, Massimiliano Marcellino, and Nicola Viegi
(2009), Anchors for Inflation Expectations, dnb Working Papers,
De Nederlandsche Bank, No. 229.
Ee, Khor Hoe, and Saktiandi Supaat (2008), “The Anchoring of
Inflation Expectations in Singapore,” bis Papers, No. 35, pp.
443-449.
Ekeblom, D. (2012), Empirical Swedish Inflation Expectations. Tech-
nical report, Seminar in Macroeconomics, Department of
Economics, Lund University.
Evans, George W., and Seppo Honkapohja (2001), Learning and
Expectations in Macroeconomics, Princeton University Press.
Galati, Gabriele, Steven Poelhekke, and Chen Zhou (2011), “Did
the Crisis Affect Inflation Expectations?,” International Journal
of Central Banking, Vol. 7, No. 1, pp. 1-27.
Gamba, Santiago, Eliana Rocío González, and Luis Fernando Melo
(2016), ¿Están ancladas las expectativas de inflación en
Colombia?, Borradores de Economía, Banco de la República,
No. 940.
Gerberding, Christina (2001), The Information Content of Survey Data
on Expected Price Developments for Monetary Policy, Discussion
Paper, Economic Research Centre of the Deutsche Bundes-
bank, No. 9/01.
Geweke, John, and Gianni Amisano (2011), “Optimal Prediction
Pools,” Journal of Econometrics, Vol. 164, issue 1, pp. 130-141,
<https://doi.org/10.1016/j.jeconom.2011.02.017>.
Giacomini, Raffaella, and Barbara Rossi (2010), “Forecast Com-
parisons in Unstable Environments,” Journal of Applied Econo-
metrics, Vol. 25, No. 4, pp. 595-620.
Giacomini, Raffaella, and Halbert White (2006), “Tests of Condi-
tional Predictive Ability,” Econometrica, Vol. 74, No. 6, pp. 1545-
1578, <https://doi.org/10.1111/j.1468-0262.2006.00718.
x>.

14
Gimeno, Ricardo, and José Manuel Marqués (2009), Extraction of
Financial Market Expectations about Inflation and Interest Rates
from a Liquid Market, Documentos de Trabajo, Banco de
España, No. 906.
Golden, Brian, and Allen Monks (2009), “Measuring Inflation
Expectations in the Euro Area,” Quarterly Bulletin, No. 1,
pp. 67-84.
Gürkaynak, Refet S., Andrew T. Levin, and Eric Swanson (2010),
“Does Inflation Targeting Anchor Long-Run Inflation Expec-
tations? Evidence from the us, uk, and Sweden,” Journal of the
European Economic Association, Vol. 8, issue 6, pp. 1208-1242,
<https://doi.org/10.1111/j.1542-4774.2010.tb00553.x>
Gürkaynak, Refet S., Andrew T. Levin, Andrew N. Marder, and Eric
Swanson (2007), “Inflation Targeting and the Anchoring of
Inflation Expectations in the Western Hemisphere,” frbsf
Economic Review, Vol. 11, 25-47.
Gürkaynak, Refet S., Brian Sack, and Eric Swanson (2005), “The
Sensitivity of Long-Term Interest Rates to Economic News:
Evidence and Implications for Macroeconomic Models,”
The American Economic Review, Vol. 95, issue 1, pp. 425-436,
<doi: 10.1257/0002828053828446>.
Hall, Stephen G., and James Mitchell (2007), “Combining Den-
sity Forecasts,” International Journal of Forecasting, Vol. 23,
issue 1, pp. 1-13, <https://doi.org/10.1016/j.ijfore-
cast.2006.08.001>.
Haubrich, Joseph, George Pennacchi, and Peter Ritchken (2012),
“Inflation Expectations, Real Rates, and Risk Premia: Evi-
dence from Inflation Swaps,” The Review of Financial Studies,
Vol. 25, issue 5, pp. 1588-1629, <https://doi.org/10.1093/
rfs/hhs003>.
Heinemann, Friedrich, and Katrin Ullrich (2006), “The Impact
of emu on Inflation Expectations,” Open Economics Review,
Vol. 17, issue 2, pp. 175-195.
Kabundi, Alain, and Eric Schaling (2013), “Inflation and Inflation
Expectations in South Africa: An Attempt at Explanation,”
South African Journal of Economics, Vol. 81, issue 3, pp. 346-
355, <https://doi.org/10.1111/saje.12007>.

15
Kapetanios, George, James Mitchell, Simon Mc Kay Price, and
Nicholas W. P. Fawcett, N. (2015), “Generalised Density
Forecast Combinations,” Journal of Econometrics, Vol. 188,
issue 1, pp. 150-165, <https://doi.org/10.1016/j.jeco-
nom.2015.02.047>.
Kiley, Michael T. (2007), Monetary Policy Actions and Long-run Inflation
Expectations, feds Working Paper 2008-03, Federal Reserve.
Lemke, Wolfgang, and Till Strohsal (2013), “What Can Break-even
Inflation Rates Tell Us About the Anchoring of Inflation
Expectations in the Euro Area?,” in Annual Conference 2013
(Duesseldorf): Competition Policy and Regulation in a Global Eco-
nomic Order, No. 79794.
Madeira, Carlos, and Basit Zafar (2015), “Heterogeneous Infla-
tion Expectations and Learning,” Journal of Money, Credit
and Banking, Vol. 47, issue 5, pp. 868-996, <https://doi.
org/10.1111/jmcb.12230>.
Maertens, Luís Ricardo and Gabriel Rodríguez (2013), “Inflation
Expectations in the Presence of Policy Shifts and Structural
Breaks: An Experimental Analysis,” The Journal of Socio-Eco-
nomics, Vol. 44, June, pp. 59-67, <https://doi.org/10.1016/j.
socec.2013.02.001>.
Mankiw, N. Gregory and Ricardo Reis (2002), “Sticky Information
Versus Sticky Prices: A Proposal to Replace the New Keynes-
ian Philips Curve,” The Quarterly Journal of Economics, Vol.
117, issue 4, pp. 1295-1328, <https://doi.org/10.1162/0
03355302320935034>.
Mehrotra, Aaron, and James Yetman (2014), Decaying Expectations:
What Inflation Forecasts Tell Us about the Anchoring of Inflation
Expectations, bis Working Papers, No. 464.
Nautz, Dieter, and Till Strohsal (2015), “Are us inflation Expec-
tations Re-anchored?,” Economics Letters, Vol. 127, pp. 6-9,
<https://doi.org/10.1016/j.econlet.2014.12.023>.
Oral, Ece, Hülya Saygili, Mesut Saygili, and S. Özge Tuncel. (2011),
“Inflation Expectations in Turkey: Evidence from Panel Data,”
Journal of Business Cycle Measurement and Analysis, oecd, Vol.
2011, issue 1, pp. 5-28, <https://doi.org/10.1787/jbcma-
2011-5kgg5k53np7c>.

16
Pagenhardt, Laura, Dieter Nautz, and Till Strohsal (2015), The (De-)
Anchoring of Inflation Expectations: New Evidence from the Euro
Area, sfb 649 Discussion Paper Series, No. 2015-044, pp. 1-23.
Pierdzioch, Christian, and Jan-Christoph Rülke (2013), “Do In-
flation Targets Anchor Inflation Expectations?,” Economic
Modelling, Vol. 35, September, pp. 214-223, <https://doi.
org/10.1016/j.econmod.2013.06.042>.
Potter, Simon M., and Joshua Rosenberg (2007), Are us Inflation
Expectations Anchored, Contained or Unmoored?, Working paper,
Federal Reserve Bank of New York.
Reid, David J. (1968), “Combining Three Estimates of Gross Do-
mestic Product,” Economica, Vol. 35, No. 14, pp. 431-444,
<doi: 10.2307/2552350>
Rossi, Barbara, and Tatevik Sekhposyan (2010), “Have Economic
Models’ Forecasting Performance for us Output Growth
and Inflation Changed over Time, and When?,” International
Journal of Forecasting, Vol. 26, issue 4, pp. 808-835, <https://
doi.org/10.1016/j.ijforecast.2009.08.004>.
Sargent, Thomas, Noah Williams, and Tao Zha (2009), “The Con-
quest of South American Inflation,” Journal of Political Econ-
omy, Vol. 117, No. 2, pp. 211-256.
Scharnagl, Michael, and Jelena Stapf, J. (2015), “Inflation, De-
flation, and Uncertainty: What Drives Euro-Area Option-
Implied Inflation Expectations, and Are they Still Anchored
in the Sovereign Debt Crisis?,” Economic Modelling, Vol. 48,
August, pp. 248-269, <https://doi.org/10.1016/j.econ-
mod.2014.11.025>.
Strohsal, Till, and Lars Winkelmann (2015), “Assessing the An-
choring of Inflation Expectations,” Journal of International
Money and Finance, Vol. 50, February, pp. 33-48, <https://
doi.org/10.1016/j.jimonfin.2014.09.001>.
Strohsal, Till, Rafi Melnick, and Dieter Nautz (2016), “The Time-
Varying Degree of Inflation Expectations Anchoring,” Journal
of Macroeconomics, Vol. 48, June, pp. 62-71, <https://doi.
org/10.1016/j.jmacro.2016.02.002>.

17
Taylor, John B. (1985), “Rational Expectations Models in Macro-
economics,” in Kenneth J. Arrow and Seppo Houkupohju
(eds.), Frontiers of Economics, Basil Blackwell Publishers, pp.
391-425.
Tian, Jing, and Heather M. Anderson (2014), “Forecast Combi-
nations Under Structural Break Uncertainty,” International
Journal of Forecasting, Vol. 30, issue 1, January-March, pp. 161-
175, <https://doi.org/10.1016/j.ijforecast.2013.06.003>.
Wright, Jonathan H. (2009), “Forecasting us Inflation by Bayesian
Model Averaging,” Journal of Forecasting, Vol. 28, No. 2, pp.
131-144, <https://doi.org/10.1002/for.1088>.

18
Formation and Measurement
of Inflation Expectations
Extraction of Inflation Expectations
from Financial Instruments

Albe r to Fue r tes

Rica rdo G im e n o
José Man uel Ma rqués

Abstract
In this paper, we estimate inflation expectations for several Latin American
countries using an affine model that takes as factors the observed inflation
and the parameters generated from zero-coupon yield curves of nominal bonds.
By implementing this approach, we avoid the use of inflation-linked securi-
ties, which are scarce in many of these markets, and obtain market measures
of inflation expectations free of any risk premium, eliminating potential bi-
ases included in other measures such as breakeven rates. Our method provides
several advantages, as we can compute inflation expectations at any hori-
zon and forward rates such as the expected inflation over the five-year period
that begins five years from today. We find that inflation expectations in the
long-run are fairly anchored in Chile and Mexico, while those in Brazil and
Colombia are more volatile and less anchored. We also find that expected in-
flation increases at longer horizons in Brazil and Chile, while it is decreas-
ing in Colombia and Mexico.
Keywords: inflation expectations, affine model, real interest rate, risk
premium.
jel classification: G12, E43, E44, C54.

A. Fuertes <[email protected]>, Banco de España, corresponding author; R.

Gimeno <[email protected]>, Banco de España, and J. M. Marqués <manuel-
[email protected]>, Banco de España.

21
1. INTRODUCTION

A
gents’ inf lation expectations are decisive when studying
changes in many of the variables shaping households’ and
firms’ decision making. One approach to obtain inflation
expectations is based on the consensus view of specialist economic
forecasters, such as the surveys of professional forecasters by the
European Central Bank and the Federal Reserve Bank of Philadel-
phia, both of which are released quarterly. Other surveys also exist,
such as the monthly University of Michigan Survey of Consumers in
the United States, which elicits information from consumers rather
than professional economic forecasters. In Latin America, several
central banks also publish surveys about inflation expectations.1 A
drawback of these surveys is that they are released relatively infre-
quently and, thus, the information received has a time lag. Moreover,
they only cover a small range of time horizons and, as identified in
the literature (Ang et al., 2007; Chan et al., 2013), there is some bias
and inertia in their responses.
An alternative way of obtaining agents’ inflation expectations is
to use prices of market-traded financial instruments employed to
hedge against inflation such us inflation-linked bonds, inflation
swaps, and inflation options. One may argue that, given that inves-
tors risk their funds when taking investment decisions based on
expected future inflation and professional forecasters do not have
any vested interest, they could provide a better forecast since they
have more skin in the game. Another advantage to this approach is
that it is possible to derive the whole probability function (Gimeno
and Ibañez, 2017). This makes it possible to estimate, for example,
the probability of the occurrence of certain extreme events or the
uncertainty of future inflation. Another additional advantage in
comparison with surveys is that changes in expectations can be ob-
served almost in real time. This makes it easier to identify the effect
of specific events or decisions on inflation expectations. Unfortu-
nately, there are not many markets of inflation-linked securities avail-
able for most countries. For example, in Latin American only a few
have inflation-linked bonds, and there are no markets for inflation

1
For example, the central banks of Chile, Colombia and Mexico pub-
lish a monthly survey about inflation expectations; the Bank of Brazil
publishes a daily survey.

22 A. Fuertes, R. Gimeno, J. M. Marqués

options at all. Another problem of obtaining inflation expectations
using this approach is the presence of various risk premia, which are
included in the prices of the underlying financial assets and which
may also vary over time. The presence of these premia may distort
the information content of these indicators, which may affect mea-
sures of agents’ inflation expectations.
Due to the lack of inflation-linked securities in Latin American
markets, we use an alternative approach developed by Gimeno and
Marques (2012) to obtain inflation expectations: An affine model
that takes as factors the observed inflation and the parameters gen-
erated in the zero-coupon yield curve estimation of nominal bonds.
Also, by implementing this approach, we obtain a measure of infla-
tion expectations free of any risk premia, since the model breaks
down nominal interest rates as the sum of real risk-free interest rates,
expected inflation, and the risk premium.
To the best of our knowledge, this is the first attempt to obtain pure
inflation expectations using nominal government bonds for Latin
American countries. We obtain government bond data for Brazil,
Chile, Colombia, and Mexico, being able to estimate the zero-coupon
yield curve and decompose that curve into the real risk-free rate, the
risk premia, and inflation expectations. We can obtain inflation ex-
pectations for all of the horizons computed in the zero-coupon yield
curve as well as forward rates such as the expected inflation over the
five-year period that begins five years from today (the 5Y5Y forward
rate). We find that inflation expectations in the long-term (5Y5Y)
seem to be anchored in Chile and Mexico, although the level of ex-
pected inflation is above the central bank target rate of 3%. On the
other hand, long-term inflation expectations in Brazil and Colombia
are more volatile and have been fluctuating over time, experiencing
a large decrease during 2017. These results may also point out that
government bond markets in Brazil and Colombia do not provide
as much information about future inflation as the other +markets.
We also find the expected inflation is currently increasing with the
horizon in Brazil and Chile, while it is decreasing in Colombia and
Mexico. For Mexico, there has been an important shock on expected
inflation after the last us presidential elections, experiencing a large
increase. None of the other countries analyzed have shown this pat-
tern, limiting the spillovers effects of the results of the us presidential
elections to inflation expectations in Mexico. Finally, we compare the
forecasting power over one year of inflation expectations obtained

Extraction of Inflation Expectations from Financial Instruments 23

using our approach with expected inflation obtained from surveys.
Our approach performs better predicting inflation for Chile, while
surveys do better for Brazil, Colombia, and Mexico.
Further analysis shows that inflation expectations from our mod-
el complement those from surveys and provide additional informa-
tion. A simple average of the expected inflation obtained using our
approach and expected inflation from surveys provides a better fit
than using only expectations from surveys for all countries but Bra-
zil. Overall there is a trade-off between the two ways of obtaining ex-
pected inflations, as surveys are less responsive to inflation shocks
and our approach produces expected inflation levels that are more
correlated with current inflation.
The paper proceeds as follows: Section 2 describes the financial
instruments from which information about inflation expectations
can be derived, analyzing their availability for Latin American mar-
kets. Section 3 summarizes the main features of the affine model we
implement to obtain inflation expectations, and Section 4 shows the
results. Section 5 concludes.

2. FINANCIAL INSTRUMENTS WITH

INFORMATION ABOUT INFLATION
EXPECTATIONS

2.1. Inflation-linked Bonds

One of the most popular metrics of inflation expectations based
on financial asset prices is the one obtained from inflation-linked
bonds (break-even inflation rates). This is calculated by compar-
ing the yield of a conventional bond (whose associated coupon and
principal payments are fixed in nominal terms), with that of an in-
flation-linked bond (indexed to a price index) of the same maturity
from the same issuer.
The inflation-linked bond market is particularly active in the
United States, where these assets (known as Treasury inflation-pro-
tected securities or tips) are issued in sufficient quantity to create
a liquid market in which price formation is fluid. However, the situ-
ation in Europe is fragmentized due to the existence of multiple is-
suers (namely the traditional issuer of treasuries for France, Italy,
and Germany, and the less frequent issuer Greece, later joined by

24 A. Fuertes, R. Gimeno, J. M. Marqués

Spain in 2014) and the use of different consumer price indices (na-
tional and European) as a reference. These factors reduce liquidity
and are an obstacle to obtaining a clear signal on the compensation
demanded by investors for the expected increases in the cost of liv-
ing. In Latin America, there are several markets of inflation-linked
bonds in countries such as Brazil, Chile, and Mexico.
Besides the lack of market depth and liquidity, an additional prob-
lem with this indicator is that it includes other components as well
as investors’ expectations about future price developments. Firstly,
given that investors are averse to inflation risk, they will demand a
premium on conventional bonds that compensates them for the risk
incurred, but not on inflation-linked bonds, as they are protected
against this risk. For this reason, the indicator does not strictly mea-
sure the level of expectations, but rather the compensation for infla-
tion that investors demand. Secondly, the different level of liquidity
of the two instruments used to obtain the indicator (generally higher
for conventional bonds than inflation-linked ones) means the yield
spread between them is also influenced by their different liquidity
premiums. As well as the aforementioned inflation-related factors,
conventional bonds include a component reflecting the expected fu-
ture course of the real interest rate, together with its associated risk
premium. Finally, it should be borne in mind that the size of the pre-
mia present in the break-even rate (inflation risk and relative liquid-
ity) may change over time, depending on changes in investors’ risk
appetite, the level of inflation risk, or market liquidity conditions.
The inflation compensation metric derived from inflation-linked
bonds may also be temporarily affected by other factors in addition
to those mentioned. Thus, for instance, changes in the supply and
demand for conventional bonds relative to inflation-linked bonds,
such as those associated with quantitative easing programs, 2 for

2
Only conventional government bonds were purchased in the Federal
Reserve Board’s first quantitative easing program. During the Federal
Reserve Board’s second quantitative easing program ( qe  II), a total
of usd 600 billion-worth of government securities was purchased, of
which 26 billion was in the form of inflation-linked bonds. The fact
that more conventional bonds are being bought than inflation-linked
bonds could push down their relative yield, and therefore depress the
inflation expectations indicator in a way that is due to a mismatch in
the supply and demand for bonds used to calculate the indicator rather
than to agents’ forecasts of future consumer price trends.

Extraction of Inflation Expectations from Financial Instruments 25

example, may cause distortions in these indicators. Given all these
drawbacks, economists have developed extensive academic litera-
ture seeking to isolate different components of the inflation expec-
tation indicators obtained from inflation-linked bonds. 3

2.2. Inflation-linked Swaps

Along with inflation-linked bonds, inflation-linked swaps (ils) are
another type of financial asset containing information about agents’
inflation expectations. In this derivative instrument, one of the
contracting parties agrees to pay the counterparty a fixed sum on a
future date in exchange for a payment linked to the future level of
a price index. For example, in the case of a one-year ils, the fixed-
rate party could agree to pay 2% of €1 million in consideration for
receiving a fraction of this nominal €1 million equivalent to the in-
crease in the cpi over this 12-month period. Contrary to the case of
inflation-linked bonds, the ils market is more liquid in Europe than
in United States (Gimeno and Ibáñez, 2017) and there are not ils
markets in Latin America, except in Brazil.
ilss are bilaterally negotiated private contracts with no interme-
diary clearinghouse. This creates the risk that the other party will
fail to meet its commitment at the end of the period, so the nego-
tiated price incorporates the corresponding premium. Neverthe-
less, the absence of cash transfers before the expiry date reduces
the size of this premium, as well as the liquidity premium, as there
is no opportunity cost relative to alternative investments (Fleming
and Sporn, 2013).
Like inflation-linked bonds, inflation swaps contain an inflation
risk premium. Therefore, they measure compensation for inflation
as well as inflation expectations. One of the main advantages of the
ils-based indicator relative to the one obtained from inflation-linked
bonds is that, since it is not necessary to compare two different bonds,
the distortions caused by ad hoc factors that affect the markets asym-
metrically are eliminated. Particularly, these indicators would not
have been directly affected by distortions linked to the implemen-
tation of central banks’ asset purchase programs.

3
See, for example, D’Amico et al. (2014) and Chernov and Mueller
(2012).

26 A. Fuertes, R. Gimeno, J. M. Marqués

2.3. Inflation-linked Options
Inflation options are contracts in which one of the parties agrees to
pay the other an amount depending on whether a price index exceeds
(cap) or falls below (floor) a given threshold (the strike rate) within a
given period. If the condition is met, the payment would be the dif-
ference, in absolute terms, between the index and the threshold.
Unlike both inflation-linked bonds and ils s, which give estimates
of the averages only at specific points in time, options can be used
together with ils s to obtain additional information such as the full
probability distribution of the future course of inflation or implied
volatility of inflation. This gives information about risk and uncer-
tainty around the expected average value. In particular, an increase
in the implied volatility suggests that agents are more concerned and
there is more uncertainty over the future course of price indices.
As in the case of ils s, options are negotiated bilaterally without
the intervention of a clearinghouse, so prices may include a counter-
party risk premium. Most of these derivatives are negotiated using
the harmonized euro area cpi, the uk rpi (Retail Price Index), or
the us cpi (Consumer Price Index), with maturities ranging from 1
to 30 years. The most liquid market is linked to the euro area index,
followed by that of the uk (see Smith, 2012). It should also be noted
that, as in the case above of the other financial instruments, option
prices also contain premiums for inflation risk, and potentially, for
liquidity risk. Currently, there are no markets for inflation options
in Latin America.
The inflation risk premium is present in all three indicators, and
the amount is the same. For its part, the liquidity risk premium is
negative in the case of the bond-based metric, as conventional bonds
are more liquid than interest-linked bonds, whereas, in the ils, the
sign of this premium is positive. The counterparty risk premium is
only present in the case of ils s and inflation options. Finally, the
estimation error may be more significant for an indicator based on
inflation-linked bonds.4

4
Unlike ils s, where the compensation for inflation is directly observ-
able from the price, the bond-based indicator requires a comparison
of the yields on inflation-linked bonds and conventional bonds. The
differences in the features of both types of bonds, beyond the fact that
in the case of inflation-linked bonds payments are linked to inflation
(such as, for example, their expiry), may distort the inflation expecta-

Extraction of Inflation Expectations from Financial Instruments 27

2.4. Inflation Expectations from Financial Instruments
in Latin America
Given the scarcity of financial instruments linked to price indexes
in Latin American, obtaining indicators of inflation expectations
from these securities is difficult and limited to a few countries. Also,
the only indicator we can obtain is the break-even rate for those mar-
kets where inflation-linked bonds and conventional bonds exist and
are liquid. This break-even rate is used as a proxy for expected infla-
tion but, as we mentioned earlier, also includes several premia such
as the risk and liquidity premia. We do not know the size of these
premia, and thus we must keep in mind that this indicator provides
only information about inflation compensation rather than pure
inflation expectations.
Unfortunately, obtaining data on break-even rates for other coun-
tries is difficult because of the lack of inflation-linked securities. Table
1 shows the availability of each type of securities for Latin American
countries. Even though there are several markets for inflation-linked
bonds, it may be the case that, for some countries, it is difficult to
obtain accurate prices, as there is either a small variety of bond ma-
turities or bond markets are relatively illiquid. In the next section,
we describe a different approach to obtain indicators about inflation
expectations without the need for data on inflation-linked securi-
ties. This approach will provide two main advantages: First, it uses
data only on conventional nominal bonds and realized inflation;
second, it makes it possible to identify the risk premia component,
obtaining a more accurate portrait of pure inflation expectations.

Table 1
INFLATION LINK ED SECURITIES

Inflation linked bonds Brazil, Chile, Mexico, Peru, Argentina,

Colombia, Bolivia, Costa Rica, Uruguay
Inflation swaps Brazil
Inflation options –

tions indicator. The indicator is also seasonal, in a way that is linked to

the behavior of inflation. To correct for these distortions, models or
adjustments are often used that are subject to potential estimation errors.

28 A. Fuertes, R. Gimeno, J. M. Marqués

3. MODELING INTEREST RATES FROM PUBLIC
DEBT MARKETS

The methodology we implement decomposes nominal interest rates

into three components from an affine model of the nominal term
structure. This methodology is related to the macro-finance liter-
ature in which authors such as Diebold et al. (2006), Diebold et al.
(2005), Carriero et al. (2006), and Ang et al. (2008) (abw) incorpo-
rate macro-determinants into a multi-factor yield curve model with
non-arbitrage opportunities. Our decomposition departs from pre-
vious approaches by extracting the risk premia from the difference
between the nominal term structure and a notional term structure
where the price of risk is set equal to zero.
We also propose an affine model where interest rates are affine
relative to a vector of factors that includes inflation rates and exog-
enously determined factors based on the Nelson-Siegel exponential
components of the yield curve (Nelson and Siegel, 1987), in a similar
vein to Carriero et al. (2006) and Diebold and Li (2006). Moreover,
in our case, we include the condition of non-arbitrage opportuni-
ties along the yield curve and take into account risk-aversion. Tak-
ing these two conditions together allows us to decompose nominal
interest rates as the sum of real risk-free interest rates, expected in-
flation, and risk premium.

3.1. The Model

Affine term structure models allow the risk premium to be separat-
ed from expectations about future interest rates. An affine model
assumes that interest rates can be explained as a linear function of
certain factors,

−1
yt ,t +k =
k
( )
Ak + BkʹX t + ut ,t +k ut � N (0,σ 2 1),

where yt ,t +k is the nominal interest rate in period t with term k, X t is

a vector of factors, Ak and Bk' are coefficients, and ut ,t +k represents
the measurement error. We also assume that X t factors follow a var
structure (in the same vein as Diebold et al., 2006):

Extraction of Inflation Expectations from Financial Instruments 29

 X t = µ + ΦX t −1 + Σεt εt � N (0, I ),

where µ  is a vector of the constant drifts in the affine variables X t ,Σ

is the variance-covariance matrix of the noise term and Φ is a matrix
of the autoregressive coefficients. To avoid arbitrage opportunities,
the values of parameters Ak and Bk' should be restricted according
to the following equation:

The consideration of risk-aversion in this framework implies some

compensation for the uncertainty of longer maturities, in which the
random shocks ε t accumulate. Coefficients that translate matrix Σ
into the risk premium are called prices of risk ( λt ) and, following
the literature, these coefficients are affine to the same factors X t ,

λt = λ0 + λ1X t ,

where λ0 is a vector, and λ1 a matrix of coefficients. If λ1 is set to be

equal to zero, then the risk premium will be constant, whereas if it is
left unrestricted, we will obtain a time-varying risk premium.
We must consider the variables that could determine the term
structure of interest rates in order to select the factors in the model.
There is ample evidence in the literature that the information con-
tent of the whole term structure could be shortened to a small num-
ber of factors. The proposal of Diebold and Li (2006) is used, with
the level ( Lt ), slope ( St ) and curvature (C t ) parameters from the
Nelson and Siegel (1987) term structure specification as factors of
an affine model. These factors can be found in most central bank
estimations of the zero-coupon yield curve. This estimation implies
that nominal interest rates can be modeled in the following equation,

1 − e −k /t  1 − e −k /τ 
yt ,t +k = Lt + St + Ct  − e −k /τ  + ut ,t +k ,
k /τ  k / τ 

30 A. Fuertes, R. Gimeno, J. M. Marqués

where τ , Lt , St , and C t are the parameters that give us the interest
rate at time t  with maturity in k  periods.
Although including a fourth factor in the model may not be nec-
essary to obtain a good fitting of the interest rate term structure, if
Nelson and Siegel’s model is considered, adding the inflation rates
allows us to take into account the yield curve information that could
be useful in forecasting inflation.

Lt 
S 
Xt =  
t

C t 
 
π t 

Once the affine model, represented by the previous equations,

has been estimated, it is possible to decompose k-period nominal
interest rates (yt ,t +k ) into real risk-free rates (Ert ,t +k ) , inflation expec-
tations (Et π t ,t +k ) and risk premia (denoted by γ t ,t +k ), according to
the following equation:

yt ,t +k = Ert ,t +k + Et [π t ,t +k ]+ γ t ,t +k .

Therefore, real risk-free rates (Ert ,t +k ) could be obtained by sub-

tracting inf lation expectations and risk premia from estimated
nominal interest rates.

4. RESULTS OF INFLATION EXPECTATIONS FROM

PUBLIC DEBT MARKETS

4.1 Yield Curve Estimation

To estimate the affine model proposed, we use monthly spot nomi-
nal interest rates for the Brazilian, Colombian, Chilean and Mexi-
can government yield curve. These data have been obtained from
a yield curve estimation that follows Diebold and Li (2006). We first
analyze the yield curve estimates using both nominal interest rates,
and inflation-indexed rates when available, to check the goodness
of fit. For the sake of comparison, Figure 1 shows the yield curve

Extraction of Inflation Expectations from Financial Instruments 31

estimates both for Mexican and Italian government bonds. The
black (gray) line represents yield curve estimates for nominal gov-
ernment bonds (inflation-indexed government bonds). The dots
represent the yield and maturity of traded bonds. Nominal yield
curve estimates provide accurate estimates for both countries while
inflation-indexed yield curve estimates only provide a good fit for
Italy. Lack of inflation-indexed bonds for different maturities, low
liquidity and low market depth make these yield curve estimates for
Mexico unreliable. We find similar problems using inflation-linked
bonds for Brazil, Chile, and Colombia. On the contrary, nominal
yield curve estimates provide a reasonable fit for all these markets,
and they will be the input to solve the affine model and obtain infla-
tion expectations for the countries we analyze. We do also estimate
the yield curve for the inflation-linked bonds in Chile. The Chilean
market is one of the most active in Latin America, and we can com-
pute the break-even rate as the difference between the estimated
yield curves from nominal bonds and inflation-linked bonds. Fig-
ure 2 shows the one-year break-even rate for Chile obtained from the
estimated yield curves. The break-even rate seems to be affected by
the liquidity premia in the inflation-linked bond market as the rate
decreases during the period when inflation rises. 5

Figure 1
YIELD CURVE ESTIMATES
NOMINAL (BLACK) VS. INFLATION LINKED BONDS (GRAY)
 (, 2007)  (, 2016)
4 8
3 7
2 6
5
1
4
0
3
−1 2
−2 1
−3 0
0 5 10 15 20 25 0 10 20 30 40

5
The break-even rate includes the spread between the liquidity premium
of the nominal and the inflation-linked bond markets. Because of that,
it decreases if the liquidity premium in the inflation-linked bond market
rises more than the premium of the nominal bond market.

32 A. Fuertes, R. Gimeno, J. M. Marqués

 Figure 2
ONE YEAR BREAK EVEN RATE FROM YIELD CURVE ESTIMATES
VS. CURRENT INFLACION FOR CHILE

Percent
6

1
0
Jun 2012

Oct 2012

Feb 2013

Jun 2013

Oct 2013

Feb 2014

Jun 2014

Oct 2014

Feb 2015

Jun 2015

Oct 2015

Feb 2016

Jun 2016

Oct 2016

Feb 2017

Jun 2017

Oct 2017
Brake-even rate Current inflation

The availability of nominal government bonds for the estimation

of the zero-coupon yield curve is different for each country, both re-
garding the number of nominal bonds used and the length of the
sample. Table 2 summarizes this information for each market.

Table 2
NOMINAL BONDS AVAILABILITY
Original bond
Number of bonds Period maturity

Brazil 104 Since Feb 2007 3 months – 11 years

Chile 15 Since July 2012 4 years – 30 years

Colombia 70 Since Feb 2005 1 year – 20 years

Mexico 47 Since May 2001 3 years – 30 years

Extraction of Inflation Expectations from Financial Instruments 33

4.2. Empirical Results
We mainly focus on the results related to inflation expectations, leav-
ing aside a deeper interpretation of the term premia and the real
yield curve. We obtain inflation expectations from the var equation.
Since vector X t includes current inflation (π t ) , expectations on this
variable can be computed from projections of the dynamics of the
affine factors in the var equation.

Et [X t +h ] = (1 + Φ + Φ 2 +  + Φh −1 )µ + Φh X t .

There are several advantages in using this method to obtain in-

flation expectations. First, there is a large degree of flexibility, as
we can estimate expectations at different horizons. Moreover, we
can also compute forward rates, allowing us to estimate, for exam-
ple, the expected inflation over the five-year period that begins five
years from today. This is a measure commonly used by central banks
to analyze the anchoring of inflation expectations in the long-run.
It is difficult to obtain these estimates in markets without inflation-
linked securities and, to the best of our knowledge, this is the first
time that these kinds of estimates are computed for Brazilian, Co-
lombian, Chilean and Mexican markets. Also, as we pointed out in
the introduction, using existing surveys on inflation expectations
provides a limited picture, as the horizons are usually short and the
frequency of publication is only monthly at best. Later we describe
the characteristics of the surveys published by the central banks of
the countries we analyze and compare the expectations obtained
from these surveys with those we obtain.
Figure 3 shows the estimates of the nominal yield and inflation
expectations over the ten-year horizon obtained from our proposed
model. The difference between the two curves represents the real
risk-free rate and the risk premium. For the sake of comparison, we
restrict the sample period to be the same for the four countries. The
results show two main features. First, inflation expectations seem
to be more anchored both in Chile and Mexico, showing less vola-
tility. Second, the level of inflation expectations is higher in Brazil,
with the other three countries showing expected rates close to or
below 4 percent.

34 A. Fuertes, R. Gimeno, J. M. Marqués

 0
2
4
6
8
10
12
14
16
18
0
2
4
6
8
10
12
14
16
18
0
2
4
6
8
10
12
14
16
18
0
2
4
6
8
10
12
14
16
18

Jul 2012 Jul 2012 Jul 2012 Jul 2012

Jan 2013 Jan 2013 Jan 2013 Jan 2013

Jul 2013 Jul 2013 Jul 2013 Jul 2013

Jan 2014 Jan 2014 Jan 2014 Jan 2014

10 Year Nominal Yield

Jul 2014 Jul 2014 Jul 2014 Jul 2014

Jan 2015 Jan 2015 Jan 2015 Jan 2015






Figure 3

Jul 2015 Jul 2015 Jul 2015 Jul 2015

Jan 2016 Jan 2016 Jan 2016 Jan 2016

Jul 2016 Jul 2016 Jul 2016 Jul 2016

Jan 2017 Jan 2017 Jan 2017 Jan 2017

Jul 2017 Jul 2017 Jul 2017 Jul 2017

Extraction of Inflation Expectations from Financial Instruments

10 Year Inflation Expectations
10 YEAR NOMINAL BOND YIELD AND INFLATION EXPECTATIONS

35
 As we previously mentioned, the model we propose allows us to com-
pute inflation expectations at different horizons. Figure 4 shows infla-
tion expectations for the one-year, five-year and ten-year horizons, as
well as the inflation targeting level established by the central bank in
each country. We can see again the different degree of anchoring by
comparing the evolution of expectations for the one-year horizon with
those for the five-years and ten-year horizons. Inflation expectations
in Brazil and Colombia show a similar pattern for all horizons while
expectations in Chile and Mexico are more volatile over the one year
horizon, showing little changes over longer horizons.
Regarding the inflation targeting levels established by the central
banks, most countries currently show inflation expectations at long
horizons within the window limits, 6 although Brazil and Colombia
have experienced recent periods where inflation expectations were
well above these limits. Both countries showed inflation expectations
above 6% before the large decreased experienced since the beginning
of 2016. On the other hand, Mexico shows long-term inflation expec-
tations slightly above the upper band of 4%, mainly due to the recent
increase in expectations after the last us presidential elections. This
effect is more apparent for the evolution of the one-year horizon, fad-
ing out at longer terms. Interestingly, it seems that the results of these
elections have barely affected inflation expectations in other countries.
For Brazil, the deep recession of 2015-2016 has affected expectations,
with a large decrease experienced since the beginning of 2016. The
path of inflation expectations changed again for Brazil at the end of
2016, with expectations turning higher at longer horizons, which sig-
nals a possible recovery. In the case of Colombia, the monetary policy
implemented by the central bank during 2016, with increases in the
policy rate from 4.5% in September 2015 to 7.75% in August 2016, have
contained inflation expectations, being now closer to the inflation tar-
get. Longer-term inflation expectations continue to show lower levels
than short-term ones for this country. Finally, Chile has experienced a
decreasing trend in short-term expectations since mid-2014 which has
been associated, first to the fall in oil prices, and since 2016 to the ap-
preciation of the Chilean peso. Although short-term inflation expecta-
tions remain below the inflation target, expected inflation at long-term
horizons is higher and have experienced little change.

6
The Bank of Brazil sets the inflation target at 4.5% with a window limit of
±1.5%. The central banks of Chile, Colombia and Mexico set the inflation
target at 3% with a window limit of ±1 percent.

36 A. Fuertes, R. Gimeno, J. M. Marqués

 1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10

Jul 2012 Jul 2012 Jul 2012 Jul 2012

Jan 2013 Jan 2013 Jan 2013 Jan 2013

1 year
Jul 2013 Jul 2013 Jul 2013 Jul 2013

Jan 2014 Jan 2014 Jan 2014 Jan 2014

5 year
Jul 2014 Jul 2014 Jul 2014 Jul 2014

Jan 2015 Jan 2015 Jan 2015 Jan 2015






Figure 4

Jul 2015 Jul 2015 Jul 2015 Jul 2015

10 year
Jan 2016 Jan 2016 Jan 2016 Jan 2016

Jul 2016 Jul 2016 Jul 2016 Jul 2016

Jan 2017 Jan 2017 Jan 2017 Jan 2017

Jul 2017 Jul 2017 Jul 2017 Jul 2017

INFLATION EXPECTATIONS AT DIFFERENT HORIZONS

Inflation target

Extraction of Inflation Expectations from Financial Instruments

37
 Figure 4 also provides information about the term structure of
inflation expectations. Expected inflation in Colombia and Mexico
is decreasing with the horizon, while in Brazil and Chile inflation
is expected to increase in the future. Figure 5 shows the term struc-
ture of inflation expectations at three different dates for all the ho-
rizons we compute, giving an idea about how inflation expectations
should evolve and how the term structure has changed since August
2016. The evolution of the term structure differs among the four
countries. For Chile, expectations from the two-year horizon have
barely changed at the three dates, experiencing a decrease over time
for short-term expectations. For Brazil, there is an overall decrease
at all horizons since August 2016, although the shape of the term
structure has changed. At the end of August 2016, the term struc-
ture showed a decreasing trend that has currently change into an
increasing one. For Mexico, the situation is the opposite, with infla-
tion expectations increasing at all horizons since August 2016, and
turning from an increasing trend to a decreasing one. The develop-
ments in the us have influenced these changes in Mexican inflation
expectations after the last presidential elections. Finally, Colombia
shows a decrease in the level of inflation expectations at all horizons,
with a decreasing trend over time at the three dates.
Being able to decompose the yield curve and extracting inflation
expectations at different horizons let us compute forward rates as
well. This is especially useful in order to analyze the anchoring of
inflation expectations over the medium and long-term. Forward
rates such as the 5Y5Y (expected inflation over the five-year period
that begins five years from today) are used by central banks to assess
the level of long-term inflation anchoring. Figure 6 shows the 2Y2Y
and 5Y5Y forward rates of inflation expectations together with the
inflation target established by each central bank. Similarly, to the
behavior of the ten-year horizon inflation expectations, the forward
rates for Chile and Mexico are more stable and hardly move over
time. The levels are above the inflation target but within the window
of ±1% for Chile and almost within that window for Mexico. These
results show that investors have almost kept unchanged the level of
long-term expected inflation for these two countries.
On the contrary, inflation anchoring for Brazil and Colombia
seems to be lower, with forward rates showing more volatility. In Bra-
zil, long-term inflation expectations are above the target level but be-
low the upper limit of ±1.5%, due to the large decrease experienced

38 A. Fuertes, R. Gimeno, J. M. Marqués

 Figure 5
TERM STRUCTURE OF INFLATION EXPECTATIONS


9
8
7
6
5
4
3
2
0 1 2 3 4 5 6 7 8 9 10
Years

9
8
7
6
5
4
3
2
0 1 2 3 4 5 6 7 8 9 10
Years

9
8
7
6
5
4
3
2
0 1 2 3 4 5 6 7 8 9 10
Years

9
8
7
6
5
4
3
2
0 1 2 3 4 5 6 7 8 9 10
Years

Aug. 31, 2016 Jan. 31, 2017 Aug. 31, 2017

Extraction of Inflation Expectations from Financial Instruments 39

since the beginning of 2016. For Colombia, there is a similar pattern,
with long-term inflation expectations currently below the target level
of 3% after the decrease in the 5Y5Y forward rate experienced since
mid-2016. The behavior of forward rates for Brazil and Colombia show
that investors seem to face more uncertainty about the expected in-
flation in the long-term for these two countries. It could be also the
case the government bond markets provide less information about
future inflation for these two countries.
These results may question the effectiveness of monetary policy
to anchor expected inflation. The results shown in Figure 5 indicate
that the central banks of Chile and Mexico have been able to anchor
long-term inflation expectations, although at levels above target,
while central bank in Brazil and Colombia face more challenges to
do so. Dincer and Eichengreen (2014) compute measures of central
bank transparency and independence for a large set of countries.
Regarding central bank transparency, among the four countries we
analyze, the central banks of Brazil and Chile were the most trans-
parent in 2010, the central bank of Colombia was less transparent
and the central bank of Mexico was the least transparent.
Their measure of central bank transparency does not seem to be
related to the level of expected inflation anchoring we observe from
our results. On the contrary, central bank independence may play
a role. According to their measure of central bank independence,
Chile and Mexico’s central banks are more independent than the
central bank of Colombia (unfortunately, they do not provide a mea-
sure of central bank independence for Brazil). In line with this result,
Gutiérrez (2003) and Jácome and Vázquez (2008) find a relationship
between central bank independence and inflation performance for
Latin American countries.7
The purpose of our analysis is to identify the inflation expecta-
tions implicit on financial markets, something that would not neces-
sarily be the best forecast for future inflation. However, we analyze
the forecast capacity of this methodology in order to compare it with
other alternatives frequently used by professional forecaster of infla-
tion trends. In this vein, we compare the information about expected

7
Gutiérrez (2003) provides the values of the central bank independence
indexes for the four countries in our study. Although we should be
careful as the indexes were calculated long time ago, Mexico and Chile
show the largest values of central bank independence.

40 A. Fuertes, R. Gimeno, J. M. Marqués

 1
2
3
4
5
6
7
8
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
2
3
4
5
6
7
8
9
Jul 2012 Jul 2012 Jul 2012 Jul 2012

Jan 2013 Jan 2013 Jan 2013 Jan 2013

Jul 2013 Jul 2013 Jul 2013 Jul 2013

2Y2Y
Jan 2014 Jan 2014 Jan 2014 Jan 2014

Jul 2014 Jul 2014 Jul 2014 Jul 2014

5Y5Y
Jan 2015 Jan 2015 Jan 2015 Jan 2015






Figure 6

Jul 2015 Jul 2015 Jul 2015 Jul 2015

Jan 2016 Jan 2016 Jan 2016 Jan 2016

Jul 2016 Jul 2016 Jul 2016 Jul 2016

Jan 2017 Jan 2017 Jan 2017 Jan 2017

Inflation target
INFLATION EXPECTATIONS OF FORWARD RATES

Jul 2017 Jul 2017 Jul 2017 Jul 2017

Extraction of Inflation Expectations from Financial Instruments

41
inflation obtained from our model with that provided by surveys.
First, as we obtain expectations from nominal government bonds,
expected inflation is derived from investor’s perceptions, comple-
menting the information from surveys which is usually obtained
from the views of economists and forecasters. Second, we can obtain
inflation expectations at different horizons and forward rates. Sur-
veys usually provide few horizons, with limited information about
long-term inflation expectations. Table 3 summarizes the informa-
tion provided by the surveys published by the central banks in the
four countries analyzed. Even though there is information about
expected inflation at different horizons in the surveys, we cannot
get all the different horizons we can compute using our proposed
methodology. The surveys do not provide forward rates either. We
next compare the forecasting accuracy of the inflation expectations
obtained from our model with those provided by surveys and a sim-
ple autoregressive process ar(1). Figure 7 shows expected inflation
obtained from surveys and our methodology as well as ex-post re-
alized inflation for the 12-months horizon. 8 Inflation expectations
obtained from surveys tend to be broadly stable over time and show
little changes and reaction.
On the other hand, inflation expectations obtained from our
model seem to be too reactive and more dependent on current infla-
tion. Expected inflation from surveys fail to react to inflation shocks
while our measures produce expectations that respond too late to
inflation shocks. The ar(1) process provides similar inflation expec-
tations to those obtained from our model although these expected
values seem smoother. The difference between the inflation expec-
tations obtained from the model and the ar(1) represents the addi-
tional information about future inflation once that we consider the
inflation expectations embedded on bond prices. In order to analyze
the forecast accuracy of the measures, we compute the mean square
error (mse) concerning ex-post realized inflation.

8
In the case of Chile, it is 11-months horizon inflation expectations
(annual change).

42 A. Fuertes, R. Gimeno, J. M. Marqués

 Table 3
SURVEYS ON INFLATION EXPECTATIONS –CENTR AL BANKS

Frequency Horizons

Brazil Daily Next 12 months; current year (t) and t+1, t+2,
t+3, t+4.

Chile Monthly Next 11 months; next 23 months; current year

(t) and t+1, t+2.

Colombia Monthly Next 12 months; next 24 months; current year

(t) and t+1.

Mexico Monthly Next 12 months; next 1-4 years; next 5-8 years.

Table 4 shows the ratio of the mse obtained using expectations

from surveys, as well as from our model and the ar(1) process, to the
mse computed using current inflation as the predicted future value
(like in a unit root process). If the ratio is lower than one, it means
that the expected values provide a better prediction of future in-
flation than assuming inflation will remain the same as today. The
three measures, inflation expectations from surveys, from the ar(1)
and our model show lower mse than the unit root prediction. Com-
paring the three measures, expected inflation from surveys shows
lower mse for Brazil and Colombia. The model is the best predictor
for Chile and the ar(1) process provides the lowest mse for Mexico.
Inflation expectations from our model provide lower mse for Chile
and Mexico than for Brazil and Colombia. It seems that our mea-
sures of expected inflation are more accurate for countries where
expectations are fairly anchored in the long-run. Our measures do
complement those from surveys in terms of predictability, provid-
ing additional forecasting power and a much richer set of expected
inflation horizons, and frequency.

Extraction of Inflation Expectations from Financial Instruments 43

 Figure 7
12-MONTHS INFLATION EXPECTATIONS FROM SURVEY
AND PROPOSED MODEL VS. REALIZED INFLATION


11
10
9
8
7
6
5
4
3
2
1
Jul 2012

Jan 2013

Jul 2013

Jan 2014

Jul 2014

Jan 2015

Jul 2015

Jan 2016

Jul 2016

Jan 2017

Jul 2017
11 
10
9
8
7
6
5
4
3
2
1
Jul 2012

Jan 2013

Jul 2013

Jan 2014

Jul 2014

Jan 2015

Jul 2015

Jan 2016

Jul 2016

Jan 2017

Jul 2017

11 
10
9
8
7
6
5
4
3
2
1
Jul 2012

Jan 2013

Jul 2013

Jan 2014

Jul 2014

Jan 2015

Jul 2015

Jan 2016

Jul 2016

Jan 2017

Jul 2017

11 
10
9
8
7
6
5
4
3
2
1
Jul 2012

Jan 2013

Jul 2013

Jan 2014

Jul 2014

Jan 2015

Jul 2015

Jan 2016

Jul 2016

Jan 2017

Jul 2017

Model Survey AR(1) Realized

Source: DataStream, Banco Central de Chile, Banco de la República - Colombia,

Banco Central do Brasil, Banco de México. Inflation expectations in 12 months for
Brazil, Colombia and Mexico. Inflation expectations in 11 months for Chile.

44 A. Fuertes, R. Gimeno, J. M. Marqués

 Table 4
EXPECTED INFLATION FORECAST ERRORS

Sample Survey1 Model1 AR(1)1

Brazil Feb 2007- 0.5833 0.8812 0.8415
Oct 2016
Chile Jul 2012- 0.7813 0.6946 0.7148
Dec 2016
Colombia Feb 2005- 0.7956 0.9354 0.8015
Nov 2016
Mexico May 2001- 0.6350 0.7078 0.6324
Nov 2016

1
Ratio of mean square error of expected inflation from surveys, an ar(1) process
and our model with respect to a naïve prediction of expected inflation equal
to current inflation. Expected inflation in 12 months for Brazil, Colombia and
Mexico; 11 months for Chile.

5. CONCLUSIONS

Agents’ inflation expectations are decisive when studying changes

in many of the variables shaping households’ and firms’ decision
making. We use a methodology to obtain inflation expectations from
nominal government bonds and realized inflation, overcoming the
problems of obtaining expected inflation using inflation-linked se-
curities. This is especially useful for markets where inflation-linked
securities are scarce and illiquid as it is the case of Latin America.
In this article, we estimate inflation expectations for Brazil, Chile,
Colombia, and Mexico. We find that inflation expectations seem to
be anchored in Chile and Mexico in the long-term (5Y5Y forward
rate), although the level of expected inflation is above the central
bank target rate of 3 percent.
On the other hand, long-term inflation expectations in Brazil and
Colombia are more volatile and have been fluctuating over time, ex-
periencing a large decrease during 2017. These results advise further
efforts from the Brazilian and Colombia central banks to anchor in-
flation expectations to make credible their inflation targets. Mexican
and Chilean central banks should be more concerned in reducing

Extraction of Inflation Expectations from Financial Instruments 45

the level of expected inflation as long-term expectations seem to be
fairly anchored and show low levels of volatility.
We also find the expected inflation is currently increasing with the
horizon in Brazil and Chile, while it is decreasing in Colombia and
Mexico. For Mexico, there has been an important shock on expect-
ed inflation after the last us presidential elections, experiencing a
large increase. None of the other countries analyzed have shown this
pattern, limiting the spillovers effects of the results of the us presi-
dential elections to inflation expectations in Mexico.
Finally, we compare the forecasting power over one year inflation
expectations obtained using our approach with expected inflation
obtained from surveys. Our approach performs better predicting
inflation for Chile, while surveys do better for Brazil, Chile, and Co-
lombia. There is a trade-off in terms of predictability as expected in-
flations from surveys is less responsive to inflation shocks, and our
approach produces inflation expectations that are more correlated
with current inflation.

References
Ang A., G. Bekaert, and M. Wei (2007), “Do Macro Variables, As-
set Markets, or Surveys Forecast Inflation Better?,” Journal of
Monetary Economics, Vol. 54, No. 4, pp. 1163-1212.
Ang A., G. Bekaert, and M. Wei (2008), “The Term Structure of
Real Rates and Expected Inflation,“ Journal of Finance, Vol.
63, pp. 797-849.
Chan, J., G. Koop, and S. Potter (2013), “A New Model of Trend
Inflation,” Journal of Business and Economic Statistics, Vol. 31,
No. 1, pp. 94-106.
Chernov, M., and P. Mueller (2012), “The Term Structure of Infla-
tion Expectations,” Journal of Financial Economics, Vol. 106,
pp. 367-394.
Carriero, A., C. A. Favero, and I. Kaminska (2006), “Financial
Factors, Macroeconomic Information and the Expectations
Theory of the Term Structure of Interest Rates,” Journal of
Econometrics, Vol. 131, pp. 339-358.

46 A. Fuertes, R. Gimeno, J. M. Marqués

Diebold, F.X., and Li, C., (2006), “Forecasting the Term Structure
of Government Bond Yields,” Journal of Econometrics, Vol.
130, pp. 337-364.
Diebold F.X., G. D. Rudebusch, and S. B. Arouba (2006), “The
Macroeconomy and the Yield Curve: A Dynamic Latent Fac-
tor Approach,” Journal of Econometrics, Vol. 131, pp. 309-338.
Diebold, F.X., M. Piazzesi, and G. D. Rudebusch (2005), “Model-
ing Bond Yields in Finance and Macroeconomics,” American
Economic Review, Vol. 95, pp. 415-420.
Dincer, N. N., and B. Eichengreen (2014), “Central Bank Transpar-
ency and Independence: Updates and New Measures,” Inter-
national Journal of Central Banking, Vol. 10, No. 1, pp 189-253.
D’Amico, S., D. H. Kim, and M. Wei (2014), Tips from tips: the In-
formational Content of Treasury Inflation-protected Security Prices,
Finance and Economics Discussion Series, Federal Reserve
Board, No. 2014-024, Washington, D.C.
Fleming, M.J., and J. R. Sporn (2013), “Trading Activity and Price
Transparency in the Inflation Swap Market,” frbny Economic
Policy Review, Vol. 19, No. 1, pp 45-57.
Gimeno, R., and A. Ibáñez (2017), The Eurozone (Expected) Infla-
tion: An Option’s Eyes View, Banco de España Working Paper,
No. 1722.
Gimeno, R., and J.M. Marqués (2012), “A Market-based Approach to
Inflation Expectations, Risk Premia, and Real Interest Rates,”
The Spanish Review of Financial Economics, No. 10, pp. 18-29.
Gimeno, R., and E. Ortega (2016), The Evolution of Inflation Ex-
pectations in the Euro Area. Banco de España Working Paper,
No. 1627.
Gutiérrez, E., (2003), Inflation Performance and Constitutional Central
Bank Independence: Evidence from Latin America and the Carib-
bean, IMF Working Paper, No. 03/53.
Jácome, L. I., and F. Vázquez (2008), European Journal of Political
Economy, Vol. 24, pp. 788-801.
Nelson, C., and A. Siegel (1987), “Parsimonious Modeling of Yield
Curves,” Journal of Business, Vol. 60, pp. 473-489.
Smith, T. (2012), “Option-implied Probability Distributions for
Future Inflation,” Bank of England Quarterly Bulletin, Vol. 52,
No. 3, pp. 224-233.

Extraction of Inflation Expectations from Financial Instruments 47

The Information Rigidities and
Rationality of Costa Rican
Inflation Expectations

Alonso Alfaro Ureña

Aarón Mora Meléndez

Abstract
Costa Rican inflation expectations cannot be characterized as rational un-
der any existing definition of the term. They cannot be categorized as adap-
tive either, since in addition to historical data on inflation, other macroeco-
nomic variables are important in explaining inflation expectations. Instead,
the sticky information model is considered a more sophisticated framework
to assess inflation expectations of Costa Rican agents. Results are based
on the Monthly Survey of Inflation and Exchange Rate Expectations elabo-
rated and published by the Banco Central de Costa Rica. This chapter col-
lects evidence to assess whether the expectations from this survey are subject
to information rigidities. Additionally, this chapter shows how a simulated
survey, based on a sticky information model, is capable of replicating features
from the observed survey.
Keywords: inflation expectations, sticky information, adaptive learning.
jel classification: C53, D84, E31, E58.

A. Alfaro Ureña <[email protected]> Department of Economic Research. Economic

Division, bccr. A. Mora Meléndez, bccr. <[email protected]>. The authors would
like to thank Olivier Coibion for his guidance, advice and helpful discussion during
the development of this project. Federico Herrera provided excellent research
assistance. The ideas expressed in this chapter are of the authors and do not necessarily
represent the view of the Banco Central de Costa Rica.

49
1. INTRODUCTION

C
onventional economic theory highlights the crucial influence
of expectations on changes in macroeconomic variables. Chang-
es in a variable affect the expectations related to its future move-
ment and these expectations also influence the variable’s underlying
path. This bilateral relation puts the problem of how agents form their
expectations into the front line of macroeconomic modeling.
Most central banks acknowledge the crucial role of expectations,
and argue that managing inflation expectations is paramount for at-
taining price stability and conducting monetary policy. The Banco
Central de Costa Rica (bccr) operates under an inflation targeting
regime, in order to accomplish its goal of a low and stable inflation lev-
el. It relies heavily on the inflation expectations of Costa Rican agents
aligning closely with monetary policy. It is necessary to understand
how inflation expectations are formed to anchor expectations to the
ones targeted by the bccr.
Until recently the research agenda on expectation formation
was eclipsed by the rational expectations ( re ) hypothesis started
by Muth (1961). This hypothesis revolutionized macroeconomic think-
ing during the seventies by incorporating the effect of expectations
into most economic models. As Thomas Sargent points out 1, the re
hypothesis allowed for the disappearance of any free parameters as-
sociated with expectations, so people’s beliefs became outputs of the
model in question. As a result, macroeconomists widely adopted the as-
sumption of re to arrive at tractable equilibrium solutions.
Nevertheless, a common critique for the re hypothesis is that it as-
sumes that people have much more information about the economy
than they really do, since it implies that agents construct expectations
and make decisions by gathering and conveying all available public in-
formation. This assumption is unrealistic and empirical studies often
reject the re hypothesis. There are three popular alternatives to the
re hypothesis: 1) agents use heterogeneous mechanisms to form their
expectations, as in Branch (2004) and Honkapohja and Mitra (2006); 2)
agents use different information sets, as in Angeletos and Lian (2016);
and 3) agents have different abilities to process information, see for
example Woodford (2001). A good survey of alternative approaches
to the specification of expectations is presented in Woodford (2013)

1
See Evans and Honkapohja (2005).

50 A. Alfaro, A. Mora
where the author presents how macroeconomic analysis under a new
Keynesian framework could be performed without relying on the
re hypothesis. Regardless, there are well developed theoretical alter-
natives to re, though many features observed in expectations survey
are not entirely taken into account by these alternatives. Authors like
Manski (2004) have pushed for more empirical studies that deepen
our knowledge of how people elicit and revise their expectations.
One approach to analyzing expectations formation has focused
on the role of information rigidities and has been supported by em-
pirical evidence, see Mankiw and Reis (2002), Woodford (2001),
and Sims (2003). In particular, Mankiw et al. (2003) depart from tra-
ditional empirical approaches to expectations measurement, which
have traditionally relied on measures of central tendency, such as the
mean or median; instead, they study the heterogeneity of inflation
expectations using statistics of dispersion. The idea is that the dis-
agreement among agents over inflation expectations can be explained
by information stickiness. They use the sticky information model
developed in Mankiw and Reis (2002) to explain the mean and dis-
persion of the United States’ inflation expectations. Under this frame-
work, just a fraction of the agents updates their expectations with
the most recent information available. This fraction is derived from
the bounded rationality associated with the cost of updating expec-
tations. Pfajfar and Santoro (2010) build on this line of work and in-
stead of using measures of central tendency, they perform percentile
analysis to study the heterogeneity, learning, and information sticki-
ness of inflation expectations.
Alfaro and Monge (2013) also document that Costa Rican infla-
tion expectations can neither be characterized as rational nor adap-
tive. If expectations were rational, the realized bias between expected
and realized inflation level could not be predicted: Costa Rican data
fails this test even with relaxed assumptions of rationality. On the
other hand, inflation expectations cannot be categorized as adap-
tive neither, since in addition to historical data on inflation, other
macroeconomic variables hold significant explanatory power for in-
flation expectations.
Alfaro and Monge (2013) note the need to evaluate more sophis-
ticated tools to model Costa Rican inflation expectations. This chap-
ter will evaluate the sticky information model to determine whether
this need is substantial. The main source of data for this research
comes from the Monthly Survey of Inflation and Exchange Rate

The Information Rigidities and Rationality 51

Expectations conducted and published by the bccr. For this chap-
ter, we used 135 months of survey observations from January 2006
to March 2017. We identify individual participants and place them
into four separate groups based on their profession. In the survey,
respondents report their 12-month expected inflation as well as ex-
pected percentage variations (to different time horizons) of the ex-
change rate between the Costa Rican colon and United States dollar.
The remainder of the chapter is organized as follows: Section 2 de-
scribes the Monthly Survey of Inflation and Exchange Rate Expec-
tations, presents its main features, and analyses the disagreement
and the realized bias or forecast error presented in the survey. Sec-
tion 3 presents the sticky information model of Mankiw et al. (2003),
gathers evidence for information rigidities in the expectations of Cos-
ta Rican agents captured in the survey as a whole and within profes-
sional groups, and simulates a sticky information model that is based
on a vector autoregressive model using Costa Rican macroeconomic
data. Finally, Section 4 discusses the findings of the paper, which show
nonconformity of the sticky information approach for the Costa Ri-
can data, as well as the work ahead for modeling Costa Rican infla-
tion expectations.

2. INFLATION EXPECTATIONS SURVEY

The bccr has conducted the Monthly Survey of Inflation and Ex-
change Rate Expectations since 2006. This survey gathers data on ex-
pected inflation for the next 12 months and the expected percentage
variation in the exchange rate between the Costa Rican colon (crc)
and the United States dollar (usd) for the next 3, 6, 12, 24, and 36
months2 . The questionnaire of the survey can be found in Annex A.
Responses to questions on inflation and exchange rate expectations
are point expectations that ask for a numerical expectation along with
the main factors that were considered to form these expectations.
The observation period starts on January 2006 and goes until March
2017, a total of 135 months. The individuals consulted in the survey
are categorized into four different groups depending on their profes-
sional expertise: 1) consulting, 2) stock market analyst, 3) academic,

2
Consultancy of the 24- and 36-month variation in the crc/usd exchange
rate started on December 2016.

52 A. Alfaro, A. Mora
and 4) business sector. The number of respondents to the survey
and its composition have changed during the observation period;
there were 27 respondents in January 2006, most of whom were stock
market analysts and by March 2017, there were 61 respondents pre-
dominantly from the business sector. Figure 1 presents the composi-
tion of the sample group during the observation period.
Two features of the survey responses stand out: first, the total num-
ber of responses has increased more than twofold since the survey
was first implemented, with a peak of 87 responses in June 20133 .
Second, the composition of responses has drastically changed in the
last years of the survey–the majority of responses have recently come
from individuals working in the business sector–. This compositional
shift has resulted from a change in the survey design from June 2012
to the present4 .
The bccr computes the 12-month expected inflation by averag-
ing the responses received during a particular month, expectations
coming from the business sector are dominant in the expectations
published, representing up to 80% of the responses since 2015. This
dominance of the business sector in the average expected inflation
can be observed in Figure 2 where the mean expectation is plotted
for the whole sample and by group.
The average expectation has clearly declined, staying in the single
digits since April 2009, and below 5% since April 2015. The behavior
exhibited by the inflation expectations has been in accordance with
the inflation target range of the bccr (3%-5%) since April 2015. In Jan-
uary 2016, even though the inflation target range was downgraded
to 2%-4%, expectations have continued to remain within the range
up until the last month in our sample, March 2017.
The alignment between the expected inflation rate and the tar-
get inflation range in recent years highlights the built-up credibility
of bccr towards society. For the thirty-year period preceding 2009,
Costa Rica experienced double-digit inflation rates, but the bccr
has seemingly regained credibility. Agents trust the bccr to steer
the inflation rate, which thereby anchors inflation expectations. De-
spite this tendency for inflation expectations to lie within the target

3
With 64 of them from the business sector.
4
The two samples were active for several months, but the aggregate
results did not differ.

The Information Rigidities and Rationality 53

 Figure 1
INFLATION EXPECTATIONS SURVEY: RESPONSES

60

40
Responses

20

0
1 4 7 10 1 4 7 10 1 4 7 10 1 4 7 10 1 4 7 10 1 4 7 10
2006 2007 2008 2009 2010 2011
Months

100

80
Responses

60

40

20

0
1 4 7 10 1 4 7 10 1 4 7 10 1 4 7 10 1 4 7 10 1
2012 2013 2014 2015 2016 2017
Months

Consulting Stock market Academic Business

Source: Own elaboration.

54 A. Alfaro, A. Mora
 Figure 2
EXPECTED INFLATION MONTHLY AVERAGE

16
14
12-months expected inflation

12
Monthly average

10
8
6
4
2

0
2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017
Survey date

Expectation Consulting Stock market Academic Business

Source: Own elaboration.

range, disagreement about inflation expectations is present in the

survey, not only between groups but also within groups5 .

2.1 Disagreement Among Expectations

Each individual in the survey sample has an identifier code and ev-
ery month that an individual responds, the observations collected
are registered with the relevant identifier (id). This way the survey
data can track respondent observations throughout the entire survey
period, allowing for comparisons in the responses over time among
individuals of the same group and within the full sample. In the survey
there are 409 identifiers that correspond to at most 409 individuals6
that respond the survey at some point during the observation period.

5
Figure 9 in the Annex, shows the increase of outliers on the expecta-
tions from the business sector in recent years.
6
Since the change in the design of the survey sample involved different
nomenclature for the identifiers, the same individual can have two
identifiers, one under the former sample and another one with the
current sample.

The Information Rigidities and Rationality 55

 Table 1
DISTRIBUTION ON THE NUMBER OF RESPONSES

Identifiers (ids) with equal or more responses

Responses (≥) Number of ids Percentage of ids
1 409 100.00
10 206 50.37
20 136 33.25
30 41 10.02
40 33 8.07
50 28 6.85
60 17 4.16
70 10 2.44
80 9 2.20
90 3 0.73

Source: Own elaboration.

The number of responses from a particular identifier range from

1 to 98, with an average of 16.46 during the 135-month observation
period. The observed distribution on the number of responses by id
is shown on Table 1. Decomposing this distribution into the four
aforementioned professional groups, we observe that the academic
and consulting groups have the highest response rates. Even though
the firm group dominates the survey responses, most of the firms’
identifiers have less than 48 responses.
Given the number of individuals participating in the survey, their
professional expertise, and background, disagreement among the in-
flation expectations can be observed on the survey. Mankiw et al.
(2003) are primarily concerned with this disagreement, which is typi-
cal in most expectations surveys and they posit that this heterogeneity
can be explained by bounded rationality, meaning that only a fraction
of the agents adjusts their expectations as new information becomes
available due to the cost associated with the adjustment.
In this context, dispersion statistics like the interquartile range
can be used to discriminate between different models of expectations

56 A. Alfaro, A. Mora
formation by pinning down their faculty to replicate features observed
on the data. Figure 3 presents the interquartile range observed every
month by group, along with the realized inflation rate for the month
that these expectations were registered. This is done to assess whether
the dispersion tends to increase when inflation is high, as has been
suggested by Ballantyne et al. (2016) and Johannsen (2014), among
others.
For the stock market analyst and academic groups, the interquar-
tile range and inflation rate attain their maximum in the last months
of 2008. For these two groups, it may seem to be a positive correlation
between the level of inflation and interquartile range during years
near the 2008-2009 financial crisis. Nonetheless, there are periods
in which the inflation rate decreases but the dispersion of the sample
expectations does not follow the same trend; the clearest example
is the dispersion within business sector responses since 2015 the in-
terquartile range has moved around 2% despite the sharp decline
in inflation. This suggests that for the Costa Rican case there is no
clear direct relation between the dispersion in inflation expectations
and the level of inflation.
A basic regression exercise between dispersion as measured by the
interquartile range and the inflation level is shown in Table 2. Regress-
ing the interquartile range by the inflation rate does not illustrate
a significant relation between the two groups: the associated coeffi-
cients are not significant when taking into account the whole survey
or individual groups.
Elliott et al. (2008) and Engelberg et al. (2009) note that disagree-
ment among inflation expectations does not necessarily indicate
that agents face different degrees of uncertainty when forming their
expectations. This is because the survey collects point predictions
from which individual distributions or probabilistic beliefs of pos-
sible outcomes for future inflation cannot be inferred. It is possible
that two forecasters who hold identical probabilistic beliefs provide
different point predictions and it is also possible that two forecasters
with different probabilistic beliefs provide the same point forecast.
When using point forecasts, we can only interpret the phrase dis-
agreement among expectations  as an acknowledgment of distinct point
forecasts; we cannot conclude anything about the uncertainty that
forecasters face.

The Information Rigidities and Rationality 57

 Figure 3
INTERQUARTILE RANGE BY GROUP


4 20

15
Interquartile range

10
2
5

0 −5
2006

2006

2007

2008

2009

2009

2010

2011

2012

2012

2013

2014

2015

2015

2016
Inflation (right axis) Consulting


6 20

15
Interquartile range

4
10

5
2
0

0 −5
2006

2006

2007

2008

2009

2009

2010

2011

2012

2012

2013

2014

2015

2015

2016

Inflation (right axis) Academic

58 A. Alfaro, A. Mora
 Figure 3 (cont.)
INTERQUARTILE RANGE BY GROUP

 
4 20

15
Interquartile range

10
2
5

0 −5
2006

2006

2007

2008

2009

2010

2011

2011

2012

2013

2014

2015

2015

2016
Inflation (right axis) Stock market


4 20

15
Interquartile range

10
2
5

0 −5
2006

2006

2007

2008

2009

2010

2011

2011

2012

2013

2014

2015

2015

2016

Inflation (right axis) Business

Source: Own elaboration.

The Information Rigidities and Rationality 59

 Table 2
REGRESSION: INTERQUARTILE RANGE AND INFLATION

Whole Stock
Coefficient survey Consulting market Academic Business
Constant 1.591c 1.209c 1.135c 1.254c 1.446c
(0.090) (0.110) (0.090) (0.138) (0.134)
Inflation −0.013 −0.007 0.003 −0.002 −0.011
(0.012) (0.015) (0.012) (0.019) (0.018)
N 135 135 135 135 135
R2 0.0087 0.0014 0.0006 0.0001 0.0029

Note: a significance level 0.1, b 0.05, c 0.01.

Source: Own elaboration.

2.2 Realized Bias

We can also perform a second descriptive analysis of the survey infla-
tion expectations focused on how well agents forecast the inflation
level. If agents can successfully predict the path of future inflation,
then the realized bias, that is the difference between the (forecasted)
expected inflation level for time t  and the realized inflation at time
t, should be close to zero.
As a result of the survey design, when 12-month expected infla-
tion is recorded at time t, its predictive power should be compared
with the realized inflation level of time t  +11, that is eleven months
later from when the observation was collected. This is because even
though agents form their expectations for each annual period, they
are consulted during the first month of the forecast period. This does
not present an issue since agents do not know the realized inflation
of the month that is consulted7. For instance, the expected inflation
of January 2006 should be compared with the inflation rate of Decem-
ber 2006 to compute the realized bias of December 2006.
With this adjustment only 124 months from January 2006 to April
2016 are used to analyze realized bias rather than all 135 months

7
The Instituto Nacional de Estadísticas y Censos (inec) of Costa Rica
publishes the inflation rate of month t until the first days of month t +1.

60 A. Alfaro, A. Mora
of the survey. The last eleven months do not yet have a realized infla-
tion level to compare to, since the last observed inflation in this paper
is March 2017. Panel A of Figure 4 compares the expected and real-
ized inflation rates, while panel B. shows the average realized bias.
Our measure of realized bias has exhibited cyclical behavior, reach-
ing its minimum at the end of 2008 and its maximum at the end of
2009. While there are months where the realized bias has been prac-
tically zero, suggesting good predictive power, it has been positive
since 2005, meaning that on average, inflation expectations have
been greater than realized inflation.
The average realized bias seems to have a general upward trend
across the entire observation period, standing above 5% during
most of 2015 and part of 2016, but decreasing since the second se-
mester of 2016. The average realized bias does not differ substantial-
ly by group–Figure 5 shows the average realized bias for each group
and also for the entire survey sample–.
As expected, the business sector has dominated recent survey re-
sults–the average bias of the business sector has largely aligned with
the average of the entire survey sample–. In addition, the average bias
has increased over the years for all four groups. Figure 5 suggests that
the differences among groups are not significant, but this can be ex-
plained as a result of using measures of central tendency such as the
average. On the other hand, valuable information can be extracted
by studying disagreement among inflation expectations via statis-
tics of dispersion. The next section explores the role of information
rigidities in explaining the heterogeneity in inflation expectations.

3. STICKY INFORMATION MODEL

Mankiw and Reis (2002) propose a model where information rigidi-

ties play a central role in the price and inflation dynamics. In their
model, only a fraction λ of agents gather, process, and optimize their
expectations with the most recent economic information available.
The parameter λ, which is exogenous to the model, can be interpret-
ed as the result of the bounded rationality associated with the cost
of adjusting to new information. This model is conceived as an alter-
native to the new Keynesian Phillips curve since it highlights the role
of information rigidities.

The Information Rigidities and Rationality 61

 Realized bias 12-month expected and realized

62
inflation

−10
−5
0
5
10
−4
−2
0
2
4
6
8
10
12
14
16
18
2006 2006
2007
2007 2007
2008
2008

Source: Own elaboration.

A. Alfaro, A. Mora
2009 2008
2009
2009 2009
2010
2010
2010

Inflation
2011
2011
Figure 4

2011 2011
2012
.  

.   

2012 2012
2013

Average realized bias

2013
2013

Expected
2014
EXPECTED AND REALIZED INFLATION

2014
2014 2014
2015
2015
2015
2016
2016
2016 2016
 Realized expectations bias Realized expectations bias Realized expectations bias Realized expectations bias

−10
−5
0
5
10
15
−10
−5
0
5
10
15

−10
−5
0
5
10
15
−10
−5
0
5
10
15
2006 2006 2006 2006
2007 2007 2007
2007
2008 2008 2008

Source: Own elaboration.

2008
2008 2008 2008
2009
2009 2009 2009
2009
2010 2010 2010
2010 2010 2010
2010
2011 2011 2011 2011

Average bias
Average bias
Average bias
Figure 5

Average bias



2012 2012 2012 2012



 
2012 2012 2012 2012
2013 2013 2013
2013
2014 2014 2014

Business
AVERAGE REALIZED BIAS BY GROUP

2014
Consulting
Consulting

Stock market
2014 2014 2014
2015
2015 2015 2015
2015
2016 2016 2016
2016 2016 2016
2016

The Information Rigidities and Rationality

63
 The sticky information Phillips  curve derived in Mankiw and Reis
(2002) concludes that the relevant expectations of the agents are those
made in the past about current conditions. Mankiw et al. (2003) fol-
low this idea and study the disagreement about inflation expecta-
tions by assuming there is information stickiness, meaning that only
a fraction of the agents generates their expectations of future infla-
tion using all available economic information. With this specification,
we can generate cross sectional samples of simulated expectations
for each period, allowing us to study the features of a simulated sur-
vey beyond measures of central tendency.
In this section, we gather evidence of information rigidities pres-
ent in the Monthly Survey on Inflation and Exchange Rate Expecta-
tions at the survey and group level. Moreover, a sticky information
model is simulated, assuming that the process used to generate ex-
pectations is an econometric model and the way that rational agents
form their expectations is through forecasts from this model. In par-
ticular, we use a vector autoregressive model with Costa Rican mac-
roeconomic data to generate 12-month inflation forecasts.

3.1 Evidence for Information Rigidities

Following Coibion and Gorodnichenko (2015), we can exploit
the conclusion from Mankiw and Reis (2002) that states that for an
economic variable x  under a sticky information model, the average
forecast across agents at time t  for time t +h, Ft xt +h , is a weighted aver-
age of the current and past rational expectation forecast such that8:

∞
1 Ft xt +h = (1 − λ )∑ λ j Et − j xt +h .
j =0

Representing rational expectations as Et xt +h = xt +h − vt +h ,t , where

vt +h ,t is the rational expectation error, which is uncorrelated with
information dated t or earlier, we can find a predicted relation be-
tween the ex post mean forecast error and the ex ante mean forecast
revision (see Coibion and Gorodnichenko, 2015, for its derivation):

8
In this equation the probabilities of an update are reparametrized so
only (1 − λ ) percent of the agents update their information sets and
acquire no new information with probability λ .

64 A. Alfaro, A. Mora
 λ
2 xt +h − Ft xt +h = ( Ft xt +h − Ft −1xt +h ) + vt +h,t .
1− λ

The relation in 2 can be applied to the data. Since it requires

the construction of a forecast revision, we will use data on the ex-
pected exchange rate variation instead of inflation expectations; only
a 12-month expected inflation is available. Under a sticky information
framework relation, 2 should be satisfied for the mean of any macro-
economic variable regardless of the frequency of t  and the horizon
h, so gathering evidence of information rigidities using the expected
exchange rate variation should be comprehensive for all expectations
in the survey. Specifically, quarterly data for the expected exchange
rate e  variation for three and six months is used to perform the fol-
lowing regression based on 2:

3 et +1 − Ft et +1 = β ( Ft et +1 − Ft −1et +1 ) + εt .

Estimates for Equation 3 at the survey and the group level are shown
in Table 3. These regressions can be used to assemble evidence for in-
formation rigidities present on the survey. Under a sticky information
model, the β coefficient in Equation 3 should be significant, which
is the case at the survey level. An advantage of the relation between
the ex post forecast error and the ex ante forecast revision on Equa-
tion 3 is that it enables us to map the estimated coefficient βˆ to an
estimate of the information rigidity parameter λ. In our case, this
gives an estimate of λ ( )
 = β 1 + β ≈ 0.1797 1.1797 ≈ 0.15237, which
suggests that 84.76% of the agents update their information sets at a
particular period and that on average an agent updates his or her in-
formation every 1.2 months.
At the group level, the estimates of Equation 3 suggest that the ev-
idence for information rigidities is stronger among some groups
compared to others. The β coefficient for Equation 3 is significant
to various degrees among the groups, with the exception of the ac-
ademic. For consultants and stock market analysts, the coefficient
is significant at a 1% level and only at a 10% level for the businesspeo-
ple. The results imply different estimates for the rate of information
acquisition λ among groups: 82.44% of the consultants, 83.61%,
of the stock market analysts, 91.91% of academics, and 91.07% of the

The Information Rigidities and Rationality 65

 Table 3
REGRESSION: EX POST MEAN FORECAST ON EX ANTE MEAN REVISION

Dependent variable
et +h–Ftet +h
Stock
Survey Consulting market Academic Business
Ftet +1–Ft –1et +1 0.1797c 0.213c 0.196c 0.088 0.098a
(0.056) (0.058) (0.056) (0.056) (0.056)
Observations 240 240 240 240 240
R2 0.041 0.054 0.049 0.010 0.013
Adjusted R2 0.037 0.050 0.045 0.006 0.009
Residual standard
20.290 20.940 19.937 20.367 21.007
error (df = 239)
F statistic (df = 1; 239) 10.192c 13.643c 12.409c 2.492 3.129a

Note: a p < 0.1; b p < 0.05; c p < 0.01.

Source: Own elaboration.

businesspeople update their expectations with the most recent infor-

mation available every period9. These results, however, show a rela-
tively low degree of information rigidity. The evidence indicates that
the sticky information assumption may not be particularly well suit-
ed to account for how the inflation expectations in the Costa Rican
economy are formed. Nevertheless, we will stick to this assumption
to evaluate how closely a model with sticky information can simulate
the data.

3.2 Simulating a Sticky Information model

In this section, we generate a simulated survey using the following al-
gorithm proposed in Mankiw et al. (2003). In this context, an agent’s
rationality is pin-downed so that we can use a vector autoregressive

9
One should keep in mind that the estimate for the academic group
is not significant and for the business group is only significant at the
10% level. The coefficients are essentially unchanged if the model is
estimated using a constant.

66 A. Alfaro, A. Mora
(var) model to generate rational forecasts10. The var model uses Cos-
ta Rican monthly data from January 1996 to March 2017 for inflation
(πt ), interest rate (it ), output gap ( yt ), an inflation index of trade
( ) ( )
partners π t C , oil prices pt oil , and annual exchange variations (et ).
The design of the var model with two lags11 is presented in 4.

4 zt = A1zt −1 + A2 zt −2 + ut ,

with

 πt 
 i 
 t 
 yt 
zt :=  C  .
 πt 
 oil 
 pt 
 et 
 

As usual A1 and A2 are 6×6 matrices of coefficients and ut  stands

for a process with a null expectation and a time invariant positive def-
inite covariance matrix. Data used comes from different sources: 1)
monthly annual inflation (π t ) is measured using the cpi; 2) the inter-
est rate (it ) is the basic passive interest rate (tasa básica pasiva, tbp);
3) the output gap ( yt ) is estimated following Hamilton (2017) using
a series of the monthly index of economic activity (índice mensual
de actividad económica, imae)12; 4) the inflation index of trade partners
( )
π t C is an index of the inflation of countries considered to be trade
partners with Costa Rica (indicador de inflación de socios comerciales)13;
( )
5) oil prices pt oil come from the monthly average of West Texas

10
We attempted unsuccessfully to estimate the degree of information
rigidity directly for inflation forecasts, using instrumental variables
similarly to Coibion and Gorodnichenko (2015).
11
Number of lags suggested by the Hannan-Quinn information criterion.
12
We regress imae series at date t +24 (to include a two-year period) on
the four most recent values as of date t. The residuals from this regres-
sion are set to be the cyclical component of the series.
13
Mainly composed by the inflation of the United States, the euro zone,
China and Central American countries.

The Information Rigidities and Rationality 67

Intermediate (wti) crude prices; and 6) the annual exchange varia-
tions (et ) are relative annual variations on the bccr’s reference bid ex-
change rate between the us dollar and the Costa Rican colon by the
end of the month. The tbp, imae, inflation index of trade partners,
and reference bid exchange rate are computed and published by the
Banco Central de Costa Rica.
The estimation of the var is done on a sample updating basis, mean-
ing that at time t  we estimate the var solely with information avail-
able up to time t −1, denoted by I t −1 = {zt −1 ,zt −2 ,…}, and done for each
month from January 2006 to March 2017. For example, for January
2006 Equation 4 is estimated using information on zt  from January
1996 up to December 2005, meaning that the initial sample size cov-
ers ten years; each subsequent month adds one observation to the
sample size and the var model is reestimated with this updated sam-
ple. Using the estimates at time t, we forecast the 12-month forward
e
inflation rate π t +12 t using the forecast for the next twelve months
form the var updated up to time t −1:

5 π te+12 t := πt −1+12 .

The updating procedure of the parameters of the var is modeled

as if the agents are econometricians who form their expectations
about the future by incorporating new information on the sample
when estimating the var.
With the var predicted values, especially for inflation {πt }, we gen-
erate cross sectional samples of expected inflation to obtain a simu-
lated survey as follows:
1) Given that the Monthly Survey of Inflation and Exchange Rate
Expectations includes data for 135 months, there will be 135
cross sectional samples, one for each t =1, …, 135.

2) The cross-sectional sample size n  is to be of 100 individuals

for all periods, n =100.

3) In the first period each individual enters the simulated survey

with the mean expectation observed from the survey in the
first month.

68 A. Alfaro, A. Mora
 4) For every t = 2, …, 135, and for each individual i = 1, …, n, a Ber-
noulli experiment with probability of success λ will be con-
ducted.
a) If the experiment is a success, individual i  at time t  will re-
port his or her expected 12-month forward inflation rate
π te+12 t using the 12-month forecast from the var model es-
timated with information up to time t −1:

6 π te+12 t := πt +11 .

b) If the experiment is a failure, π te+12 t is set to the previous

known expected value for individual i.

{
5) The previous steps give for each period t  a series π ie,t +12 t }
for i = 1, …, n. For each series the mean and the interquartile
range (iqr) are recorded.

6) The value of λ is selected to minimize the difference between14

the simulated mean expectation and the observed mean ex-
pectation from the survey.
Running the previous algorithm gives the results presented in Fig-
ure 6: panel A, shows the generated average expectation, the observed
average from the survey, and the realized inflation level at the sur-
vey date. We found the value of λ to be 0.17, meaning that only 17%
of the agents in the simulated sample adjust their expectation with
the most recent information, suggesting that an agent updates his or
her information set every 5.9 months on average. The simulated mean
expectations fit relatively well with the observed mean expectation
from the survey, especially at the beginning and the end of the sam-
ple. The correlation between these two series is 91.15%. In the three
months of 2017 included in the survey the observed mean expecta-
tions were 3.60% for January, 3.78% for February and 3.86% for March;
while the simulated mean values are 3.23%, 3.25% and 3.23% respec-
tively, illustrating the simulation’s ability to replicate the real survey.

14
We compute the mean of square differences between the simulated
and observed series.

The Information Rigidities and Rationality 69

 Interquartile range (%) Mean expected inflation (%)

70
3
4
5

0
1
2
6
7
8
0
2
4
6
8
10
12
14
18
2006 2006
2007 2007
2007 2007
2008 2008
2008 2008
2009 2009

Source: Own elaboration.

A. Alfaro, A. Mora
2009 2009
2009 2009
2010 2010
2010 2010
2011 2011
2011 2011

Observed
Observed
2011 2011
2012 2012
Figure 6

2012 2012
2013 2013
.  FL

.   

2013 2013
2014 2014
2014 2014
2014 2014

Simulated
Simulated
2015 2015
STICKY INFORMATION MODEL SIMULATION

2015 2015
2016 2016
2016 2016
2016 2016
 On the other hand, the simulated series for the interquartile range
has a correlation of only 22.55% with the series from the survey. From
panel B of Figure 6 we observe that simulated iqrs are close to the
real iqrs only in the second half of the survey. This is due to a depar-
ture from the original algorithm in Mankiw et al. (2003) where λ
is selected to maximize the correlation between the simulated series
of iqrs and the survey series. Since we are interested in the mean ex-
pectation, our simulation was modified to put more emphasis on rep-
licating the mean expectation.
The evidence of this simulated model also suggests that the sticky
information assumption may not be appropriate. The value of the
parameter λ required to match the dynamic of the mean forecast
implies dynamics of disagreement that vary significantly from those
found in the data.

4. CONCLUSIONS

This chapter builds on existing characterizations of Costa Rican in-

flation expectations by considering information rigidities in the ex-
pectation formation process. Our results are based on the Monthly
Survey of Inflation and Exchange Rate Expectations. We analyze
its panel structure to identify individual respondents and their groups
of professional expertise (consulting, stock market, academic, and busi-
ness). We found a set of stylized facts that describe the survey: 1) re-
sponses are dominated by business sector respondents, implying that
the mean expectations from the survey primarily reflect the mean
expectation of the business sector; 2) since April 2015 the mean ex-
pected inflation rate is within the inflation target range of the bccr,
(currently 2%-4%), suggesting that inflation expectations have been
anchored by the bccr’s credibility and monetary policy; 3) different
groups have differing expectations and feature a positive interquartile
range over time; 4) there is no clear relation between the dispersion
of inflation expectations and the inflation level, neither at the survey
nor group level; 5) on average agents, from the survey have positive
forecast errors or realized bias, meaning that agents tend to expect
greater inflation than in reality.
Because of these stylized facts, and the existing literature on Cos-
ta Rican inflation expectations, we proposed to test for information
rigidities on the expectation formation process. We found some

The Information Rigidities and Rationality 71

evidence suggesting that agents in the survey are subject to informa-
tion stickiness and that only a fraction of agents form their expecta-
tions with the most recent information available. At the group level,
we found that information rigidities are most prominent in the con-
sulting and stock market analyst groups and less prominent in the
academic and the business groups. However, the magnitude of the ri-
gidity is not large enough to support the claim that the sticky informa-
tion model is well suited to account for what we observe in the data.
Additionally, a simulated inflation expectations survey was gener-
ated using a sticky information algorithm and a vector autoregressive
model to pin down the rationality of agents. This survey captured in-
formation on the inflation level, interest rates, output gap, inflation
levels of trade partners, oil prices and annual exchange rate variations.
The simulated survey replicated the mean expected inflation from
the survey fairly well. Nevertheless, the level of stickiness required
to match the data is low, and implies dynamics of disagreement that
vary significantly from those found in the data.
Our findings show nonconformity of the sticky information ap-
proach for survey data along several dimensions, such as the Costa Ri-
can data. We show that there is no correlation found between the level
of inflation and the amount of disagreement among agents, the in-
formation rigidities for forecasts of exchange rates are much lower
than what is needed to account for forecasts of inflation and finally,
the value of needed to match dynamics of mean forecasts of infla-
tion does not yield predictions for dynamics of disagreement that
conform to those of the data.
Further work to deepen our knowledge about the expectation
formation process of Costa Rican agents may consider the litera-
ture on the effects of learning on expectation formation. Moreover,
we could redefine some questions in the survey to assess the proba-
bility beliefs of the respondents instead of point expectations. This
would elicit information about the uncertainty agents’ face when
forming their expectations.

72 A. Alfaro, A. Mora
ANNEX

Annex A. Monthly Inflation and Exchange

Rates Expectations Survey

Banco Central de Costa Rica, Economic Division

Monthly Survey on Inflation and Exchange Rate Expectations
July 2017
We appreciate your responses between July 10 and July 24

Respondent code:

1. What is your expected inflation rate, measure by the consumer

price index, for the period between July 1, 2017 and June 30, 2018
(12 months)?
Answer: (%)

2. Mention, in order of importance, the variables you take into con-

sideration to form your expected inflation for the 12-month period:
i)
ii)
iii)
vi)
v)

3. The reference bid rate calculated by the Banco Central de Costa

Rica for June 30, 2017 was of 567.09 colones for us dollar. What is you
your expected level for the reference bid exchange rate on the fol-
lowing dates?
3.1 On September 30, 2017 (3 months):
3.2 On December 31, 2017 (6 months):
3.3 On June 30, 2018 (12 months):
3.4 On June 30, 2019 (24 months):
3.5 On June 30, 2020 (36 months):

4. Please detail the elements considered to form your exchange rate

expectations in the short and long run:
Short run (3, 6 and 12 months):

The Information Rigidities and Rationality 73

 i)
ii)
iii)

Long run (24 and 36 months)

i)
ii)
iii)

5. How do you consider that the general economic conditions for pri-
vate production activities will evolve in the next six months in con-
trast with the past six months? (Please check one box)
Will improve []
Will be the same []
Will deteriorate []
Explain why:

6. How do you label the current conditions for firms to invest in the
country? (Please check one box)
Good conditions []
Bad conditions []
Not sure []

Contact: [email protected]
Telephone: (506) 2243-3312. Fax: (506) 2243-4559
The Department of Economic Research makes readily available doc-
uments elaborated on topics related to: inflation, monetary policy,
financial stability, etc. If you want to subscribe, go to the following
address: http://www.bccr.fi.cr/suscripcion/default.aspx

74 A. Alfaro, A. Mora
 Annex B. Expected Inflation, Responses
and Dispersion by Group

Figure 7
INFLATION EXPECTATIONS SURVEY, RESPONSES BY GROUP

18
14
12
10
Responses

6
4

2
0
2006
2006
2007
2007
2007
2008
2008
2008
2009
2009
2009
2010
2010
2010
2011
2011
2011
2012
2012
2012
2013
2013
2013
2014
2014
2014
2015
2015
2015
2016
2016
2016
2017
2006

Consulting Stock market Academic Business

Source: Own elaboration.

Figure 8
DISPERSION OF EXPECTED INFLATION BY GROUP



20

15

10

0
Jan 1, 2006

Jan 1, 2007

Jan 1, 2008

Jan 1, 2009

Jan 1, 2010

Jan 1, 2011

Jan 1, 2012

Jan 1, 2013

Jan 1, 2014

Jan 1, 2015

Jan 1, 2016

Jan 1, 2017

The Information Rigidities and Rationality 75

 0
5
10
15
20
0
5
10
15
20
0
5
10
15
20

76
Jan 1, 2006 Jan 1, 2006 Jan 1, 2006

Jan 1, 2007 Jan 1, 2007 Jan 1, 2007

Jan 1, 2008 Jan 1, 2008 Jan 1, 2008

Source: Own elaboration.

A. Alfaro, A. Mora
Jan 1, 2009 Jan 1, 2009 Jan 1, 2009

Jan 1, 2010 Jan 1, 2010 Jan 1, 2010

Jan 1, 2011 Jan 1, 2011 Jan 1, 2011




Jan 1, 2012 Jan 1, 2012 Jan 1, 2012

 
Figure 8 (cont.)

Jan 1, 2013 Jan 1, 2013 Jan 1, 2013

Jan 1, 2014 Jan 1, 2014 Jan 1, 2014

Jan 1, 2015 Jan 1, 2015 Jan 1, 2015

DISPERSION OF EXPECTED INFLATION BY GROUP

Jan 1, 2016 Jan 1, 2016 Jan 1, 2016

Jan 1, 2017 Jan 1, 2017 Jan 1, 2017

REFERENCES

Alfaro, Alonso, and Carlos Monge (2013), Expectativas de Inflación

para Costa Rica, Documento de Investigación, No. D1-01-2013,
Banco Central de Costa Rica.
Angeletos, George-Marios, and Chen Lian (2016), Forward Guidance
without Common Knowledge, nber Working Paper, No. 22785,
<doi: 10.3386/w22785>.
Ballantyne, Alexander, Christian Gillitzer, David Jacobs, and Ewan
Rankin (2016), Disagreement about Inflation Expectations, Re-
search Discussion Paper, No. rdp 2016-02, Reserve Bank of
Australia.
Branch, William A. (2004), “The Theory of Rationally Hetero-
geneous Expectations: Evidence from Survey Data on
Inflation Expectations,” The Economic Journal, Vol. 114,
No. 497, July, pp. 592-621, <https://doi.org/10.1111
/j.1468-0297.2004.00233.x>.
Coibion, Olivier, and Yuriy Gorodnichenko (2015), “Information
Rigidity and the Expectations Formation Process: A Simple
Framework and New Facts,” The American Economic Review,
Vol. 105, No. 8, August, pp. 2644-2678, <https://www.jstor.
org/stable/43821351>.
Elliott, Graham, Ivana Komunjer, and Allan Timmermann (2008),
“Biases in Macroeconomic Forecasts: Irrationality or Asym-
metric Loss?,” Journal of the European Economic Association, Vol.
6, No. 1, March, pp. 122-157, <https://doi.org/10.1162/
JEEA.2008.6.1.122>.
Engelberg, Joseph, Charles F. Manski, and Jared Williams (2009),
“Comparing the Point Predictions and Subjective Probability
Distributions of Professional Forecasters,” Journal of Business
& Economic Statistics, Vol. 27, No. 1, pp. 30-41, <https://doi.
org/10.1198/jbes.2009.0003>.
Evans, George W., and Seppo Honkapohja (2005), “An Interview
with Thomas J. Sargent,” Macroeconomic Dynamics, Vol. 9, is-
sue 4, September, pp. 561-583, <https://doi.org/10.1017/
S1365100505050042>.

The Information Rigidities and Rationality 77

Hamilton, James D. (2017), Why you Should Never Use the Hodrick-
Prescott Filter, nber Working Paper, No. 23429, <doi: 10.3386/
w23429>.
Honkapohja, Seppo and Kaushik Mitra (2006), “Learning Stability
in Economies with Heterogeneous Agents,” Review of Economic
Dynamics, Vol. 9, issue 2, April, pp. 284-309, <https://doi.
org/10.1016/j.red.2006.01.003>.
Johannsen, Benjamin K. (2014), “Inflation Experience and In-
flation Expectations: Dispersion and Disagreement within
Demographic Groups,” feds Working Paper, No. 2014-2089
Mankiw, N. Gregory, and Ricardo Reis (2002), “Sticky Information
versus Sticky Prices: A Proposal to Replace the New Keynes-
ian Phillips Curve,” The Quarterly Journal of Economics, Vol.
117, issue 4, pp. 1295-1328, <https://doi.org/10.1162/0
03355302320935034>.
Mankiw, N. Gregory, Ricardo Reis, and Justin Wolfers (2003),
“Disagreement about Inflation Expectations,” nber Macro-
economics Annual, Vol. 18, pp. 209-248, <https://www.jstor.
org/stable/3585256>.
Manski, Charles F. (2004), “Measuring Expectations,” Econometrica,
Vol. 72, issue 5, July, pp. 1329-1376, <https://doi.org/10.
1111/j.1468-0262.2004.00537.x>.
Muth, John F. (1961), “Rational Expectations and the Theory of
Price Movements,” Econometrica, Vol. 29, issue 3, July, pp.
315-335, <doi: 10.2307/1909635>.
Pfajfar, Damjan, and Emiliano Santoro (2010), “Heterogeneity,
Learning and Information Stickiness in Inflation Expecta-
tions,” Journal of Economic Behavior & Organization, Vol. 75, is-
sue 3, September, pp. 426-444, <https://doi.org/10.1016/j.
jebo.2010.05.012>.
Sims, Christopher A. (2003), “Implications of Rational Inattention,”
Journal of Monetary Economics, Vol. 50, issue 3, April, pp. 665-
690, <https://doi.org/10.1016/S0304-3932(03)00029-1>.
Woodford, Michael (2001), Imperfect Common Knowledge and the Ef-
fects of Monetary Policy, nber Working Paper, No. 8673, <doi:
10.3386/w8673>.
Woodford, Michael (2013), “Macroeconomic Analysis without the
Rational Expectations Hypothesis,” Annual Review of Econom-
ics, Vol. 5, issue 1, pp 303-346, <https://doi.org/10.1146/
annurev-economics-080511-110857>.

78 A. Alfaro, A. Mora
Formation and Evolution of Inflation
Expectations in Paraguay

Pablo Agustín Alonso Méndez

Abstract
The establishment of the inflation targeting regime in Paraguay is relatively
recent, however, the results have been satisfactory. This is because based on the
observed data, from the implementation of this framework, it has been possible
not only to reduce inflation levels, but also align inflation expectations along
the medium-term inflation target. This chapter seeks to identify the main deter-
minants of the formation of inflation expectations in Paraguay since the adop-
tion of the inflation targeting regime. This work bases the analysis on the results
obtained from the expectations surveys conducted by the country’s monetary au-
thority. The evolution of inflation should be an important factor to consider.
Moreover, for the correct functioning of the expectations channel, it is essen-
tial that the monetary authority has sufficient credibility. A credibility index
has been constructed to capture the effect of the credibility that the Banco Central
del Paraguay has acquired during the inflation targeting regime. To guarantee
the robustness of our results, the model we use has been estimated by three econo-
metric methods: ordinary least squares (ols), fully modified ols (fmols),

Pablo Alonso <[email protected]>, Director of Analysis and Research Department,

Banco Central del Paraguay (bcp). This research was developed within the framework
of Joint Research Program 2017 of cemla coordinated by the Banco de la República
(Colombia). The author thanks counseling and technical advisory provided by the Fi-
nancial Stability and Development (fsd) Group of the Inter-American Development
Bank in the process of writing this document. The author also appreciates the valuable
comments of Miguel Mora, Chief Economist of Economic Studies, and Samuel Cañete,
Head of the Division of Financial and Fiscal Analysis, both from bcp, and of the direc-
tors of the Economic Research departments. The opinions expressed in this chapter
are those of the author and do not reflect the views of cemla, the fsd group, the Inter-
American Development Bank or the Banco Central del Paraguay.

79
and the generalized method of moments (gmm). According to the outcomes
of all these methods, inflation expectation formation in Paraguay is deter-
mined mainly by the inflation expectation of the previous month. In addition,
the annual inflation information of the previous month is significant at the
time of forming expectations.
Furthermore, the credibility index presents an expected negative sign, as in-
flation expectations have effectively aligned around the medium-term infla-
tion target since the implementation of the inflation targeting regime. The ex-
change rate was not significant in the regressions. This could partly be due
to a relatively low pass-through of the exchange rate to total inflation, espe-
cially in the last few years.
Keywords: inflation targeting, inflation expectations, monetary policy.
jel classification: E31, E52, E58.

1. INTRODUCTION

T
he Banco Central del Paraguay (central bank of Paraguay, bcp)
officially adopted the inflation targeting regime to regulate
its monetary policy in May 2011. Prior to this policy frame-
work, Paraguay exhibited marked levels of volatility even though
there were no historical records of high inflation periods. Under
the inflation targeting regime, volatility and inflationary levels have
been reduced. These inflationary levels fostered uncertainty in eco-
nomic agents when forming their inflation expectations. All this
was reflected in the fact that these expectations showed considerable
variability, in accordance with the results obtained in the expecta-
tions surveys of economic variables carried out by the central bank
on a monthly basis.
The main purpose of this chapter is to try to identify some of the
determinants that Paraguay’s economic agents consider when form-
ing their inflationary expectations. In view of the results of the sur-
vey, a series of factors that may influence the expectations formation
of those who answered the survey have been considered. To do this,
simple econometric regressions are carried out, and the results
of these can be considered a first attempt to find the determinants
of inflation expectations in Paraguay. In addition, the regressions
highlight the importance of the establishment of the inflation target-
ing framework, not only in reducing inflation levels and their vola-
tility, but also lowering inflation expectations. Furthermore, it can

80 P. Alonso
be affirmed that the bcp has managed to gain significant credibility
with respect to the handling of the monetary policy in its attempt
to maintain a low and stable inflation. This is reflected in the cred-
ibility index, which shows the alignment of expectations around
the inflation target since the establishment of the inflation target-
ing regime.
Inflation expectations play a critical role in the process of price
formation in the market. In addition, the decisions of households
and firms depend heavily on the real return that could be expected
on the savings and investments they make. Therefore, central banks
closely monitor the development of inflation expectations in order
to implement their monetary policy in a successful manner.
The results of the empirical model of this chapter show that the es-
tablishment of the inflation targeting scheme has helped to anchor
expectations around the target, and that the dispersion of these ex-
pectations has been adjusted within the inflation range. Further-
more, this dispersion has been reduced with the decrease of the range
during the consolidation process of the inflation targeting regime.
The first part of this chapter contains a brief narrative of mon-
etary policy in Paraguay, highlighting their main characteristics,
and delineates the most important results obtained from it, espe-
cially since the implementation of the inflation targeting framework.
Next, the importance of inflation expectations in monetary policy,
in general and specifically in Paraguay, is highlighted. Subsequently,
after a description of the characteristics of the data according to the
results of the economic variables survey, an estimation model of in-
flation expectations determinants in Paraguay is shown. The main
outcomes of the model show the robustness of the results through
different methodologies of estimating. Finally, in the last section
some conclusions and final comments are presented.

2. MONETARY POLICY IN PARAGUAY

Throughout its history, the Paraguayan economy has not displayed

significant macroeconomic imbalances, such as severe fiscal deficits
or hyperinflationary episodes. The average growth of the gross do-
mestic product (gdp) has been placed at relatively acceptable levels,
although it has presented periods of high volatility. In regard to pric-
es, inflation in Paraguay has been characterized by moderate levels,

Formation and Evolution of Inflation Expectations in Paraguay 81

 Figure 1
INFLATION AND REAL GDP GROWTH
Percentage

50

40

30

20

10

−10
1960
1962
1964
1966
1968
1970
1972
1974
1976
1978
1980
1982
1984
1986
1988
1990
1992
1994
1996
1998
2000
2002
2004
2006
2008
2010
2012
2014
2016
Inflation GDP

Source: Banco Central del Paraguay

unlike most countries of the region (Figure 1). Likewise, the main
problem regarding inflation has been its volatility. The macroeco-
nomic performance of Paraguay can be attributed in part to the
sound management of monetary policy. This is reflected partly in the
fact that the guarani, the local currency of Paraguay, has not been
modified since its inception, thus making it one of the oldest curren-
cies in the region. The relatively prudent management of fiscal policy
has contributed, to certain extent, to keeping inflation at a low level.
As pointed out in the document Política monetaria en Paraguay:
Metas de inflación, un nuevo esquema (bcp, 2013), the design of mon-
etary policy in Paraguay has considered the existence of a relation
between the growth of money supply and inflation. Historically,
this design has adopted a monetary policy scheme of intermediate
objectives, in this case, setting targets for the growth of a specific
monetary aggregate. Thus, the Central Bank used its instruments
to control the money supply’s growth to a level compatible with
the inflation objective, which was based on the achievement of low

82 P. Alonso
inflation, using the quantitative theory of money as a conceptual
framework reference.
Regarding economic activity, in general, the average growth of the
Paraguayan economy has been acceptable, even though it has been
characterized by its volatility. While the expansion of the economy
was quite significant in the 1970s, mainly due to the construction
of the Itaipu hydroelectric dam, there was a period of slowdown in the
1980s and 1990s. In this weakened situation and as a consequence
of a weak financial system, and the fragility of the regulatory and su-
pervisory frameworks, between 1995 and 1998, there were episodes
of large financial crises. In this period, economic authorities needed
a comprehensive reorganization of monetary and financial policy,
which was attained through the enactment of important laws that
allowed a much more stringent regulatory framework for financial
institutions.1
In 2002, the Argentine economy fell into a deep crisis, causing
the abandonment of the convertibility regime to which that country’s
exchange rate policy was subordinated. This episode also affected
the Paraguayan economy. Despite the bcp’s effort to curb capital
outflows and exchange rate depreciation through sharp increases
in the interest rates of monetary regulation instruments, the second
financial crisis occurred towards the end of 2002, although of small-
er magnitude than the first one.
Despite these episodes of crisis, the enactment of the aforemen-
tioned regulatory laws for the financial system allowed the bcp to fo-
cus more on the achievement and maintenance of low and stable
inflation, driving its monetary policy of intermediate objectives,
under a monetary aggregates framework.
As of 2004, the bcp began to lay the foundations for the establish-
ment of an inflation targeting framework, albeit in an experimental
way. Thus, the central bank modernized its monetary policy opera-
tional instruments with the establishment of a medium-term infla-
tion target with a tolerance range. Under this scheme, it was possible
to reduce the average inflation rate in the period from 2000 to 2010
to a single digit level.2

1
The Law No. 489 of the bcp and the Law No. 861 “General of Banks,
Finance, and other Credit Institutions.”
2
In that period average inflation was 8.1%, while in the 1990-2000 period
it was 15.1 percent.

Formation and Evolution of Inflation Expectations in Paraguay 83

 With a more consolidated and orderly monetary policy framework,
the bcp formally adopted the inflation targeting regime in May 2011,
establishing a target of five percent annually with a tolerance range
of +/−2.5 percentage points (pp). After the establishment of the in-
flation targeting regime, lower levels of inflation and volatility were
recorded. For this reason, monetary authorities decided to reduce
the tolerance range to +/−2 pp at the beggining of 2014, and at the end
of that year, they also announced the reduction of the inflation tar-
get to 4.5% annually, which would apply in 2015 and 2016. In order
to achieve its objective of maintaining low and stable inflation, at the
beginning of 2017, the Central Bank announced a new reduction
of the medium-term target to a rate of four percent annually, main-
taining the tolerance range of +/−2 pp.
From the establishment of the inflation targeting regime, in the
2011-2016 period, average inflation was recorded at 3.9%. With these
results and with the efforts of the monetary authorities to not only
maintain low levels of inflation, but also reach a significant degree
of credibility, inflationary expectations were aligned to values around
the inflation target with less variability over the years.

3. INFLUENCE OF EXPECTATIONS ON INFLATION

Economics is a social science that somehow attempts to explain hu-

man behavior, so the perceptions of economic agents on the future
evolution of a wide range of economic indicators are important.
Therefore, an interesting challenge for monetary authorities is to
try to interpret these perceptions in order to implement coherent
policies that help guide them towards clear and precise objectives.
Thus, it is in the macroeconomic field and particularly the theory
of monetary policy, where expectations have become a powerful
analytical tool.
Under the inflation targeting framework, the transmission mech-
anism of inflationary expectations is crucial for the achievement of a
medium-term inflation target. The effectiveness of the expectations
channel depends on the credibility of the central bank. Therefore,
establishing a systematic and transparent decision-making process
in monetary policy is key in facilitating the process of price forma-
tion and private expectations.

84 P. Alonso
 The achievement of the objectives proposed by the central bank,
its transparency and communication increase its credibility, which
contributes to that the expectations remain anchored to the tar-
get in the policy horizon. When a central bank has built a credible
and transparent reputation, a monetary policy decision aimed at con-
trolling inflation keeps inflation expectations anchored to the target.
Therefore, in the face of an expectation of controlled inflation, deci-
sions to adjust prices and wages will be made in line with the inflation
target announced by the central bank.
Taking into account that the objective of clear and transparent
communication is to give signals about the implications of monetary
policy decisions, in general terms, the expectations channel may have
a more rapid impact on the achievement of the inflation target com-
pared to others transmission mechanisms that act with a greater lag.
This makes the expectations channel an important and timely chan-
nel for the effectiveness of monetary policy.
Since the implementation of the inflation targeting regime, the Ban-
co Central del Paraguay has made a great effort to improve its cred-
ibility. As mentioned above, Paraguay’s main problem has not been
high levels of inflation, but rather high volatility. Since the formal es-
tablishment of the inflation targeting scheme by the bcp, not only have
inflation levels been reduced, but, above all, their volatility has been
reduced (Figure 2). Likewise, it has been verified in the expectations
data that there has been a decrease both in their levels and their vol-
atility given the decrease in observed inflation rates. This suggests
that the bcp has managed to increase its credibility in recent years.
As mentioned above, an interesting fact that has been observed
with the implementation of the inflation targeting regime is the re-
duction of inflation expectations (average or median) to levels closer
to the target (Figure 3 and 4). Additionally, the dispersion has been re-
duced, mainly because of the reduction of the tolerance range in 2014.
The reduction of the tolerance range can be proven through tradi-
tional statistics of variability, such as the standard deviation and the
coefficient of variation (Figure 5), which effectively show a reduction
(on average) in recent years, coinciding with the reduction of toler-
ance bands.
Finally, it was run, as an additional test, a simple model of the vola-
tility statistics with respect to a dummy variable that takes the value
of 1 if there is a reduction in the band. The variable is significant with
an expected negative sign. In summation, these results suggest that

Formation and Evolution of Inflation Expectations in Paraguay 85

 Figure 2
ANNUAL INFLATION AND INFLATION EXPECTATIONS
FOR YEAR T AND T+1
Percentage
16

14

12

10

0
Apr2006
Oct2006
Apr2007
Oct2007
Apr2008
Oct2008
Apr2009
Oct2009
Apr2010
Oct2010
Apr2011
Oct2011
Apr2012
Oct2012
Apr2013
Oct2013
Apr2014
Oct2014
Apr2015
Oct2015
Apr2016
Oct2016
Apr2017
Oct2017
Inflation expectation year t Inflation expectation year t+1
Inflation (annual)
Source: Banco Central del Paraguay
Figure 3
DISPERSION OF INFLATION EXPECTATIONS FOR YEAR T¹
Percentage

14

12

10

0
Jan2011
Apr2011
Jul2011
Oct2011
Jan2012
Apr2012
Jul2012
Oct2012
Jan2013
Apr2013
Jul2013
Oct2013
Jan2014
Apr2014
Jul2014
Oct2014
Jan2015
Apr2015
Jul2015
Oct2015
Jan2016
Apr2016
Jul2016
Oct2016
Jan2017
Apr2017
Jul2017
Oct2017

Bands Median Target

¹ The different dots correspond to the respondents in each period, which for ethical
reasons cannot be identified individually.
Source: Banco Central del Paraguay

86 P. Alonso
 Figure 4
DISPERSION OF INFLATION EXPECTATIONS FOR YEAR T+1
Percentage
12

10

0
Jan2011
Apr2011
Jul2011
Oct2011
Jan2012
Apr2012
Jul2012
Oct2012
Jan2013
Apr2013
Jul2013
Oct2013
Jan2014
Apr2014
Jul2014
Oct2014
Jan2015
Apr2015
Jul2015
Oct2015
Jan2016
Apr2016
Jul2016
Oct2016
Jan2017
Apr2017
Jul2017
Oct2017
Bands Median Target
Source: Banco Central del Paraguay

Figure 5
STANDARD DEVIATION AND COEFFICIENT OF VARIATION
Percentage

    

2.5 35

30
2.0
25
1.5 20

1.0 15

10
0.5
5

0.0 0
Jan2011
Aug2011
Mar2012
Oct2012
May2013
Dec2013
Jul2014
Feb2015

Jan2011
Aug2011
Mar2012
Oct2012
May2013
Dec2013
Jul2014
Feb2015
Apr2016
Nov2016
Jun2017

Apr2016
Nov2016
Jun2017

Source: author’s calculations.

Formation and Evolution of Inflation Expectations in Paraguay 87

the reduction of the band contributed to decreasing the dispersion
of the expectations of the economic agents.

4. EMPIRICAL MODEL FOR PARAGUAY

In the bcp, the expectations of the main macroeconomic variables

are obtained with monthly frequency–as of April 2006, from the Eco-
nomic Variables Survey (eve). In its beginning, the eve was mainly
focused on representatives of some of the country’s banks. Currently,
this survey is aimed at agents representing different economic sec-
tors that include banks and financial companies, risk rating agencies,
brokerage firms, consulting firms, independent analysts, economic
organizations, and universities. The number of respondents amounts
to 34, of which, taking into account banks and financial companies,
they comprise 22 representatives of financial institutions.
The eve is divided into four blocks that include questions related
to the expectations of economic agents with respect to total inflation,
measured by the variation of the consumer price index, the evolu-
tion of the nominal exchange rate (guarani versus the United States
dollar), gdp growth, and the trajectory of the monetary policy rate.
The set of questions corresponds to the expectations of the vari-
ables mentioned at different periods: for the end of the current month
and the following, the current year, the next 12 months, the follow-
ing year, and for the monetary policy horizon (which comprises be-
tween 18 and 24 months).
Considering that inflation expectations constitute an important
tool for the bcp in the management of monetary policy under the in-
flation targeting scheme, this chapter aims to identify the main vari-
ables that affect the formation of inflation expectations.

4.1 Data Features

Taking into account the structure of the eve surveys in relation to the
expectations of the economic variables studied, the survey is de-
signed to obtain information on the perspectives of the economic
agents for the current year and for the following year. Thus, the sur-
vey data provide information for fixed event forecasts, which, to a cer-
tain extent, are limitations when estimating an econometric model.

88 P. Alonso
 In order to identify the main determinants of the process of form-
ing expectations, it is necessary to have a series of fixed horizon infla-
tion expectations. To carry out an approximation of fixed horizon
forecasts from the fixed event forecasts of the eve, we follow the work
of Dovern et al. (2012), in which this approximation is made as a
weighted average of fixed-event forecasts as follows:

12 − (m − 1) fe m − 1 fe
1 Fyf0,hm ,12 (x ) = Fy 0,m ,y 0 (x ) + F (x )
12 12 y 0,m ,y 0+1

where Fyfe0,m ,y 0 (x ) is the fixed-event forecast of the variable x for

the current year (y0) made in the month m of the year y0; Fyfe0,m ,y 0+1(x )
is the fixed-event forecast of the variable x for the following year (y 0 +
fh
1) made in the month m of the year y0; and Fy 0,m ,12 (x ) is the fixed hori-
zon twelve-month-ahead forecast made in the month m of the year y0.3
For example, the inflation expectation made in October 2014
for the time period between October 2014 and October 2015 is approx-
fe fe
,10,2014 (π ) ,10,2015 (π ),
imated by the sum of F2014 and F2014 and weighted
by 3 12 and 9 12, respectively.
In this section, we identify some variables that determine infla-
tion expectations in Paraguay, according to empirical literature
related to the subject, and as consider some characteristics of the
Paraguayan economy.
Taking into account that price formation has certain persistence
in its adjustment process, for a certain period, the expectations
of the recent past period should also be considered, since, in these
expectations, agents are acquiring more information about events
that may affect those expectations. In addition, the evolution of in-
flation should be an important factor to consider, since this evolu-
tion provides significant information when determining the future
evolution of prices.
On the other hand, the establishment of the inflation targeting
regime in Paraguay has been an important factor in the formation
of inflation expectations, since it has led to a significant structural
change in Paraguayan monetary policy, thus constituting an an-
chor that serves as a guide for the formation of these expectations
(Figure 6). According to the observed inflation data, which were re-
duced both in levels and in variability, and the inflation targeting

Formation and Evolution of Inflation Expectations in Paraguay 89

 Figure 6
INFLATION EXPECTATIONS FOR YEAR T AND T+1,
AND 12 MONTHS FORWARD
Percentage
12

10

0
Apr2006
Oct2006
Apr2007
Oct2007
Apr2008
Oct2008
Apr2009
Oct2009
Apr2010
Oct2010
Apr2011
Oct2011
Apr2012
Oct2012
Apr2013
Oct2013
Apr2014
Oct2014
Apr2015
Oct2015
Apr2016
Oct2016
Apr2017
Inflation expectation year t Inflation expectation year t+1
Inflation expectation 12 month-ahead
Source: Banco Central del Paraguay

framework, the monetary policy in Paraguay has achieved important

credibility with economic agents. In part, this is reflected in the fact
that when effective inflation data were adjusted around the target
after the implementation of the inflation targeting scheme, expec-
tations were also adjusted to the inflation target determined by the
Banco Central del Paraguay.
For the correct functioning of the expectations channel, it is essen-
tial that the monetary authority has sufficient credibility. Economic
agents must trust that the central bank will do everything necessary
to achieve price stability and its inflationary objective in the medi-
um term. Credibility would be able to neutralize, in part, the effects
of economic shocks on prices that are transmitted through the chan-
nel of expectations.
In this sense, to try to capture the effect of the credibility that
the Banco Central del Paraguay has acquired during the inflation
targeting regime, a credibility index has been constructed following

90 P. Alonso
the work of Mendonça (2007), in which it is assumed that the cen-
tral bank is able to guide inflation expectations towards the target
and reaffirm its commitment to the inflation ranges. Thus, when
expectations are equal to the inflation target the credibility index
is equal to one, and decreases when expectations move away from
the target. In cases where inflation expectations are located outside
the inflation target bands, the index is equal to zero (see Annex).
Finally, it may be thought that a priori changes in the nominal ex-
change rate (guarani-dollar) should influence the formation of in-
flation expectations of economic agents on the cost side of imported
goods (and inputs), especially when considering that Paraguay is a
relatively open economy.4 A similar analysis could be made when
considering variations in oil price, since this product directly affects
the price of fuels, an important input for any production process.

4.2 Estimation of the empirical model

To guarantee the robustness of our results, the model we use has been
estimated by three econometric methods: ordinary least squares
(ols), fully modified ols (fmols), and the generalized method
of moments (gmm). 5 The fmols method assumes the existence of a
cointegration relation between the variables, while the gmm method
is created to avoid potential endogeneity problems with some regres-
sors using ols. The model has been estimated in monthly frequency.
In accordance with the aforementioned information and taking into
account some characteristics of the Paraguayan monetary policy,
the estimated model is as follows:

2 π te = α 0 + α1π te−1 + α2π t −1 + α 3 ∆nert −1 + α 4 ∆oilt −1 +

+α 5credt −1 + α 6credt −1 * π te−1 + a7dummytIT + εt ,

where π te is the inflation expectation for twelve months ahead; π t −1

is the annual inflation of period t – 1; ∆nert −1 is the annual variation
of the nominal exchange rate (guarani-dollar); ∆oilt −1 is the an-
nual variation in the price of oil; cred is a variable that measures

Formation and Evolution of Inflation Expectations in Paraguay 91

the credibility of the central bank,6 and dummytIT represents the pe-
riod since the implementation of the inflation targeting regime.
According to our regressions’ outcomes, inflation expectation
formation in Paraguay (twelve-month-ahead) is determined mainly
by the inflation expectation of the previous month (Table 1). In addi-
tion, the annual inflation information of the previous month is sig-
nificant at the time of forming expectations.
On the other hand, the credibility index presents an expected neg-
ative sign, as inflation expectations have effectively aligned around
the medium-term inflation target since the implementation of the
inflation targeting scheme.
Changes in the exchange rate and the price of oil were not sig-
nificant in the inflation expectation formation process. This could
partly be due, to a relatively low pass-through of the exchange rate
to total inflation, especially in the last few years.7 Likewise, the oil
price reduction in international markets has influenced the decrease
of fuel prices in the local market.
Since the establishment of the inflation targeting scheme, both
the level of inflation and its volatility have decreased. This behavior
is also reflected in the results of the surveys, in which it is observed
that inflation and its expectations present an important variability.
The credibility achieved by the monetary authority has been essen-
tial in ensuring that expectations are adjusted to the inflationary
objective of the medium term.
On the other hand, as of May 2011, the estimate of a dummy vari-
able reflects the change in the monetary policy regime. In addition,
it is proven that under the inflation targeting regime inflation ex-
pectations have been adjusted downward, as observed inflation data
were aligned around the inflation target.

As previously indicated, since January 2014, the fluctuation bands

have been reduced from +/− 2.5 pp to +/− 2 pp with respect to the in-
flation target. To test if the lower band has had a greater effect on in-
flation expectations, in the base equation, a dummy variable equal
to 1 has been introduced since the period in which the decrease

6
This index was constructed according to the work of Mendonça (2007),
whose criterion is described in the Annex.
7
See Banco Central del Paraguay (2015, recuadro 1).

92 P. Alonso
 Table 1
ESTIMATED EQUATIONS FOR INFLATION EXPECTATIONS
Dependent variable: inflation expectations (12 month-ahead)

Models
ols fmols gmm

Sample 2006M05- 2006M05- 2006M05-

2017M12 2017M12 2017M12

Constant 2.41 1.84 2.57

(0.0000) (0.0005) (0.0071)

0.53 0.62 0.49

π te−1
(0.0000) (0.0000) (0.0002)

π t −1 0.16 0.13 0.17

(0.0000) (0.0000) (0.0000)

∆nert −1 –0.0003 –0.0006 –0.0003

(0.9570) –0.0008 –0.0015

∆oilt −1 –0.0009 –0.0008 0.0015

(0.5948) (0.5962) (0.3086)

credt −1 –1.791 –1.904 –2.0967

(0.0028) (0.0011) (0.0061)

0.27 0.31 0.32

credt −1 * π te−1
(0.0047) (0.0006) (0.0022)

–0.31 –0.23 –0.30

dummy IT
(0.0049) (0.0276) (0.0224)

Adjusted R2 0.92 0.92 0.92

Note: p-values are in parenthesis.

Source: author’s calculations.

Formation and Evolution of Inflation Expectations in Paraguay 93

 Table 2
ESTIMATED EQUATIONS FOR INFLATION
EXPECTATIONS FOR BANDS REDUCTION
Dependent variable: inflation expectations (12 month-ahead)

Models
ols fmols gmm

Sample 2006M05- 2006M05- 2006M05-

2017M12 2017M12 2017M12

Constant 2.67 2.19 2.92

(0.0000) (0.0005) (0.0000)

0.47 0.55 0.43

π te−1
(0.0000) (0.0000) (0.0000)

π t −1 0.19 0.18 0.20

(0.0000) (0.0000) (0.0000)

∆nert −1 0.0055 0.0047 0.0020

(0.3606) (0.3744) (0.7441)

∆oilt −1 –0.0014 –0.0019 0.0029

(0.3760) (0.1929) (0.0639)

credt −1 –0.502 –0.640 –0.7481

(0.4371) (0.2710) (0.1869)

0.03 0.07 0.07

credt −1 * π te−1
(0.7699) (0.4350) (0.4295)

–0.20 –0.18 –0.20

dummy IT
(0.0598) (0.0564) (0.0552)

–0.57 –0.489 –0.570

dummy bands
(0.0001) (0.0001) (0.0000)

Adjusted R2 0.93 0.93 0.93

Note: p-values in parenthesis.

Sources: author’s calculations.

94 P. Alonso
in the range occurred. In this regard, interesting results are observed
in all the estimation methodologies, as they show that the lower in-
flationary range has had an impact on getting inflation expectations
adjusted to this new range (Table 2). This also shows that the bcp
has had a significant influence on the credibility of economic agents
in achieving the inflation goal under the inflation targeting regime.8
On the other hand, an exercise was carried out that reflects the be-
havior of the inflation expectations of the group of respondents
categorized as financial entities (banks and financial companies).
The results show that the expectations of the financial agents follow
a similar pattern to the base equation (Table 3).

5. CONCLUSION

The implementation of an inflation targeting regime is relatively

recent and because of this, economic agents have a learning curve
with respect to the functioning of monetary policy transmission
mechanisms and with respect to other macroeconomic variables
that are relevant to explaining inflation. In the case of the Para-
guayan economy, finding an econometric model that helps deter-
mine the main factors of inflation expectations is not a trivial task.
The establishment of the inflation targeting framework has led
to an important structural change in the conduct of monetary pol-
icy in Paraguay. On top of helping reduce inflation levels and their
volatility, this framework has also helped guide the inflation expec-
tations of the economic agents through the nominal anchor of the
medium-term inflation target.
Considering that the formation of prices is characterized by a
change in persistence, it is reasonable to think that both the data
of the observed inflation rate and that of their expectations in a pre-
vious period are important determinants at the time that economic
agents define their expectations of inflation in the current period.
The observed trajectory of the inflation data shows that the imple-
mentation of the inflation targeting scheme has been satisfactory.

8
The introduction of the band dummy variable diminishes the signifi-
cance from the credibility index. This could be due to the fact that both
variables reflect greater credibility in the inflation targeting scheme,
so that the two variables cannot be together in the same base equation.

Formation and Evolution of Inflation Expectations in Paraguay 95

 Table 3
ESTIMATED EQUATIONS FOR INFLATION
EXPECTATIONS OF FINANCIAL ENTITIES
Dependent variable: inflation expectation (12 month-ahead)
ols
Sample 2011M02-2017M12
Constant 1.82
(0.0045)
0.55
π te−1
(0.0000)
π t −1 0.14
(0.0001)
∆nert −1 –0.0046
(0.4850)
∆oilt −1 0.0010
(0.6300)
credt −1 –2.333
(0.0019)
0.46
credt −1 * π te−1
(0.0006)
Adjusted R2 0.89
Note: p-values in parenthesis.
Sources: author’s calculations.

This proves that the bcp has achieved significant credibility in its
purpose of keeping inflation low and stable around the inflation
target. Therefore, the alignment of inflation expectations around
the target can be attributed to an increase in credibility.
It should be noted that the reduction in inflationary bands also
reflects an adjustment of inflation expectations around the target, at-
testing likewise to greater credibility of economic agents in the man-
agement of monetary policy under the inflation targeting scheme.
In addition, when the respondents are grouped in the category of fi-
nancial entities, it is observed that the expectations of these agents
follow a pattern similar to that observed in the base equation.

96 P. Alonso
ANNEX

Annex A. Credibility Index

1 if π te − π * 
 
 1 
1 − lower  e *
* π t − π 
if π lower < π te 
 π − π 
 .
 1 upper e 
1 − upper  e *
* π t − π 
if π > πt 
 π −π 
0 if π te ≥ π upper or π te ≤ π lower 


References
Banco Central de Chile (2007), La política monetaria del Banco Central
de Chile en el marco de metas de inflación.
Banco Central de Reserva del Perú (2016), El impacto de la meta de
inflación sobre la inflación.
Banco Central del Paraguay (2003), Política monetaria en Paraguay:
metas de inflación, un nuevo esquema.
Banco Central del Paraguay (2015) Informe de política monetaria de
septiembre de 2015, <https://www.bcp.gov.py/informe-de-
politica-monetaria-i14>.
Banco Central do Brasil (2015), Inércia inflacionária e determinantes
das expectativas de inflação.
Cerisola, Martin, and R. Gaston Gelos (2005), What Drives Infla-
tion Expectations in Brazil? An Empirical Analysis, imf Working
Paper, No. wp/05/109, June.
Coibon, Olivier, and Yuriy Gorodnichenko (2012), Information Rigid-
ity and the Expectations Formations Process: A Simple Framework and
New Facts, imf Working Paper, No. wp/12/296, December.

Formation and Evolution of Inflation Expectations in Paraguay 97

De Mendonça, Helder F. (2007), “Towards Credibility from In-
flation Targeting: The Brazilian Experience,” Applied Eco-
nomics, Vol. 39, issue 20, pp. 2599-2615, <https://doi.
org/10.1080/00036840600707324>.
Departamento de Programación Monetaria y Estudios Económicos
(2017), Informe de Política Monetaria, Banco Central de la
República Dominicana.
Dovern, Jonas, Ulrich Fritsche, and Jiri Slacalek (2009), Disagree-
ment among Forecasters in G7 Countries, Working Paper Series,
No. 1082, European Central Bank.
Goldman Sachs International (2013), Market Inflation Expectations:
Evolution and Determinants, European Economics Analyst,
New York.
Hoggarth, Glenn (1997), Introducción a la política monetaria, Ensayos,
No. 54, Centro de Estudios Monetarios Latinoamericanos,
pp. 1-20.
Hurtarte Aguilar, Guisela (2005), What Drives Inflation Expectations
in Guatemala, and What Do They Imply for Monetary Policy Deci-
sions?, Working Paper.
Instituto Nacional de Estadística (1995), La elaboración del Índice
de Difusión de Empleo, Boletín Trimestral de Coyuntura, No.
58, Spain, December.
Jarociński, Marek, and Bartosz Maćkowiak, B (2013), Granger-
Causal-Priority and Choice of Variables in Vector Autoregressions,
Working Paper Series, No. 1600, European Central Bank.
Muñoz Salas, Evelyn, and Carlos Torres Gutiérrez (2006), Un mo-
delo de formación de expectativas de inflación para Costa Rica,
Documento de Investigación, No. die-03-2006-di/r, Banco
Central de Costa Rica.
Rubli Kaiser, Federico (2006), “La relevancia de las expectativas
para los fenómenos monetarios,” Economía Informa, No.
341, pp. 29-39.
Sabrowski, Henry (2008), Inflation Expectation Formation of German
Consumers: Rational or Adaptive?, Working Paper Series in
Economics, No. 100, University of Lüneburg.
Soybilgen, Barış, and Ege Yazgan (2017), “An Evolution of Infla-
tion Expectations in Turkey,” Central Bank Review, Vol. 17,
issue 1, March, pp. 31-38, <https://doi.org/10.1016/j.
cbrev.2017.01.001>.

98 P. Alonso
The Degree of Inflation
Expectation Anchoring
Anchoring of Inflation Expectations
in Latin America

Rocío Gondo
James Yetman

Abstract
We use inflation survey data from Consensus Economics to assess how firmly
inflation expectations are anchored in Latin America. Following the method-
ology proposed by Mehrotra and Yetman (2018), we model inflation forecasts
using a decay function, where forecasts monotonically diverge from an esti-
mated anchor towards recent actual inflation as the forecast horizon short-
ens. Our results suggest that most countries do have an inflation anchor,
with the estimated weight of the anchor increasing through time, indicating
more strongly anchored expectations. This is consistent with the improving
credibility of central banks’ monetary policy management over our sample
period (1993-2016). For countries with formal inflation targets, our results
indicate that inflation targeting regimes are generally credible, with estimated
anchors lying within the inflation target range for all countries in the most
recent sample that we consider.
Keywords: inflation expectations, inflation anchoring, decay function.
jel classification: E31, E58.

R. Gondo <[email protected]>, researcher at the Economic Research De-

partment of the Banco Central de Reserva del Perú, and J. Yetman <james.yetman@
bis.org>, principal economist at the bis Representative Office for Asia and the Pacific.
Julieta Contreras and Berenice Martínez provided excellent research assistance. This
paper is part of cemla’s joint research project on Inflation Expectations and their
Degree of Anchoring. The authors thank, without implication, other project partici-
pants and especially our advisor for this paper, Olivier Coibion, for their comments.
The views expressed here are those of the authors and are not necessarily shared
by the Bank for International Settlements or the Banco Central de Reserva del Perú.

101
1. INTRODUCTION

M
onetary policy effectiveness, and especially the achievement
of price stability, can be greatly assisted when inflation expec-
tations are well anchored. In many models of inflation, for ex-
ample, volatile inflation expectations directly increase the volatility
of inflation outcomes. In Latin America, with a history of repeated
episodes of high inflation, many countries have adopted inflation
targeting (it) as a framework to support a move to low and stable
inflation and provide for better anchoring of inflation expectations.
Some of these countries have adopted a schedule of decreasing targets
over time with a view to gradually reducing inflation.
Challenges of inflation control for central banks in the region
remain. In 2015-2016, some countries experienced inflation rates
above the top of their target ranges, mainly commodity exporters
who experienced large currency depreciations. In the cases of Co-
lombia and Peru inflation expectations appear to have become de-
anchored to some extent, with high inflation persisting (see Figure
1). Monetary policy tightening actions were taken in response to these
developments, with their central banks raising policy rates by 3.25%
and 1%, respectively.
The goal of this paper is to assess whether or not countries have
an inflation expectations anchor and, if they do, how strongly in-
f lation expectations are anchored. For economies with formal
it frameworks, we also examine whether the anchor is consistent
with the central bank’s target. We define an inflation anchor as the
expected level of inflation in the absence of any shocks to the econ-
omy. It should be noted that the inflation anchor is not necessar-
ily equal to the inflation target for countries with an it framework.
For each country, first, we evaluate whether there is an anchor
for inflation expectations and, if so, how the anchor has evolved over
time. Second, we analyze how well identified the inflation anchor is,
using the standard deviation of the estimated anchor as an indica-
tor of the degree of anchoring. Third, we compare the anchoring
of inflation expectations between countries in the region that have
inflation targets with such anchoring in those that do not.
We model inflation forecasts using a decay function, where fore-
casts monotonically diverge from the estimated anchor towards re-
cent actual inflation as the forecast horizon shortens. We estimate

102 R. Gondo, J. Yetman

 Figure 1
INFLATION EXPECTATIONS PUBLISHED BY CENTRAL BANKS

1 2
5 5

4 4
3.6 3.0
Percentages

Percentages
3 3
2.9
2 2

1 1

0 0
2014

2015

2016

2014

2015

2016

2017

2014

2015

2016

2014

2015

2016

2017
Central bank survey Expectations by:
Target Financial institutions
Economic analysts
Nonfinancial companies
1
Expectations for 12-month inflation. 2 Expectations of current year (December) 12-
month inflation.
Sources: National data.

this relationship for each country over eight-year rolling samples

using maximum likelihood, obtaining parameter estimates that
define the decay function and the anchor.
Our results suggest that most countries do have an inflation an-
chor, although in some countries (including Argentina and Venezu-
ela), the degree of anchoring declined in recent periods. For most
countries, we observe a pattern of increasing anchoring of inflation
expectations, consistent with the improved credibility of central
banks’ monetary policy management. This result stands in contrast
with the results of Davis and Mack (2013), who found a low degree
of anchoring of inflation expectations for Latin America compared
with other regions, using a Phillips curve regression on core inflation.
In it countries, inflation expectations appear to be well anchored.
In addition, we find that the estimated anchors are generally con-
sistent with their inflation targets; in the most recent sample that
we examine, our estimated inflation anchors lie within the infla-
tion target range for all countries with formal inflation targets. This

Anchoring of Inflation Expectations in Latin America 103

result is consistent with the results in De Carvalho et al. (2006), where
they find that the inflation anchor does not differ statistically from
the inflation target for Brazil, Chile, and Mexico. For countries that
adopted it after 2009, the estimated anchor is slightly higher than
the target, but this might be due to the rolling sample containing
some years before the adoption of the regime.
We then consider some second-stage regressions based on these
estimates, focusing on the estimated weight on the anchor at a two-
year horizon, to explore what is driving our results. We show that
it and low levels of inflation persistence help explain strongly an-
chored inflation expectations.
Moreover, we find that inflation-targeting countries generally have
more precisely estimated inflation expectations anchors. Capistran
and Ramos-Francia (2010) report similar results: Countries with
it show a lower dispersion of long-run inflation expectations, espe-
cially in the case of emerging market countries. Similarly, for a sam-
ple of 15 advanced countries, Cecchetti and Hakkio (2009) find that
the adoption of it reduces the dispersion of inflation expectations.
In addition to the papers already cited, our work is related to models
of inflation expectations extracted from financial data. For instance,
Gurkaynak et al. (2007) find that iters such as Canada and Chile have
better anchored long-run inflation expectations than the United
States (us), using break-even inflation rates from nominal and in-
flation-indexed bonds. For Latin America, De Pooter et al. (2014),
using both survey-based and financial market-based data, find that
inflation expectations have become better anchored over the past
decade in Brazil, Chile, and Mexico. Focusing on Colombia, Espi-
nosa-Torres et al. (2017) find that inflation expectations, obtained
through break-even inflation measures, have remained anchored
to values inside the inflation target range in the period following
the Great Financial Crisis. Finally, for Brazil, Vicente and Guillen
(2013) find that break-even inflation is an unbiased predictor of fu-
ture inflation at short horizons, but is actually negatively correlated
with inflation outcomes at 24- and 40-month horizons.
The paper is organized as follows. Section 2 provides a short de-
scription of the estimation methodology. Section 3 describes the data.
Section 4 discusses the results. Section 5 then concludes.

104 R. Gondo, J. Yetman

2. METHODOLOGY

Following the methodology proposed by Mehrotra and Yetman

(2018), we model inflation forecasts using a decay function, where
forecasts diverge monotonically from an estimated anchor towards
recent actual inflation as the forecast horizon shortens. This frame-
work makes full use of the multiple-horizon dimension of the data
to provide a measure of the level of the inflation anchor.
The functional form used to model inflation expectations is based
on the cumulative density function of the Weibull distribution. This
functional form assumes that, as the forecast horizon shortens, in-
flation expectations become increasingly sensitive to newly arriving
information about inflation outcomes.
Given the observed behavior of inflation forecasts from the mean
and median data from Latin American Consensus Forecasts, we mod-
el the expectations process for each country as follows:1

1 f (t ,t − h ) = α (h )π * + (1 − α (h ) )π (t − h ) + ε (t ,t − h ) ,

where f (t ,t − h ) is the forecast of inflation for year t at horizon h ;

h is the number of months before the end of year being forecasted;
α (h ) is the weight on the anchor (which follows a decay function); π *
is the inflation anchor; π (t − h ) is the observed inflation at the time
that the forecast is made; and ε (t ,t − h ) is a residual term.
We assume that the decay function α (h ) follows a Weibull cumu-
lative density function:2

1
We parametrize the model to separately identify the anchor and the
coefficients indicating the weight on the anchor. If there is a link
between the two (for example, adopting an inflation target leads to a
change in both the anchor and how strongly inflation is anchored), our
estimation allows for this possibility but does not impose it. As such, it
may be possible to improve the efficiency of the estimation approach
taken here.
2
Our results are conditional on the decay function. Mehrotra and Yetman
(2018) demonstrate that, provided inflation follows an autoregressive
process, a monotonically decreasing decay function should fit inflation
expectations.

Anchoring of Inflation Expectations in Latin America 105

  h  c

2 α (h ) = 1 − exp  −    .
 b 
 

The two parameters to estimate from the decay function are b

and c . Higher values of b result in a smaller weight on the inflation
anchor at short horizons, whereas higher values of c provide more
curvature, and a more rapid decline the weight on the inflation an-
chor as the horizon shortens.
The variance of the residual is ε (t ,t − h ) modeled as a function
of the forecast horizon h :

3 V ( ε (t ,t − h ) ) = exp (δ 0 + δ1h + δ 2h 2 ).

The use of the exponential function here ensures that the fitted
values of the variance are positive for any values of the parameters
defining the variance ( δ 0 , δ 1 and δ 2). Note that, aside from this re-
striction, our modeling assumptions for the variance are very flexible:
It can be increasing or decreasing in the forecast horizon, or even
follow a u-shaped (or inverse u-shaped) pattern across horizons.
Forecasts made at different horizons for the inflation outcome
in a given year t are likely to be correlated, and more strongly so the
closer the two horizons are. Therefore, the correlation between the re-
sidual at two different horizons h and k is modeled as:

4 corr ( ε (t ,t − h ),ε (t ,t − k ) ) = φ0 + φ1 h − k .

We estimate the set of parameters {π * ,b,c,δ 0 ,δ1,δ 2 ,φ0 ,φ1} by maxi-

mum likelihood, economy by economy, based on eight-year rolling
samples. Given the high degree of non-linearity of the model, we use
100 different sets of starting values in each case to ensure convergence
to a global maximum. We then choose the estimates with the highest
log-likelihood function value for which the parameters of the decay
function are identified.

106 R. Gondo, J. Yetman

3. DATA

We use data on mean or median inflation forecasts from Latin Amer-

ican Consensus Forecasts. Our preference is median forecasts, con-
structed based on the full panel of inflation forecasts available from
Consensus Economics at a monthly frequency. Medians are less af-
fected by outlier forecasts than means, and may, therefore, be less
vulnerable to data errors, for example. However, for some coun-
tries, forecaster-level data only becomes available partway through
our sample. For other countries, only average forecasts are available
for the full sample. Where we cannot construct median forecasts,
we use mean forecasts instead.
Our sample covers 18 countries in the region, as listed in Ta-
ble 1. The economies in our sample account for more than 95%
of gdp for Latin America and the Caribbean in 2015 at market ex-
change rates. This sample includes countries with and without it re-
gimes, those that achieved low and stable inflation rates, and others
where inflation has stayed relatively high and volatile.

Table 1
LIST OF COUNTRIES AND SAMPLE

Data Inflation Data Inflation

available target available target
from adopted from adopted
Argentina 1993 Guatemala 2009 2005
Bolivia 1993 Honduras 2009
Brazil 1990 1999 Mexico 1990 2001
Chile 1993 1999 Nicaragua 2009
Colombia 1993 1999 Panama 1993
Costa Rica 1993 2005a Paraguay 1993 2011
Dominican 1993 2012 Peru 1993 2002
Republic
Ecuador 1993 Uruguay 1993 2007
El Salvador 2009 Venezuela 1993

a
Transition to an explicit it regime started in 2005 with the announcement
of an annual inflation target.

Anchoring of Inflation Expectations in Latin America 107

 Arguably, there may be better inflation forecast datasets that
could be used to answer this question, at least for some of the econo-
mies in our sample. For example, Consensus Economics’ inflation
forecasts are typically based on the annual average inflation rate,
whereas most inflation targets are defined in terms of year-on-year
inflation. Hence, central bankers are likely to care more strongly
about anchoring in terms of year-on-year inflation, rather than an-
nual average inflation. Offsetting this, we expect that measures of an-
choring are likely to be highly correlated across the two measures.
Further, using Consensus data, we are able to focus on a larger cross-
section of countries, covering a longer period for many economies
than would be possible with forecasts from other sources. The fore-
cast surveys are also constructed using consistent methodology (in
terms of variable definition and the timing of the forecasts, for ex-
ample), so the results are likely to be comparable across countries.
Table 1 shows the availability of data for each country, including
the starting date and the year of adoption of an it regime, where ap-
plicable. Note that data availability is limited to bi-monthly for some
economies in the early part of the sample, with monthly forecasts
only published beginning in 2002. In these cases, we ensure that
the contribution of the missing observations to the likelihood func-
tion is set to zero.
Figure A.1 in the Annex shows the evolution of inflation forecasts
for each country in the sample. For countries that have had it re-
gimes for an extended period (displayed in Figure A.1, Section A:
Brazil, Chile, Colombia, Mexico, and Peru), longer-horizon fore-
casts are more strongly anchored than for other countries in the
sample. In particular, two-year-ahead inflation forecasts are close
to the inflation target and the dispersion between the inflation fore-
casts for different years is quite small. In this set of countries, infla-
tion forecasts only start to deviate from the target around 12 months
ahead of the date being forecast, when observed inflation outcomes
become more informative about the path of inflation.
The second group of countries (displayed in Figure A.1, Section B:
Costa Rica, the Dominican Republic, Guatemala, Paraguay, and Uru-
guay) adopted it more recently. For longer-horizon forecasts, e.g.,
24 months ahead, we observe a wide dispersion in inflation forecasts
across time, but a declining trend in the initial forecast point after
the adoption of it.

108 R. Gondo, J. Yetman

 The last subset of countries is those without an explicit inflation
target throughout our sample (see Figure A.1, Section C). These
countries tend to show the largest dispersion between inflation fore-
casts at both short and long horizons.

4. RESULTS

We estimate our non-linear model by maximum likelihood using

eight-year rolling samples. For each sample, we consider a large set of
different starting values to ensure convergence to the global maxi-
mum. We consider that an inflation anchor exists if the estimated
weight on the anchor at 24 months is higher than 0.10. Below this
threshold, the estimated anchor tends to be very volatile and highly
dependent on starting values, which we interpret as indicating that
there is no inflation anchor.

4.1 Decay Function

Figure 2 shows the estimated decay functions for all the countries
in the sample, using the most recent rolling sample of 2009-2016.
The figures show that the weight on the anchor is high–generally
above 0.7–for all horizons longer than 12 months for all countries
in our sample, with the exception of Argentina (which is barely vis-
ible in the bottom left corner of the right-hand panel). We generally
observe a sharp decline in the weight assigned to the inflation anchor
in horizons shorter than six months, when forecasters have more in-
formation about realized inflationary shocks that are likely to con-
tinue to influence inflation through to the inflation outcome being
forecast. Qualitatively, there does not seem to be a large difference
between countries with it in our sample and other Latin American
countries in terms of the estimated decay functions.
With respect to the evolution through time, Figure 3 shows the es-
timated weight on the anchor at a horizon of two years (i.e., α ( 24 ) ),
the longest horizon for which we use the Consensus Forecast data. 3
We include all countries for which there are multiple rolling sam-
ples (i.e., forecasts are available before 2009). These results suggest

3
Consensus Forecasts also publishes average forecasts at longer horizons,
of up to ten years, for some economies in our sample, but these are
only available twice per year.

Anchoring of Inflation Expectations in Latin America 109

 Figure 2
DECAY FUNCTIONS 2009-2016
FL     15 
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
22 20 18 16 14 12 10 8 6 4 2 0
Brazil Colombia Peru Chile Mexico
FL     15 
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
22 20 18 16 14 12 10 8 6 4 2 0
Costa Rica Dominican Republic Guatemala
Paraguay Uruguay
  FL 
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
22 20 18 16 14 12 10 8 6 4 2 0
Argentina Bolivia Ecuador Honduras
Nicaragua Panama El Salvador
Notes: The horizontal axis represents the forecast horizon, defined as the number
of months before the end of the calendar year being forecasted.
The graph does not include the decay function for Venezuela because the last available
rolling sample is 2008–2015.
Sources: Authors’ calculations.

110 R. Gondo, J. Yetman

 Figure 3
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 24)
.        15 
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015


1.0

0.8
0.6
0.4

0.2
0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 111

 Figure 3 (cont.)
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 24)
.        15 
   
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

112 R. Gondo, J. Yetman

 Figure 3 (cont.)
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 24)
.   FL 
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016


1.0

0.8
0.6
0.4

0.2
0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 113

that the degree of anchoring of long-run inflation expectations
has generally increased over the sample, most notably for some of the
economies with inflation targets (Chile, Colombia, Peru in Panel A,
and Paraguay and Uruguay in Panel B).4 In the most recent rolling
sample, the weight on the anchor exceeds 0.7 for all economies except
Argentina and Venezuela. Similar results are observed at other ho-
rizons too (see Figure A.2 in the Annex for anchoring at a 12-month
horizon, for example).
Table 2 displays the key estimated parameters for the most recent
rolling sample, 2009-2016. We report an estimated inflation anchor
for all economies, including those for which this is poorly identified
in the data. There is a wide variety of parameter estimates across
countries. We note that Venezuela has a much higher estimated an-
chor than any of the other economies (at over 28%), and Argentina
and Venezuela have much less precisely estimated anchors than
the other countries in the sample, consistent with relatively weakly
anchored inflation expectations for these countries.
Regarding the parameters that govern the shape of the decay func-
tion, most countries show a very low degree of curvature (i.e., low es-
timates of c ), which means that the weight on the anchor remains
high even as the forecast horizon shortens, as shown in Figure 2.

4.2 Estimated Inflation Anchors

Figure 4 shows the evolution of the estimated inflation expectations
anchors, for the same set of countries displayed in Figure 3. Solid
lines correspond to the point estimate of the anchor, while dashed
lines represent the 95% confidence interval. Gray regions illustrate
inflation target ranges where applicable.
Section A of the figure presents the results for countries that have
had it for more than 15 years. Since the adoption of it, all these coun-
tries show a reduction in their anchor towards the inflation target.

4
In our modeling of inflation expectations, we are implicitly assuming
that changes in inflation persistence reflect changes in the anchor-
ing of inflation expectations. To the extent that declining inflation
persistence reflects changed price-setting mechanisms that results
from greater anchoring of inflation expectations, this assumption is
warranted (see Section 4.3). But there may be other, more mechani-
cal sources of changes in inflation persistence–such as changes in the
sectoral composition of the economy–that could bias our results.

114 R. Gondo, J. Yetman

 Table 2
ESTIMATION RESULTS, 2009-2016

b c π* s.e.( π *)

Argentina 24.60 59.56 5.39 0.411

Bolivia 4.20 0.62 6.00 0.027

Brazil 6.37 0.38 4.88 0.028

Chile 2.58 0.55 2.98 0.004

Colombia 11.84 0.58 3.45 0.012

Costa Rica 3.39 0.49 6.09 0.043

Dominican Republic 0.35 0.25 5.84 0.044

Ecuador 6.25 0.72 4.17 0.015

El Salvador 3.47 0.33 3.06 0.014

Guatemala 6.62 0.59 7.83 0.032

Honduras 2.85 0.53 6.97 0.034

Mexico 1.29 0.29 3.54 0.006

Nicaragua 2.90 0.36 7.21 0.025

Panama 2.53 0.36 3.81 0.022

Paraguay 0.89 0.86 5.10 0.027

Peru 0.02 0.06 2.55 0.016

Uruguay 1.45 0.52 6.67 0.026

Venezuela1 29.64 2.39 28.35 0.328

1
For Venezuela, results are for 2008-2015, since data are not available for 2016.

Anchoring of Inflation Expectations in Latin America 115

Moreover, for all countries except Brazil, estimates of the anchor
are quite stable from one rolling sample to the next towards the lat-
ter end of the rolling samples.
The confidence bands (constructed from the standard devia-
tion of the estimated anchor) indicate that the estimated anchors
are generally tightly estimated. 5 Chile displays the most tightly esti-
mated anchor across the rolling samples, whereas Colombia shows
an increasing degree of tightness after the adoption of the inflation
target, consistent with improving credibility.
Figure 4, Section B, shows the results for the more recent iters.
These countries, except for Uruguay, show a decreasing trend in their
anchors. In the case of Costa Rica, this is consistent with their de-
creasing inflation target. In the case of Uruguay, the inflation tar-
get has remained at 5% since its adoption, but estimated inflation
appears to be diverging from it towards the upper bound of the tar-
get range of 7%, at the same time as actual inflation has been close
to 7%. This group of countries also shows a tightly estimated anchor
for most countries and rolling samples; for Uruguay, the confidence
band visibly narrows as time goes by.
For countries that are not iters, displayed in Figure 4, Section C,
there is generally more dispersion in both the estimated anchors
and their trends. Ecuador has a stable estimated anchor of 4%, where-
as Venezuela has many rolling samples without an identifiable an-
chor. The degree of tightness of the inflation anchor is, in general,
lower for this group of countries too.
The degree of tightness of the inflation anchor exploits informa-
tion from dispersion across the time series and horizons. We could
also complement the estimation by further exploiting information
on the standard deviation across forecasters for each country, al-
though the availability of data would reduce the sample of countries.
Thus, we leave this to future work.
One caveat with the data used in the analysis is that inf lation
forecasts have a maximum horizon of two years, which might not be

5
The estimated confidence intervals for the inflation anchor depend
on the functional form of the decay function. However, for a sample of
advanced and emerging countries, Mehrotra and Yetman (2018) find
that the Weibull-based decay function fits the data better than more
restrictive forms, and more general forms do not increase explanatory
power markedly.

116 R. Gondo, J. Yetman

 Figure 4
EVOLUTION OF ESTIMATED INFLATION ANCHOR1
.        15 
 
6.0 4
5.5 3
5.0 2
4.5 1
4.0 0
3.5 −1
3.0 −2
1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
10 12.5

8 10.0
6 7.5
4 5.0

2 2.5
0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015


8

6 Anchor
95% confidence
4
Inflation target range
2

0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1
Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 117

 Figure 4 (cont.)
EVOLUTION OF ESTIMATED INFLATION ANCHOR1
.        15 
   
12 8

10 6

8 4

6 2

4 0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
20 12

15 10
8
10
6
5
4
0 2
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015

Anchor 95% confidence Inflation target range

1
Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

118 R. Gondo, J. Yetman

 Figure 4 (cont.)
EVOLUTION OF ESTIMATED INFLATION ANCHOR1
.    
 
12 8

7
8
6
4
5

0 4
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
25 5

20 4
15 3
10 2

5 1
0 0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016


40

35
30
Anchor
25 95% confidence
20
15
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1
Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 119

long enough to capture long-run inflation expectations.6 We test this
by plotting the longer-term Consensus inflation forecasts for six-to-ten
years ahead, for the countries for which these are available, against
the estimated anchors. Figure 5 shows that six-to-ten year ahead fore-
casts are highly correlated with the estimated anchor, with Venezu-
ela being the main outlier, regardless of whether we take a particular
sample period or the average. This is consistent with the results dis-
played in Mehrotra and Yetman (2018) for a larger sample of countries.

4.3 Effect of it
Next, we focus on the sample of countries with it and analyze wheth-
er or not the estimated anchor is consistent with the inflation tar-
get. By doing so, we are assessing whether our results are consistent
with these countries building credibility for their it monetary pol-
icy frameworks. 7 We focus on the average across all rolling samples
where a country has an it framework. Table 3 shows that the estimated

6
On the other hand, long-horizon forecasts (e.g., six-to-ten years ahead)
might relate to outcomes too far into the future to be useful for monetary
policy purposes. For monetary policy setting, the most relevant horizon is
related to the frequency with which most prices and wages are adjusted,
and hence has the greatest impact on inflation dynamics. Thus, one could
imagine wage and price-setting decisions being influenced by inflation
expectations that are anchored by a level of expected inflation that dif-
fers from expectations of long-run inflation (if, for example, forecasters
anticipated that the monetary policy framework might be adjusted in a
few years). In that case, six-to-ten year ahead inflation expectations might
not be relevant for explaining inflation dynamics, but they could still
be important for other economic decisions such as deciding to invest
in fixed assets or determining long-term savings goals.
7
The anchor of inflation expectations could become more consistent
with the inflation target, even if the central bank is not building cred-
ibility, e.g., if inflation moves towards the target for reasons unrelated
to monetary policy or the inflation target is adjusted endogenously to
track inflation. In the former case, these effects are likely to be transitory
(so are mitigated against in part by our use of rolling samples). With re-
spect to the latter case, we see limited evidence of inflation targets being
adjusted strategically in response to deviations of inflation from target
in the inflation targeters that we examine: Inflation targets are either
constant over most of the 2009-2016 period (Brazil, Chile, Colombia,
Guatemala, Mexico, Peru, and Uruguay) or follow a consistent declin-
ing path as inflation targets become more established over time (Costa
Rica, the Dominican Republic, and Paraguay).

120 R. Gondo, J. Yetman

 Figure 5
RELATION BETWEEN ESTIMATED INFLATION ANCHOR AND LONG-TERM
FORECAST FROM CONSENSUS ECONOMICS

 -  

30
Average long-term forecast

25

20
15

10

0
0 5 10 15 20 25
Inflation anchor

    (1996-2016)

30
Average long-term forecast

25

20

15
10

0
0 5 10 15 20 25
Inflation anchor

2009-2016  

10
Average long-term forecast

0
0 1 2 3 4 5 6 7 8 9
Inflation anchor

Note: Sample of countries with long-term forecasts from Consensus Economics

includes Argentina, Brazil, Chile, Colombia, Mexico, Peru and Venezuela.
Sources: Consensus Economics©; authors’ calculations.

Anchoring of Inflation Expectations in Latin America 121

 Table 3
ESTIMATED ANCHOR AND INFLATION TARGET, 2009-2016

Estimated Inflation Estimated Inflation

anchor target1 anchor target1
Brazil 4.88 4.5 Guatemala 7.83 4.5
Chile 2.98 3.0 Mexico 3.54 3.0
Colombia 3.45 3.3 Paraguay2 5.10 4.8
Costa Rica 6.09 5.1 Peru 2.55 2.0
Dominican 5.84 4.6 Uruguay 6.67 5.0
Republic2

1
The inflation target is the simple average of the annual inflation target for each
country in the given sample. 2 For countries that adopted it later than 2009 such
as the Dominican Republic and Paraguay, the sample starts in 2012 and 2011,
respectively.

anchor is quite close to the average midpoint value of the inflation

target in each country, and inside the range of +/– 1 percentage
point for most countries. The gap between the two is wider in the
case of the most recent iters (such as Guatemala and the Domini-
can Republic) but, in those cases, the rolling sample includes years
before the adoption of it, so a wider deviation does not necessarily
indicate a lack of central bank credibility.
We also estimate a modified version of our model only for countries
with it. Instead of estimating the anchor, we consider the midpoint
value of the inflation target π T (t ) and add a parameter d to capture
deviations from the target.

f (t ,t − h ) = α (h )(π T (t ) + d ) + (1 − α (h ))π (t − h ) + ε (t ,t − h ).

A simple test with a null hypothesis of d =0 is then a test of wheth-

er the inflation target was credible or not. Note that, in cases where
central banks have time-varying inflation targeting, we capture this
with our π T (t ), as we then use different values of the target for dif-
ferent years.

122 R. Gondo, J. Yetman

 Table 4
ESTIMATION RESULTS WITH INFLATION TARGET, 2009-2016

b c d s.e.(d)
Brazil 6.37 0.38 0.38 0.028
Chile 2.58 0.56 –0.02 0.004
Colombia 11.27 0.74 0.35 0.014
Costa Rica 3.31 0.56 0.97 0.030
Dominican Republic 5.86 0.46 0.12 0.014
Guatemala 9.01 0.45 1.46 0.027
Mexico 1.29 0.29 0.54 0.005
Paraguay 1.34 1.17 0.10 0.017
Peru 0.32 0.12 0.55 0.016
Uruguay 1.45 0.52 1.67 0.025

Note: Uruguay has a target range of +/–2 percentage points; all other countries
have a target range of +/–1 percentage point.

Table 4 shows the results of these estimations, for the most recent
eight-year rolling sample. These confirm that the anchors of infla-
tion expectations are in line with the inflation target range in all
countries: within a +/−1 percentage point range in all cases except
for Guatemala and Uruguay, the latter of which has an inflation tar-
get range of +/−2 percentage points. That is, we cannot reject the hy-
pothesis that inflation expectations are anchored by the inflation
targets for most countries.
In order to complement the comparison between countries with
and without inflation targets, we further examine whether it im-
proves the anchoring of expectations. To do this, we perform a sec-
ond step panel estimation. We regress the weight of the anchor (α (h ))
for each country for each eight-year rolling sample on a set of country
characteristics. The set of regressors includes: 1) a dummy variable
that takes the value of 1 for countries with it for the full rolling sample
during the rolling sample; 2) the number of years since the adoption
of the it regime; 3) mean inflation; 4) inflation variability, measured
by the standard deviation of inflation; 5) inflation persistence, based

Anchoring of Inflation Expectations in Latin America 123

on an estimated ar(1) coefficient in a regression on annual inflation
that includes a constant; and 6) real gdp per capita.
The results, shown in Table 5, indicate that, aside from an inter-
cept, only the coefficients for inflation persistence and the it dummy
are statistically significant. it is associated with an increase in the
degree of anchoring of inflation expectations by 0.25, whereas coun-
tries with less inflation persistence are associated with an increase
in the degree of anchoring (the coefficient of –0.768 indicates that
a decrease in inflation persistence from 0.9 to 0.8 corresponds to an
increase in anchoring of 0.08). We obtain similar results when we re-
peat the regression with weights at shorter horizons, such as one year.8
One way to interpret these results is that, even when we control for in-
flation persistence, which is negatively correlated with the it dummy
and anchoring, we still find that it is associated with a significant
increase in the anchoring of inflation expectations.
Table 6 displays second step estimation results where the depen-
dent variable is the estimated standard error of the anchor. Here,
the number of years since the adoption of it and the persistence
of inflation are marginally statistically significant, but the it dum-
my is insignificant.

5. CONCLUSIONS

In this paper, we modeled inflation expectations from Consensus

Forecasts to assess inflation expectations anchoring in Latin Amer-
ica. Our results suggest that most countries do have an inflation an-
chor, and that expectations have become more tightly anchored
through time, consistent with the improving credibility of central
banks’ monetary policy management.
For countries with it, we find that inflation targets are generally
credible, in the sense that the estimated anchors lie within the in-
flation target range for all countries in the most recent sample that
we estimate. Also, the adoption of it is generally associated with
an improvement in the degree of anchoring of expectations, both

8
At a forecast horizon of 12 months, being under an it regime is associated
with an increase in the degree of anchoring of inflation expectations
by 0.25, and a 0.1 drop in inflation persistence is associated with an
increase in the degree of anchoring by 0.09.

124 R. Gondo, J. Yetman

 Table 5
SECOND STEP ESTIMATION RESULTS
Dependent variable: inflation anchor weight ( h = 24 )

Coefficient Standard error

it dummy 0.245 c
0.0617
Years under it 0.00682 0.01385
Inflation mean 4.39e-04 6.41e-04
Inflation standard deviation 4.55e-03 4.34e-03
Inflation ar(1) coefficient –0.768 b
0.343
gdp per capita 4.18e-06 7.30e-06
Constant 1.37c 0.322
R squared within 0.280
Between 0.002
Overall 0.107
F-statistic 4.20

Note: a, b, c indicates statistical significance at the 10%, 5%, and 1% levels,

respectively.

Table 6
SECOND STEP ESTIMATION RESULTS
Dependent variable: standard error of the inflation anchor

Coefficient Standard error

it dummy –1.67e-03 13.2e-03
Years under it –6.03e-03 a
2.92e-03
Inflation mean 2.97e-04 1.95e-04
Inflation standard deviation 5.42e-04 6.25e-04
Inflation ar(1) coefficient 0.0971a 0.0522
gdp per capita 3.68e-06 2.22e-06
Constant –0.0733 0.0551
R squared within 0.176
between 0.0007
overall 0.004
F-statistic 2.58

Note: a, b, c indicates statistical significance at the 10%, 5%, and 1% levels,

respectively.

Anchoring of Inflation Expectations in Latin America 125

in terms of the weight on the anchor increasing and the anchor be-
ing more precisely identified by the data.
In future work, it would be possible to investigate inflation expecta-
tions anchoring further by focusing on the cross-sectional dispersion
of forecasts. For example, Yetman (2017) focuses on forecaster-level
data for Canada and the usa , while Hattori and Yetman (2017) con-
duct a similar exercise for Japan. However, for Latin America, similar
data are only available from Consensus Economics for a limited subset
(seven) of the countries that we study, and the number of forecasters
for most of those countries is limited relative to those other studies.

126 R. Gondo, J. Yetman

ANNEX
Figure A.1
INFLATION FORECASTS AT DIFFERENT HORIZONS
.   FL     15 
 
12 6

10 5

8 4

6 3

4 2

2 1
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1

 
8 6

6 5

4 4

2 3

0 2
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1


5

4 2010 2014
2011 2015
3 2012 2016
2013 2017
2

1
23 21 19 17 15 13 11 9 7 5 3 1
Notes: Horizontal axis represents the forecast horizon, defined as the number of
months before the end of the calendar year being forecast. Dots represent the
realized inflation at the end of year t.
Source: Consensus Economics ©; national data.

Anchoring of Inflation Expectations in Latin America 127

 Figure A.1 (cont.)
INFLATION FORECASTS AT DIFFERENT HORIZONS
.   FL     15 
   
10 9.0

8 7.5

6 6.0

4 4.5

2 3.0

0 1.5

−2 0.0
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1

 
8 8

7 7

6 6

5 5

4 4

3 3

2 2
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1


10

9
2010 2014
8
2011 2015
7 2012 2016
6 2013 2017

4
23 21 19 17 15 13 11 9 7 5 3 1
Notes: Horizontal axis represents the forecast horizon, defined as the number of
months before the end of the calendar year being forecast. Dots represent the
realized inflation at the end of year t.
Source: Consensus Economics ©; national data.

128 R. Gondo, J. Yetman

 Figure A.1 (cont.)
INFLATION FORECASTS AT DIFFERENT HORIZONS
.   FL 
 
50 10

40 8

30 6

20 4

10 2

0 0
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1

  
7.5 8

6.0 6

4.5 4

3.0 2

1.5 0

0.0 −2
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1
2010 2011 2012 2013 2014 2015 2016 2017
Notes: Horizontal axis represents the forecast horizon, defined as the number of
months before the end of the calendar year being forecast. Dots represent the
realized inflation at the end of year t.
Source: Consensus Economics ©; national data.

Anchoring of Inflation Expectations in Latin America 129

 Figure A.1 (cont.)
INFLATION FORECASTS AT DIFFERENT HORIZONS
.   FL  (.)
 
10 10

8 8

6 6

4 4

2 2
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1

 
8 600

6 450

4 300

2 150

0 0
23 21 19 17 15 13 11 9 7 5 3 1 23 21 19 17 15 13 11 9 7 5 3 1
2010 2011 2012 2013 2014 2015 2016 2017
Notes: Horizontal axis represents the forecast horizon, defined as the number of
months before the end of the calendar year being forecast. Dots represent the
realized inflation at the end of year t.
Source: Consensus Economics ©; national data.

130 R. Gondo, J. Yetman

 Figure A.2
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 12)
.        15 
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1990-1997
1992-1999
1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015


1.0

0.8
0.6
0.4

0.2
0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 131

 Figure A.2 (cont.)
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 12)
.        15 
   
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1994-2001
1996-2003
1998-2005
2000-2007
2002-2009
2004-2011
2006-2013
2008-2015
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

132 R. Gondo, J. Yetman

 Figure A.2 (cont.)
ESTIMATED WEIGHT ON INFLATION ANCHOR (h = 12)
.   FL 
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016
 
1.0 1.0

0.8 0.8
0.6 0.6
0.4 0.4

0.2 0.2
0.0 0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016


1.0

0.8
0.6
0.4

0.2
0.0
1993-2000
1995-2002
1997-2004
1999-2006
2001-2008
2003-2010
2005-2012
2007-2014
2009-2016

Notes: Horizontal axis displays the eight-year rolling sample. Periods where no line is
displayed correspond to rolling samples for which no anchor can be identified.
Source: Authors’ calculations.

Anchoring of Inflation Expectations in Latin America 133

References
Capistrán, Carlos, and Manuel Ramos-Francia (2010), “Does Infla-
tion Targeting Affect the Dispersion of Inflation Expecta-
tions?,” Journal of Money, Credit, and Banking, Vol. 42, No. 1,
pp. 113-134.
Cecchetti, Stephen G., and Craig Hakkio (2009), Inflation Targeting
and Private Sector Forecasts, No. W15424, National Bureau of
Economic Research.
Davis, Scott, and Adrienne Mack (2013), Cross-country Variation in
the Anchoring of Inflation Expectations,  Staff Papers, No. 21,
Federal Reserve Bank of Dallas, October.
De Carvalho, Fabia A., and Maurício Soares Bugarin (2006), “Infla-
tion Expectations in Latin America,” Economía, Vol. 6, No.
2, pp. 101-145.
De Pooter, Michel, et al. (2014), “Are Long-term Inflation Expecta-
tions Well Anchored in Brazil, Chile, and Mexico?,” Interna-
tional Journal of Central Banking, Vol. 10, No. 2, pp. 337-400.
Espinosa-Torres, Juan Andrés, Luis Fernando Melo-Velandia, and
José Fernando Moreno-Gutiérrez (2017), “Expectativas de
inflación, prima de riesgo inflacionario y prima de liquidez:
una descomposición del break-even inflation para los bonos
del Gobierno colombiano,” Revista Desarrollo y Sociedad, Vol.
78, pp. 315-365.
Gürkaynak, Refet S., et al. (2007), “Inflation Targeting and the
Anchoring of Inflation Expectations in the Western Hemi-
sphere,” Economic Review-Federal Reserve Bank of San Francisco,
pp. 25-47.
Hattori, Masazumi, and James Yetman (2017), “The Evolution of
Inflation Expectations in Japan,” Journal of the Japanese and
International Economies, Vol. 46, pp. 53-68.
Mehrotra, Aaron, and James Yetman (2018), “Decaying Expecta-
tions: What Inflation Forecasts Tell Us about the Anchoring
of Inflation Expectations,” International Journal of Central
Banking, Vol. 14, No. 5, December, pp. 55-101.
Vicente, José Valentim Machado, and Osmani Teixeira de Carv-
alho Guillen (2013), “Do Inflation-linked Bonds Contain
Information about Future Inflation?,”  Revista Brasileira de
Economia, Vol. 67, No. 2, pp. 251-260.

134 R. Gondo, J. Yetman

Yetman, James (2017), “The Evolution of Inflation Expectations
in Canada and the us,” Canadian Journal of Economics/Revue
canadienne d’économique, Vol. 50, No. 3, pp. 711-737.

Anchoring of Inflation Expectations in Latin America 135

The Time-Varying Degree of
Inflation Expectation
Anchoring in Bolivia
Mauricio Mora Barrenechea
Juan Carlos Heredia Gómez
David Esteban Zeballos Coria

Abstract
This chapter analyzes the time-varying degree of inflation expectations an-
choring in Bolivia and, more precisely, whether inflation expectations have
been in line with the inflation objectives announced by the Banco Central
de Bolivia (central bank of Bolivia, bcb) and if they have become better an-
chored over time. Two considerations are particularly relevant in this regard.
First, the main sources of information are the bcb survey and Focus Econom-
ics survey, which only have data for short- and medium-term inflation expec-
tations. Second, monetary policy in Bolivia is under a monetary-targeting
regime, so bcb projections represent the main references. The anchoring de-
gree analysis of short-term inflation expectations was performed considering
bcb projections, while the medium-term analysis used an implicit inflation
target. In both cases, the results indicate there is a high degree of anchoring

M. Mora <[email protected]>, Research Economist; J. Heredia <jheredia@bcb.

gob.bo>, Research Economist; and D. Zeballos <[email protected]>, Head, Cen-
tral Banking Research Department, Banco Central de Bolivia (bcb). This research
was undertaken within the framework of cemla’s Joint Research Program 2017, coor-
dinated by the Banco de la República (Colombia). The paper was originally published
as part of the working paper series of the Inter-American Development Bank (idb).
The authors acknowledge the advice and technical advisory provided by the Financial
Stability and Development (fsd) Group of the idb in the process of writing this docu-
ment. We are grateful to our academic advisor, Professor Olivier Coibion (University
of Texas Austin and nber) for his advice and suggestions in the development of the
present research. We also thank Sergio Lago Alves (Banco Central do Brasil) for his
comments and suggestions at the xxii Annual Meeting of the Central Bank Research-
ers Network and Sandra Vásquez Willcarani (bcb) for her research assistance. Addi-
tionally, we thank John Dunn Smith (idb), Sophia Vargas (cemla) and Diego García
(cemla) for their editing suggestions. The opinions expressed in this publication
are those of the authors and do not reflect the views of cemla, the fsd group, the In-
ter-American Development Bank or the Banco Central de Boliviaor its authorities.

137
of inflation expectations in Bolivia, especially during the last four years.
This study considers information from July 2005 to June 2017, with monthly
frequency.
Keywords: inflation expectations, anchoring degree, monetary-targeting
regime, bcb projections, time-varying parameters model.
jel classification: E31, E52, E58, C32.

1. INTRODUCTION

T
he analysis of the behavior of the expectations of inflation
of economic agents has been heavily studied in the past, espe-
cially with regards to the degree of anchoring of expectations,
understood as the ability of monetary policymakers to manage infla-
tion expectations (King, 2005). Theoretical literature and monetary
policymakers agree that the anchoring of inflation expectations is of
high importance in maintaining price stability, and expectations
by private agents play an important role in macroeconomics since
they can be a determinant of macroeconomic performance. Infla-
tion expectations not only reflect private agents’ perceptions about
future inflation, but also directly impact current and future inflation.
Relatedly, a central bank should focus on the management of pri-
vate expectations through communication for two reasons (Hubert,
2015). First, the expectations channel is one of the subtlest channels
of monetary policy, because it depends on private agents’ interpreta-
tion. As King (2005) notes, “because inflation expectations matter
to the behavior of the households and firms, the critical aspect of mon-
etary policy is how decisions of the central bank affect those expecta-
tions.” Second, given the delay between policy actions and their real
effects on macroeconomic variables, central bank communication
provides policymakers with a way to promptly affect private expec-
tations to shorten the transmission lag of monetary policy.
According to Blinder et al. (2008), central bank communication
can take different forms: statements, minutes, interviews, speech-
es, or internal macroeconomic forecasts. We will focus on the latter
instrument of communication because monetary policy in Bolivia
is under a monetary-targeting regime. However, although the Ban-
co Central de Bolivia(bcb, for its acronym in Spanish) does not have
an explicit inflation target, its active communication policy and pro-
jections, announced twice per year in its Monetary Policy Report,

138 M. Mora, J.C. Heredia, D. Zeballos

become important reference points for agents at the time of form-
ing their expectations.
Since the inflation expectations of private agents are not generally
known, they can by approximated by: i) surveys of inflation expecta-
tions of professional forecasters or households and ii) market-based
measures of inflation expectations. In the present document, we use
information from the survey conducted by the bcb for the period
between July 2005 and June 2017. This is a monthly survey of expec-
tations for the rates of inflation (among other variables) for several
short-term horizons. Additionally, we use information from the Latin
Focus Consensus Forecast report of Focus Economics to gather data
regarding medium-term inflation expectations in Bolivia.
There are not many studies that analyze the degree of anchoring
of expectations in Bolivia. We can mention the work of Cerezo and He-
redia (2013), who found that there was a greater degree of anchoring
of inflation expectations in recent years than between 2008 and 2010.
Nevertheless, they also found that expectations were not rational,
suggesting that expectations reflect backward-looking behavior.
The main objective of this paper is to analyze the time-varying de-
gree of inflation expectations anchoring in Bolivia. More precisely,
we aim to assess whether inflation expectations have been in line
with the inflation objectives announced by the bcb, and if they have
become better anchored. The anchoring degree analysis of short-
term inflation expectations was performed considering the bcb
projections, while the medium-term analysis used an implicit infla-
tion target. In both cases, the results indicate there is a high degree
of anchoring of inflation expectations in Bolivia, especially during
the last four years.
In the next section, there is a brief analysis about the behavior
of inflation expectations in Bolivia and their stability. Subsequently,
we show the results of the estimated models, analyzing the behavior
of short-term inflation expectations with respect to the bcb projec-
tions, past inflation and other variables that could affect the forma-
tion of expectations. Then, the results of the analysis of medium-term
expectations are presented. Finally, we present our conclusions.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 139

2. INFLATION EXPECTATIONS IN BOLIVIA

In order to evaluate the evolution of the degree of anchoring of in-

flation expectations in Bolivia, we consider data from the survey
conducted by the bcb for the period between July 2005 and June
2017.1 This monthly survey contains information of the expectations
of economic analysts, academics, members from financial sector
and private business in Bolivia about the future behavior of eco-
nomic variables of interest for bcb authorities such as inflation, ex-
change rate, gdp growth, trade balance, and fiscal balance, among
others. In the case of inflation expectations, the survey focuses on: i)
monthly inflation expected by the end of current month, ii) year-on-
year inflation expected by the end of current year, iii) year-on-year
inflation expected by the end of next calendar year and, iv) one year-
ahead inflation expectations.
It is important to mention that, unlike surveys available in other
countries, the bcb survey does not take into account long-term in-
flation expectations (e.g., five years-ahead expectations). Certainly,
this issue restricts, to a certain extent, the variety of econometric
analyses that can be implemented. Moreover, in Bolivian financial
markets, no inflation-indexed bonds are traded, a feature that makes
it impossible to estimate break-even inflation rates for this economy,
which are a measure of inflation expectations widely used in topi-
cal literature.
Our analysis will be focused on approximately the last 12 years.
During this period, important shocks (mainly foreign and supply-
side shocks) hit the Bolivian economy and affected domestic infla-
tion behavior. These shocks, along with some developments observed
in monetary markets and the macroeconomic framework and chang-
es in the dynamics of the local economy, may have affected the de-
gree of anchoring of inflation expectations.
Between 2007 and 2008, the Bolivian economy went through an in-
flationary process triggered especially by a shock in international
food and energy prices, reaching double-digit inflation rates not ob-
served since the beginning of the previous decade. In this period,
expectations of agents were significantly exacerbated, with median
inflation expectations placing themselves above observed inflation
rates. Subsequently, a process of disinflation took place associated

1
Information for previous periods is not available.

140 M. Mora, J.C. Heredia, D. Zeballos

with the global financial crisis in 2009, an episode characterized
by a high degree of uncertainty about the performance of the world
economy, with effects on Bolivian economic activity. Within this
setting, inflation expectations followed a downward trend as well,
although their decline was more moderate (Figure 1a).
In the period 2010-2011, new inflationary upsurges were noticed,
although of smaller scale and persistence with respect to previous
years. In this period, the main explanatory factors were a new re-
bound in the international prices of commodities and an increase
in domestic prices caused by speculative activities after the Govern-
ment temporarily readjusted fuel prices.2 Beginning in 2012, the be-
havior of inf lation was characterized by moderate f luctuations,
exhibiting a downward trend during the last two years. In recent
years, temporary hikes can be observed in the behavior of inflation,
which are explained by increases of the prices of some foods, whose
supply was affected by adverse weather events (like frosts, floods
and droughts, among others). The trajectory of inflation expectations
reflected a path similar to that of inflation between 2005 and 2011,
although from 2012 onward it displayed stable behavior, with a me-
dian generally above observed inflation (Figure 1b).
The stability of inflation expectations is an important issue to con-
sider, since it represents an initial approximation to its anchorage.
A useful way to measure stability is through its degree of dispersion3
(disagreement or uncertainty). Less dispersion can be interpreted
as a signal of a better anchoring of inflation expectations.4 For this
purpose, we chose the cross-sectional standard deviation of infla-
tion expectations (Figure 2). A higher degree of dispersion can be
observed between mid-2007 and early 2011. 5 Afterwards, the degree

2
It is important to note that fuels are subsidized in Bolivia. In December
2010, the government decided to withdraw the subsidy which gener-
ated an environment of uncertainty, causing expectations of inflation
to increase. Although the measure was eliminated shortly, important
second-round effects were generated during the following months.
3
Although, the dispersion of expectations in a survey is a measure
of heterogeneity of beliefs rather than a measure of uncertainty ( imf,
2016), both tend to move together (Gürkaynak and Wolfers, 2007).
4
Dovern, Fritsche and Slacalek (2009), Capistran and Ramos-Francia
(2010), Siklos (2013), and Ehrmann (2015).
5
During this period, Bolivian economy went through different circum-
stances that caused strong inflationary pressures: increased international

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 141

 0
2
4
6
8
10
12
14
16
18

0
5
10
15
20
25
30
35
Jun 2005 Jun 2005
Oct 2005 Oct 2005

Percentage
Percentage
Feb 2006 Feb 2006
Jun 2006 Jun 2006
Oct 2006 Oct 2006
Feb 2007 Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009

Max-Min range
Oct 2009 Oct 2009
Feb 2010 Feb 2010
Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011

142 M. Mora, J.C. Heredia, D. Zeballos

Jun 2011 Jun 2011
Figure 1

Oct 2011 Oct 2011

Feb 2012 Feb 2012
Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
.  ,   

Oct 2013 Oct 2013

Note: National Statistics Institute and Central Bank of Bolivia.

Feb 2014 Feb 2014

.  -  

Jun 2014 Jun 2014
Oct 2014 Oct 2014
Feb 2015 Feb 2015
Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016

Median inflation expectations

Oct 2016
Feb 2017 Feb 2017
Jun 2017
EVOLUTION OF HEADLINE INFLATION AND INFLATION EXPECTATIONS

Jun 2017
 Figure 2
CROSS-SECTIONAL STANDARD DEVIATION OF ONE YEAR-AHEAD
INFLATION EXPECTATIONS

Percentage
7

0
Jun 2005
Oct 2005
Feb 2006
Jun 2006
Oct 2006
Feb 2007
Jun 2007
Oct 2007
Feb 2008
Jun 2008
Oct 2008
Feb 2009
Jun 2009
Oct 2009
Feb 2010
Jun 2010
Oct 2010
Feb 2011
Jun 2011
Oct 2011
Feb 2012
Jun 2012
Oct 2012
Feb 2013
Jun 2013
Oct 2013
Feb 2014
Jun 2014
Oct 2014
Feb 2015
Jun 2015
Oct 2015
Feb 2016
Jun 2016
Oct 2016
Feb 2017
Jun 2017
Source: Authors’ calculations based on  data.

of dispersion tended to moderate, with a slight rebound between

2013 and 2014. 6 Except for those years, a lower degree of uncer-
tainty about rates of inflation expected by economic agents can be
observed beginning in 2012. Hence, the trajectory of expectations
observed in recent years suggests a strengthening of their degree
of anchoring over time.
Inflation expectations in Bolivia seem to be more homogeneous
in recent years. This homogeneity may reflect the existence of a com-
mon reference point that is taken into account by economic agents
while forming their inflation expectations. One of these possible

commodity prices, economic acceleration, regulated price adjustments

and others. All these factors created an environment of uncertainty
regarding the future level of prices.
6
In 2013 and 2014 inflationary pressures were observed due to the rise
in prices of some foods because adverse weather events reduced agri-
cultural supply in local markets.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 143

 Figure 3
INFLATION EXPECTATIONS, HEADLINE INFLATION
AND BCB PROJECTION

Percentage
16
14
12
10
8
6
4
2
0
2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017

Projection range BCB projection

Headline inflation Inflation expectations

Note: Inflation expectations are computed as the mean of inflation expectations for a
given year.  projections and the projection range are computed as the average of
the inflation projections announced at the beginning and middle of the year.
Source: National Statistics Institute and Central Bank of Bolivia.

reference points is the inf lation projection of the Central Bank

announced in its Monetary Policy Report twice per year. Between
2005 and 2011, headline inflation and inflation expectations ended
the year above the bcb projection, except for in 2009, and, in some
cases, even above the projected range (Figure 3). The shocks noted
above generated an environment of uncertainty, making it difficult
for the bcb and private agents to project inflation. It seems that dur-
ing this time economic agents mainly considered past headline in-
flation or possibly other variables to formulate their expectations.
In 2012, this situation changed, a result of the expectations of the
agents landing closer to the bcb projection, especially between 2015
and 2017. This could indicate that there is a significant degree of an-
choring of expectations in recent years. This item will be studied
empirically in the next section of the paper.

144 M. Mora, J.C. Heredia, D. Zeballos

3. EMPIRICAL ANALYSIS OF THE SHORT TERM

While this study focuses mainly on assessing the anchoring of short-

term inflation expectations over time, it should nonetheless be noted
that the behavior of short-term expectations is also relevant to poli-
cymakers. According to Łyziak and Paloviita (2016), the credibility
of a central bank should not only be measured in terms of its abil-
ity to anchor long-term expectations, but also in terms of its abili-
ty to affect short and medium-term expectations, since these have
an important role in wage adjustments and price-setting by firms.
In addition, another point that must be emphasized is that in Bo-
livia, monetary policy is not based on an inflation-targeting regime.
On the contrary, the monetary regime of Bolivia is one of mone-
tary-targeting. However, although the bcb does not have an explic-
it inflation target, its active communication policy and projections
announced twice per year in its Monetary Policy Report become
important reference points for agents at the time of forming their
expectations.
In a similar vein, the work of Anderson and Maule (2014) assess-
es the anchoring of short-term inflation expectations in the United
Kingdom considering the Bank of England’s inflation projections
as one of its determinants. Likewise, Hubert (2015) showed that
the projections of the European Central Bank play an important
role in the formulation of short and medium-term expectations
in the Eurozone.
In this context, an econometric model is estimated to analyze
the evolution of the degree of anchoring of inflation expectations.
Before we start, two aspects must be considered. First, most of the
surveys contain “fixed-event” (FE) information (i.e., information
always points to a single moment, like the end of the current or next
calendar year) on the expectations of different variables, so they con-
stitute an abundant source of information. Notwithstanding their
availability, this paper requires the use of “fixed-horizons” (FH) vari-
ables (i.e., those that keep an n horizon, such as 12 months ahead)
with the purpose of working with econometric models because fore-
casting horizons of FE forecasts (or expectations) vary from month
to month (the horizon shrinks as time passes).
We, therefore, employ a technique that allows us to use the FE in-
formation. Following Dovern, Fritsche and Slacalek (2009), we create

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 145

a FH variable as a weighted average of FE forecasts; the weights are de-
termined by the number of months forecasted in both the current
and subsequent years. Denote Fy 0,m ,yo (x ) as the FE forecast of variable
fe

x for year Y0 made in month m of year Y0 and Fy 0,m ,y1 (x ) the FE forecast
fe

of variable x for year Y1 made in month m of year Y0. Then Fy 0,m ,12 (x )
fh

represent the FH forecast 12 months ahead made in month m of year Y0.

We approximate the FH forecast for the next 12 months as an average
of the forecast for the current and next calendar year weighted by their
share in forecasting horizon:

12 − m + 1 m −1
Fy 0,m ,12 (x ) = * Fy 0,m ,yo (x ) + * Fy 0,m ,y1 (x ) (1)
fh fe fe
1
12 12

According to Winkelried (2017), a survey that registers FE expecta-

tions for horizons Y0 and Y1 does contain information for expectations
at any intermediate horizon; for instance, expectations for 12 months
ahead are implicitly contained in current and next year forecast. There-
fore, the inflation expectation obtained with this technique (Figure 4a)
is equal to the inflation expectation one year ahead shown in Figure 1b.
This technique was also used with the information from the bcb pro-
jection for the current and next calendar year (Figure 4b).
A second point we should consider is the effect of new inflation infor-
mation on the formulation of economic agents’ expectations. Accord-
ing to Hubert (2015), the effects of central bank inflation projections
on private agents are stronger at the beginning of each year than at the
end, when much more information is available on the actual behavior
of inflation. Consequently, this document mainly considers the projec-
tions announced by the bcb at the beginning of each year. However,
a second variable was created to reflect the bcb projection, which also in-
cludes updates of the projection announced after the first semester of ev-
ery year, mainly with the purpose of performing robustness analysis.7

7
Annex 1 presents the evolution of the bcb projection for the current
and next calendar year separated, and the bcb inflation projections con-
structed using the technique of equation (1) that includes the updates
at middle of each year.

146 M. Mora, J.C. Heredia, D. Zeballos

 0
2
4
6
8
10
12
14
16
18
20

0
1
2
3
4
5
6
7
8
9
Jun 2005
Jun 2005 Oct 2005
Oct 2005 Percentage

Percentage
Feb 2006
Feb 2006 Jun 2006
Jun 2006 Oct 2006
Oct 2006 Feb 2007
Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010
Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011
Jun 2011 Jun 2011
Figure 4

Oct 2011 Oct 2011

Source: Authors’ calculations based on  data.

Feb 2012 Feb 2012
Jun 2012 Jun 2012

.   

Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
Oct 2013 Oct 2013
Feb 2014 Feb 2014
Jun 2014 Jun 2014
.   -  

Oct 2014 Oct 2014

Feb 2015 Feb 2015
FIXED HORIZONS VARIABLES FOR SHORT TERM

Jun 2015 Jun 2015

Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017
Jun 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 147

Jun 2017
3.1 bcb Projection against Headline Inflation
In this section, the specification of the model is based on the method-
ology applied by Łyziak and Paloviita (2016), who estimate different
models to measure the degree of anchoring of inflation expectations
for the Euro Zone. The specified equation is as follows:

2 π te|t +n = γ proj π tproj π

+n + γ π t −1 + µt (2 )

where:

γ proj + γ π = 1

e
where πt t +n represents the inflation expectations in period t for
proj
the horizon; t +n;πt+n is the inflation projection for the horizon;
t +n;πt−1 represents observed inflation lagged one period and n
is equal to 12 months. Additionally, an error term (µt ) is included
in the equation. Note that, by construction, the sum of the coefficients
of the model must be equal to one. If the coefficient γ proj reaches
a value as close as possible to one, it would reflect a significant de-
gree of anchoring of expectations.
According to Strohsal, Melnick and Nautz (2015), the central
bank’s credibility can be gained, but it can also be lost. As a conse-
quence, the degree of inflation expectations anchoring might not be
constant over time. Meanwhile, Orphanides (2015) once pointed
out that inflation expectations are well anchored until they are not.
This means that the degree of anchoring can change over time,
so using a model with constant parameters may not be the best op-
tion. In that sense, in the present document a time-varying param-
eter model is estimated, in line with other works such as Demertzis,
Marcellino and Viegi (2012) and Strohsal, Melnick and Nautz (2015).
In the name of simplification, we assume that the state param-
eters follow a random walk process. We use the Kalman filter (Kal-
man, 1960) to compute the one-step ahead estimates of the means
and variances 8 of the states by maximum likelihood.

8
During the estimation, the variances parameters are expressed in expo-
nential form to ensure that the variances themselves are non-negative.

148 M. Mora, J.C. Heredia, D. Zeballos

 The results for this first estimation showed that the coefficient
γ proj attained a value close to 0.80, which implies that there is a sig-
nificant degree of anchoring of short-term expectations in Bolivia
(Table 1). On the other hand, the coefficient γ π for lagged inflation
is significant at 10 percent.

Table 1
RESULTS FROM TIME-VARYING PARAMETER MODEL 1

Projections Past Inflation

Coefficient 0.80 0.20
rmse 0.13 0.12
z-Statistic 6.28 1.66
p-value (0.00) (0.09)

A strength of state-space models is that they permit observe the evo-

lution of the different coefficients over time. It can be seen that the val-
ue of the coefficient γ proj was negative between mid-2005 and late
2008 (Figure 5a), in line with the overshooting of expectations that
took place then. In this period the anchoring degree of expecta-
tions was null. Later, an improvement in the degree of anchoring
of expectations can be observed as of 2009,9 reaching values near
0.6 until mid-2010, when it fell again because of a new inflationary
rebound. The bcb projections coefficient reflected stable behavior
around 0.25 from 2012 until mid-2014. In July 2014 this coefficient
begins important growth, reaching 0.80 in the last two years under
consideration.

9
It is also interesting to note that the degree of anchoring of expectations
did not decline in time of the international financial crisis, something
that was analyzed in different documents such as Galati, Poelhekke
and Zhou (2011), Autrup and Grothe (2014), and Nautz and Strohsal
(2015). However, this does not imply that in that period there was a
greater degree of central bank credibility.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 149

 −0.8
−0.6
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2

−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
1.6
1.8
Jun 2005
Jun 2005 Oct 2005

Percentage
Oct 2005

Percentage
Feb 2006
Feb 2006 Jun 2006
Jun 2006 Oct 2006
Oct 2006 Feb 2007
Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010

Note: Smoothed coefficient ± 2RMSE.

Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011

150 M. Mora, J.C. Heredia, D. Zeballos

Jun 2011 Jun 2011
Figure 5

Oct 2011 Oct 2011

Feb 2012 Feb 2012
Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
.    (γ proj)

Oct 2013 Oct 2013

Feb 2014 Feb 2014
Jun 2014 Jun 2014

.     (γπ)

Oct 2014 Oct 2014
EVOLUTION OF COEFFICIENTS IN MODEL 1

Feb 2015 Feb 2015

Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017
Jun 2017
Jun 2017
 In the case of headline lagged inflation (Figure 5b), the highest
values were observed between 2007 and 2008 when it reached values
higher than one, which shows the exacerbation of expectations dur-
ing this time. Later, values tended to decrease and seemingly lose
importance in the formulation of agents’ expectations.
Annex 2 contains the results using the updated bcb projection
under this specification. The results obtained are similar to those
found with Model 1; there also exists a significant degree of anchoring
of short-term expectations with respect to updated bcb projections.
These first results showed that short-term inflation expectations
are anchoring,10 since the bcb projection had a bigger impact on eco-
nomic agents than headline inflation. However, information from
other variables may affect the formulation of expectations.

3.2. bcb Projection against Other Variables

Economic agents are exposed to a great diffusion of local and inter-
national information, especially in light of advances in communi-
cation. This means that the behavior of other variables may affect
the formulation of private agents’ expectations. Relatedly, there
exists a strand of literature that investigates how inflation expecta-
tions respond to macroeconomic news (Beechey and Wright, 2009,
and Beechey, Johannsen and Levin, 2011), though with a long-term
focus. Since short-term inflation expectations respond to observed
inflation, they should be more sensitive to changes in other variables.
With the objective of analyzing the effects of information from other
variables on the behavior of inflation expectations, in this section
we make estimates with different models, including a broad set of ex-
ternal variables in addition to bcb projections and observed inflation.

3 +n + β2π t −1 + βm X t ( L ) + µt
π te|t +n = β1π tproj (3)

e
Once again πt t+n represents one year-ahead inflation expectation;
πtproj
+n is the bcb inflation projection for the horizon t + n; where n is
equal to 12 months and πt−1 represents observed inflation lagged
one period. We include X t , which represents the battery of different

10
This result does not imply that inflations expectations are rational; that
issue is not analyzed in this study.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 151

external variables used to estimate the models; some of them will
be introduced with lags. Additionally, an error term (µt ) is includ-
ed in the equation.
In order to guide our selection of external variables, we follow
the works of Celasun, Gelos and Prati (2004), Cerisola and Gel-
os (2005), Bevilaqua, Mesquita and Minella (2007), and Carrasco
and Ferreiro (2013). The variables chosen were output gap,11 one year-
ahead expectations of nominal depreciation,12 and expectations
of fiscal balance in percent of gdp.13
We also incorporate other variables that may be related to the
characteristics of the Bolivian economy, such as shocks from climatic
events14 (as food represents an important part of the cpi in Bolivia,
nearly 28 percent) and external shocks15 (as previously noted, the Bo-
livian economy was exposed to major external shocks during the last
decade)16. In the case of inflation expectations and bcb projections,
we use the variables created in the previous section.

11
The information was obtained from the Global Index of Economic
Activity (igae, for its acronym in Spanish) which represents a proxy
variable of economic activity in monthly frequency, minus its trend
value (where the trend is approximated through a Hodrick-Prescott
filter).
12
Most of the documents use movements in the nominal exchange rate.
However, in Bolivia, the exchange rate has been fixed since 2011, and it
is an important variable since it works as a nominal anchor. For this
reason, we use economic agents’ expectations of future depreciation.
13
We use expectations of fiscal balance as a proxy of the primary fiscal
balance in order to have a variable with monthly data. For this case
and the expectations of nominal depreciation we use the information
from the bcb survey employing the technique of equation (1).
14
We employ the Multivariate enso (El Niño/Southern Oscillation)
Index (mei) of the United States National Oceanic and Atmospheric
Administration (noaa) as a proxy variable to reflect the changes in the
weather condition.
15
The Food Price Index of the International Monetary Fund (imf) was con-
sidered. International food price shocks have a significant impact
on inflation in Bolivia because of the high share of food in the country’s
cpi.
16
We also use other variables like igae growth YoY, economic agents’
expectations of economic growth and the imf international energy
price index; none of these, however, showed satisfactory results.

152 M. Mora, J.C. Heredia, D. Zeballos

 As in the previous section, for the estimation we use time-vary-
ing parameter models with different specifications, and we suppose
that the state parameters of all the variables follow a random walk
process. The results of the different models’ specifications can be
observed in Table 2.

Table 2
DETERMINANTS OF SHORT-TERM INFLATION EXPECTATIONS

Model 2 Model 3 Model 4 Model 5

bcb projection 0.74 0.75 0.74 0.73
(0.00) (0.00) (0.00) (0.00)
Inflation (t−1) 0.29 0.29 0.28 0.28
(0.06) (0.07) (0.08) (0.09)
Nominal depreciation 0.20 0.21 0.21 0.20
expectations (0.76) (0.75) (0.75) (0.77)
International food 0.01 0.01 0.01 0.004
price index (t−1) (0.77) (0.75) (0.75) (0.81)
Output gap (t−2) 0.29 0.30
(0.36) (0.35)
Climatic events −0.03 −0.06
(0.91) (0.82)
Fiscal Balance/gdp −0.03
Expectations (0.43)

Note: The values in parentheses represent the p-values.

We created four different models, and in each one the bcb pro-
jection remained the most important explanatory variable with
coefficients around to 0.74, close to those obtained in Section 3.1.
Also, lagged inflation was significant (at 10 percent) in all models,
with a coefficient near 0.28. The remaining variables were not sta-
tistically significant. The least relevant were the international price

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 153

food index,17 expectations of the fiscal balance in percent of gdp,
and the climatic event variable.18 The lagged output gap displayed
a high coefficient, but it was not significant.19
The evolution of the coefficients of the bcb projection and head-
line inflation is similar to that found in Section 3.1 (Figure 6). It can
be observed that headline inflation had a greater impact on inflation
expectations between 2005 and 2010, while bcb projections had a
greater effect in recent years. The effect of bcb projections at the
beginning of the sample, however, are around 0.45 (in the model
used in Section 3.1, the coefficient was close to 0 during this period).
Meanwhile, the coefficient of observed inflation was near 0.65 (in
the results of previous model, it was near 1).
It seems that the inclusion of other variables simply tended to re-
duce the explanatory value of observed inflation over inflation expec-
tations. Most of the additional variables also work as determinants

17
During 2007-2008 and 2010-2011, international food prices rose expo-
nentially, so national producers decided to sell most of their production
to foreign markets, generating a shortage in local markets. This caused
an increase in the prices of some foods (like sugar) or inputs (such
as soybeans that are important for poultry farms), which translated into
an inflationary process. However, in recent years international food
prices have fallen and shown less dynamism; in addition, limits were
applied to exports in order to ensure supply to local markets. These
factors may have diminished the index’s relationship with local food
prices, so this variable turned out to be not significant in the formula-
tion of expectations.
18
The sign of the coefficient of climatic events was negative in the models.
Since the mei was used as a proxy variable, when it presents negative
values it denotes the presence of the La Niña phenomenon. This phe-
nomenon can generate heavy rains, floods and landslides, especially
in the eastern part of Bolivia, where most of the agricultural produc-
tion is located. Therefore, it can be inferred that when the La Niña
phenomenon occurs, the inflation expectations of economic agents
would increase, although not significantly. This variable’s lack of sig-
nificance is possibly explained by the fact that the effects of climatic
events generally affect food prices for no longer than three months;
prices subsequently decrease as supply normalizes in local markets.
Economic agents thus do not expect there to be a constant rise in prices
in following months.
19
It is worth mentioning that, unlike the rest of the variables, the igae
information is available to the general public with a greater lag time.
In that sense, the output gap entered the model with a lag of two
periods.

154 M. Mora, J.C. Heredia, D. Zeballos

 0.0
0.2
0.4
0.6
0.8
1.0
1.2

−0.2
0.0
0.2
0.4
0.6
0.8
1.0
Jun 2005
Jun 2005 Oct 2005
Oct 2005 Percentage

Percentage
Feb 2006
Feb 2006 Jun 2006
Jun 2006 Oct 2006
Oct 2006 Feb 2007
Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010
Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011
Jun 2011 Jun 2011
Figure 6

Oct 2011 Oct 2011

Note: Smoothed coefficient ± 2RMSE of Model 4.

Feb 2012 Feb 2012
Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
.   

Jun 2013 Jun 2013

Oct 2013 Oct 2013
Feb 2014 Feb 2014

.    

Jun 2014 Jun 2014
Oct 2014 Oct 2014
EVOLUTION OF COEFFICIENTS IN MODEL 2

Feb 2015 Feb 2015

Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017
Jun 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 155

Jun 2017
of headline inflation; this could be the reason why none of them
are significant, since their impacts are already contained in the path
of the inflation. The evolution of this last variable reflects the im-
pacts of imported inflation, demand pressures or climatic events.
Therefore, the agents maybe only need to see the path of inflation,
which already includes a lot of additional underlying information.
A special analysis deserves depreciation expectations, although
these were found to be non-significant, there was a time when they
had a more relevant role. The exchange rate in Bolivia has been un-
der a crawling-peg regime since the late 1980s, and during the 1990s
the local currency was continually depreciated in order to maintain
the country’s external competitiveness. This caused a significant
process of dollarization (Berg and Borensztein, 2000), and a high
pass-through effect (Laguna, 2010). In addition, in such a situation
the population becomes accustomed to seeing depreciation as a
normal process of the economic system (Humérez and De la Barra,
2007). However, this pattern changed radically after 2006. In 2007
and 2008 the local currency appreciated in order to mitigate the ef-
fects of the external environment on internal prices (Figure 7b). This
measure had the effect of reducing expectations of inflation (Fig-
ure 7a), illustrating the important role of exchange policy in main-
taining price stability.
Since 2011 the exchange rate has remained stable in order to an-
chor expectations and contain external inflationary pressures. This
may have caused agents to stop considering the exchange rate as a rel-
evant variable for the formation of their expectations in recent years.
The inclusion of other variables did not affect the previous re-
sults from Section 3.1, and it supports the possibility that short-term
inflation expectations are anchoring in Bolivia. However, it would
be good to analyze whether bcb announcements have effects on the
inflation expectations of a longer horizon, such as the medium term.

4. EMPIRICAL ANALYSIS IN THE MEDIUM TERM

Although, our main analysis has been done with the bcb survey and,
therefore, with short-term information; there are other sources
where anyone can find information on the expectations of economic
agents. Most of the research papers on this topic consider data from
international private companies that conduct surveys on different

156 M. Mora, J.C. Heredia, D. Zeballos

 −1.5
−1.0
−0.5
0.0
0.5
1.0
1.5

−10
−8
−6
−4
−2
0
2
4
6
8
10
Jun 2005
Jun 2005 Oct 2005
Percentage

Oct 2005

Percentage
Feb 2006
Feb 2006 Jun 2006
Jun 2006 Oct 2006
Oct 2006 Feb 2007
Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010
Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011
Jun 2011 Jun 2011
Figure 7

Oct 2011 Oct 2011

Note: Smoothed coefficient ± 2RMSE of Model 2.

Feb 2012 Feb 2012
Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
EXCHANGE RATE IN BOLIVIA

Oct 2013 Oct 2013

Feb 2014 Feb 2014
Jun 2014 Jun 2014
Oct 2014 Oct 2014
.      2

Feb 2015 .         Feb 2015
Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 157

Jun 2017
Jun 2017
variables in a large number of countries. In this case we choose to use
the information provided by the Latin Focus Consensus Forecast20
report from Focus Economics.21 While the large sample size allows
us to study the expectations of private agents, we chose this database
mainly because it offers information not only on forecasts for the
current and next calendar year, but also for years further ahead.22
In order to compare the information offered by the Focus Econom-
ics survey with the bcb survey, we use the technique from equation
(1) in Section 3 to transform the data of inflation expectations for the
current and next calendar year. The series obtained reflect similar
behavior in general terms (Figure 8). Between 2007-2008 and 2010-
2011 both series show an increase, although one of less magnitude
in the case of Focus Economics expectations. Since 2012, both series
have stabilized, except for a slight increase in bcb expectations be-
tween 2013 and 2014, and from 2015 on they present similar values.
By performing a cross correlation analysis considering the whole sam-
ple (July 2005 - June 2017), a high level of correlation (0.92) was ob-
tained. Therefore, the Focus Economics information on inflation
expectations can be considered a complement to bcb survey data.
The forecast information of interest in the Focus Economics sur-
veys, conducted with a monthly frequency, is that from April 2010.23
We gathered information for the current year, the next calendar year,
and the third, fourth, and fifth years ahead, so we have data on infla-
tion expectations up to five years ahead. Although the information

20
The Latin Focus Consensus Forecast report is a monthly publication,
which contains macroeconomic projections from nearly 200 different
sources. It covers approximately 30 macroeconomic indicators per coun-
try for a five-year forecast horizon including economic activity (GDP),
industrial production, business confidence, consumer confidence,
inflation, monetary policy decisions and exchange rate movement.
21
Focus Economics is a company that has information on economic fore-
casts for many key indicators in 127 countries. Its reports draw on many
economic and commodities price forecasts and on economic analysts
around the world.
22
There exist other institutions that provide information about economic
forecast; one of the most famous is Consensus Economics. Nevertheless,
in the case of Bolivia its report has only forecast information for the
current and next calendar year of the variables of interest for the pres-
ent document.
23
There exists forecast information for the current and next calendar
year for a longer period, but, not for the rest of the years.

158 M. Mora, J.C. Heredia, D. Zeballos

 Figure 8
EVOLUTION OF INFLATION EXPECTATIONS FROM BCB SURVEY
AND FOCUS ECONOMICS SURVEY

Percentage
20
18
16
14
12
10
8
6
4
2
0
Jun 2005
Oct 2005
Feb 2006
Jun 2006
Oct 2006
Feb 2007
Jun 2007
Oct 2007
Feb 2008
Jun 2008
Oct 2008
Feb 2009
Jun 2009
Oct 2009
Feb 2010
Jun 2010
Oct 2010
Feb 2011
Jun 2011
Oct 2011
Feb 2012
Jun 2012
Oct 2012
Feb 2013
Jun 2013
Oct 2013
Feb 2014
Jun 2014
Oct 2014
Feb 2015
Jun 2015
Oct 2015
Feb 2016
Jun 2016
Oct 2016
Feb 2017
Jun 2017
BCB expectations FE expectations

Source: Authors’ calculations based on  and Focus Economics data.

is on fixed-event variables, in order to work with these data we also

convert them into fixed-horizon variables using the technique from
equation (1) in Section 3. We end with information on inflation ex-
pectations for the current year (first), the next calendar year (sec-
ond), and the third and fourth years (Figure 9a). The last years would
be used to study the degree of anchoring in the medium term.24
The four variables show high values between 2011 and the begin-
ning of 2012, and later they reflect more moderate behavior, similar
to that observed with expectations from the bcb survey. A rebound

24
Although most of the literature defines the medium term as beginning
with the fifth year ahead (see, Carrasco and Ferreiro, 2013; imf, 2016),
this document defines the medium term as beginning with the second
year ahead, like Łyziak and Paloviita, 2016.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 159

can be observed by the end of 2015 for all cases, except the first year.
In the last six months, the inflation expectations at the second, third,
and fourth years stabilize around 4.78 percent, while the expecta-
tions for the present year (first year) fall to 4.31 percent.
In the case of bcb projections, we have the projections for the
current and next calendar year from the Monetary Policy Reports.
The bcb does not undertake projections for longer periods in their
reports, which poses a challenge for analyzing the degree of anchor-
ing in the medium term. To deal with this issue, we use an implicit
inflation target as a reference for inflation expectations in the me-
dium term. 25 We considered the level of inflation that is normally
used in the medium-term projections for internal analysis in the
bcb. In this case, it would be precisely 5 percent,26 which is in line
with the projections made for the Economic and Social Develop-
ment Plan 2016–2020 for Bolivia. As in the previous case, we take
fixed-event variables and use equation (1) to change them to fixed-
horizon variables (Figure 9b).
With the variables prepared, the first step was to analyze the be-
havior of short-term inflation expectations (current year) in order
to compare the results with those obtained with the expectations
from the bcb survey in Section 3.127 with equation (2). The results
show an important role of headline inflation, especially in 2007,
2008, and 2011 (Figure A5b). Nevertheless, since 2012 the coeffi-
cient of bcb projections (degree of anchoring) has reflected an up-
ward trend with slight fluctuations, reaching a value of 0.83 at the
end of the sample (Figure A5a). The results have the same observed
pattern as those obtained in Section 3.1, showing a greater degree
of anchorage in recent years. This shows the importance the bcb’s
projections acquired in the last few years, not only for local economic
agents but also for foreign forecasters.
In order to compare the results from the degree of anchoring
of inflation expectations in the short term and medium term, we use
the same time-varying parameter model from equation (2) with
the same assumptions from the previous section. We introduce

25
There exist research papers that have used implicit inflation targets
such as Mumtaz and Theodoridis (2017).
26
Also, this level has been used as reference for the next calendar year’s
projections in the bcb Monetary Policy Report since 2015.
27
The results of Model 6 can be found in Annex 3.

160 M. Mora, J.C. Heredia, D. Zeballos

 2
3
4
5
6
7
8

0
1
2
3
4
5
6
7
8
9
Jun 2010
Jun 2005 Percentage
Oct 2005 Oct 2010

Percentage
Feb 2006
Jun 2006 Feb 2011
Oct 2006
Feb 2007 Jun 2011
Jun 2007 Oct 2011

First year
Oct 2007
Feb 2008 Feb 2012

Current year
Jun 2008
Oct 2008 Jun 2012
Feb 2009
Jun 2009 Oct 2012
Oct 2009 Feb 2013
Feb 2010
Jun 2010 Jun 2013
Oct 2010
Oct 2013

Second year
Feb 2011
Jun 2011
Feb 2014
Figure 9

Oct 2011
Feb 2012 Jun 2014
Jun 2012

Next calendar year

Oct 2012 Oct 2014
Feb 2013
Jun 2013 Feb 2015
Oct 2013

.     

Jun 2015
Third year
Feb 2014
Jun 2014 Oct 2015
.     

Oct 2014

Note: Authors’ calculations based on Focus Economics and  data.

Feb 2015 Feb 2016
Jun 2015
FIXED HORIZONS VARIABLES FOR MEDIUM TERM

Oct 2015 Jun 2016

Feb 2016
Jun 2016 Oct 2016
Oct 2016 Feb 2017
Forth year

Medium term year

Feb 2017
Jun 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 161

Jun 2017
the inflation expectations by year horizon with the respective bcb pro-
jection; for example, the bcb projection for the first and second year
will be included in the models with the inflation expectations for the
current and next calendar year, respectively. Meanwhile, the implicit
inflation target will be introduced into the models with inflation expec-
tations for the third and fourth years. Thus, we have four models, whose
results are in Table 3.

Table 3
DEGREE OF ANCHORING OF INFLATION EXPECTATIONS
IN THE SHORT TERM AND MEDIUM TERM

Model 6 Model 7 Model 8 Model 9

First year Second year

(current year) (next year) Third year Fourth year
bcb projection 0.83 0.90 0.91 0.93
(implicit target) (0.00) (0.00) (0.00) (0.00)
Past inflation 0.17 0.10 0.09 0.07
(0.09) (0.19) (0.45) (0.65)

Note: The values in parentheses represent the p-values.

The results show a greater degree of anchoring in the medium term

than in the short term, in line with the results of Carrasco and Ferreiro
(2013), Strohsal, Melnick and Nautz (2015) or imf (2016). The coefficient
of past inflation becomes smaller and not significant in the second, third,
and fourth years. Meanwhile the degree of anchoring (coefficient of bcb
forecast) is stronger in recent years; it is a difference of almost 10 per-
centage points between the coefficients in the first and fourth years.
The coefficients for the first and second years reflect more volatile be-
havior over the time (Figure 10). In all of these cases, an improvement
in the degree of anchoring can be seen since 2012, with higher or lower
fluctuations. The degree of anchoring of inflation expectations is gen-
erally greater in the medium term than in the short term.
The bcb does not publish an inflation target for medium-term. Never-
theless, as Strohsal, Melnick and Nautz (2015) mentioned, inflation tar-
gets do not have to be officially announced to be effective. Many central

162 M. Mora, J.C. Heredia, D. Zeballos

 Figure 10
COEFFICIENTS OF BCB PROJECTIONS FOR DIFFERENT YEAR HORIZONS

Percentage
1.1
1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
Mar 2010
Jun 2010
Sep 2010
Dic 2010
Mar 2011
Jun 2011
Sep 2011
Dic 2011
Mar 2012
Jun 2012
Sep 2012
Dic 2012
Mar 2013
Jun 2013
Sep 2013
Dic 2013
Mar 2014
Jun 2014
Sep 2014
Dic 2014
Mar 2015
Jun 2015
Sep 2015
Dic 2015
Mar 2016
Jun 2016
Sep 2016
Dic 2016
Mar 2017
Jun 2017
1st year 2nd year 3rd year 4th year

Source: Authors’ compilation based on  data.

banks, including the European Central Bank or the u.s. Federal Re-
serve, do not publish official inflation targets but are able to commu-
nicate the level of their inflation objective to the markets.
Although inflation expectations appear to be well anchored in the
medium term with respect to past inflation, there is a strand of lit-
erature that postulates that long-term (medium-term) expectations
should not respond to changes in short-term inflation expectations
either ( Jochmann, Koop and Potter, 2010; Łyziak and Paloviita,
2016). In that sense, we additionally create a model to study if there
is a relationship between medium-term and short-term inflation ex-
pectations using the information from Focus Economics.
If medium-term inflation expectations are well anchored, they
should not respond to changes from short-term inflation expecta-
tions. In this case, following the work of Strohsal, Melnick and Nautz

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 163

 e
(2015), medium-term inflation expectations28 ( π m ,t ) are a function
e
of observed inflation ( π t −1 ), short-term expectations29 ( π s ,t ) and the
*
implicit inflation target ( π ):

4 π me ,t = α1π t −1 + α2π se,t −1 + α 3π * + t (4 )

where:
α1 + α2 + α 3 = 1

If α1 > 0 it means that medium-term inflation expectations follow

past inflation. If α2 > 0 , the information from short-term inflation
expectations is relevant for the medium term. With these consider-
ations, medium-term inflation expectations will show a greater de-
gree of anchorage as long as the value of α 3 is close to 1. For inflation
expectations to be perfectly anchored it is necessary that α1 = α2 = 0 .
As in the previous cases, a time-varying parameter model is used
with monthly data from April 2010 to June 2017. The state parameters
follow a random walk process for simplification and variances param-
eters are expressed in exponential form. The Kalman filter is used
to compute the one-step ahead estimates of the means and variances
of the states by maximum likelihood. The results of the estimation
are shown in Table 4.

Table 4
DEGREE OF ANCHORING OF INFLATION EXPECTATIONS
IN THE MEDIUM TERM, MODEL 10

Past inflationn Short-term expectations bcb implicit target

alpha 1 alpha 2 alpha 3

0.03 0.25 0.71
(0.82) (0.17) (0.00)

Note: The values in parentheses represent the p-values.

28
As a reference of medium-term we choose the inflation expectations
for the fourth year of Focus Economics.
29
As a reference of short-term we choose the inflation expectations for the
first year of Focus Economics, in order to work with the same survey
sample.

164 M. Mora, J.C. Heredia, D. Zeballos

 0.0
0.2
0.4
0.6
0.8
1.0
1.2
−0.2
−0.1
0.0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0

Jun 2010 Jun 2010 Jun 2010

Oct 2010 Oct 2010 Oct 2010
Feb 2011 Feb 2011 Feb 2011
Jun 2011 Jun 2011 Jun 2011
Oct 2011 Oct 2011 Oct 2011
Feb 2012 Feb 2012 Feb 2012
Jun 2012 Jun 2012 Jun 2012
Oct 2012 Oct 2012 Oct 2012
Feb 2013 Feb 2013 Feb 2013
Jun 2013 Jun 2013 Jun 2013
Oct 2013 Oct 2013 Oct 2013
Feb 2014 Feb 2014 Feb 2014
Figure 11

Jun 2014 Jun 2014 Jun 2014

Note: Smoothed coefficient ± 2RMSE of Model 10.

Oct 2014 Oct 2014 Oct 2014
Feb 2015 Feb 2015 Feb 2015
.   

Jun 2015 Jun 2015 Jun 2015

.    

Oct 2015 Oct 2015 Oct 2015
. -  

Feb 2016 Feb 2016 Feb 2016

Jun 2016 Jun 2016 Jun 2016
EVOLUTION OF COEFFICIENTS IN MODEL 10

Oct 2016 Oct 2016 Oct 2016

Feb 2017 Feb 2017 Feb 2017
Jun 2017 Jun 2017 Jun 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 165

 The past inflation coefficient (Figure 11a) shows erratic behav-
ior over time, reaching its highest values during 2010 and the end of
2015, in the last months its value decreased to 0.03, a low and insig-
nificant value. The short-term expectations coefficient (Figure 11b)
displays a value of about 0.25 for the whole sample, being almost
constant. However, it is not significant; the effect that this variable
could have on medium-term expectations seems to be already res-
cued with the information of past inflation so it does not present
any significant changes to its behavior.
Finally, the bcb implicit target coefficient (Figure 11c) exhibits
an upward trend, similar to those observed in other models, with
a temporary fall between the second quarter of 2014 and the third
quarter of 2015. This coefficient rose from 0.34 in mid-2010 to 0.71
in mid-2017. Under this specification, medium-term inflation expec-
tations reflect a high degree of anchoring since past inflation ceased
to be significant and short-term inflation expectations did not have
a significant effect throughout the analysis period.

5. SOME CONSIDERATIONS
REGARDING THE RESULTS

The results obtained show that there could be a significant degree

of anchoring of inflation expectations in Bolivia, both in the short
and medium-term, mainly since 2014. In the case of short-term ex-
pectations, it is quite noticeable that bcb’s projections have great-
er effect than observed inflation and other variables, unlike other
studies that indicate that past inflation has a high relevance in this
time horizon (Łyziak and Paloviita, 2016). However, in the medium
term (fourth year), as expected, there is a greater degree of anchor-
ing than in the short term (first year). It is also remarkable consider-
ing this result was obtained with two different samples (bcb survey
and Focus Economics survey).
This behavior indicates a significant improvement in the degree
of credibility of the bcb, and it could be associated with several fac-
tors. These include the adoption of a more active role by the mon-
etary authority (with a higher degree of intervention in the money
market and a more active communication policy), a stable macro-
economic environment, and the progress made in the process of fi-
nancial de-dollarization.

166 M. Mora, J.C. Heredia, D. Zeballos

 Figure 12
PERCENTAGE OF DOLLARIZATION OF FINANCIAL LOANS

100
90
80
70
60
50
40
30
20
10
0
Feb 1998
Oct 1998
Jun 1999
Feb 2000
Oct 2000
Jun 2001
Feb 2002
Oct 2002
Jun 2003
Feb 2004
Oct 2004
Jun 2005
Feb 2006
Oct 2006
Jun 2007
Feb 2008
Oct 2008
Jun 2009
Feb 2010
Oct 2010
Jun 2011
Feb 2012
Oct 2012
Jun 2013
Feb 2014
Oct 2014
Jun 2015
Feb 2016
Oct 2016
Jun 2017
Source: Central Bank of Bolivia.

During the 1990s and the first five years of the 2000s, almost all of
the loans and deposits in the financial system were denominated
in u.s. dollars because people in Bolivia had greater confidence in the
dollar to carry out their daily transactions. This situation can be at-
tributed to the constant depreciations during this period, which
led to a loss of the value of the local currency. In 2006, when the Bo-
livian appreciated, the degree of financial dollarization in Bolivia
began to decrease. This aspect, with other measures applied by the
local authorities, allowed the de-dollarization process to accelerate.
This in turn created a more favorable environment for monetary pol-
icy and a greater role for the bcb in local economic activity. While
97 percent of loans were made in dollars at the beginning of 1998,
by mid-2017 this figure had fallen to 2.7 percent (Figure 12). In the
same period, deposits in dollars declined from 92.7 percent to 15.6
percent. These developments apparently helped to create a more
predictable environment for economic agents.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 167

6. CONCLUSIONS

This study with different specifications of time-varying parameters

models shows that a high degree of anchoring of inflation expectations
in Bolivia could exist. Our main analysis was performed considering
information from the bcb survey, which was complemented with data
from Focus Economics survey. Considering the limitations of these
data sources, our study focuses mainly on the analysis of the short
and medium-term expectations, obtaining good results in both cases.
The results show that the bcb’s projections, presented in its Mon-
etary Policy Report have a significant effect on short-term inflation
expectations, unlike other studies that indicate that past inflation
has a high relevance in this time horizon (Łyziak and Paloviita, 2016).
The anchoring of short-term inflation expectations for central banks
is not of less importance since these have a relevant role in wage ad-
justments and price setting by firms. It is remarkable that we found
a high level of anchoring degree with two different samples (bcb sur-
vey and Focus Economics survey).
In the case of medium-term inflation expectations, we use an im-
plicit inflation target of five percent for time horizons longer than
two years. Also, we use information from Focus Economics, which
has data on inflation expectations up to five years ahead. Follow-
ing the work of Łyziak and Paloviita (2016) and Strohsal, Melnick
and Nautz (2015), we found that past inflation and short-term expec-
tations do not have a significant impact. Meanwhile, the implicit tar-
get would be the main reference for the formulation of medium-term
inflation expectations.
This research paper represents a first step in understanding the be-
havior of inflation expectations in Bolivia. There are not many studies
that have analyzed their conduct or how they react to the announce-
ments made by the bcb about the future trajectory of inflation. Since
2006, the bcb has actively participated in press conferences, semi-
nars and presentations in order to forge a closer relationship with
the population in general (academics, experts, students, reporters,
and others). The results of this paper show that the bcb’s projections
may have excerted a greater influence on agents’ inflation expecta-
tions in recent years. However, more studies should be carried out to
understand and evaluate better the capacity of the bcb to anchor
the inflation expectations of the Bolivian population.

168 M. Mora, J.C. Heredia, D. Zeballos

 0
1
2
3
4
5
6
7
8
9
10
0
1
2
3
4
5
6
7
8
9
10
Jun 2005 Jun 2005
Oct 2005 Oct 2005
Feb 2006 Feb 2006
Jun 2006 Jun 2006
Oct 2006 Oct 2006
ANNEXES

Feb 2007 Feb 2007

Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008

Note: Central Bank of Bolivia.

Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009
Annex 1. bcb Projections

Oct 2009
Feb 2010 Feb 2010
Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011
Jun 2011 Jun 2011
Oct 2011 Oct 2011
Figure A.1

Feb 2012 Feb 2012

Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
ORIGINAL BCB PROJECTIONS

Oct 2013 Oct 2013

.       

Feb 2014 Feb 2014

.        

Jun 2014 Jun 2014
Oct 2014 Oct 2014
Feb 2015 Feb 2015
Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017
Jun 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 169

Jun 2017
 0
2
4
6
8
10
12
Jun 2005
Oct 2005
Feb 2006
Jun 2006
Oct 2006
Feb 2007
Jun 2007
Oct 2007
Feb 2008
Jun 2008
Oct 2008
Feb 2009

Source: Central Bank of Bolivia.

Jun 2009
Oct 2009
Feb 2010
Jun 2010
Oct 2010
Feb 2011

170 M. Mora, J.C. Heredia, D. Zeballos

Jun 2011
Oct 2011
Figure A.2

Feb 2012
Jun 2012
Oct 2012
Feb 2013
Jun 2013
Oct 2013
Feb 2014
Jun 2014
Oct 2014
Feb 2015
Jun 2015
Oct 2015
Feb 2016
Jun 2016
Oct 2016
Feb 2017
UPDATED BCB PROJECTION (INFLATION BY THE END OF CURRENT YEAR)

Jun 2017
 −0.5
0.0
0.5
1.0
1.5
2.0
−0.8
−0.6
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Jun 2005 Jun 2005
Oct 2005 Oct 2005
Feb 2006 Feb 2006
Annex 2

Jun 2006 Jun 2006

Oct 2006 Oct 2006
Feb 2007 Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010

Note: Smoothed coefficient ± 2RMSE.

Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011
Jun 2011 Jun 2011
Oct 2011 Oct 2011
Figure A.3

Feb 2012 Feb 2012

Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013

.    (γπ)

Oct 2013 Oct 2013
Feb 2014 Feb 2014
.     (γ proj)

Jun 2014 Jun 2014

Oct 2014 Oct 2014
Feb 2015 Feb 2015
Jun 2015 Jun 2015
Oct 2015 Oct 2015
EVOLUTION OF COEFFICIENTS IN ALTERNATIVE MODEL 1

Feb 2016 Feb 2016

Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 171

Jun 2017 Jun 2017
 −0.6
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
−0.4
−0.2
0.0
0.2
0.4
0.6
0.8
1.0
1.2
1.4
Jun 2005 Jun 2005
Oct 2005 Oct 2005
Feb 2006 Feb 2006
Annex 3

Jun 2006 Jun 2006

Oct 2006 Oct 2006
Feb 2007 Feb 2007
Jun 2007 Jun 2007
Oct 2007 Oct 2007
Feb 2008 Feb 2008
Jun 2008 Jun 2008
Oct 2008 Oct 2008
Feb 2009 Feb 2009
Jun 2009 Jun 2009
Oct 2009 Oct 2009
Feb 2010 Feb 2010

Note: Smoothed coefficient ± 2RMSE.

Jun 2010 Jun 2010
Oct 2010 Oct 2010
Feb 2011 Feb 2011

172 M. Mora, J.C. Heredia, D. Zeballos

Jun 2011 Jun 2011
Oct 2011 Oct 2011
Figure A.4

Feb 2012 Feb 2012

Jun 2012 Jun 2012
Oct 2012 Oct 2012
Feb 2013 Feb 2013
Jun 2013 Jun 2013
Oct 2013 Oct 2013

.    (1− γπ)

Feb 2014
.     (γ proj)

Feb 2014
Jun 2014 Jun 2014
EVOLUTION OF COEFFICIENTS IN MODEL 6

Oct 2014 Oct 2014

Feb 2015 Feb 2015
Jun 2015 Jun 2015
Oct 2015 Oct 2015
Feb 2016 Feb 2016
Jun 2016 Jun 2016
Oct 2016 Oct 2016
Feb 2017 Feb 2017
Jun 2017 Jun 2017
References
Anderson, G., and B. Maule. (2014). “Assessing the Risk to Infla-
tion from Inflation Expectations.” Bank of England Quarterly
Bulletin 2014 Q2.
Autrup, S.L., and M. Grothe. (2014). “Economic Surprises and Infla-
tion Expectations: Has Anchoring of Expectations Survived
the Crisis?” European Central Bank Working Paper 1671.
Frankfurt, Germany: European Central Bank.
Beechey, M.J., B.K. Johannsen and A.T. Levin. (2011). “Are Long-
Run Inflation Expectations Anchored More Firmly in the
Euro Area Than in the United States?” American Economic
Journal: Macroeconomics 3(2): 104-129.
Beechey, M.J., and J.H. Wright. (2009). “The High-Frequency Im-
pact of News on Long-Term Yields and Forward Rates: Is It
Real?” Journal of Monetary Economics 56(4): 535–544.
Berg, A., and E. Borensztein. (2000). “The Choice of Exchange
Rate Regime and Monetary Target in Highly Dollarized
Economies.” imf Working Paper 00/29. Washington, dc,
United States: International Monetary Fund.
Bevilaqua, A., M. Mesquita and A. Minella. (2007). “Brazil: Taming
Inflation Expectations.” Working Paper 129. Brasilia, Brazil:
Banco Central do Brasil.
Blinder, A., M. Ehrmann, M. Fratzscher, J. De Haan and D.J. Jan-
sen. (2008). “Central Bank Communication and Monetary
Policy: A Survey of Theory and Evidence.” Journal of Economic
Literature 46(4): 910-45.
Capistran, C., and M. Ramos-Francia. (2010). “Does Inflation
Targeting Affect the Dispersion of Inflation Expectations?”
Journal of Money, Credit and Banking 42: 113–134.
Carrasco, C., and J. Ferreiro. (2013). “Inflation Targeting and In-
flation Expectations in Mexico.” Journal of Applied Economics
45(23): 3295-3304.
Celasun, R., G. Gelos and A. Prati. (2004). “Obstacles to Disinfla-
tion: What is the Role of Fiscal Expectations?” imf Working
Paper WP/04/111. Washington, dc, United States: Inter-
national Monetary Fund.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 173

Cerezo, S., and J.C. Heredia. (2013). “La Encuesta de Expectativas
Económicas del bcb: Una Evaluación de la Información Con-
tenida y Racionalidad para la Inflación.” Revista de Análisis
19(2): 103-130.
Cerisola, M., and R.G. Gelos. (2005). “What Drives Inflation Expec-
tations in Brazil? An Empirical Analysis.” imf Working Paper
WP/05/109. Washington, dc, United States: International
Monetary Fund.
Demertzis, M., M. Marcellino and N. Viegi. (2012). “A Credibility
Proxy: Tracking us Monetary Developments.” be Journal of
Macroeconomics 12(1).
Dovern, J., U. Fritsche and J. Slacalek. (2009). “Disagreement
among Forecasters in G7 Countries.” European Central
Bank Working Paper 1082. Frankfurt, Germany: European
Central Bank.
Ehrmann, M. (2015). “Targeting Inflation from Below: How Do
Inflation Expectations Behave?” International Journal of Central
Banking 11(4): 213-249.
Galati, G., S. Poelhekke and C. Zhou. (2011). “Did the Crisis Af-
fect Inflation Expectations?” International Journal of Central
Banking 7(1):167–207.
Gürkaynak, R.S., and J. Wolfers. (2007). “Macroeconomic Deriva-
tives: An Initial Analysis of Market-Based Macro Forecasts,
Uncertainty, and Risk.” In: J.A. Frankel and C.A. Pissarides,
editors. nber International Seminar on Macroeconomics 2005.
Cambridge, United States: mit Press.
Hubert, P. (2015). “ECB Projections as a Tool for Understanding
Policy Decisions.” Journal of Forecasting 34: 574–587.
Humérez, J., and V. H. De la Barra. (2007). “Nivel de Dolarización,
Conflictos Sociales, Impuesto a las Transacciones Finan-
cieras y Diferencial de Tipo de Cambio.” Revista de Análisis
Económico 22: 34-58.
International Monetary Fund - imf. (2016). “Chapter 3: Global
Disinflation in an Era of Constrained Monetary Policy.” In:
World Economic Outlook (weo): Subdued Demand: Symptoms and
Remedies. Washington, dc, United States: imf.

174 M. Mora, J.C. Heredia, D. Zeballos

Jochmann, M., G. Koop and S.M. Potter. (2010). “Modeling the
Dynamics of Inflation Compensation.” Journal of Empirical
Finance 17(1): 157 – 167.
Kalman, R. (1960). “A New Approach to Linear Filtering and Pre-
diction Problems.” Transactions of the asme–Journal of Basic
Engineering 82 (Series D): 35-45.
King, M. (2005). “Monetary Policy: Practice ahead of Theory.” Mais
Lecture. Bank of England Quarterly Bulletin Summer 2005.
Kwiatkowski, D., P. Phillips, P. Schmidt and Y. Shin (1992). “Testing
the Null Hypothesis of Stationarity against the Alternative
of a Unit Root.” Journal of Econometrics 54: 159–78.
Laguna M. (2010). “Características de la Inflación Importada en
Bolivia: ¿Puede Contenerse con Política Cambiaria?” Revista
de Análisis 11: 77-109.
Łyziak, T., and M. Paloviita. (2016). “Anchoring of Inflation Ex-
pectations in the Euro Area: Recent Evidence Based on
Survey Data.” European Central Bank Working Paper 1945.
Frankfurt, Germany: European Central Bank.
Mumtaz, H., and K. Theodoridis. (2017). “The Federal Reserve’s
Implicit Inflation Target and Macroeconomic Dynamics: A
SVAR Analysis.” Economics Working Paper 2017/015. Lan-
caster, United Kingdom: Lancaster University.
Nautz, D., and T. Strohsal. (2015). “Are US Inflation Expectations
Re-Anchored?” Economics Letters 127: 6–9.
Orphanides, A. (2015). “Fear of Liftoff: Uncertainty, Rules, and
Discretion in Monetary Policy Normalization.” Federal Reserve
Bank of St. Louis Review 97(3): 173-96.
Siklos, P. (2013). “Sources of Disagreement in Inflation Forecasts:
An International Empirical Investigation.” Journal of Interna-
tional Economics 90: 218–231.
Strohsal, T., R. Melnick and D. Nautz. (2015). “The Time-Varying
Degree of Inflation Expectations Anchoring.” Available at:
https://ssrn.com/abstract=2609394 or http://dx.doi.
org/10.2139/ssrn.2609394.
Winkelried, D. (2017). “Inferring Inflation Expectations from
Fixed-Event Forecasts.” International Journal of Central Bank-
ing. 13(2): 1-31.

The Time-Varying Degree of Inflation Expectation Anchoring in Bolivia 175

Expectations Anchoring Indexes for Brazil
Using Kalman Filter: Exploring
Signals of Inflation Anchoring in the
Long Term

Fernando Nascimento de Oliveira

Wagner Piazza Gaglianone

Abstract
Our objective in this paper is to build expectations anchoring indexes for in-
flation in Brazil that are fundamentally driven by the monetary authority’s
capacity to anchor long-term inflation expectations vis-à-vis short-run infla-
tion expectations. The expectations anchoring indexes are generated from a
Kalman filter, based on a state-space model that also takes into account fiscal
policy dynamics. The model’s signals are constructed using inflation expecta-
tions from the Focus survey of professional forecasters, conducted by the Banco
Central do Brasil, and from the swap and federal government bond markets,
which convey daily information of long-term inflation expectations. Although
varying across specifications, the expectations anchoring indexes that we pro-
pose tend to display a downward trajectory, more clearly in 2009, and show a
recovery starting in 2016 until the end of the sample (mid-2017).
Keywords: credibility index, inflation expectation, inflation anchoring,
Kalman filter, Banco Central do Brasil.
jel classification: E50, E52, E58.
F. Nascimento de Oliveira <[email protected]>, Research Depart-
ment, Banco Central do Brasil and ibmec/rj, and W. P. Gaglianone <wagner.gaglia-
[email protected]>, Research Department, Banco Central do Brasil. This research
was undertaken within the framework of cemla’s Joint Research Program 2017
coordinated by the Banco de la República, Colombia. The authors are especially grate-
ful for the helpful comments and suggestions given by Olivier Coibion. The authors
benefited from comments made by seminar participants in the workshop on inflation
expectations, their measurement and degree of anchoring at cemla headquarters
(Mexico City, September 14-15, 2017). We also thank Carlos Viana de Carvalho, André
Minella and Bernardus Doornik for comments made at the Second Network of Eco-
nomic Research of the Banco Central do Brasil. The views expressed in the papers
are those of the authors and do not necessarily reflect those of the Banco Central
do Brasil.

177
1. INTRODUCTION

W
ell-anchored inflation expectations are fundamental for the
conduct of monetary policy. Properly anchoring inflation ex-
pectations requires the central bank to be regarded as cred-
ible, that is, economic agents should be confident that the central
bank will react to the various shocks that affect the economy to main-
tain price stability.
Cukierman and Meltzer (1986) stressed that the future objectives
of central banks depend on inflation expectations. In this sense,
a credible commitment to an explicit inflation objective helps to an-
chor inflation expectations to the desired level. This anchoring
contributes to delivering price stability, which is the main objective
of central banks.
In turn, Blinder (2000) sent questionnaires to 127 heads of cen-
tral banks around the world asking their opinion on the importance
of central bank credibility. The answers showed clearly that credibil-
ity matters in practice. A credible central bank is one that can make
a believable commitment to low inflation policy and has complete
dedication to price stability. This will make disinflation less costly
and decrease the sacrifice ratio.
Nonetheless, building credibility is costly and takes repeated
successes to establish. Moreover, credibility evolves in asymmetric
fashion and can be lost rapidly, depending on the perception by eco-
nomic agents that the central bank is able (or not) to achieve its ob-
jectives. As famously put by Benjamin Franklin: “It takes many good
deeds to build a good reputation, and only one bad one to lose it.”1
Central banks have imperfect control over inflation in the short
run. As Gomme (2006) remarked, current inflation provides a noisy
signal of a central bank’s long-term intentions, and therefore of its
type. According to the author, a central bank is credible when the pub-
lic assigns a high probability of low inflation-type to the central bank.
In this context, a central bank will lose credibility when this prob-
ability decreases. The credibility of central banks is very much con-
cerned with people’s beliefs about what the central bank will do in
the future.

1
See Isaacson (2004).

178 F. Nascimento de Oliveira, W. P. Gaglianone

 On the other hand, central bank credibility is a latent variable2
and, consequently, it is not easy to measure in practice. One possibil-
ity is to look for measures that reflect the capacity central banks have
to anchor inflation expectations. In the literature, this is done mostly
by looking at how closely short-run expectations match the central
bank’s explicit or implicit inflation target (see Bordo and Siklos,
2015). The problem with these measures, in our view, is that other
signals can exist in the economy that may also help to give an idea
of how well inflation expectations are anchored.
Figure 1 compares the consensus inflation forecast in Brazil (ho-
rizon of one year) with the inflation target and respective tolerance
bands. Based on these series, Figure 2 shows the evolution of some
credibility indexes (hereafter cis) for the Banco Central do Brasil
(bcb) from January 2002 to June 2017. The measures are, respec-
tively, ci- ck (Cecchetti and Krause, 2002), ci-m (Mendonça, 2004)
and ci-ms (Mendonça and Souza, 2009).
These indexes measure deviations of short-run inflation expec-
tations from bcb’s inflation target. 3 For instance, note that at the
end of 2002, before the presidential election, these indexes had a
substantial decline in credibility. This fact can be related to an ex-
ogenous shock to bcb: the uncertainty about the policy regime with
a likely victory of the presidential candidate Lula, which triggered
the country sovereign risk premium (embi+br) to sharply rise during
this period. This was a situation completely out of bcb’s control.4
Also, note that Figure 2 shows a very volatile ci-m, considering
the whole sample, indicating a fast loss and recovery of credibility.
The other indexes show different behavior of credibility: ci-ck varies
very little, while ci-ms looks constant almost all the time. In fact,

2
The international literature on credibility indexes of central banks
is vast. They are many theoretical as well as empirical papers on the
subject. See, for example, Gomme (2006), Svenson (1993), Clarida
and Waldman (2007), Ceccheti and Krause (2002), Kaseeream (2012)
and Bordo and Siklos (2015).
3
Other papers also build credibility indexes for the Banco Central
do Brasil focusing on deviations of short-term inflation expectations
from inflation target, such as Teles and Nemoto (2005), Sicsú (2002),
Nahon and Meurer (2005), and Lowenkron and Garcia (2007).
4
Note that ci-m decreases substantially during the subprime crisis, which
like Lula’s election is also exogenous to bcb. At the end of the period,
ci-m shows a steep credibility recovery that also seems counterfactual.

Expectations Anchoring Indexes for Brazil using Kalman Filter 179

 Figure 1
SURVEY-BASED INFLATION EXPECTATIONS, INFLATION TARGET
AND TOLERANCE BANDS

12

10

0
Jan 2002
Mar 2003
May 2004
Jul 2005
Sep 2006
Nov 2007
Jan 2009
Mar 2010
May 2011
Jul 2012
Sep 2013
Nov 2014
Jan 2016
Mar 2017
Consensus inflation expectations Inflation target
(cross-sectional average)

Note: Average inflation expectations (Focus survey) with forecast horizon of one year.
Inflation targets and tolerance bands from <http://www.bcb.gov.br/pec/metas/In-
flationTargetingTable.pdf>.

Figure 2
CREDIBILITY INDEXES FROM THE LITERATURE

1.0

0.8

0.6

0.4

0.2

0.0
Jan 2002
Mar 2003
May 2004
Jul 2005
Sep 2006
Nov 2007
Jan 2009
Mar 2010
May 2011
Jul 2012
Sep 2013
Nov 2014
Jan 2016
Mar 2017

CI-MS CI-M CI-CK

Note: CI-CK means Cecchetti and Krause (2002), CI-M denotes Medoça (2004) and
CI-MS m e a n s M e n d o ç a a n d S o u z a ( 2 0 0 9 ) . I n fl a t i o n e x p e c t a t i o n s a re t h e
survey-based cross-sectional average expectations with fixed horizon of one year.

180 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 3
DAILY NUMBER OF SURVEY PARTICIPANTS THAT REPORT INFLATION
FORECAST FOR THE CURRENT AND THE FOLLOWING CALENDAR YEARS
(END-OF-YEAR FIXED-EVENT FORECAST)

200

180 Calendar year 1

Calendar year 2
160 Calendar year 3
Calendar year 4
140
Calendar year 5
120

100

80

60

40

20

0 Jan 2011

April 2015
Jan 2002

Jun 2004

Aug 2006

Oct 2008

Mar 2013

Jun 2017
Source: Banco Central do Brasil and authors’ calculations.

the credibility dynamics implied by these indexes seem not to appro-

priately represent the dynamics of mean and standard deviation in-
flation expectations measured in fixed horizons and taken from bcb’s
daily survey of expectations (Focus), presented in Figure 5. The first
graph shows that the cross-sectional mean of inflation expectations
with a forecast horizon of four years–a measure of long-term expec-
tations–has much less volatility than the one-year (short-term) infla-
tion expectations. Not only that but in the run-up to Lula’s election
and the subprime crises, the four-year expectations varied much less
than the one-year counterpart. The second graph of Figure 5 shows
a similar dynamic pattern for the short-run (one year) and long-run
(four years) standard deviation of inflation expectations. 5

5
There are other papers in the literature that build credibility indexes
for the bcb taking different approaches from those that look at short-
term deviations of inflation expectations from the target. This is the

Expectations Anchoring Indexes for Brazil using Kalman Filter 181

 In practice, one should examine a variety of signals to construct
a measure that really reflects the ability of central banks to anchor
inflation expectations (see Demertzis et al., 2012). We think that
the problem with most traditional cis available in the literature
is that they focus on the short-run deviations of inflation expectations
from the inflation target. In contrast, we construct in this paper ex-
pectations anchoring indexes (hereafter, eais) that are specifically
designed to measure the degree of anchoring of long-term inflation
expectations vis-à-vis the short-run.
The bottom-line of our argument is that a central bank is credible
if it has the capability to properly anchor long-run inflation expec-
tations. The extent of long-term inflation anchoring will serve as a
proxy for anchoring. If the central bank is credible and anchors long-
term inflation expectations, then the long-run expectations will be-
come less responsive to short-run economic news.6 This means that
in the presence of a negative or positive short-term shock to infla-
tion, economic agents believe the central bank will take appropriate
countervailing actions to keep inflation on target in the long run.
Our view is in line with Demertzis et al. (2012) and Buono and For-
mai (2016). Demertzis et al. point out that the credibility of the cen-
tral bank decouples long-run inflation expectations from short-run
expectations. Buono and Formai notice that inflation expectations
are anchored when movements in short-run expectations do not af-
fect movements in the long term.7
To build expectations anchoring indexes for inflation in Brazil
that decouple long-term from short-term inflation expectations,
we also need to incorporate explicitly in our approach some measure

case of Garcia and Guillén (2011), Leal et al. (2012), Issler and Santos
(2017), and Val et al. (2017).
6
Bernanke (2007) describes inflation anchoring in the following man-
ner: “…“anchored” to mean relatively insensitive to incoming data. So,
for example, if the public experiences a spell of inflation higher than
their long-run expectation, but their long-run expectation of inflation
changes little as a result, then inflation expectations are well anchored.
If, on the other hand, the public reacts to a short period of higher-
than-expected inflation by marking up their long-run expectation
considerably, then expectations are poorly anchored”.
7
For other empirical papers with definitions of credibility, see Davis
(2012), Levieuge et al . (2015) and Dimitris et al . (2016). For theoretical
papers with definitions of central bank credibility, see Barro and Gor-
don (1983), Walsh (1995) and Blackburn and Christensen (1989).

182 F. Nascimento de Oliveira, W. P. Gaglianone

of fiscal policy. The reason is that, in some periods in Brazil, per-
ceptions about fiscal policy and fiscal sustainability seemed to have
played an important role in explaining inflation expectations. If we
do not control for that, processes of deanchoring of expectation
may be attributed to the bcb’s policies and not to broader economic
policies. In emerging countries where the public debt is high (in
terms of gdp) and with short average maturity, periods of fiscal
dominance may occur.
As Sargent and Wallace (1981) argue, under fiscal dominance,
the monetary authority faces the constraints imposed by the de-
mand for government bonds. If the fiscal authority cannot finance
its deficits solely by new bond sales, then the monetary authority
is forced to create money and tolerate additional inflation. Although
such a monetary authority might still be able to control inflation
over the long run, it is less capable than a monetary authority un-
der a no fiscal dominance situation. Blanchard (2004) argues that
fiscal dominance describes the situation of the Brazilian economy
in 2002 and 2003.
In periods of fiscal dominance, there may be a reversal of the tra-
ditional roles of monetary and fiscal policies: central banks are in-
clined to reduce interest rates when inflation rises, the opposite
of their standard response, in order to guarantee the stability and sol-
vency of debts and deficits. Therefore, in such periods even a cred-
ible central bank may find difficulty in keeping long-term inflation
expectations unaffected by short-term shocks on inflation or short-
term inflation expectations.
Our objective in this paper is to build eais for bcb that are fun-
damentally driven by the capacity the bcb has to anchor long-term
inflation expectations vis-à-vis short-run expectations. The eais will
be constructed from a Kalman filter, based on a linear state-space
model that also takes into account fiscal policy dynamics. The sig-
nals of the state-space model will give information on the anchoring
of long-term inflation expectations.
There are many possible signals of long-term inflation anchor-
ing in the literature, 8 based on nonparametric or parametric ap-
proaches. We use as many signals as possible from all sources that
are available. In this sense, we have disaggregated daily data (from
January 2002 to June 2017) of inflation expectations from the Focus

8
See Natoli and Sigalotti (2017).

Expectations Anchoring Indexes for Brazil using Kalman Filter 183

survey of professional forecasters conducted by the bcb. From this
survey, we extracted 17 signals. We also have market data of nomi-
nal federal government bonds (Letras do Tesouro Nacional, hereafter
ltn) and inflation-indexed bonds (Notas do Tesouro Nacional, here-
after ntn-b) from April 2005 to June 2017. Finally, we have informa-
tion on swaps of fixed interest rate instruments against inflation
from January 2005 to June 2017. From the bond and swap markets,
we extracted 14 signals.
We contribute to the literature in several manners. Firstly, as far
as we know, this is the first paper to use a large number of signals
of long-term inflation expectation anchoring, coming from both
surveys and market data. Secondly, we focus on long-term inflation
expectations, unlike the great majority of empirical papers on the
subject in Brazil.9 We can update our eais on a daily basis with disag-
gregated and aggregated data obtained through surveys or through
market information. By construction, our eais give a prompt idea
of how well the long-term inflation expectations are anchored, which
is very important in the implementation of monetary policy, espe-
cially in an inflation targeting regime.
In the third place, we take into account both fiscal policy and mon-
etary policy when estimating the state-space model using our sur-
vey and market data for long-term inflation expectation anchoring
compared to short-run inflation expectations. Finally, the disaggre-
gated confidential survey data of the bcb –an essential part of our
database–is unique and enables us to have a much better grasp of in-
flation expectations of economic agents in Brazil, and hence of bcb’s
ability to anchor them.
The rest of the paper is organized as follows: Section 2 describes
the data; Section 3 presents the empirical analyses, and Section
4 concludes.

2. DATA

We have survey and market data. In the former case, we have data
from January 2002 to June 2017. In the latter case, we have data from
April 2005 to June 2017.

9
See Gaglianone (2017) for a recent survey of applied research on infla-
tion expectations in Brazil.

184 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 4
SURVEY DATA: CROSS-SECTIONAL MEAN, MEDIAN, STANDARD DEVIATION
AND INTER-QUARTILE RANGE OF INDIVIDUAL SURVEY-BASED
INFLATION FORECASTS (FIXED EVENTS)
Raw data from the focus survey (calendar-year forecasts)

14
Mean forecast (1 year)
Mean forecast (2 year)
Mean forecast (3 year)
12 Mean forecast (4 year)
Mean forecast (5 year)

10

2
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

3.5
Std. deviation (1 year)
Std. deviation (2 year)
3.0
Std. deviation (3 year)
Std. deviation (4 year)
2.5 Std. deviation (5 year)

2.0

1.5

1.0

0.5

0.0
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

Source: Banco Central do Brasil and authors’ calculations.

Expectations Anchoring Indexes for Brazil using Kalman Filter 185

 Figure 4 (cont.)
SURVEY DATA: CROSS-SECTIONAL MEAN, MEDIAN, STANDARD DEVIATION
AND INTER-QUARTILE RANGE OF INDIVIDUAL SURVEY-BASED
INFLATION FORECASTS (FIXED EVENTS)
Raw data from the focus survey (calendar-year forecasts)

14
Median forecast (1 year)
Median forecast (2 year)
Median forecast (3 year)
12 Median forecast (4 year)
Median forecast (5 year)

10

2
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

4.5
Inter-quartile range (1 year)
4.0 Inter-quartile range (2 year)
Inter-quartile range (3 year)
3.5 Inter-quartile range (4 year)
Inter-quartile range (5 year)
3.0

2.5

2.0

1.5

1.0

0.5

0.0
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

Source: Banco Central do Brasil and authors’ calculations.

186 F. Nascimento de Oliveira, W. P. Gaglianone

 Our survey data are proprietary, with confidential information
at the individual level and publicly available data at the aggregate
level. The data were obtained from the Focus survey organized by the
bcb, collected every workday by the bcb.10 We have the distribution
of inflation expectations for every workday.
We have unbalanced panel data of survey inflation expectations.
The number of registered institutions that take part in the survey
is 277 in our sample. The number of workdays in our sample is 3,781.
The average number of institutions that report inflation forecasts
is 83 for the forecast horizon of one year and 48 for the four-year
horizon.
Figure 3 presents the number of institutions that forecast infla-
tion every workday for one year up to five years. As can be seen, there
are some workdays on which very few institutions reported. This
is particularly relevant in the case of forecasts for four or five years.
In addition, for each end-of-year inflation, the number of institu-
tions reporting forecasts increases as long as the forecast horizon
diminishes. To avoid problems in our estimations, we consider that
when there were fewer than 10 institutions reporting on a certain
workday, we repeat the forecasts of the previous workday in which
there were more than 10 institutions reporting for the same period.
Raw information on inflation expectations pertains to fixed events
(e.g., end-of-year inflation forecasts for the current and following
years); see Figure 4. We transform them to fixed-horizon inflation
expectations by linear interpolation using the daily (decreasing)
forecast horizon of the fixed-event inflation forecasts; see Figure 5.
Since the longest horizon of inflation forecasts available in the Focus
survey involves the five-year-ahead forecast (calendar year), we em-
ploy the inflation expectations for the following four and five calen-
dar years to build the interpolated forecast with a maximum fixed
horizon of four years.
On the other hand, there is no inflation target set for such long ho-
rizons. Since the beginning of the inflation targeting regime in 1999
and up to the inflation target announced for 2019, the inflation target

10
Nowadays, the bcb releases on the internet the micro data of the Focus
survey of expectations, in a panel data with fake IDs (i.e., the identity
of the survey participants is preserved and the disclosed database only
contains anonymous participants). For more details, see the website:
http://dadosabertos.bcb.gov.br/dataset/expectativas-mercado/
resource/23f6c983-f9bd-48f8-a889-72def3ae17c8

Expectations Anchoring Indexes for Brazil using Kalman Filter 187

 Figure 5
SURVEY DATA: CROSS-SECTIONAL MEAN, MEDIAN, STANDARD DEVIATION
AND INTER-QUARTILE RANGE OF INDIVIDUAL SURVEY-BASED
INFLATION FORECASTS (FIXED HORIZONS)
Transformed data (fixed-horizons forecasts)

12
Mean forecast (1 year)
11 Mean forecast (4 year)

10

3
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

2.5
Std. deviation (1 year)
Std. deviation (4 year)
2.0

1.5

1.0

0.5

0.0
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

Source: Banco Central do Brasil and authors’ calculations.

188 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 5 (cont.)
SURVEY DATA: CROSS-SECTIONAL MEAN, MEDIAN, STANDARD DEVIATION
AND INTER-QUARTILE RANGE OF INDIVIDUAL SURVEY-BASED
INFLATION FORECASTS (FIXED HORIZONS)
Transformed data (fixed-horizons forecasts)

12
Median forecast (1 year)
11 Median forecast (4 year)

10

3
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

4.0
Inter-quartile range (1 year)
3.5 Inter-quartile range (4 year)

3.0

2.5

2.0

1.5

1.0

0.5

0.0
Jan 2002 Jun 2004 Aug 2006 Oct 2008 Jan 2011 Mar 2013 Apr 2015 Jun 2017

Source: Banco Central do Brasil and authors’ calculations.

Expectations Anchoring Indexes for Brazil using Kalman Filter 189

 Figure 6
MARKET DATA: BREAKEVEN INFLATION
BEI, percentage 12 months

10
9
8
7
6
5
4
3
2
Apr 2005

Aug 2006

Dec 2007

Apr 2009

Aug 2010

Dec 2011

Apr 2013

Aug 2014

Dec 2015

Apr 2017
 swap 1y  swap 4y  bonds 1y  bonds 4y

Source: Anbima, B3 and authors’ calculations.

and tolerance bands had been set up to June of year t for the calendar
year t+2. Nowadays, the new target is announced up to June of year t
for the calendar year t+3.11 Since many signals depend on the infla-
tion target, and since our longest forecast horizon is four years, we as-
sume that the inflation target four years ahead is equal to the target
set for the calendar year t+2 (or t+3, whenever available).
In the case of market data, we have publicly available information
on federal government bonds and swaps of fixed interest rate against
inflation and a coupon from April 2005 to June 2017. The former
are obtained from Anbima (Brazilian Financial and Capital Mar-
ket Association) and the latter are registered by b3 (a Brazilian com-
pany that operates securities, commodities and futures exchange,
among others, previously known as bm&fbovespa). Federal govern-
ment bonds are nominal bonds (ltns) and inflation-indexed bonds

11
See <https://www.bcb.gov.br/pec/metas/InflationTargetingTable.
pdf>.

190 F. Nascimento de Oliveira, W. P. Gaglianone

(ntn-bs). The yields of these bonds for different maturities are cal-
culated by fitting ltn and ntn-b with the Nelson-Siegel-Svensson
functional form.
The difference between yields of the same maturity of ltn s
and ntn-bs is known as breakeven inflation (hereafter bei). Accord-
ing to Shen (2006): “An increase in the breakeven rate is sometimes
viewed as a sign that market inflation expectations may be on the
rise. For example, the fomc frequently refers to the yield spread as a
measure of ‘inflation compensation’ and considers the yield spread
an indicator of inflation expectations in policy deliberations.”12
In this paper, we use bei series as proxies of market inflation expec-
tations. It is important to note that these measures are embedded
with a liquidity premium as well as an inflation risk premium that
might distort it from pure measures of inflation expectations.
Swaps of inf lation plus a coupon against fixed interest rates
are registered by b3. The bcb collects workday information in this
respect. The difference between fixed rate and coupon gives bei s
of swaps. One advantage of bei s coming from swaps–compared
to beis from federal government bonds–is that they have very low li-
quidity premiums.13 Figure 6 shows the dynamics of bei from swaps
and federal government bonds with maturities of one and four years.
In both Figures 5 and 6, it is easy to observe that four-year survey
inflation expectations and four-year beis have lower variance and are
more persistent than one-year inflation expectations and one-year
beis, respectively.
As for an indicator of high frequency fiscal policy, we use work-
day expectations of primary balance as a percentage of gdp. These
data are also collected from the Focus survey. We use in our empiri-
cal analyses the one-year ahead expectations. The raw data on the
expectations are for fixed events and we transform them for a fixed
horizon by linear interpolation in exactly the same way as we do
for inflation expectations.
Figure 7 shows the dynamics of this series. As can be seen, there
is a clear turning point in fiscal expectations in our sample. Until

12
fomc means the Federal Open Market Committee of the U.S. Federal
Reserve.
13
We have yields for fixed-interest bonds with maturities of one, three
and ten years. We interpolate linearly the three- and ten-year yields
to get the four-year yields that we used to construct bei s for the swap
market.

Expectations Anchoring Indexes for Brazil using Kalman Filter 191

 Figure 7
CONSENSUS SURVEY-BASED EXPECTATIONS
OF PRIMARY FISCAL BALANCE (zt)
Percentage of GDP, forecast horizon of 12 months

−1

−2

−3
Jan 2002

Jul 2003

Jan 2005

Jul 2006

Jan 2007

Jul 2007

Jan 2008

Jul 2009

Jan 2010

Jul 2010

Jan 2017
Source: Banco Central do Brasil, Focus survey, cross-section average expectations.

2009, the expectations were relatively stable around a primary sur-

plus of 4% of gdp. From mid-2009 until mid-2012, expectations fluc-
tuated near a primary surplus of 3% of gdp. However, from mid-2012
on there was clear deterioration of these expectations, reaching a pri-
mary balance of -2% of gdp in the beginning of 2017.

3. EMPIRICAL ANALYSIS

Our method to construct the expectations anchoring indexes

can be summarized as follows:
1) we build a set of normalized (i.e., zero mean and unit variance)
signals from both survey and market data; 2) we employ factor analy-
sis to summarize the panel data information of signals into a single
“common factor” series that contains the core dynamics of long-
term inflation expectation anchoring with respect to the short-run

192 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 8
SURVEY SIGNALS
Exponential smoothing, half-life of one year

6
5
4
3
2
1
0
−1
−2
−3
−4
Sep 2002

Jan 2004

May 2005

Sep 2006

Jan 2008

May 2009

Sep 2010

Jan 2012

May 2013

Sep 2014

Jan 2016

May 2017
s1 s2 s3 s4 s5 s6 s7 s8

6
4
2
0
−2
−4
−6
Sep 2002

Jan 2004

May 2005

Sep 2006

Jan 2008

May 2009

Sep 2010

Jan 2012

May 2013

Sep 2014

Jan 2016

May 2017

s9 s10 s11 s12 s13 s14 s14 s15 s16

Expectations Anchoring Indexes for Brazil using Kalman Filter 193

inflation expectations; 3) we estimate a state-space model using
a Kalman filter to build two separate states for monetary policy cred-
ibility and fiscal stance; and 4) we employ a logit transformation to set
the scale of states into the [0;1] interval.
We next describe the signals of long-term inflation anchoring
that we used in the paper.

3.1 Signals of Long-term Inflation Anchoring

Some of our signals are based on recursive correlations or recursive
regressions. In these cases, we used a training sample of six months
(126 workdays) in order to generate the first signal observation.
Moreover, we treated the observations of our recursive analyses
in three different ways: each observation was weighted by exponen-
tially smoothed weights with a half-life of one or two years,14 or by
using a rolling window of three years. Moreover, all the signals that
we used to build our eais were normalized z-scores (i.e., with zero
mean and standard deviation equal to 1).

3.1.1 Signals from Survey Data

Table 1 lists the signals that we extracted from the bcb survey. We built
signals based on recursive Pearson correlation and recursive ordi-
nary least squares (ols) of mean and median four-year inflation ex-
pectations against one-year inflation expectations. We also built
signals based on recursive correlations and recursive ols between
the standard deviation and inter-quartile range of four- and one-
year inflation expectations. In the case of regressions, our signals
are the slope coefficients of the regressors related to one-year infla-
tion expectations.
We built a signal based on the estimation of time-varying var as
in Demertzis et al. (2012). The estimation is based on Stock and Wat-
son (1996). The coefficients vary through time like random walks.
The coefficient of interest is the one that measures the elasticity
of four-year inflation expectations in relation to one-year inflation
expectations.

14
In other words, for a given sample, a weight equal to 1 is attached
to the most recent observation. After a half-life period (e.g., 1 year
=252 workdays), the weight exponentially decays to 0.5.

194 F. Nascimento de Oliveira, W. P. Gaglianone

 We built two signals based on the evolution of the distribution
of the four-year inflation expectations. One signal is equal to 0 if the
median of the distribution is equal to the inflation target and 1 oth-
erwise. The other signal is equal to 0 on workday t if the distribution
on this day is equal to the distribution on workday t-21 (previous
month) and 1 otherwise, based on the Kolmogorov-Smirnov test.15
We built another signal based on Nautz and Strohsal (2015). The au-
thors estimate by ols a multiple regression between long-term in-
flation expectations and lag of long-term inflation expectations
and surprises in macroeconomic variables. We tested for the possi-
bility of structural breaks between the dependent variable and the
regressors that measure macroeconomic surprises according to An-
drews (1993) and Quandt (1960)16. We used as macroeconomic vari-
ables levels of the nominal foreign exchange rate (R$/US$), embi+br
and the yield of the 360 days interest rate swap. We considered a sur-
prise in these macroeconomic variables when the value of the series
is higher (or lower) than the mean of the series plus (minus) one stan-
dard deviation. Our coefficient of interest is the one related to the
nominal foreign exchange rate.
We built a signal based on recursive logistic regressions, with equal
weights for the time series observations, such as in Natoli and Siga-
lotti (2017). The model estimates the probability that four-year in-
flation expectations will be higher or lower than the 75% percentile
of the workday distribution of this series (the dependent variable is 1
if it is higher and 0 if it is lower). This probability is estimated given
that the one-year inflation expectations were higher or lower than
the 75% percentile of the distribution of the same workday of this se-
ries (the regressor is 1 if it is higher and 0 if it is lower). Our coefficient
of interest is the one related to the one-year inflation expectations.
Figures 8, 9 and 10 show the evolution of the signals above–nor-
malized z-scores with zero mean and standard deviation equal to 1–
of recursive regressions estimated with exponentially smoothed

15
See Massey (2012).
16
In this paper, we employ the idea behind the Quandt-Andrews test,
in which a single Chow (1960) breakpoint test is performed for every
observation between two dates. The test statistics from those Chow tests
are used to build dummy variables representing the different regimes
between breakpoints.

Expectations Anchoring Indexes for Brazil using Kalman Filter 195

weights with a half-life of one or two years or using weights from
a rolling window of three years.

3.1.2 Signals from Market Data

In the case of market data, we built signals based on beis of one year
and four years obtained in the swap and bond markets. Several of the
signals were obtained in exact ways described in the previous sec-
tion. We included two different signals from the survey signals: one is
the difference between bei and the inflation target and the other
one is the square of this difference. Table 2 lists the market signals
and Figures 11, 12 and 13 show the evolution of the market signals.

3.1.3 Selection of Signals Based on Correlation Analysis

We have a total of 31 signals: 17 are selected from survey data and 14
are selected from market data. To obtain our benchmark eais that
we present in Section 3.4, we select from these 31 signals the ones
whose correlations are less than 0.7. Table 3 shows the correlation
matrix of the selected signals. As a result, the following 14 signals
were selected: S3, S9, S12, S13, S14, S15, S17, SM3, SM4, SM7, SM8,
SM9, SM12, and SM14.

3.2 Factor Analysis

Next, we employ factor analysis (fa) to extract common factors from
the set of signals chosen. There are many ways suggested in the liter-
ature to combine the set of signals into a single indicator (e.g., equal
weights or pca–principal component analysis). We adopt the factor
analysis setup,17 since our goal here is to build a single time series
that reflects long-term anchoring of inflation expectations (in re-
spect to short-run inflation expectations) by extracting common
movements from the set of selected signals.
To do so, we use the principal factors as the factor extraction meth-
od and the ordinary correlation for covariance analysis. The idea is to

17
Factor analysis (fa) and principal component analysis (pca) are similar
statistical techniques in the sense that both generate linear combina-
tions of the original series. However, pca is used to retain the maximum
amount of information from data in terms of total variation, whereas
fa accounts for common variance. Thus, fa is often employed to build
factors (latent variables), while pca is often used in data reduction
frameworks. See Johnson and Wichern (1992) for further details.

196 F. Nascimento de Oliveira, W. P. Gaglianone

 Table 1
SIGNALS CONSTRUCTED FROM SURVEY-
BASED INFLATION EXPECTATIONS

Group Signals Description

1 S1 cross-section mean forecast long run - inflation target
1 S2 cross-section median forecast long run - inflation target
1 S3 cross-section standard deviation (forecast long run -
inflation target)
1 S4 cross-section inter-quartile range (forecast long run -
inflation target)
2 S5 recursive Pearson correlation between (cross-section
mean) short and long run inflation expectations
2 S6 recursive Pearson correlation between (cross-section
median) short and long run inflation expectations
2 S7 recursive Pearson correlation between (cross-section
std. dev.) short and long run inflation expectations
2 S8 recursive Pearson correlation between (cross-section
inter-quartile range) short and long run expectations
3 S9 recursive ols regression with (cross-section mean)
short and long run inflation expectations
3 S10 recursive ols regression with (cross-section median)
short and long run inflation expectations
3 S11 recursive ols regression with (cross-section std. dev.)
short and long run inflation expectations
3 S12 recursive ols regression with (cross-section inter-
quartile range) short and long run inflation
expectations
4 S13 binary variable from the hypothesis test (Ho:
median expectation = inflation target) for the long
run expectations
4 S14 binary variable from the hypothesis test Ho:
distr(t) = distr(t–21) for the long-run cross-section
distribution
5 S15 Nautz and Strohsal (2015), fx-rate slope from
ols (median expectation, macro shocks)
6 S16 Natoli and Sigalotti (2017), slope from logit
regression, median inflation expectations (short,
long)
7 S17 Demertzis et al . (2012), time-varying var, median
inflation expectations (short, long)

Expectations Anchoring Indexes for Brazil using Kalman Filter 197

obtain a vector of loadings that maximizes the cumulative commu-
nality using a number of n factors. This way, each considered signal
(sit) can be decomposed into a common component and an idiosyn-
cratic component:

1 sit = Λi Ft + εit

The common component captures the bulk of the covariation

between s it and the other signals, whereas the idiosyncratic term af-
fects only s it by assumption. Thus, it is simply a scaled common factor
(Ft), which is estimated using the entire set of signals. The long-term
inflation-anchoring indicator is defined to be this common factor.
We adopt here a parsimonious model with two factors (n =
2), since alternative models with more factors, in general, deliver es-
timations with higher uniqueness and lower communality (in the ad-
ditional variables and/or factors) in relation to a model with fewer
factors.18
As a result, the first factor accounts for 37% of the total variance
of the set of 14 selected signals, whereas the first and second factors
together represent 55% of the fraction of total variance.19 Next, we use
those figures to build a combined single factor, as a linear combi-
nation of the two original factors, as follows: Ft =F1,t * 0.37/0.55 + (1-
0.37/0.55) * F2,t. Table 4 summarizes the factor loadings and Figure
14 shows the factors in the baseline case.

3.3 State-space Model

We build our expectations anchoring indexes based on the maxi-
mum likelihood estimation of a linear state-space model as described
in the system of Equations 2-3, presented next. The idea is to disen-
tangle the fiscal policy effect from the common factor Ft, constructed

18
We use the parsimonious number of two factors since they account
for more than half of the fraction of total variance of the set of signals.
Nonetheless, there are many alternative factor selection tools avail-
able in the literature, such as the ones proposed by Bai and Ng (2002)
or Alessi, Barigozzi and Capasso (2010).
19
These figures are computed using the eigenvalues obtained in the
solution of each factor’s linear combination, as explained in Jolliffe
(2002).

198 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 9
SURVEY SIGNALS
Exponential smoothing, half-life of two years

6
5
4
3
2
1
0
−1
−2
−3
−4
Sep 2002

Jan 2004

May 2005

Sep 2006

Jan 2008

May 2009

Sep 2010

Jan 2012

May 2013

Sep 2014

Jan 2016

May 2017
s1 s2 s3 s4 s5 s6 s7 s8

6
4
2
0
−2
−4
−6
Sep 2002

Jan 2004

May 2005

Sep 2006

Jan 2008

May 2009

Sep 2010

Jan 2012

May 2013

Sep 2014

Jan 2016

May 2017

s9 s10 s11 s12 s13 s14 s14 s15 s16

Expectations Anchoring Indexes for Brazil using Kalman Filter 199

 −6
−4
−2
0
2
4
6
−4
−3
−2
−1
0
1
2
3
4
5
6

s9
Sep 2002 Sep 2002

s1
Jan 2004 Jan 2004

s10
May 2005 May 2005

s2

s11
Sep 2006 Sep 2006

s3
Jan 2008 Jan 2008

s12
May 2009 May 2009

s4

s13
Figure 10

Sep 2010 Sep 2010

s5

200 F. Nascimento de Oliveira, W. P. Gaglianone

SURVEY SIGNALS

Jan 2012 Jan 2012

s14
s6
May 2013 May 2013

s14
Rolling window weights, window of three years

Sep 2014 Sep 2014

s7

s15
Jan 2016 Jan 2016
s8
May 2017 May 2017

s16
 Figure 11
MARKET SIGNALS
Exponential smoothing, half-life of one year

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
sm1 sm2 sm3 sm4 sm5 sm6 sm7

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017

sm8 sm9 sm10 sm11 sm12 sm13 sm14

Expectations Anchoring Indexes for Brazil using Kalman Filter 201

 Figure 12
MARKET SIGNALS
Exponential smoothing, half-life of two years

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
sm1 sm2 sm3 sm4 sm5 sm6 sm7

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017

sm8 sm9 sm10 sm11 sm12 sm13 sm14

202 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 13
MARKET SIGNALS
Rolling window weights, window of three years

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
sm1 sm2 sm3 sm4 sm5 sm6 sm7

−2

−4
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017

sm8 sm9 sm10 sm11 sm12 sm13 sm14

Expectations Anchoring Indexes for Brazil using Kalman Filter 203

 Figure 14
FACTORS FROM LONG-TERM INFLATION EXPECTATION ANCHORING
Baseline ES2y

3.5

2.5

1.5

0.5

−0.5

−1.5

−2.5
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
F1,t F2,t Ft (combined factor)

Table 2
SIGNALS CONSTRUCTED FROM BREAKEVEN INFLATION (bei) MARKET DATA

Signals Description
sm1 slope from recursive ols regression, bei four years against bei one year (swaps)
sm2 recursive correlation between bei four years and one year (swaps)
sm3 Nautz and Strohsal (2015), fx-rate slope from ols (bei 4y swaps, macro shocks)
sm4 Natoli and Sigalotti (2017), slope from logit regression, ∆ bei swaps (1y, 4y)
sm5 (bei 4y swaps-inflation target)
sm6 (bei 4y swaps-inflation target)2
sm7 Demertzis et al . (2012), time-varying var, bei swaps (1y, 4y)
sm8 slope from recursive ols regression, bei four years against bei one year (bonds)
sm9 recursive correlation between bei four years and one year (bonds)
sm10 Nautz and Strohsal (2015), fx-rate slope from ols (bei 4y bonds, macro shocks)
sm11 Natoli and Sigalotti (2017), slope from logit regression, ∆ bei bonds (1y, 4y)(bei
4y bonds-inflation target)
sm12 (bei 4y bonds-inflation target)
sm13 (bei 4y bonds-inflation target)2
sm14 Demertzis et al . (2012), time-varying var bei bonds (1y, 4y)

204 F. Nascimento de Oliveira, W. P. Gaglianone

in the previous section, and build a filtered anchoring indicator from
the state-space model:

2 xt = Axt −1 + Bt ,

3 yt = Cxt + Dvt ,

where xt = [ct ; ft ; ot ] ’ is a vector of states and yt = [ zt ; Ft ; 1t ] ’ is a vector of ob-

servable variables, and εt and vt are uncorrelated Gaussian residuals.
First, ct is the monetary policy (espectations anchoring) state of in-
terest, f t is a state designed to capture the fiscal stance dynamics,
and ot is an auxiliary state to include the intercepts in the equations.
In turn, zt is the consensus expectation (Focus survey) of the pri-
mary fiscal balance as a percentage of gdp, one-year ahead, Ft is the
long-term anchoring factor and 1t is a constant series with unit values
to play the role of the intercept. The matrices A, B, C, and D are 3 x 3
null matrices, except for eight parameters estimated by maximum
likelihood (ml) within a standard Kalman filter.

 θ1 0 0 1 0 0  0 θ3 θ4 θ8 0 0 
    C = θ θ θ   
4 A = 0 θ2 0 ; B = 0 1 0  ;  5 6 7 and D = 0 0 0 
0 0 1  0 0 0  0 0 1  0 0 0 

Note that the state ot = 1t plays the role of the intercept and states
ct = θ1ct −1 + ε1,t and ft = θ2 ft −1 + ε 2,t are ar(1) processes with zero mean.
On the other hand, the observable fiscal expectation (zt ) is driven
by the fiscal state ( ft ) plus an intercept and the idiosyncratic shock
v1,t . The long-term anchoring factor Ft is decomposed into two states,
ct and ft , which are designed to capture, respectively, the dynamics
of monetary and fiscal policies.

5 zt = θ3 ft + θ 4 + θ8v1,t ,

6 Ft = θ5ct + θ6 ft + θ7

Expectations Anchoring Indexes for Brazil using Kalman Filter 205

 Table 3
CORRELATION MATRIX OF SELECTED SIGNALS (SURVEY AND MARKET)

S3 S9 S12 S13 S14 S15 S17 SM3 SM4 SM7 SM8 SM9 SM12 SM14
S3 1.00
S9 0.30 1.00
S12 –0.56 –0.62 1.00
S13 0.35 0.26 –0.22 1.00
S14 0.16 0.00 0.01 0.08 1.00
S15 0.00 0.26 –0.31 –0.11 –0.21 1.00
S17 0.48 0.69 –0.61 0.15 0.00 0.46 1.00
SM3 –0.07 –0.04 –0.22 0.03 –0.33 0.51 –0.15 1.00

206 F. Nascimento de Oliveira, W. P. Gaglianone

SM4 –0.34 –0.43 0.39 0.00 –0.16 0.07 –0.68 0.54 1.00
SM7 –0.08 –0.34 0.38 0.12 –0.05 –0.52 –0.51 0.10 0.47 1.00
SM8 –0.25 –0.57 0.66 0.11 –0.05 –0.26 –0.69 0.16 0.60 0.61 1.00
SM9 0.08 –0.08 0.29 0.38 0.15 –0.65 –0.31 –0.48 0.01 0.47 0.52 1.00
SM12 0.30 0.25 –0.63 –0.01 0.16 0.13 0.34 –0.21 –0.41 –0.57 –0.66 –0.23 1.00
SM14 –0.49 –0.40 0.56 0.14 0.07 –0.38 –0.68 –0.22 0.54 0.36 0.53 0.52 –0.15 1.00

Note: Only signals with pairwise absolute correlation below 0.7 are selected for the ES2y baseline case.
 Figure 15
MONETARY POLICY CREDIBILITY STATE (ct), FISCAL POLICY STATE (ft),
EXPECTATION OF PRIMARY FISCAL BALANCE (zt) AND LONG-TERM
AND ANCHORING FACTOR (Ft)

5 0.0
4 −0.2
3 −0.4
2 −0.6
1 −0.8
0 −1.0
−1 −1.2
−2 −1.4
−3 −1.6
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
zt (left scale) ct (left scale)
Ft (left scale) ft (right scale)

The following restrictions are employed in the ml estimation:

such that increases
in the states ct and f t represent a better anchored expectations
state and a better fiscal stance, respectively. Also note, from (5), that
the fiscal expectations series zt is not linked to the monetary policy
credibility state–which is a restriction adopted to properly identify
the model parameters–and that there is no residual in (6) to guar-
antee that all the dynamics observed in the common factor Ft are ei-
ther driven by the monetary policy state or by the fiscal policy state.20

20
This assumption, in principle, could be relaxed by including an error
term with zero mean and low variance (set as initial condition in the
Kalman filter estimation).

Expectations Anchoring Indexes for Brazil using Kalman Filter 207

 Table 4
FACTOR MODEL LOADINGS (BASELINE ES2Y)

Signal Loadings F1 Loadings F2 Communality Uniqueness

S3 –0.47 0.32 0.33 0.67
S9 –0.65 0.15 0.45 0.55
S12 0.80 –0.01 0.65 0.35
S13 –0.02 0.31 0.10 0.90
S14 –0.03 0.34 0.11 0.89
S15 –0.47 –0.68 0.69 0.31
S17 –0.87 0.11 0.77 0.23
SM3 0.02 –0.83 0.69 0.31
SM4 0.67 –0.50 0.70 0.30
SM7 0.67 0.11 0.47 0.53
SM8 0.86 –0.06 0.74 0.26
SM9 0.50 0.71 0.75 0.25
SM12 –0.62 0.20 0.42 0.58
SM14 0.74 0.19 0.59 0.41
Notes: Sample from September 28, 2005, to June 2, 2017 (2,916 workdays).
Unrotated loadings and prior communalities via squared multiple correlation.
The variation explained by the first factor is 37%, whereas the first and second
factors explain 55% of total variance.

As is well known, the model described in the system of equations 2-3

has only one global maximum, so initial conditions of the state vari-
able do not have any influence on its estimation by maximum likeli-
hood, except maybe on the number of interactions until convergence
is reached. 21 Finally, the eai is defined as the logit-transformed 22
smoothed Kalman filtered state ct . Table 5 presents the Kalman
filter parameter estimates and Figure 15 exhibits the states and ob-
servable variables in the baseline case.
We should stress that the results obtained from the reduced-form
model represented by equations (1) to (6) hinge on the assessment

21
We limit to 1,000 the number of interactions of the maximum likeli-
hood estimations. In all estimations presented in this paper, maximum
likelihood converged before reaching the limit of interactions. For the
Kalman filter, we considered the expectation of initial state vector equal
to zero.
22
To guarantee the EAI to be inside the [0;1] interval.

208 F. Nascimento de Oliveira, W. P. Gaglianone

that the expectations anchoring indexes concerning monetary pol-
icy have been disentangled from fiscal policy. Our strategy to imple-
ment such separation of policies is based on a standard state-space
model using survey and market data. We acknowledge that the simpli-
fied setup, due to several modelling assumptions, might not entirely
purge the fiscal policy outlook from the proposed expectations an-
choring index.23 The empirical results next presented should be in-
terpreted with this caveat in mind.

3.4 Baseline eais

Our baseline eais are the ones in which we used both signals from
survey and market data (total of 14 signals), selected with correlation
analysis (see Section 3.1.3). We create three versions of these indexes
depending on whether the signals are constructed from recursive
correlations (or regressions) weighting the observations with expo-
nentially smoothed weights with a half-life of one or two years or us-
ing a rolling window of three years (see Figure 16).
Because we have market data only starting from 2005, the baseline-
eais start then. Overall, they indicate that in the beginning of the
sample (2005-2008), the degree of expectations anchoring showed
a reasonably high and stable pattern. In other words, market infla-
tion expectations reflected the commitment of the bcb to keep in-
flation at the center of the inflation target.
When the subprime crisis hit Brazil’s economy, the expectations
anchoring indexes dropped and only started to improve again in the
second quarter of 2013, when a contractionist monetary cycle (in-
creases in the Selic interest rate) took place. By the end of the sample
(mid-2017), the eais reached similar levels to those observed in the
beginning of the sample, reflecting the bcb clear objective to curb

23
For instance, the single fiscal expectations series, coupled with an au-
toregressive structure assumed for the fiscal state ft, might not properly
capture the core standpoint of fiscal policy. Alternative approaches
to tackle this issue could consider, for instance, a state-space model con-
taining an entire block of equations (instead of a single one) to model
the fiscal policy in a disaggregate way. On the other hand, the set
of observable variables could include data from credit default swaps
and/or real interest rates (e.g., long-maturity forwards) or even risk
premium estimates using satellite term-structure models.

Expectations Anchoring Indexes for Brazil using Kalman Filter 209

inflation with the help of fiscal measures that intended to signal bet-
ter public debt dynamics.

3.5 Robustness Analyses

We conduct a robustness analysis in three main dimensions. First,
we create two other groups of eais based only on survey data or on mar-
ket data. Each one is divided into three other groups, again depending
on whether the signals are created from recursive correlations (or re-
gressions) in which observations are weighted by exponential smooth-
ing with a half-life of one or two years or a rolling window of three years.
Figures 17 and 18 show the evolution of these eais.
The dynamics of survey-eais are similar to the baseline ones, with
one important difference. Survey eais obtained with rolling windows
are more volatile (in particular, after 2006) when compared to the oth-
er survey eais. We do not have a precise explanation for this. Howev-
er, we suspect that this may have to do with the fact that we use binary
survey signals, which may have had a greater impact on this eai due
to the rolling windows.
As a second robustness exercise, we estimate and remove from
the breakeven inflation (bei) series the risk premium, which is expect-
ed to be nontrivial, particularly in the short run. To do so, we regress
each bei series against an intercept and the cross-section interquar-
tile range constructed from the survey-based inflation expectations
data (using the same forecast horizon). For instance, in the case of the
bei from swaps with one-year maturity, we use the following regression:

7 BEI swap 1y (t ) = a + b * IQR 1y (t ) + e (t ).

The risk premium series is proxied by *IQR1y(t), whereas the bei

series without risk premium is given by â+e(t).24 In the case of bei from
bonds, we include an additional regressor to account for liquidity pre-
mium (given by the ratio between the market value of ntn-bs and ltns
outstanding). Figure 19 shows the original bei series and those

24
The advantage of our approach is that the estimated risk premium
is “model-free” in the sense that it is not grounded on a specific theoreti-
cal model, but instead is solely based on survey data at the micro level.

210 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 16
EXPECTATIONS ANCHORING INDEX
Baseline

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
Baseline (RW) Baseline (ES1y) Baseline (ES2y)

Notes: ES1y and ES2y denote the exponenially smoothed weights with half-life of one
year and 2 years, respectively, and rw means rolling window weights (window of three
years). Only signals with pairwise absolute correlation below 0.7 are selected for the
baseline case. The following signals are selected: S3, S9, S12, S13, S14, S15, S17, SM3,
SM4, SM7, SM8, SM9, SM12 and SM14.

Table 5
KALMAN FILTER ESTIMATION OF THE EXPECTATIONS
ANCHORING INDEX (BASELINE ES2Y)

Parameter Estimate S.E.

θ1 0.9897 0.0004 a
θ2 0.9900 0.0004 a
θ3 5.7601 0.0682 a
θ4 5.8999 0.0669 a
θ5 1.5670 0.0105 a
θ6 1.0880 0.0552 a
θ7 0.2627 0.0016 a
θ8 0.0004 0.0546
Note: Sample from September 28, 2005, to June 2, 2017 (2,916 observations).
“a” indicates statistical significance at 1% level. Only signals with pairwise
absolute correlation below 0.7 are selected for the ES2y baseline case.
The following signals are selected: S3, S9, S12, S13, S14, S15, S17, SM3, SM4,
SM7, SM8, SM9, SM12, and SM14.

Expectations Anchoring Indexes for Brazil using Kalman Filter 211

without the risk premium. Figure 20 presents the effect of the risk
premium extraction in the expectations anchoring index construct-
ed with market data. They show similar dynamics to our baseline eais.
The third robustness check consists of using a different method
in the factor analysis. Instead of extracting two factors, we employ
here the minimum average partial (map) criterion for selecting
the number of factors. In the baseline case, the method suggests
a single factor, which is used as Ft in model (2)-(3). Figure 21 presents
the expectations anchoring index obtained from the single factor us-
ing map; with a very similar trajectory compared to the baseline eai.

4. CONCLUSION

According to Blinder (1998): “In the real world, credibility is not

created by incentive compatible compensation schemes or by rig-
id precommitment. Rather, it is painstakingly built up by a history
of matching deeds to words.”
Our objective in this paper is to build expectations anchor-
ing indexes for inflation in Brazil that are essentially driven from
the bcb’s ability to anchor long-term inflation expectations. The eais
are smoothed Kalman filtered maximum likelihood estimates from
a linear statespace model, which also includes expected fiscal dynam-
ics from survey data. The model signals give information on the de-
gree of long-term inflation expectation anchoring.
We derive our eais from surveys of inflation expectations and from
market data. Although varying across specifications, the expecta-
tions anchoring indexes that we propose tend to display a downward
trajectory, more clearly in 2009, and show a recovery starting in 2016
until the end of the sample (mid-2017).
Future extensions of the paper could include other signals of long-
term inflation anchoring. We also think that our method can be
extended to the creation of eai s for other central banks around
the world, despite different data on long-term inflation expectations
from those we have in Brazil and used in this paper.

212 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 17
EXPECTATIONS ANCHORING INDEX
Market signals

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
Market (RW) Market (ES1y) Market (ES2y)

Notes: ES1y and ES2y denote the exponenially smoothed weights with half-life of one
year and two years, respectively, and  means rolling window weights (window of
three years).

Figure 18
CREDIBILITY INDEX
Survey signals

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Sep 2002

Jan 2004

May 2005

Sep 2006

Jan 2008

May 2009

Sep 2010

Jan 2012

May 2013

Sep 2014

Jan 2016

May 2017

Survey (RW) Survey (ES1y) Survey (ES2y)

Notes: ES1y and ES2y denote the exponentially smoothed weights with half-life of one
year and two years, respectively, and  means rolling window weights (window of three
years).

Expectations Anchoring Indexes for Brazil using Kalman Filter 213

 Figure 19
MARKET DATA: BREAKEVEN INFLATION
AND RISK PREMIUM EXTRACTION
, percentage 12 months

10

−2
Apr 2005

Aug 2006

Dec 2007

Apr 2009

Aug 2010

Dec 2011

Apr 2013

Aug 2014

Dec 2015

Apr 2017
BEI swaps 1y BEI swaps 4y

BEI swaps 1y (no risk premium) BEI swaps 4y (no risk premium)

10

−2
Apr 2005

Aug 2006

Dec 2007

Apr 2009

Aug 2010

Dec 2011

Apr 2013

Aug 2014

Dec 2015

Apr 2017

BEI swaps 1y BEI swaps 4y

BEI swaps 1y (no risk premium) BEI swaps 4y (no risk premium)

214 F. Nascimento de Oliveira, W. P. Gaglianone

 Figure 20
EXPECTATIONS ANCHORING INDEX AND THE EFFECT
OF RISK PREMIUM EXTRACTION FROM MARKET DATA

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017
Market (Es2y) Market (ES2y), no risk premium

Figure 21
EXPECTATIONS ANCHORING USING A DIFFERENT METHOD
TO CONSTRUCT THE COMMON FACTOR Ft

1.0
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0.0
Sep 2005

Nov 2006

Jan 2008

Mar 2009

May 2010

Jul 2011

Sep 2012

Nov 2013

Jan 2015

Mar 2016

May 2017

Baseline (ES2y) Baseline (ES2y), single-factor (map)

Note: The single-factor comes from the “minimum average partial” criterion for
selecting the number of factors.

Expectations Anchoring Indexes for Brazil using Kalman Filter 215

References
Alessi, L., M. Barigozzi, and M. Capasso (2010), “Improved Penaliza-
tion for Determining the Number of Factors in Approximate
Factor Models,” Statistics & Probability Letters, Vol. 80, Nos.
23-24, pp. 1806-1813.
Andrews, D. W. K. (1993), “Tests for Parameter Instability and
Structural Change with Unknown Change Point,” Economet-
rica, Vol. 61, No. 4, pp. 817-858.
Bai, J., and S. Ng (2002), “Determining the Number of Factors in
Approximate Factor Models,” Econometrica, Vol. 70, No. 1,
pp. 191-221.
Barro, R. and D. Gordon (1983), “A Positive Theory of Monetary
Policy in a Natural Rate Model,” Journal of Political Economy,
91(4), pp. 589-610.
Bernanke, S., Ben. (2007), “Inflation Expectations and Inflation
Forecasting. Speech at the Monetary Economics Workshop
of the National Bureau of Economic Research Summer In-
stitute,” Cambridge, Massachusetts.
Blackburn, K., and M. Christensen (1989), “Monetary Policy and
Policy Credibility: Theories and Evidence,” Journal of Economic
Literature, 27(1), pp. 1-45.
Blanchard, Olivier (2004), Fiscal Dominance and Inflation Targeting:
Lessons from Brazil, nber Working Papers, No. 10389.
Blinder, A.S. (1998), Central Banking in Theory and Practice, The
mit Press, Cambridge.
Blinder, A. S. (2000), “Central-bank Credibility: Why Do We Care?
How Do We Build It?,” American Economic Review, Vol. 90,
No. 5, pp. 1421-1431.
Bolle, de Monica, and Nikola Mirkov (2017), “Fiscal Dominance
in High-frequency Data,” mimeo.
Bordo, D. M., and L. P. Siklos(2015), Central Bank Credibility:
An Historical and Quantitative Exploration, nber Working
Paper, No. 20824.
Buono, Ines, and Sara Formai (2016), The Evolution of the Anchoring
of Inflation Expectations, Questioni di Economia e Finanza
(Occasional Papers), No. 321, Economic Research and In-
ternational Relations Area, Bank of Italy.

216 F. Nascimento de Oliveira, W. P. Gaglianone

Cecchetti, S. G., and S. Krause (2002), “Central Bank Structure,
Policy Efficiency and Macroeconomic Performance: Explor-
ing Empirical Relationships,” Federal Reserve Bank of St. Louis
Economic Review, Vol. 84, No. 4, pp. 47-59.
Chow, G. C. (1960), “Tests of Equality between Sets of Coefficients
in Two Linear Regressions,” Econometrica, Vol. 28, No. 3, pp.
591-605.
Clarida, R., and D. Waldman (2007), Is Bad News About Inflation Good
News for the Exchange Rate?, nber Working Paper, No. 13010.
Cukierman, A., and A. Meltzer (1986), “A Theory of Ambiguity,
Credibility, and Inflation under Discretion and Asymmetric
Information,” Econometrica, Vol. 54, No. 5, pp. 1099-1128.
Davis, J. Scott (2012), Re-establishing Credibility: The Behavior of Infla-
tion Expectations in the Post-Volcker United States, Working Paper
117, Federal Reserve Bank of Dallas.
D’Amico, S., D. H. Kim, and M. Wei (2016), Tips from tips: the In-
formational Content of Treasury Inflation-protected Security Prices,
Finance and Economics Discussion Series, No. 2014-024,
Federal Reserve Board.
Demertzis, M., M. Marcellino, and N. Viegi (2012), “A Credibility
Proxy: Tracking us Monetary Developments,” The B. E. Journal
of Macroeconomics, Vol. 12, No. 1.
Dimitris, C., G. Dimitris, T. Jappelli, and M. Rooij (2016), Trust
in the Central Bank and Inflation Expectations, Working Paper
537, De Nederlandsche Bank.
Gaglianone, W. P. (2017), Empirical Findings on Inflation Expectations
in Brazil: A Survey, Working Paper, No. 464, Banco Central
do Brasil.
Garcia, M. G. P., and D. A. Guillén (2011), “Expectativas desagre-
gadas, credibilidade do Banco Central e cadeias de Markov,”
Revista Brasileira de Economia, Vol. 68, No. 2, pp. 197-223.
Gomme, P. (2006), “Central Bank Credibility,” Economic Commentary,
Federal Reserve Bank of Cleveland, August 1.
Isaacson, W. (2004), Benjamin Franklin: An American Life, Simon
and Schuster Editors.
Issler, J. V., and A. F. Santos (2017), “Central Bank Credibility and
Inflation Expectations: A Microfounded Forecasting Ap-
proach,” mimeo., Getulio Vargas Foundation.

Expectations Anchoring Indexes for Brazil using Kalman Filter 217

Johnson, R. and D. Wichern (1992), Applied Multivariate Statisti-
cal Analysis. Upper Saddle River, Prentice-Hall.
Jolliffe, I. T. (2002), Principal Component Analysis, Second edition,
Springer, New York.
Kaseeream, I. (2012), Essays on the Impact of Inflation Targeting in
South Africa, University of Zululand Working Paper.
Leal, A. R., A. N. Ranciaro, and C. A. O. Tejada (2012), “Credibi-
lidade não linear para as metas de inflação no Brasil,” 41°
Encontro Nacional de Economia, Anpec, 10-13 December,
2013.
Levieuge, G., Y. Lucotte, S. Ringuedé (2015), Central Bank Credibility
and the Expectations Channel: Evidence based on a new credibility
index, Working Paper 209, Narodowy Bank Polski.
Lowenkron, A., and M. G. P. Garcia (2007), Monetary Policy Cred-
ibility and Inflation Risk Premium: A Model with Application to
Brazilian Data, Textos para Discussão, No. 543, Department
of Economics puc-Rio (Brazil).
Massey, F.J. Jr. (2012), “The Kolmogorov-Smirnov Test for Good-
ness of Fit,” Journal of American Statistical Association, Vol. 46,
No. 256, pp. 68-78.
Mendonça, H. F. (2004), “Mensurando a Credibilidade do Regime
de Metas Inflacionárias no Brasil,” Revista de Economia Política,
Vol. 24, No. 3, pp. 344-350.
Mendonça, H.F., and G. Souza (2007), “Credibilidade do regime
de metas para inflação no Brasil,” Pesquisa e Planejamento
Econômico, Vol. 37, 247-282.
Mendonça, H. F., and G. Souza (2009), “Inflation Targeting Cred-
ibility and Reputation: The Consequences for the Interest
Rate,” Economic Modelling, Vol. 26, No. 6, pp. 1228-1238.
Nahon, B., and N. Meurer (2005), “A relação entre a credibili-
dade do banco central e a inflação no brasil do regime de
metas inflacionárias,” Encontro de Economia da Região
Sul-anpec-Sul.
Natoli, F., and L. Sigalotti (2017), A New Indicator of Inflation Ex-
pectations Anchoring, Working Paper, No. 1996, European
Central Bank.
Nautz, D., and T. Strohsal (2015), “Are us Inflation Expectations
Re-anchored?,” Economics Letters, Vol. 127, pp. 6-9.

218 F. Nascimento de Oliveira, W. P. Gaglianone

Quandt, R. E. (1960), “Tests of Hypotheses that a Linear System
Obeys Two Separate Regimes,” Journal of the American Statisti-
cal Association, Vol. 55, No. 290, pp. 324-330.
Sargent, T. J., and N. Wallace (1981), “Some Unpleasant Monetar-
ist Arithmetic,” Quarterly Review, Fall, Federal Reserve Bank
of Minneapolis.
Shen, P. (2006), “Liquidity Risk Premia and Breakeven Inflation
Rates,” Economic Review, Federal Reserve Bank of Kansas City,
second quarter, pp. 29-54.
Sicsú, J. (2002), “Expectativas inflacionárias no regime de metas
de inflação: uma análise preliminar do caso brasileiro,”
Economia Aplicada, Vol. 6, No. 4, pp. 703-711.
Stock, J. H., and M. W. Watson (1996), “Evidence on Structural
Instability in Macroeconomic Time Series Relations,” Journal
of Business and Economic Statistics, Vol. 14, No. 1, pp. 11-30.
Svensson, L. E. (1993), The Simplest Test of Inflation Target Credibility,
nber Working Paper, No. 4604.
Teles, V. K., and J. Nemoto (2005), “O Regime de Metas de Infla-
ção do Brasil é crível?,” Revista Brasileira de Economia, Vol.
59, No. 3, pp. 483-505.
Val, F. F., W.P. Gaglianone, M.C. Klotzle, and A. C. F. Figueiredo
(2017), Estimating the Credibility of Brazilian Monetary Policy
Using Forward Measures and a State-space Model, Working Paper,
No. 463, Banco Central do Brasil.
Walsh, C. (1995), “Optimal Contract for Central Bankers,” American
Economic Review85(1), pp. 150-167.

Expectations Anchoring Indexes for Brazil using Kalman Filter 219

 Inflation Forecasting and
Its Performance Evaluation
Forecasting Inflation in Argentina
Lorena Garegnani
Mauricio Gómez Aguirre

Abstract
During the year 2016, the Banco Central de la República Argentina has be-
gun to announce inflation targets. In this context, providing the authorities
of good estimates of relevant macroeconomic variables turns out to be crucial
to make the pertinent corrections in order to reach the desired policy goals. This
paper develops a group of models to forecast inflation for Argentina, which
includes autoregressive models, and different scale Bayesian var s (bvar),
and compares their relative accuracy. The results show that the bvar model
can improve the forecast ability of the univariate autoregressive benchmark’s
model of inflation. The Giacomini-White test indicates that a bvar performs
better than the benchmark in all forecast horizons. Statistical differences
between the two bvar model specifications (small and large-scale) are not
found. However, looking at the rmses, one can see that the larger model
seems to perform better for larger forecast horizons.
Keywords: Bayesian vector autoregressive, forecasting, prior specification,
marginal likelihood, small-scale and large-scale models.
jel classification: C11, C13, C33, C53.

Lorena Garegnani <[email protected]>, Principal Analyst of Research Depart-

ment at the Banco Central de la República Argentina, and Mauricio Gómez Aguirre
<[email protected]>, Principal Analyst of Research Department
at the Banco Central de la República Argentina. This research was developed within
the framework of cemla’s Joint Research Program 2017 coordinated by the Banco
Central de la República, Colombia. The authors thank counseling and technical ad-
visory provided by the Financial Stability and Development Group (fsd) of the Inter-
American Development Bank in the process of writing this document. The opinions
expressed in this publication are those of the authors and do not reflect the views
of cemla, the fsd group, the Inter-American Development Bank, or the Banco Cen-
tral de la República Argentina.

223
1. INTRODUCTION

S
everal long-term nominal commitments such as labor contracts,
mortgages and other debt are widespread features of modern
economies. Forecasting how the general price level will evolve
over the life of a commitment is an essential part of private sector
decision-making.
The existence of long-term nominal obligations is also among
the primary reasons economists generally believe that monetary
policy is not neutral, at least over moderate horizons.
Central banks aim is to keep inflation stable, and perhaps also
to keep output near an efficient level. With these objectives, the New
Keynesian model makes explicit that optimal policy will depend
on optimal forecasts (e.g., Svensson, 2005), and further, that policy
will be most effective when it is well understood by the public.
Under inflation targeting the central banks generally released fore-
casts in quarterly Inflation Reports in a way to be more transparent
in their actions. The costs and benefits of transparency are widely
debated, but the need for a central bank to be concerned with infla-
tion forecasting is broadly agreed. In short, inflation forecasting is of
foremost importance to households, businesses, and policymakers.
During the year 2016, the Banco Central de la República Argen-
tina (bcra) has begun to announce inflation targets. In this context,
providing the authorities of good estimates of relevant macroeco-
nomic variables turns out to be crucial to make the pertinent cor-
rections in order to reach the desired policy goals.
A standard tool in macroeconomics that is widely employed in fore-
casting is vector autoregressive (var) analysis. var s are flexible time
series models that can capture complex dynamic relationships among
macroeconomic aggregates. However, their dense parameterization
often leads to unstable inference and inaccurate out-of-sample fore-
casts, particularly for models with many variables, due to the estima-
tion uncertainty of the parameters.
Litterman (1980) and Doan, Litterman, and Sims (1984) have pro-
posed to combine the likelihood function (the data) with some infor-
mative prior distributions (the researcher’s belief about the values
of coefficients) to improve the forecasting performance of var mod-
els, introducing a Bayesian approach into var modeling.

224 L. Garegnani, M. Gómez

 In any Bayesian inference, a fundamental yet challenging step
is prior specification, which influences posterior distributions of the
unknown parameters and, consequently, the forecasts (Geweke,
2005). Fortunately, the literature has proposed some methodologies
to set how informative the prior distributions should be.
Regarding prior selection, Litterman (1980) and Doan, Litter-
man, and Sims (1984) set the tightness of the prior by maximizing
the out-of-sample forecasting performance of a small-scale model.
Many authors follow this strategy, such as Robertson and Tallman
(1999) and Wright (2009), and Giannone et al. (2014), who minimize
the root mean square error (rmse) of the forecasts.
On the other hand, Banbura et al. (2008) propose to control
the overfitting caused by the considerable number of variables in the
model, by selecting the shrinkage of the coefficients in such a way
as to give an adequate fitting in-sample. Within this second selection
strategy, we can find authors such as Giannone et al. (2012), Bloor
and Mathenson (2009), Carriero et al. (2015) and Koop (2011).
Banbura, Giannone, and Reichlin (2008) showed that, by ap-
plying Bayesian var methodology, they were able to handle large
unrestricted var s models and therefore they demonstrated that
var framework can be applied to empirical problems that require
the analysis of more than a few sets of time series. The authors showed
that a Bayesian var is a viable alternative to factor models or panel
var s for analysis of large dynamic systems.
This paper develops a group of models to forecast inflation for Ar-
gentina, which includes autoregressive models, and different scale
Bayesian var s (bvar), and compares their relative accuracy.
The paper is organized as follows: Section 2 presents the method-
ological aspects related to the application of Bayesian analysis in a
var framework, Section 3 presents a brief description of the data,
Section 4 goes through the empirical results, and finally, Section
5 concludes.

2. BAYESIAN var METHODOLOGY

A var model has the following structure

1 yt = c + B1 yt −1 + + B p yt − p + ε t ,

Forecasting Inflation in Argentina 225

 ε N (0, Σ )
where yt is a n × 1 vector of endogenous variables, t is a
n × 1 vector of exogenous shocks, c is a n × 1 vector of constants, B1
to Bp are n × n matrices, and Σ is n × n covariance matrix.
The bvar coefficients are a weighted average of the prior mean
(researcher’s belief) and the maximum likelihood (ml) estimators
(inferred from the data), with the inverse covariance of the prior
and the ml estimators as weights.
C on sider t he fol low i ng p o ster ior d i st r ibut ion for t he
var coefficients

2 β Ω ∼ N (β 0 ,Ω −1ξ )

where the vector β 0 is the prior mean (whose elements will represent
the coefficient in Equation 1, the matrix Ω is the known variance
of the prior, and ξ is a scalar parameter controlling the tightness
of the prior information. Even though Ω could have many shapes,
gamma and Wishart distributions are frequently used in the litera-
ture, since they ensure a normally distributed posterior.1
The conditional posterior of β can be obtained by multiplying
the prior by the likelihood function. The posterior takes the form

3 β (
, y ∼ N βˆ (ξ ),Vˆ(ξ ) , )
where

4 ( )
βˆ (ξ ) ≡ vec βˆ (ξ ) ,

and

5 Bˆ (ξ ) ≡ (x ʹx Σ −1 + ( ξ)−1) −1( x ʹy Σ −1 + ( ξ )−1 β 0 ),

1
If the posterior distributions are in the same family as the prior prob-
ability distribution, the prior and posterior are then called conjugate
distributions.

226 L. Garegnani, M. Gómez

 6 Vˆ(ξ ) ≡ (x ʹx Σ −1 + ( ξ )−1 )−1.

Vectors y and x represent observed data while β 0 is a matrix where

each column corresponds to the prior mean of each equation.
It is important to note that if we choose a large value for ξ , the prior
will have little weight into the posterior. This translates to large vola-
tility of the prior and not enough information coming from the prior.
On the other hand, if the ξ is set to a small value (i.e., close to zero),
the prior becomes more informative and the posterior mean moves
towards the prior mean. To see this point, we can express 5 as follows:

7 Bˆ (ξ ) ≡ ˆ[ −1
0 β 0 +(
−1
⊗ x ) y]

and

ˆ =[ −1 −1 −1 .
0 + Σ ⨂ x ʹx ]
8

If the second element between brackets in Equation 7 is multiplied

by (x ʹx )−1(x ʹx ), we obtain the following equations:

9 Bˆ (ξ ) ≡ ˆ[ −1
0 β 0 ]+
ˆ[Σ −1⨂ x ʹx (x ʹx )−1 x ʹ y]

10 Bˆ (ξ ) ≡ ˆ[ −1
0 β 0 ]+
ˆ[Σ −1 x ʹx (x ʹx )β ]
ols

As can be seen, the posterior is a weighted average between the pri-

or and the ordinary least square (ols) estimators,2 where the weights
are the reciprocal of the prior covariance matrix and the reciprocal
of the ols covariance matrix respectively. As a result, if the informa-
tion contained in the data is good enough to describe the process

2
The ols estimators of a var coincide exactly with the ml estimators
conditional on the initial values.

Forecasting Inflation in Argentina 227

behind it, the posterior will move towards the ols estimators. How-
ever, it is important to underscore that, even if the available series
are adequate to describe the data generating process, the researcher
could still formulate a hypothesis about the distribution of the pa-
rameters based on his own beliefs. That would imply ignoring the in-
formation contained in the data, and usually that kind of decisions
are based on strong beliefs.
The issue mentioned in the last paragraph demonstrates the need
to be cautious about choosing the prior mean and the hyperpriors.
In the following subsections, these aspects are discussed in more
detail.

2.1 Level or Growth Rate

It is unclear a priori whether transforming variables into their growth
rates can enhance the forecast performance of a bvar model. On one
hand, the level specification can better accommodate the existence
of long-run (cointegrating) relationships across the variables, which
would be omitted in a var in differences. On the other hand, Cle-
ments and Hendry (1996) have shown that in a classical framework,
differencing can improve forecasting performance in the presence
of instability.
There has been little effort in the bvar literature to compare speci-
fications in levels versus differences. Carriero et al. (2015) work with
this specific topic and found that models in growth rates generally
yield more accurate forecasts than those obtained from the models
in levels. However, we can find both approaches in the literature.
Following the Litterman (1986) tradition, some authors considered
bvar s with variables in levels (e.g., Banbura et al., 2008; Giannone
et al., 2014, and Giannone et al., 2012). Other authors used bvar s with
variables in differences or growth rates (e.g., Clark and McCracken,
2007, and Del Negro et al., 2004).
As mentioned above, there is no apparent reason to opt for series
in levels or in differences to work with; nevertheless, choosing a rep-
resentation ex-ante, gives us information about the characteristics
of the prior distribution (values of the mean prior). For example,
working with variables in differences implies that the persistence
of those variables should be low, and that one should impose a num-
ber close to zero as a prior mean of the first lag, denoting low persis-
tence in the series.

228 L. Garegnani, M. Gómez

 Since it is a good practice to start with some idea about the value that
the prior could take and encouraged by the evidence found by Carri-
ero et al. (2015), we have opted to work with variables in differences.
In the next subsection, we will treat the variance of the prior as an-
other aspect of prior distribution.

2.2 Choice of Hyperparameters and Lag Length Strategy

To select the hyperparameters and the lag length we will follow
the strategy suggested by Banbura et al. (2008), Carriero, et al. (2015)
and Giannone et al. (2012). Suppose, that a model is described by a
likelihood function p ( y|θ) and a prior distribution pγ (θ ), where θ
is a vector parameter of the model and γ is a vector of hyperparam-
eters affecting the distribution of all the priors of the model. It is
natural to choose these hyperparameters by interpreting the model
as a hierarchical one, i.e. replacing pγ (θ ) with p(θ γ ) and evaluating
their posterior (Berger, 1985; Koop, 2003). In this way, the posterior
can be obtained by applying Bayes’ law

11 p(γ y) ≈ p ( y γ )p (γ ),

where p(γ ) is the density of the hyperparameters and p( y γ ) is the

marginal likelihood. In turn, the marginal likelihood is the density
that comes from the data when the hyperparameters change–in oth-
er words, the marginal likelihood can be obtained after integrating
out the uncertainty about the parameters in the model,

12 p (γ y) = ∫ p ( y θ , γ )p (θ γ )d θ .

For every conjugate prior, the density p(γ y) can be computed

in closed form. To obtain the Bayesian hierarchical structure, it is
necessary to obtain the distribution of p(θ ) by integrating out the
hyperparameters

13 p (θ ) = ∫ p (θ , γ )π (γ )d γ .

Forecasting Inflation in Argentina 229

 More precisely, we can find different values of the prior distribu-
tion from different hyperparameter values and, in this way, we can
represent the posterior as:

14 p (θ , γ y) = p ( y θ , γ )p (θ , γ )p (γ ).

The marginal likelihood should be sufficient to discriminate

among models; in this sense, we can choose models with differ-
ent hyperparameters and different likelihood specification (more
precisely, lags length structure). To make this point operational,
we estimate different models, following Giannone et al. (2012),
who introduce a procedure allowing to optimize the values of the hy-
perparameters that maximize the value of the marginal likelihood
of the model. This implies that the hyperparameter values are not
set a priori but are estimated.
Then the marginal likelihood can be estimated for every com-
bination of hyperparameter values within specified ranges and for
different lag length structures, and the optimal combination is re-
tained as the one that maximizes that value.

2.3 Comparison Strategy

In this subsection, we present some details about our strategy for mod-
el comparison. We will mention the steps that we will follow to do
it and then give more details about the predictive ability tests used
for comparison:
a) Estimate a univariate ar model.

b) Compute the relative rmse to the ar from (a).

c) Compute the relative rmse to the bvar . 3

d) Run the test of Giacomini and White (2006) to compare both

models.
Our benchmark is a univariate model. This means that we have
at hand different statistical measures that cover both the frequentist

3
The mean of the predictive density is considered.

230 L. Garegnani, M. Gómez

and the Bayesian approaches. While frequentist literature tends
to compare the forecasts with actual values, Bayesian literature com-
pares the realized values with the whole posterior predictive density.
The testing methodology of Giacomini and White (2006) con-
sists on evaluating relative forecast accuracy with a Diebold-Maria-
no (1995) like test, but with one central difference: The size of the
in-sample estimation window is kept fixed, instead of expanding.
Using the sample observations available at time t, forecasts of yt +τ
are produced for different t for given τ periods into the future, with
rolling windows of estimation with the two models that are being
compared. The sequences of forecasts are then evaluated accord-
ing to some loss function and then the difference of forecast losses
is computed. This way, a time series of differences in forecast losses
∆Lt +τ (θˆ ) that depends on the estimated parameters is constructed.
The test then consists on a Wald test on the coefficients of the regres-
sion of that series against a constant, the unconditional version of the
test in Equation 15, or against other explanatory variables, the con-
ditional version in Equation 16:

15 ∆Lt +τ (θˆ ) = µ + ε t ,

16 ∆Lt +τ (θˆ ) = β ʹ X t + ε t .

Standard errors may be calculated using the Newey-West covari-

ances estimator, controlling for heteroskedasticity and autocorrela-
tion. In this paper, the unconditional version is used.
The Giacomini-White test4 has many advantages: It captures the ef-
fect of estimation uncertainty on relative forecast performance, it al-
lows for comparison between either nested or non-nested models,
and, finally, it is quite easy to compute.

4
See chapter 17 of the book by Hashimzade and Thornton (2013) for a
detailed discussion about this test.

Forecasting Inflation in Argentina 231

2.4 Model Specification
We follow Banbura et al. (2008) and analyze two var models that in-
corporate variables of special interest, including indicators of real
economic activity, consumer prices, and monetary variables. We con-
sider the following two alternative models:
Small-scale model. This is a small monetary var including three
key variables:
a) Prices: We used the consumer price index constructed by the
Instituto Nacional de Estadística y Censos de la República Argenti-
na (indec). After December 2006 until July 2012, the previous
series is linked with the evolution of the consumer price index
provided by the Instituto Provincial de Estadísticas y Censos
de San Luis and, after July 2012, series is again linked with
the evolution of the consumer price index of the city of Bue-
nos Aires. 5

b) Economic activity: We used a monthly economic activity indi-

cator known as emae (Estimador Mensual de Actividad Eco-
nómica) published by the indec. The emae is based on the
value added for each activity at a base price plus net taxes (wi-
thout subsidies), and it uses weights provided by Argentina’s
National Accounts (2004). It tries to replicate quarterly gdp at
a monthly frequency.

c) Interest rate: We used data from the bcra on 30 to 59-day fixed

term deposit rates.
Large-scale Model. In addition to the variables included in the small-
scale model, this version also includes the rest of the variables in the
data set. These are detailed in the next section.
In September 2016, Argentina transitioned to an inflation tar-
geting regime. This could generate a structural break in the mean
and variance. To account for this possible change in the mean of the

5
From December 2006 to October 2015, the index by the indec pre-
sented severe discrepancies with provincial and private price index,
and hence was discarded for that period.

232 L. Garegnani, M. Gómez

process, we incorporate a dummy variable in both specifications
(Marcelino and Mizon, 2000).6
As we compare models of different sizes, we need a strategy on how
to choose the shrinkage hyperparameter as models become larger.
As the dimension increases, we want more shrinkage, as suggested
by the analysis in De Mol et al. (2008) to control for overfitting. We set
the tightness of the prior for the model to have better in-sample fit;
in this way, we are shrinking more in a larger dimension model.

3. DATA

Our data set is composed of a group of 16 monthly macroeconomic

variables of Argentina available on a monthly frequency. Sources
of the series, the transformations did on them and their stationar-
ity characteristics are described in the Annex.

4. RESULTS

4.1 Estimation of the bvar Model

4.1.1 The Optimal Hyperparameters
We work with a Normal-Wishart bvar specification. In this type
of specification, there are two hyperparameters and two param-
eters. We estimate the overall tightness λ 1, lag decay λ 3 , and the
lag length as we have described in Section 2.2, and then we impose
the value of the prior mean (the autoregressive coefficient) equal
to zero as discussed earlier.
The hyperparameter of the overall tightness λ 1 is the standard
deviation of the prior of all the coefficients in the system other than
the constant. In other words, it determines how all the coefficients
are concentrated around their prior means.
λ
The term λ 3 is a decay factor and 1 (L 3 ) controls the tightness
on lag L relative to the first lag. Since the coefficients of higher order

6
In the Annex, we show the posterior estimation of the whole sample
to see the effect of this. We controlled the change in the mean due the
transition to an inflation targeting regime and indeed obtained a sig-
nificant coefficient in both models.

Forecasting Inflation in Argentina 233

lags are more likely to be close to zero than those of lower order lags,
the prior for the standard deviations of the coefficients decrease
as the lag length increases. The values usually used in the literature
are 1 or 2, so we settle for λ 3 = 2.
The prior variance of the parameters of ββ̂(ξ ) is set according to:

 1   λ 2
17 σ ij2 =  2   λ13 
 σ j  L 
 

where σ 2j denotes the ols residual variance of the autoregressive

coefficient for variable j, λ1 is an overall tightness parameter, L is
the current lag, and λ 3 is a scaling coefficient controlling the speed
at which coefficients for lags greater than 1 converge to 0.
For exogenous variables, we define the variances as:

σ 2 = ( λ1λ4 )
2
18

The results for the hyperparameters and prior means of the small
and the big scale model are shown in Table 1. All the hyperparameters
are equal for both type of models except for the hyperparameter λ 1.
The characteristics of our hyperparameters after the optimiza-
tion procedure is as follow:

Table 1
LIST OF HYPERPARAMETER VALUES

Hyperparameters values Large-scale model Small-scale model

Autoregressive coefficient: 0 0

Overall tightness (λ1 ). 0.05 0.23

Lag decay (λ3 ): 2 2

Exogenous variable tightness 1 1

Lag length 1 1

234 L. Garegnani, M. Gómez

 The hyperparameter λ1 is equal to 0.05 for the large-scale model
while the hyperparameter λ1 for the small-scale is 0.23. From a prac-
tical point of view, this means that the true value of the coefficients
estimated (posterior) is probably to be farther from the prior mean
in the small-scale model than in the large-scale one.
Another aspect to consider about λ1 is the fact that this hyperpa-
rameter impacts on the distribution of the parameters of lagged en-
dogenous and exogenous variables of each equation in the system.
In this sense, with more shrinkage, for example, it is less probable
that the posterior coefficients of the lagged endogenous and exog-
enous variables depart from the prior.
As can see in Table 1, the posterior coefficients of the variables
in the large-scale model are less probable to depart from the prior
than the small-scale ones. Models with lots of variables will tend
to have a better in-sample fit even when λ1 is set to loose value.
The posteriors obtained for the small- and the large-scale mod-
el of the inflation equation in each type of model are shown in the
Annex.

4.1.2 Forecasting Exercise

Our forecasting exercise is conducted in the following way. We esti-
mate the hyperparameters considering the whole sample, through
the maximization of the marginal likelihood; and then, we compute
the forecasts.
As we mentioned before, the data set goes from January 2004
to July 2017. We compute one-, three- and six-step-ahead forecasts
with rolling windows. The size of the estimation sample is the same
for each forecast horizon. Out-of-sample forecast accuracy is mea-
sured in terms of rmse of the forecasts. Therefore, we obtained three
rmses for each model.
Relative forecast accuracy is analyzed in Table 2, by computed
the different combinations of rmse ratios. On average, the bvar pres-
ents better accuracy than the benchmark independently of the fore-
cast horizon. For immediate horizons, the small-scale model slightly
outperforms the larger one, but the large-scale model outperforms
the small one for further forecast horizons.

Forecasting Inflation in Argentina 235

 Table 2
RELATIVE FORECAST ACCURACY

One-step-ahead Three-steps-ahead Six-steps-ahead

Ratio Ratio Ratio
Ratio Ratio large Ratio Ratio large Ratio Ratio large
small large model- small large model- small large model-
model- model- small model- model- small model- model- small
benchmark benchmark model benchmark benchmark model benchmark benchmark model

0.77 0.90 1.69 0.78 0.77 1.02 0.87 0.82 0.94

In the next subsection, we analyze these results with a Giacomi-

ni-White test.

4.2 Forecast Evaluation

To evaluate the predictive performance of the different models,
we used the tests described earlier. Each column of Table 3 contains
the probability value of Giacomini-White test statistic for the differ-
ent models.

Table 3
GIACOMINI-WHITE TEST

Large bvar vs. Small bvar vs. Difference between

Forecast horizon benchmark benchmark bvar models

One-step-
0.03 0.01 0.29
ahead

Three-steps-
0.00 0.00 0.49
ahead

Six-steps-
0.09 0.05 0.41
ahead

236 L. Garegnani, M. Gómez

 The result of the Giacomini-White test shows that, at a 5% of sig-
nificance level, the large bvar model outperforms the benchmark
for one step and three steps ahead forecast horizon, while the small
bvar outperforms the benchmark at a 5% significance level for all
forecast horizons. The last column of the table shows the Giacomini-
White test applied to the differences in predictive ability between
the small- and large-scale bvar models, but in this case, the differ-
ences are not significant for all forecast horizons.

5. CONCLUSIONS

This paper assesses the performance of Bayesian var to forecast in-

flation in Argentina. We considered a Normal-Wishart bvar specifi-
cation for a small- and a large-scale model of differentiated variables
setting the prior mean according to standard recommendations
in previous studies. The overall tightness hyperprior and the
lag length of the different models were set by optimization of the
marginal likelihood. We found that large-scale models have nar-
rower priors, giving more weight to the priors mean than small-
scale models.
Overall, the results show that the bvar model can improve the fore-
cast ability of the univariate autoregressive benchmark’s model of in-
flation. The Giacomini-White test indicates that a bvar performs
better than the benchmark in all forecast horizons. Statistical differ-
ences between the two bvar model specifications (small and large-
scale) are not found. However, looking at the rmses, one can see that
the larger model seems to perform better for larger forecast horizons.

Forecasting Inflation in Argentina 237

Annex A. Data Characteristics

Table A.1
LIST OF ENDOGENOUS VARIABLES

Source Description Transf. Characteristics

1 indec emae log sa Unit-root

2 indec cpi inflation – – Trend

Core cpi inflation (ex.

3 indec seasonal and regulated)
– – Trend

4 indec Industrial employment log sa Unit-root

5 indec Construction employment log sa Unit-root

6 indec Retail trade employment log sa Stationary

7 bcra M2 monetary aggregate log sa Unit-root

Multilateral nominal
8 bcra exchange rate
log – Unit-root

9 bcra 30 to 59-day deposit rate – – Unit-root

Imports of intermediate
10 indec goods
log sa Unit-root

11 indec Total exports log sa Unit-root

12 utdt Consumer confidence index – – Unit-root

13 indec Monthly supermarket sales log sa Unit-root

14 afcp Cement sales log sa Unit-root

15 minem Asphalt sales log – Stationary

16 merval Stock market index log – Unit-root

238 L. Garegnani, M. Gómez

Annex B. Results characteristics

Table B.1
SMALL bvar CHARACTERISTICS

Endogenous variables: Inflation, interest rate, real activity

Exogenous variables: Constant, dummy 2016-11

Estimation sample: July 2004 to July 2017

Sample size (omitting initial

156
conditions):

Number of lags included

1
in regression:

Prior: Normal-Wishart

Autoregressive coefficient: 0

Overall tightness: 0.23

Lag decay: 2

Exogenous variable tightness: 1

Table B.2
SMALL bvar INFLATION EQUATION COEFFICIENT VALUES

Median sd lb ub
inf(–1) 0.468 0.066 0.338 0.598
I(–1) 0.901 0.640 –0.356 2.157
Y(–1) 2.631 3.500 –4.237 9.499
Constant 0.280 0.071 0.140 0.420
d112016 –0.197 0.144 –0.479 0.086

Sum of squared residuals: 91.05

R-squared: 0.291
Adj. R-squared: 0.272

Forecasting Inflation in Argentina 239

 Table B.3
LARGE bvar CHARACTERISTICS

Inflation, interest rate, real activity, multilateral

exchange rate, industrial employment,
cement sales, asphalts sales, imports
Endogenous variables of intermediate goods, total exports, M2,
core inflation, construction employment,
consumer confidence index, supermarket
sales, stock market index

Exogenous variables Constant, dummy 2016-11

Estimation sample July 2004 to July 2017

Sample size 156

Number of lags 1

Prior Normal-Wishart

Autoregressive
0
coefficient

Overall tightness 0.05

Lag decay 2

Exogenous variable
1
tightness

240 L. Garegnani, M. Gómez

 Table B.4
LARGE BVAR INFLATION EQUATION COEFFICIENT VALUES

Median sd lb ub
inf(–1) 0.145 0.045 0.057 0.234

I(–1) 0.436 0.407 –0.362 1.235

Y(–1) 1.177 2.131 –3.005 5.359

E(–1) 7.261 3.431 0.528 13.994

empi(–1) 16.644 11.611 –6.143 39.431

cem(–1) –0.680 0.556 –1.771 0.410

asph(–1) 0.083 0.411 –0.723 0.888

imp(–1) 0.125 0.477 –0.810 1.061

exp(–1) 0.091 0.491 –0.873 1.055

M2(–1) 4.093 2.410 –0.637 8.823

infc(–1) 0.183 0.047 0.091 0.275

empc(–1) –1.452 2.933 –7.207 4.303

icc(–1) –0.011 0.013 –0.036 0.013

sup(–1) 2.243 1.322 –0.351 4.837

stk(–1) 0.133 1.110 –2.045 2.310

Constant 0.056 0.039 –0.021 0.132

d112016 –0.014 0.042 –0.096 0.067

Sum of squared residuals: 89.33

R-squared: 0.304
Adj. R-squared: 0.224

Forecasting Inflation in Argentina 241

References
Banbura M., D. Giannone, and L. Reichlin (2008), Large Bayesian
vars, Working Paper Series, No. 996, European Central Bank.
Blake, Andrew, and Haroon Mumtaz (2012), Center of Central Bank-
ing Studies Technical Handbook: Applied Bayesian Econometrics for
Central Bankers, Technical Handbook, No. 4, Bank of England.
Berger, J. (1985), Statistical Decision Theory and Bayesian Analysis,
Springer Series in Statistics.
Bloor, C., and T. Matheson (2009), Real-time Conditional Forecasts with
Bayesian vars: An Application to New Zealand, Reserve Bank
of New Zealand Discussion Paper Series, No. dp2009/02,
Reserve Bank of New Zealand.
Carriero A., T. Clark, and M. Marcellino (2015), “Bayesian vars:
Specification Choices and Forecast Accuracy,” Journal of Ap-
plied Econometrics, Vol. 30, Issue 1, pp. 46-73.
Clark, T., and M. McCracken (2007), Forecasting with Small Macro-
economic vars in the Presence of Instabilities, Finance and Eco-
nomics Discussion Series, No. 2007-41, Board of Governors
of the Federal Reserve System.
Clements, Michael P., and F. Hendry David (1996), “Multi-step
Estimation for Forecasting,” Oxford Bulleting of Economics and
Statistics, Vol. 58, Issue 4, pp. 657-684.
De Mol, C., D. Giannone, and L. Reichlin (2008), “Forecasting
Using a Large Number of Predictors–Is Bayesian Regression
a Valid Alternative to Principal Components?,” Journal of
Econometrics, Vol. 146, Issue 2, pp. 318-328.
Del Negro, Marco, F. Schorfheide, F. Smets, and R. Wouters (2004),
On the Fit and Forecasting Performance of New Keynesian models,
frb Atlanta Working Paper, No. 2004-37, Federal Reserve
Bank of Atlanta.
Diebold, F. X., and R. S. Mariano (1995), “Comparing Predictive
Accuracy,” Journal of Business and Economic Statistics, Vol. 13,
No. 253-263.
Doan, T., R. Litterman, and C. A. Sims (1984), “Forecasting and
Conditional Projection Using Realistic Prior Distributions,”
Econometric Reviews, No. 3, pp. 1-100.

242 L. Garegnani, M. Gómez

Koop, Gary (2011), Forecasting with Medium and Large Bayesian
vars, sire Discussion Papers, No. 2011-38, Scottish Institute
for Research in Economics.
Geweke J. (2005), Contemporary Bayesian Econometrics and Statistics,
Wiley, Hoboken.
Gelman, A. (2006), “Prior Distributions for Variance Parameters
in Hierarchical Models (Comment on Article by Browne
and Draper),” Bayesian Analysis, Vol. 1, No. 3, pp. 515-534.
Giacomini, R., and H. White (2006), “Test of Conditional Predic-
tive Ability,” Econometrica, Vol. 74, pp. 1545-1578.
Giannone, D., M. Lenza, and G. E. Primiceri (2012), Prior Selection
for Vector Autoregressions, ecares Working Papers, No. 2012-
002, Universite Libre de Bruxelles.
Giannone, D., M. Lenza, D. Momferatou, and L. Onorante (2014),
“Short-term Inflation Projections: A Bayesian Vector Autore-
gressive Approach,” International Journal of Forecasting, Vol.
30, No. 3, pp. 635-644.
Gneiting, T., and A. Raftery (2007), “Strict Proper Scoring Rules,
Prediction, and Estimation,” Journal of the American Statistical
Association, Vol. 102, No. 477.
Gneiting, T. (2011), “Making and Evaluating Point Forecasts,” Jour-
nal of the American Statistical Association, Vol. 106, pp. 746-762.
Hashimzade, N., and M. A. Thornton (2013), Handbooks of Re-
search Methods and Applications in Empirical Macroeconomics,
Ed. Edward Elgar.
Koop, G. (2003). Bayesian Econometrics, Ed. Wiley.
Litterman, R. B. (1979), Techniques of Forecasting Using Vector Autore-
gressions, Ph. D. thesis, University of Minnesota.
Litterman, R. B. (1980), A Bayesian Procedure for Forecasting with Vec-
tor Autoregression, Working Paper, Massachusetts Institute of
Technology, Department of Economics.
Litterman, R. B. (1986), “Forecasting with Bayesian Vector Autore-
gressions: Five Years of Experience,” Journal of Business and
Economic Statistics, Vol. 4, No. 1, pp. 25-38.
Mao, W. (2010), Bayesian Multivariate Predictions, PhD Thesis, Uni-
versity of Iowa.

Forecasting Inflation in Argentina 243

Marcellino, M., and G. Mizon (2000), “Modelling Shifts in the
Wage-price and Unemployment-inflation Relationships in
Italy, Poland and the uk,” Economic Modelling, Vol. 17, pp.
387-413.
Robertson, J. C., and E. W. Tallman (1999), “Vector Autoregressions:
Forecasting and Reality,” Economic Review, No. Q1, pp. 4-18.
Svensson, Lars E.O. (2005), “Monetary Policy with Judgment:
Forecast Targeting,” International Journal of Central Banking,
No. 1, pp. 1-54.
Waggoner, D. F., and T. Zha (1999), “Conditional Forecasts in
Dynamic Multivariate Models,” The Review of Economics and
Statistics, Vol. 81, No. 4, pp. 639-651.
Warne, A., G. Coenen, and K. Christoffel (2013), Marginalised Predic-
tive Likelihood Comparisons with Application to dsge, dsge-var,
and var Models, Technical Report, Center for Financial Stud-
ies.
Wright, J. H. (2009), “Forecasting us Inflation by Bayesian Model
Averaging,” Journal of Forecasting, Vol. 28, No. 2, pp. 131-144.

244 L. Garegnani, M. Gómez

MIDAS Modeling for Core
Inflation Forecasting

Luis Libonatti

Abstract
This paper presents a forecasting exercise that assesses the predictive poten-
tial of a daily price index based on online prices. Prices are compiled using
web scraping services provided by the private company PriceStats in coopera-
tion with a finance research corporation, State Street Global Markets. This
online price index is tested as a predictor of the monthly core inflation rate
in Argentina, known as “resto IPCBA” and published by the Statistics Office
of the City of Buenos Aires. Mixed frequency regression models offer a conve-
nient arrangement to accommodate variables sampled at different frequen-
cies and hence many specifications are evaluated. Different classes of these
models are found to produce a slight boost in out-of-sample predictive perfor-
mance at immediate horizons when compared to benchmark naïve models
and estimators. Additionally, an analysis of intra-period forecasts, reveals
a slight trend towards increased forecast accuracy as the daily variable ap-
proaches one full month for certain horizons.
Keywords: MIDAS, distributed lags, core inflation, forecasting.
JEL Classification: C22, C53, E37.

The author is affiliated with the Central Bank of Argentina. E-mail: luis.libonatti@
bcra.gob.ar. This research was undertaken within the framework of cemla’s Joint
Research Program 2017 coordinated by the Central Bank of Colombia. The author
gratefully acknowledges counseling and technical advisory provided by the Financial
Stability and Development (fsd) Group of the Inter-American Development Bank
in the process of writing this document. The opinions expressed in this publication
are those of the author and do not reflect the views of cemla, the fsd group,
the InterAmerican Development Bank or the Central Bank of Argentina.

245
1. INTRODUCTION

F
orecasting inf lation has become increasingly important
in Argentina as it is essential for economic agents to adjust
wages and prices—particularly in recent years—in a context
of high and volatile inflation. Having timely updates about the future
trajectory of the inflation rate is essential for conducting monetary
policy, specially, since the Central Bank is transitioning towards
an inflation targeting regime. Recent developments in the use of
“big data” have greatly facilitated tracking macroeconomic vari-
ables in real-time. A remarkable example is the construction of on-
line price indexes that are sampled daily, rather than monthly, as it
is standard for traditional price indexes from statistical offices.
The question naturally arises of whether this information can help
predict the future trajectory of traditional consumer price indexes.
Ghysels et al. (2004) introduced a regression framework that allows
for the exploitation of time series sampled at different frequencies,
known in the literature as Mixed Data Sampling (midas) regression
models. The methodology reduces to fitting a regression model to a
low-frequency variable using high-frequency data as regressors. As it
will be shown later, this technique closely resembles distributed
lag models. This paper employs this methodology to assess whether
the combination of price series sampled at different frequencies
is an effective tool for improving forecast accuracy compared to na-
ïve models, using the online price index constructed by PriceStats
in cooperation with State Street Global Markets.
The rest of the paper is organized as follows. In the next section,
a brief introduction to midas models is presented. In the third sec-
tion, existing theoretical research on midas regressions as well
as some applications in forecasting inflation are briefly reviewed.
In the fourth section, the forecasting exercise is described, and the
results are discussed. And finally, the fifth section concludes.

2. MIDAS REGRESSION MODELS

midas regression models propose a data-driven method to aggre-

gate high frequency variables into lower-frequency predictors. They
provide an alternative to the well-known “bridge” approach (Schum-
acher, 2016) in which high frequency variables are aggregated with

246
equal weights (flat aggregation).1 Ghysels et al. (2004) suggested com-
bining yt ,a low frequency process, and , a high frequency process
that is observed a discrete and fixed number of times m each time
a new value of yt  is observed, in a plain regression equation,

( m −1)
2.1 yt = ∑ θ j xt − j m + ut ,
( j = 0)

or more compactly,

2.2

where is a 1×m row vector that collects all the

corresponding to period t and is the vector
of weight coefficients. 2 Each j high frequency observation
within the low frequency period t enters the model linearly as a vari-
able accompanied by its specific weight, totaling m explanatory
variables and m weights, plus an error term. The high frequency sub-
index needs to be represented in terms of the low frequency index
t by noting that for since m is fixed, where
would be the most recent observation. This structure actually

conceals a high frequency lag polynomial

so that is similar in fashion to a distributed lags model.
To provide a clearer perspective, it is perhaps easier to intro-
duce matrix notation. Defining as the matrix

that groups all the x t vectors together; the collection

of the low frequency observations of size and
the residuals of the same length as y, it is possible to unveil a simple
multiple regression equation,

1
In fact, this can be considered a special case of a midas regression.
2
This equation may also include constants, trends, seasonal terms or other
low frequency explanatory variables.

midas Modeling for Core Inflation Forecasting 247

 2.3

Indeed, this problem can be solved by ordinary least squares

(ols) and this method will produce consistent coefficient estimates.
Equation (2.1) is usually referred to as the unrestricted midas regres-
sion model (u-midas). 3 However, an inconvenience arises when m,
the length of the vector is large relative to the sample size T, as is
usually the case in midas regressions. When this occurs, the mod-
els suffer from parameter proliferation and ols induces poor esti-
mates and consequently, poor forecasts. A straightforward way to
overcome this deficiency is to impose restrictions on the coefficients
of the high frequency lag polynomial and restate each as a func-
tion of some q hyperparameters and its subindex j (its position with-
in the low frequency lag polynomial) in such a way that Each
is redefined as where the vector γ is the collection
of q hyperparameters that characterize the weight function
Equation (2.1) is transformed to,

2.4

where λ is an impact parameter and the weights are normalized

so that they sum up to unity. Ghysels et al. (2004) initially recom-
mended what is known as the exponential Almon polynomial as a
candidate for weight function as it allows for many different shapes
and depends only on a few parameters. This is an exponentiated
version of an Almon lag polynomial, which is well known in the dis-
tributed lags literature,4

3
Foroni et al. (2015) present a detailed assessment of this strategy.
4
See for example the book by Judge et al. (1985).

248
 2.5

Another conventional candidate is the beta probability density,

2.6

with and
Parameterization as in equation (2.5) has proved to be quite pop-
ular and has become the standard among researchers, particularly
when q=2.
The introduction of constrained coefficients has many far-reach-
ing implications. The model turns nonlinear and lacks a closed form
solution. It is necessary to resort to nonlinear least squares and ap-
proximate the solution by numerical optimization routines. Addi-
tionally, the constraints are highly likely to introduce a bias in each
However, based on Monte Carlo simulations, when the sample
size is small relative to the number of parameters, Ghysels et al. (2016)
argue that both, parameter estimation precision and out-of-sample
forecast accuracy, gained by the increase in degrees of freedom,
far offset the effects of the bias generated by misspecified constraints.
midas models are generally intended as a direct forecasting tool
since this could prove to be more robust against misspecification
(Marcellino et al., 2006). This implies that estimation additionally
depends on the time displacement of the variables, and the
forecast horizon, 5 The direct strategy requires estimation
of as many models as per pair (d, h) is required. If T Y is the time in-
dex of latest yt available for estimation, and TX is the time index of the
latest available for both estimation and forecasting, then d can

be defined as Setting
a forecast can be computed with,

2.7

5
How many periods into the future it is necessary to forecast.

midas Modeling for Core Inflation Forecasting 249

 The “nowcast” can be retrieved when d=–1 and h=1. Note also that,
the fact that d is a rational number implies that it is possible to gen-
erate intra-period forecasts.
To arrive at equation (2.7), it is first necessary to estimate,

2.8

and then compute with the estimated parameters, and

and the vector
It is possible to extend the midas model by allowing for more than
m high frequency regressors. For example, by including pX lags of the
vector xt totaling high frequency variates where
the midas-dl model is formed,

2.9

or equivalently,

2.10

In matrix notation, this can be represented by,

2.11

 θ 0, 0 
x L x1−( m−1)/ m L x1− pX L x1− pX −( m−1)/ m  M 
 y1   1    u1 
 y   x2 L x2−( m−1)/ m L x2− pX L x2− pX −( m−1)/ m   θ0,m−1   u1 
 2   M   
 M = M O M O M O M   +  M .
    
 yT −1   xT −1 L xT −1−( m−1)/ m L xT − pX −1 L xT − pX −1−( m−1)/ m  θ pX ,0  uT −1 

 
 yT    
x L xT −( m−1)/ m L xT − pX L xT − pX −( m−1)/ m   M   uT 
 T  
θ pX ,m−1 

250
 If different weight functions for each θ r in equation (2.9), then
the multiplicative or aggregates-based midas model is obtained (Ghy-
sels et al., 2016). On the contrary, employing a single weight function
for all m × L X coefficients vectors θ r is also possible. The first meth-
od allows for greater flexibility but at the cost of more parameters
to estimate, so this possibility will not be considered, as this may not
be convenient for a very short sample size.
Other possible extensions include constructing high frequency
factors (Marcellino and Schumacher, 2010), incorporating cointegra-
tion relations (Miller, 2013), integrating Markov switching (Guérin
and Marcellino, 2013), estimating multivariate models (Ghysels
et al., 2007), using infinite polynomials (Ghysels et al., 2007) or add-
ing low frequency autoregressive augmentations (Ghysels et al.,
2007; Clements and Galvão, 2008; Duarte, 2014), for example. Fo-
roni and Marcellino (2013) provide a comprehensive survey of pos-
sible extensions in a recent survey about mixed frequency models.

3. LITERATURE REVIEW

Clements and Galvão (2008) were among the first to study applica-
tions of midas regressions to macroeconomic variables. In their pa-
per, they forecast u.s. real quarterly output growth in combination
with three different monthly variables: i) industrial production, ii)
employment growth, and iii) capacity utilization. They find a slight
increase in out-of-sample forecast accuracy with both vintage and re-
vised data compared to two benchmarks models, an autoregression
and an adl model in particular, for short-term horizons. They also
derive and assess a model with autoregressive dynamics introduced
as a common factor shared by the low and the high-frequency lag poly-
nomials. Based on comments by Ghysels et al. (2007), they argue that
including an autoregressive term in a standard midas model, as in
the next equation,

3.1 yt = φ yt −1 + λW ( L1/ m ; γ ) xt + ut ,

induces a seasonal response from yt to xt irrespective of wheth-

er xt exhibits a seasonal pattern. They suggest further restricting

midas Modeling for Core Inflation Forecasting 251

the model by adding a common lag polynomial shared between yt
and xt ,

3.2 (1 − φ L) yt = λ (1 − φ L)W ( L(1 m ) ; γ ) xt + ut ,

so that when writing the model in distributed lag representation,

the polynomial in L cancels out, eliminating the spurious season-
al response. A multi-step generalization of (3.2) for h -step-ahead
forecasts would be,

3.3 (1 − φ Lh ) yt = λ (1 − φ Lh ) W ( L1/ m ; γ ) xt + ut .

Armesto et al. (2010) analyze the performance of midas models

for the us economy for four different variable combinations: i) quar-
terly gdp growth and monthly employment growth; ii) monthly cpi in-
flation and daily Fed funds rate; iii) monthly industrial production
growth and a measure of term spread, and iv) employment growth
and again a measure of term spread. They contrast the results of flat
aggregation, the exponential Almon polynomial and a step weight
function, but are unable to find a dominant model specification. They
provide detailed results for one-step-ahead intra-period forecasting
performance of the models, computed by accumulating leads6 as the
high frequency variable approaches a full low frequency period.
They find an erratic pattern for the root mean square forecast error
(RMSFE) of the models as a function of the leads included in the re-
gression. Thus, in a real-time setting, which intra-period forecasts
could be the most accurate would not be trivial.
Monteforte and Moretti (2013) develop midas models to forecast
the euro area harmonized price index inflation. They put forward
a two-step approach involving low and high frequency variables.
In the first place, they estimate a generalized dynamic factor model
(Forni et al., 2000) for the inflation rate based on a set of variables,

6
In this instance “lead” refers to an observation of the high-frequency
predictor that corresponds to the same temporal period of the low fre-
quency variable.

252
and then they extract a common component and separate that into
a long-run and a cyclical, or short-run, component. The second step
consists in fitting the model of Clements and Galvão (2008) to cap-
ture short-term dynamics and use financial time series as high fre-
quency regressors, in addition to the long-run component previously
estimated as well as other low frequency variables. They design three
midas models, M1, M2 and M3, each with different high frequen-
cy regressors: i) M1 includes the short-term interest rate, changes
in interest rate spread and oil future prices; ii) M2 uses changes in the
wheat price, oil future quotes and the exchange rate; and finally, iii)
M3 consists of long-term rates, changes in the interest rate spreads,
and changes in the short-term rate. They contrast the out-of-sample
performance in terms of rsmfe of these models against the equations
for the inflation rate of two different low frequency vector autore-
gressions, and univariate random walks, autoregressions and autore-
gressive-moving average models. They compute all the intra-period
forecasts for the midas models and the monthly average of these daily
forecasts, and compare this average to all the low frequency models.
All the analysis is conducted for one-month-ahead and two-month-
ahead forecasts. They find on average a 20% reduction in forecast
error dispersion. The authors also provide a final empirical exercise
by using forecast combinations with the midas models and the in-
flation rate implied by financial derivatives, but this approach does
not produce any significant gains.
Duarte (2014) discusses in detail the implications of autoregressive
augmentations in midas regression models and diverse ways to incor-
porate them. She explores the out-of-sample performance of midas
models with autoregressive augmentations with no restrictions, with
an autoregressive augmentation with a common factor restriction,
and models with autoregressive augmentations with no restrictions
and a multiplicative scheme to aggregation. She then compares these
models to the same models but without the autoregressive compo-
nent, and to two low frequency benchmark models, a low frequency
autoregression and multiple regression model. She computes fore-
casts for quarterly euro area gdp growth based on three different
series: i) industrial production, ii) an economic sentiment indicator
and iii) the Dow Jones Euro Stoxx index. She disregards the seasonal
spikes impulse responses as the relevant impulse responses, as she
argues that it is not possible to single out a particularly relevant im-
pulse response for a mixed-frequency process since responses vary

midas Modeling for Core Inflation Forecasting 253

depending on when the shocks occur within the low-frequency pro-
cess. Although there is no superior model among all tested, Duarte
finds once again that there are sizable gains compared to the bench-
marks at all horizons.
Breitung and Roling (2015) propose a “nonparametric” midas
model to forecast monthly inflation rates using a daily predictor.
Instead of imposing any particular polynomial parameterization,
the nonparametric approach consists on enforcing some degree
of smoothness to the lag distribution by minimizing a penalized
least squares cost function,

3.4 S (θ ) = ( y − X θ )′ ( − X θ ) + ηθ ′D ′Dθ

where D is a ( m − 1) × ( m + 1) matrix such that

1 −2 1 0 L 0
0 1 −2 1 L 0
3.5 D= , # (3.5)
M O O O O M
 
0 L L 1 −2 1

and η is a pre-specified smoothing parameter. They refer to this

estimator the Smoothed Least Squares estimator, and its structure
closely resembles the well-known Hodrick-Prescott filter. If η is not
known, they suggest solving for the η that minimizes the Akaike
Information Criterion. Their target variable is the harmonized in-
dex of consumer prices for the euro area and they use a commodity
price index as a high frequency regressor. They compare their model
against the unconditional mean and the parametric midas model
(exponential Almon weights) for two different forecast horizons.
They conclude that the commodity index paired with the nonpara-
metric midas results in a reasonably good one-month-ahead fore-
casts. Additionally, the authors conduct a Monte Carlo experiment
and compare their model to four parametric midas alternatives: i)
the exponential Almon polynomial, ii) a hump shaped function, iii)
a declining linear function, and iv) a sinusoidal function. They find
that the nonparametric method performs on par with the parametric
competitors.

254
4. DATA, EXERCISE, AND RESULTS

The out-of-sample predictive performance of an online price index

will be analyzed to forecast the core inflation rate in real-time. To be
more specific, this will be assessed using many different midas spec-
ifications discussed in the previous sections and these estimations
will be compared with benchmark single frequency naïve models
and estimators. midas turn out to be intuitive for this purpose since
the monthly inflation rate can be approximately decomposed as the
aggregation of daily inflation rates of the corresponding month,
when evaluated in logarithmic differences, π tm ≈ ∑ ( log pτd − log pτd−1 )
τ ∈t

Atkeson and Ohanian (2001), Stock and Watson (2007) and Faust
and Wright (2009) have shown that simple benchmarks are not eas-
ily beaten by more sophisticated models (at least in the case of the
US economy), and so these could serve as a good starting point
to gauge the predictive power of the daily series.

4.1 Data
The online price index is compiled by the company PriceStats in co-
operation with State Street Global Markets, a leading financial re-
search corporation. PriceStates is a spin-off company that emerged
from the Billion Prices Project at mit, founded by professors Alber-
to Cavallo and Roberto Rigobón. It is the first company, institution,
or organization to apply a big data approach to produce real-time
(daily) price indexes to track general price inflation and other re-
lated metrics. Essentially, they collect daily data of prices from on-
line retailers by “web scraping” (i.e. recording price information
contained inside specific HyperText Markup Language tags in the
retailers’ websites) and aggregate the data by replicating the meth-
odology of a traditional consumer price index, as is done by Nation-
al Statistics Offices with offline prices. Cavallo (2013) goes through
the methodology and provides comparisons between online and of-
fline price indexes for Argentina, Brazil, Chile, Colombia, and Ven-
ezuela. He concludes that online price indexes can track the dynamic
behavior of inflation rates over time fairly well with the exception
of Argentina. In fact, the construction of online price indexes was ini-
tially motivated by the desire to provide the public with an alternate
measure of the inflation rate in Argentina because from the years

midas Modeling for Core Inflation Forecasting 255

2007 to 2015 there were large discrepancies between the official price
indexes compiled by the National Institute of Statistics and Census
(indec) and price indexes compiled by provincial statistics offices
or those compiled by private consultants. Throughout the rest of the
paper, this price index will be referred to as the State Street PriceS-
tats Index (ssps). Data for Argentina is available since November 1,
2007 with a three-day publication lag.
A provincial price index that raised itself to prominence in recent
years is the consumer price index compiled by the General Depart-
ment of Statistics and Censuses of the Government of the Autono-
mous City of Buenos Aires, known as ipcba . Although this index
only takes into account the territory of the City of Buenos Aires (with
a population close to 3 million), it should be reasonable to expect
that price dynamics in the Buenos Aires Metropolitan Area (which
encompasses a much larger population, close to 14 million or 1/ 3
of the total population of Argentina) share most of its features with
the pricing structure of the City of Buenos Aires, resulting from ar-
bitrage by reason of geographical proximity, as this should prevent
large distortions, at least in nonregulated markets. A more restricted
version of the index is also published, called “resto ipcba” (ripcba)
witch serves as a measure of core inflation. Compared to the headline
version, it excludes products with strong seasonal patterns and regu-
lated prices (e.g. public utility services) and represents 78.15% of the
headline index. ripcba is available from July 2012 onward and is
released monthly, with approximately a two-week publication lag.
These two indexes, as well as other provincial private and public
price indexes, are closely monitored by the monetary authorities
as well as the general public. This is particularly true for INDEC’s
recently introduced National Consumer Price Index. As the name
implies, this is the only index with full national coverage. However,
this index so far consists of less than two years of data points and this
limits the possibility of drawing any relevant inferences.
Inflation in Argentina in recent years has been high, unstable
and volatile, particularly from 2012 to most of 2016 when Argentina
experienced high monetization of fiscal deficits, strict capital controls
and two major devaluations of the currency.7 The average monthly

7
The last one coinciding with the lifting of the majority of the capital
controls in December 2015 and a subsequent transition to a flexible
exchange rate regime and inflation targeting.

256
 Figure 1
COMPARISON BETWEEN rIPCBA INFLATION AND SPSS INFLATION
AGGREGATED TO MONTHLY FREQUENCY

300 9
270 8
240 7

Monthly Inflation Rate (%)

Cumulative Inflation Rate (%)

210 6
180 5
150 4
120 3
90 2
60 1
30 0
0 −1

2013 2014 2015 2016 2017 2018 2013 2014 2015 2016 2017 2018
Period Period

ripcba ssps

inflation rate has been fluctuating around 2.2% for ripcba and 2.1%
for the monthly aggregated ssps series, with coefficients of variation
at 35% and 49% respectively. This should pose a significant chal-
lenge for economists’ ability to formulate accurate forecasts. Figure
1 illustrates the comparison between these two indexes and provides
a quick glimpse at the potential predictive power of the high-frequen-
cy index. Overall and for the scope of this work, ripcba is available
from July 2012 to December 2017 (66 data points) while ssps ranges
from November 1, 2007 to December 31, 2017 (3,714 data points).

4.2 Forecasting Exercise

The midas specifications tested were the midas-dl, the unrestrict-
ed autoregressive midas-dl (midas-adl), and the autoregressive
midas-dl with the common factor restriction (midas-adl-cf).
All midas specifications were evaluated with several high frequency

midas Modeling for Core Inflation Forecasting 257

regressors equal to m × L X ,8 with L X ∈ {1, 2, 3}, and forecasts were com-
puted for horizons h ∈ {1, 2, 3} over a 36-observation evaluation sample,
spanning from January 2015 to December 2017, and an 18 obser-
vation subsample from July 2016 to December 2017 (a period with
a more stable inflation rate), using recursive (expanding) windows.
midas-adl-cf models included quadratic and cubic variations of the
standard Almon polynomial and the exponential version, as well as the
Beta probability density function. midas-adl models further added
flat aggregation (equal weights); and finally, midas-dl models add-
ed the nonparametric (np) model described in Section 3. Forecast
combinations of the various midas models with equal weights were
also considered. In addition, all these models were compared to two
benchmarks: i) the low-frequency unconditional mean and ii) a low-
frequency first order autoregression.9
In a first stage, the models were estimated with a balanced dataset.
In other words, there is exact frequency matching: m daily observa-
tions from the same month or LX groups of m daily observations from
the same months correspond to a specific low-frequency monthly ob-
servation of the dependent variable. In total, two sets of rmsfe were
computed, one corresponding to the large sample and the other to a
reduced subsample. For all forecast horizons, d was set to d = −1 .
A second stage involved estimating intra-period forecasts for the
best selected LX for each forecast horizon based on the results from
the large sample of the first stage and briefly analyzing the stability
of the forecasts as more recent information is incorporated in the
models. When intra-period forecasts were computed, d is a fraction
in the interval [−1, 0) . More specifically, d = −1 + i / m for i in 1,,m
where m is the frequency. Forecasts from the autoregression and the
uncondditional mean remained the same throughout the month.
To account for the fact that ssps is an irregularly spaced series,
the frequency was assumed fixed at m = 28 , and so days 29 , 30 and 31
of each month are discarded. Daily inflation rates were first computed
with the full dataset and then the observations beyond day 28 of each
month were discarded.

8
First order midas-adl-cf models include m ×  L X + min ( L X , h ) .
high-frequency regressors since the common factor restriction increases
the number of variates depending on the forecast horizon and the num-
ber of high frequency lags.
9
A detailed list of the models can be found in Appendix A.

258
 Estimation was conducted in R with the midasr package developed
by Ghysels et al. (2016) while optimization was performed with three
routines included in optimx10 for nonlinear models or with the lm
function from the stats package for linear ones. Models that require
optimx were solved simultaneously with three optimization routines
(ucminf, nlminb and Nelder-Mead) for each model, forecast horizon
h , number of high frequency regressors LX , and out-of-sample pe-
riod. Only the best solution was kept. The algorithm was initialized
taking the hypothesis of equal weights and a null impact parameter
as starting conditions. This strategy delivered reasonable results
empirically and serves as a check on whether the high-frequency re-
gressors are actually relevant.

4.3 Empirical Results

Tables 1 and 2 summarize the main results of the first stage. In gen-
eral, for h = 1 (nowcasts), larger values of LX produce better results
while this tends to reverse when forecasting further into the future,
i.e. h = 3 . For h = 2 , the results are ambiguous and indicate that
LX = 2 or LX = 3 perform best. All three classes of midas models
exhibit similar performance irrespective of the inclusion of the au-
toregressive term or how it is incorporated. For all h , most midas
models for at least some LX are able to produce a small gain at around
10% when compared to the autoregression and a larger 25% against
the unconditional mean.11 The smaller sample greatly amplifies these
results. Note that for each h, there is a flat aggregation model that
performed very well and, at times, even better than standard midas
models, but overall, there is not a single midas model that system-
atically outperforms the rest. The forecast combination tested does
not seem to improve over any particular midas model.
Figures (2)-(4) condense the main findings of the second stage.
Forecasts for h = 1 display a clear trend towards better accuracy
as the high frequency variable reaches a full low frequency period.
In day 1 to day 28 point to point comparison, the rmsfe is reduced
by approximately 20% and particularly, in the second half of the
month, the models start to surpass the accuracy of the autoregression

10
A comprehensive description about this package can be found in Nash
and Varadhan (2011).
11
Tables with rmsfe ratios are presented in Appendix B.

midas Modeling for Core Inflation Forecasting 259

by up to 15% at most for some days. The improved performance,
when evaluated in the subsample, suggests that it is even possible
to obtain even better results as the inflation rate stabilizes. Similar
behavior, although less evident, is observed for forecasts for period
h = 3 in the case of midas-dl models. Forecasts for horizon h = 2 dis-
play a rather erratic pattern excepting the flat aggregation midas-dl
and midas-adl models.
Figure 5 zooms in on the evolution of all intra-period forecasts
for selected models, either h = 1 , h = 2 or h = 3 . Despite the intra-pe-
riod forecasts evidencing some volatility within the month, this does
not seem to be a major concern as inflation stabilizes at the end of
the sample. Additionally, note that forecasting further into the fu-
ture yields a dynamic closer to the unconditional mean of the whole
process. In the future, these results could be used as a training sample
from which to compute inverse mean square error weights and per-
form forecast combinations, which could prove to be effective in miti-
gating intra-period forecast volatility.
Although the results look promising, they should be interpreted
with caution. The predictive ability of the models was tested with
the methodology by Giacomini and White (2006)12 and both the un-
conditional and the conditional versions of the test were examined.
The midas models were evaluated against the two naïve benchmarks,
modeling the difference in forecast accuracy as a constant (uncondi-
tional) and also as a first order autoregression (conditional). The results
do not indicate that the difference in forecast accuracy is significant
(at 0.05) for most midas models. However, since the “large” out-of-
sample evaluation set actually constitutes a small sample by literature
standards, the result of the tests cannot be taken as final. As more ob-
servations become available, the tests could be updated with a larger
sample to arrive at a more robust conclusion.

12
This is similar to the standard test by Diebold and Mariano (1995).
The key difference lies in that the estimation sample size is kept fixed
instead of ever expanding, as this allows to better incorporate estimation
uncertainty and to compare nested models.

260
 Table 1
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, rmsfe
h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )

midas-dl 0.630 0.627 0.564 0.790 0.717 0.712 0.740 0.721 0.751
midas-adl 0.578 0.589 0.564 0.775 0.729 0.724 0.760 0.741 0.765
midas-adl-cf 0.592 0.625 0.561 0.785 0.722 0.719 0.748 0.733 0.768

Almon ( q = 3 )
midas-dl 0.660 0.623 0.571 0.819 0.731 0.720 0.755 0.724 0.757
midas-adl 0.609 0.609 0.574 0.805 0.745 0.731 0.770 0.739 0.770
midas-adl-cf 0.617 0.636 0.576 0.827 0.741 0.719 0.762 0.734 0.777

Exp. Almon ( q = 2 )
midas-dl 0.705 0.646 0.566 0.816 0.755 0.749 0.803 0.863 0.837
midas-adl 0.627 0.633 0.632 0.775 0.768 0.766 0.839 0.840 0.835
midas-adl-cf 0.639 0.629 0.557 0.765 0.745 0.885 0.875 0.831 0.860

Exp. Almon ( q = 3 )
midas-dl 0.731 0.648 0.661 0.834 0.826 0.742 0.809 0.785 0.778

midas Modeling for Core Inflation Forecasting 261

midas-adl 0.628 0.633 0.645 0.822 0.834 0.814 0.827 0.792 0.807
midas-adl-cf 0.663 0.661 0.563 0.812 0.863 0.866 0.824 0.800 0.824
 Beta

262
midas-dl 0.668 0.624 0.574 0.768 0.694 0.701 0.747 0.730 0.746
midas-adl 0.571 0.622 0.568 0.728 0.707 0.716 0.732 0.750 0.748
midas-adl-cf 0.614 0.619 0.558 0.739 0.697 0.704 0.740 0.736 0.737
Flat
midas-dl 1.158 0.617 0.568 0.745 0.673 0.690 0.713 0.736 0.745
midas-adl 0.944 0.592 0.568 0.733 0.694 0.746 0.729 0.766 0.777
Nonparametric
midas-dl 0.623 0.629 0.567 0.782 0.717 0.717 0.718 0.721 0.755
EW Forecast Combination
midas-dl 0.657 0.621 0.569 0.780 0.719 0.710 0.738 0.741 0.750
midas-adl 0.585 0.595 0.570 0.765 0.729 0.728 0.763 0.760 0.764
midas-adl-cf 0.609 0.627 0.563 0.768 0.734 0.754 0.770 0.756 0.770
Autoregression

p =1 0.619 0.619 0.619 0.757 0.757 0.757 0.757 0.757 0.757

Unconditional Mean

y 0.790 0.790 0.790 0.800 0.800 0.800 0.806 0.806 0.806

Notes: The evaluation sample comprises 36 data points, from January 2015 to December 2017. Characters in bold indicate the best
number of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number
of variables, LX , and forecast horizon, h .
 Table 2
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, rmsfe

h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )
midas-dl 0.555 0.521 0.427 0.646 0.541 0.538 0.568 0.548 0.547
midas-adl 0.460 0.482 0.430 0.608 0.547 0.544 0.579 0.562 0.569
midas-adl-cf 0.508 0.528 0.422 0.625 0.548 0.543 0.586 0.574 0.573

Almon ( q = 3 )
midas-dl 0.564 0.560 0.430 0.683 0.572 0.546 0.580 0.547 0.559
midas-adl 0.465 0.523 0.435 0.642 0.581 0.551 0.578 0.561 0.584
midas-adl-cf 0.533 0.569 0.433 0.663 0.581 0.549 0.596 0.581 0.601

Exp. Almon ( q = 2 )
midas-dl 0.568 0.523 0.432 0.656 0.540 0.539 0.574 0.561 0.561
midas-adl 0.470 0.453 0.453 0.600 0.538 0.535 0.568 0.572 0.578
midas-adl-cf 0.581 0.531 0.425 0.620 0.533 0.821 0.669 0.585 0.575

Exp. Almon ( q = 3 )
midas-dl 0.620 0.549 0.430 0.566 0.551 0.533 0.555 0.559 0.559

midas Modeling for Core Inflation Forecasting 263

midas-adl 0.469 0.460 0.456 0.640 0.651 0.540 0.565 0.561 0.565
midas-adl-cf 0.613 0.545 0.424 0.720 0.657 0.634 0.573 0.579 0.579
 Beta

264
midas-dl 0.598 0.491 0.444 0.592 0.506 0.523 0.587 0.570 0.583
midas-adl 0.489 0.437 0.441 0.550 0.513 0.529 0.578 0.585 0.569
midas-adl-cf 0.548 0.492 0.432 0.568 0.511 0.525 0.580 0.570 0.571
Flat
midas-dl 0.542 0.515 0.417 0.610 0.501 0.526 0.539 0.577 0.594
midas-adl 0.540 0.479 0.405 0.577 0.515 0.547 0.544 0.580 0.604
Nonparametric
midas-dl 0.551 0.539 0.432 0.633 0.543 0.540 0.554 0.548 0.557
EW Forecast
Combination
midas-dl 0.488 0.522 0.429 0.621 0.533 0.533 0.562 0.555 0.561
midas-adl 0.417 0.462 0.426 0.597 0.544 0.534 0.565 0.566 0.573
midas-adl-cf 0.541 0.527 0.426 0.624 0.551 0.592 0.595 0.573 0.575
Autoregression

p =1 0.552 0.552 0.552 0.675 0.675 0.675 0.692 0.692 0.692

Unconditional Mean
y 0.697 0.697 0.697 0.691 0.691 0.691 0.685 0.685 0.685

Notes: The evaluation sample comprises 18 data points, from July 2016 to December 2017. Characters in bold indicate the best
number of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number
of variables, LX , and forecast horizon, h .
 Figure 2
EVOLUTION OF THE RMSFE FOR HORIZON h =1 WITHIN A MONTH
FOR SELECTED MODELS WITH LX = 3

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7



0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

-- (2015.01:2017.12) -- (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

 (1) Flat Almon (q=2)

  Almon (q=3)

midas Modeling for Core Inflation Forecasting 265

 Figure 3
EVOLUTION OF THE RMSFE FOR HORIZON h =2 WITHIN A MONTH
FOR SELECTED MODELS WITH LX = 3

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7



0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

-- (2015.01:2017.12) -- (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

 (1) Flat Almon (q=2)

  Almon (q=3)

266
 Figure 4
EVOLUTION OF THE RMSFE FOR HORIZON h =3 WITHIN A MONTH
FOR SELECTED MODELS WITH LX = 2

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7



0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

- (2015.01:2017.12) - (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

-- (2015.01:2017.12) -- (2016.07:2017.12)

1.0 1.0
0.9 0.9
0.8 0.8
0.7 0.7




0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0 5 10 15 20 25 0 5 10 15 20 25
Day within month Day within month

 (1) Flat Almon (q=2)

  Almon (q=3)

midas Modeling for Core Inflation Forecasting 267

 Figure 5

EVOLUTION OF INTRA-PERIOD FORECASTS FOR SELECTED MODELS AND FORECAST HORIZONS

268
-- - 
7
6 Observed Value
5
-- Almond (q=2)
4
3 - Nonparametric
2
1

Monthly Inflation Rate (%)

0
01-2015 04-2015 07_2015 10-2015 01-2016 04-2016 07-2016 10-2016 01-2017 04-2017 07-2017 10-2017 01-2018
Date

-- - 

7
6 Observed Value
5
- Flat
4
3 - Flat
2
1

Monthly Inflation Rate (%)

0
01-2015 04-2015 07_2015 10-2015 01-2016 04-2016 07-2016 10-2016 01-2017 04-2017 07-2017 10-2017 01-2018
Date

-- - 

7

%)
 1

Mon
0
01-2015 04-2015 07_2015 10-2015 01-2016 04-2016 07-2016 10-2016 01-2017 04-2017 07-2017 10-2017 01-2018
Date

-- - 

7
6 Observed Value
5
- Almon (q=2)
4
3 - Nonparametric
2
1

Monthly Inflation Rate (%)

0
01-2015 04-2015 07_2015 10-2015 01-2016 04-2016 07-2016 10-2016 01-2017 04-2017 07-2017 10-2017 01-2018
Date

midas Modeling for Core Inflation Forecasting 269

5. CONCLUSION

For some particular midas specifications, there is a slight improve-

ment compared to the low-frequency benchmark autoregression
and the unconditional mean. In principle, this would imply that
high-frequency online price indices have a good potential to fore-
cast future behavior of consumer inflation for immediate horizons
in Argentina, but these results are still not robust. This could serve
as a useful complementary tool to assess the out-of-sample perfor-
mance of perhaps more sophisticated models. Future research could
focus on building an alternative variable such as a daily financial
factor as suggested by Monteforte and Moretti (2013) or comparing
with measures of market expectations in order to further validate
the findings of this paper.

ANNEX

Appendix A: midas Specifications

The full set of specifications of the models is detailed below. All mod-
els were estimated with L X ∈ {1, 2, 3} , h ∈ {1,2,3} and d as explained
in subsection 4.2. The subscript ( d , h ) on parameter estimates de-
noting dependence on d and h has been suppressed for simplicity.

midas-dl:

m ×LX −1
 2 
A.1 πˆTrIPCBA
+h = αˆ + λˆ ∑  ∑ γ s j s πTSSPS
−d − j m
j =0  s =1 

m ×LX −1
 3 
A.2 πˆTrIPCBA
+h
ˆ
= αˆ + λ ∑  ∑ γˆs j s πTSSPS
−d − j m
j =0  s =1 

 
e ∑ s =1 s
2
m ×LX −1 γ js
  SSPS
A.3 ˆ
πTrIPCBA
+h = αˆ + λˆ ∑  m ×L −1 2 γ j s πT −d − j m
j =0  ∑ j =0 X e ∑ s =1 s 
 

270
  
e ∑ s =1 s
3
m ×LX −1 γ js
  SSPS
A.4 ˆ
πTrIPCBA
+h = αˆ + λˆ ∑  m ×L −1 3 γ j s πT −d − j m
j =0  ∑ j =0 X e ∑ s =1 s 
 

A.5   

A.6   

A.7   

midas-adl:

A.8     

A.9     


A.10    



A.11    


 

A.12    
 

midas Modeling for Core Inflation Forecasting 271

 A.13    

midas-adl-cf:

A.14      

A.15      


A.16     



A.17     


 

A.18     
 

Other:

A.19   

A.20 

272
 Appendix B: Additional Tables
Table B.1
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, RATIO TO rmsfe OF AUTOREGRESSION ×100
h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )
midas-dl 101.8 101.4 91.2 104.2 94.7 93.9 93.1 90.8 94.5
midas-adl 93.4 95.1 91.1 102.3 96.2 95.6 95.6 93.2 96.2
midas-adl-cf 95.7 101.0 90.6 103.6 95.3 94.9 94.1 92.2 96.6

Almon ( q = 3 )
midas-dl 106.6 100.7 92.2 108.1 96.5 95.0 95.0 91.2 95.2
midas-adl 98.3 98.4 92.7 106.3 98.4 96.5 96.9 93.0 96.9
midas-adl-cf 99.8 102.7 93.0 109.2 97.8 94.9 95.8 92.3 97.7

Exp. Almon ( q = 2 )
midas-dl 113.8 104.4 91.5 107.7 99.7 98.8 101.0 108.5 105.3
midas-adl 101.3 102.2 102.1 102.4 101.4 101.1 105.5 105.7 105.1
midas-adl-cf 103.2 101.6 90.0 100.9 98.4 116.8 110.1 104.6 108.2

Exp. Almon ( q = 3 )

midas Modeling for Core Inflation Forecasting 273

midas-dl 118.2 104.7 106.8 110.1 109.1 97.9 101.8 98.8 97.9
 101.4 102.3 104.2 108.5 110.1 107.4 104.0 99.7 101.5

274
midas-adl
midas-adl-cf 107.1 106.8 90.9 107.3 114.0 114.4 103.7 100.6 103.7
Beta
midas-dl 108.0 100.8 92.7 101.4 91.6 92.6 94.0 91.8 93.8
midas-adl 92.3 100.5 91.7 96.1 93.4 94.5 92.1 94.4 94.2
midas-adl-cf 99.2 100.1 90.2 97.6 92.0 92.9 93.1 92.6 92.7
Flat
midas-dl 187.1 99.8 91.8 98.3 88.8 91.1 89.7 92.6 93.8
midas-adl 152.6 95.6 91.8 96.7 91.6 98.5 91.7 96.4 97.8
Nonparametric
midas-dl 100.6 101.6 91.7 103.3 94.6 94.6 90.4 90.7 95.0
EW Forecast
Combination
midas-dl 106.2 100.4 92.0 102.9 95.0 93.7 92.9 93.3 94.3
midas-adl 94.5 96.1 92.2 100.9 96.3 96.1 96.0 95.6 96.2
midas-adl-cf 98.4 101.4 90.9 101.4 96.9 99.5 96.9 95.1 96.9
Unconditional Mean

y 127.7 127.7 127.7 105.6 105.6 105.6 101.4 101.4 101.4

Notes: The evaluation sample comprises 36 data points, from January 2015 to December 2017. Characters in bold indicate the best number
of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number of variables, LX ,
and forecast horizon, h .
 Table B.2
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, RATIO TO rmsfe OF AUTOREGRESSION×100
h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )
midas-dl 100.6 94.4 77.4 95.7 80.1 79.6 82.2 79.2 79.0
midas-adl 83.4 87.4 77.9 90.1 81.0 80.6 83.8 81.2 82.4
midas-adl-cf 92.0 95.7 76.5 92.6 81.2 80.5 84.8 82.9 82.8

Almon ( q = 3 )
midas-dl 102.3 101.5 78.0 101.2 84.8 80.8 83.9 79.2 80.8
midas-adl 84.2 94.7 78.9 95.1 86.0 81.6 83.6 81.1 84.4
midas-adl-cf 96.6 103.2 78.5 98.3 86.0 81.3 86.1 84.1 86.9

Exp. Almon ( q = 2 )
midas-dl 103.0 94.8 78.3 97.1 80.0 79.9 83.1 81.2 81.2
midas-adl 85.3 82.1 82.1 88.9 79.7 79.2 82.1 82.8 83.5
midas-adl-cf 105.4 96.2 77.1 91.9 79.0 121.6 96.8 84.6 83.2

Exp. Almon ( q = 3 )
midas-dl 112.4 99.5 77.9 83.8 81.6 79.0 80.3 80.8 80.8

midas Modeling for Core Inflation Forecasting 275

 85.0 83.4 82.7 94.8 96.4 80.0 81.7 81.2 81.7

276
midas-adl
midas-adl-cf 111.1 98.8 77.0 106.6 97.3 94.0 82.9 83.7 83.7
Beta
midas-dl 108.5 89.1 80.5 87.7 74.9 77.5 84.9 82.4 84.3
midas-adl 88.6 79.2 80.0 81.5 75.9 78.4 83.5 84.6 82.3
midas-adl-cf 99.4 89.3 78.3 84.1 75.7 77.7 83.8 82.4 82.5
Flat
midas-dl 98.3 93.4 75.6 90.3 74.2 77.9 77.9 83.4 85.9
midas-adl 98.0 86.8 73.4 85.5 76.3 81.0 78.7 83.8 87.3
Nonparametric
midas-dl 100.0 97.8 78.3 93.8 80.4 80.0 80.1 79.2 80.6
EW Forecast
Combination
midas-dl 88.4 94.7 77.7 91.9 78.9 78.9 81.2 80.2 81.2
midas-adl 75.5 83.9 77.3 88.4 80.6 79.1 81.8 81.9 82.8
midas-adl-cf 98.1 95.5 77.3 92.4 81.6 87.7 86.0 82.9 83.1
Unconditional Mean

y 126.3 126.3 126.3 102.4 102.4 102.4 99.0 99.0 99.0

Notes: The evaluation sample comprises 18 data points, from July 2016 to December 2017. Characters in bold indicate the best number
of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number of variables, LX ,
and forecast horizon, h .
 Table B.3
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, RATIO TO rmsfe OF UNCONDITIONAL MEAN×100

h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )
midas-dl 79.7 79.4 71.4 98.7 89.6 88.9 91.8 89.5 93.2
midas-adl 73.1 74.5 71.4 96.9 91.1 90.5 94.2 91.9 94.9
midas-adl-cf 75.0 79.1 71.0 98.1 90.3 89.8 92.8 90.9 95.3

Almon ( q = 3 )
midas-dl 83.5 78.9 72.2 102.4 91.3 89.9 93.7 89.9 93.9
midas-adl 77.0 77.1 72.6 100.6 93.2 91.3 95.5 91.7 95.6
midas-adl-cf 78.1 80.4 72.9 103.4 92.6 89.8 94.5 91.0 96.3

Exp. Almon ( q = 2 )
midas-dl 89.1 81.7 71.7 101.9 94.4 93.6 99.6 107.0 103.8
midas-adl 79.4 80.1 79.9 96.9 96.0 95.7 104.0 104.2 103.6
midas-adl-cf 80.8 79.5 70.5 95.6 93.1 110.6 108.5 103.1 106.7

Exp. Almon ( q = 3 )
midas-dl 92.5 82.0 83.6 104.3 103.3 92.7 100.3 97.4 96.6

midas Modeling for Core Inflation Forecasting 277

midas-adl 79.4 80.1 81.6 102.7 104.2 101.7 102.6 98.3 100.1
 83.8 83.7 107.9 108.3 102.2 99.2 102.2

278
midas-adl-cf 71.2 101.6
Beta
midas-dl 84.5 79.0 72.6 96.0 86.7 87.6 92.6 90.5 92.5
midas-adl 72.3 78.7 71.8 91.0 88.4 89.4 90.8 93.0 92.8
midas-adl-cf 77.6 78.4 70.6 92.4 87.1 88.0 91.8 91.3 91.4
Flat
midas-dl 146.5 78.1 71.9 93.1 84.1 86.2 88.4 91.3 92.5
midas-adl 119.5 74.9 71.9 91.6 86.7 93.3 90.5 95.0 96.4
Nonparametric
midas-dl 78.8 79.6 71.8 97.8 89.6 89.6 89.1 89.4 93.7
EW Forecast Combination
midas-dl 83.2 78.6 72.1 97.5 89.9 88.7 91.6 92.0 93.0
midas-adl 74.0 75.3 72.2 95.6 91.2 90.9 94.6 94.3 94.8
midas-adl-cf 77.1 79.4 71.2 96.0 91.7 94.2 95.5 93.7 95.5
Autoregression

p =1 78.3 78.3 78.3 94.7 94.7 94.7 98.6 98.6 98.6

Notes: The evaluation sample comprises 36 data points, from January 2015 to December 2017. Characters in bold indicate the best
number of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number
of variables, LX , and forecast horizon, h .
 Table B.4
OUT-OF-SAMPLE PREDICTIVE PERFORMANCE, RATIO TO rmsfe OF UNCONDITIONAL MEAN×100
h =1 h =2 h =3

LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3 LX = 1 LX = 2 LX = 3

Almon ( q = 2 )
midas-dl 79.6 74.8 61.3 93.5 78.2 77.7 83.0 80.0 79.8
midas-adl 66.1 69.2 61.7 88.0 79.1 78.7 84.6 82.0 83.2
midas-adl-cf 72.9 75.8 60.6 90.4 79.2 78.5 85.6 83.8 83.7

Almon ( q = 3 )
midas-dl 81.0 80.3 61.8 98.8 82.8 78.9 84.7 79.9 81.6
midas-adl 66.7 75.0 62.4 92.9 92.9 84.0 79.7 84.5 81.9
midas-adl-cf 76.5 81.7 62.2 95.9 84.0 79.4 87.0 84.9 87.7

Exp. Almon ( q = 2 )
midas-dl 81.5 75.0 62.0 94.8 78.1 78.0 83.9 82.0 82.0
midas-adl 67.5 65.0 65.0 86.8 77.8 77.3 82.9 83.6 84.4
midas-adl-cf 83.5 76.2 61.0 89.7 77.1 118.7 97.7 85.4 84.0

Exp. Almon ( q = 3 )
midas-dl 89.0 78.8 61.7 81.8 79.6 77.2 81.1 81.6 81.6

midas Modeling for Core Inflation Forecasting 279

midas-adl 67.3 66.0 65.5 92.6 94.1 78.1 82.5 82.0 82.5
 87.9 78.2 60.9 104.1 95.0 91.8 83.7 84.5 84.5

280
midas-adl-cf
Beta
midas-dl 85.9 70.5 63.7 85.6 73.2 75.7 85.7 83.2 85.2
midas-adl 70.2 62.7 63.3 79.6 74.1 76.5 84.3 85.5 83.1
midas-adl-cf 78.7 70.7 62.0 82.1 73.9 75.9 84.6 83.2 83.3
Flat
midas-dl 77.8 73.9 59.8 88.2 72.4 76.1 78.7 84.2 86.8
midas-adl 77.6 68.7 58.1 83.5 74.5 79.1 79.4 84.6 88.2
Nonparametric
midas-dl 79.1 77.4 62.0 91.5 78.5 78.1 80.9 80.0 81.4
EW Forecast Combination
midas-dl 70.0 75.0 61.5 89.7 77.0 77.0 82.0 81.0 82.0
midas-adl 59.8 66.4 61.2 86.3 78.7 77.2 82.6 82.7 83.6
midas-adl-cf 77.6 75.6 61.2 90.2 79.7 85.6 86.8 83.7 84.0
Autoregression

p =1 79.2 79.2 79.2 97.6 97.6 97.6 101.0 101.0 101.0

Notes: The evaluation sample comprises 18 data points, from July 2016 to December 2017. Characters in bold indicate the best number
of variables, LX , for each model and forecast horizon, h . Characters in italics indicate the best model for each number of variables,
LX , and forecast horizon, h
References
Armesto, M.T., K.M. Engemann and M.T. Owyang (2010), “Fore-
casting with Mixed Frequencies,” Federal Reserve Bank of St.
Louis Review, 92(6), pp. 521–36.
Atkeson, A., and L.E. Ohanian (2001), “Are Phillips Curves Useful
for Forecasting Inflation?,” Federal Reserve Bank of Minneapolis
Quarterly Review, 25(1), pp. 2–11.
Breitung, J., and C. Roling (2015), “Forecasting Inflation Rates Us-
ing Daily Data: A Nonparametric midas Approach,” Journal
of Forecasting, 34(7), pp. 588–603.
Cavallo, A. (2013), “Online and Official Price Indexes: Measuring
Argentina’s Inflation,” Journal of Monetary Economics, 60(2),
pp. 152–165.
Clements, M.P., and A.B. Galvão (2008), “Macroeconomic Forecast-
ing with Mixed-Frequency Data: Forecasting Output Growth
in the United States,” Journal of Business & Economic Statistics,
26(4), pp. 546–554.
Diebold, F.X., and R.S. Mariano (1995), “Comparing Predictive
Accuracy,” Journal of Business & Economic Statistics, 13(3),
pp. 134–144.
Duarte, C. (2014), Autoregressive Augmentation of MIDAS Regressions,
Banco de Portugal Working Paper 201401, Lisbon, Portugal,
Banco de Portugal.
Faust, J., and J.H. Wright (2009), “Comparing Greenbook and Reduced
Form Forecasts Using a Large Realtime Dataset,” Journal of Busi-
ness & Economic Statistics, 27(4), pp. 468–479.
Forni, M. et al. (2000), “The Generalized Dynamic-Factor Model:
Identification and Estimation,” Review of Economics and Sta-
tistics, 82(4), pp. 540–554.
Foroni, C., and M. Marcellino (2013), A Survey of Econometric Methods
for Mixed-Frequency Data, Norges Bank Working Paper 2013-
06, Oslo, Norway: Norges Bank.
Foroni, C., M. Marcellino and C. Schumacher (2015), “Unrestricted
Mixed Data Sampling (midas): midas Regressions with Unre-
stricted Lag Polynomials,” Journal of the Royal Statistical Society:
Series A (Statistics in Society), 178(1), pp. 57–82.

midas Modeling for Core Inflation Forecasting 281

Ghysels, E., V. Kvedaras and V. Zemly (2016), “Mixed Frequency
Data Sampling Regression Models: The R Package midasr,”
Journal of Statistical Software, 72(4), pp. 1–35.
Ghysels, E., P. Santa-Clara and R. Valkanov (2004), The MIDAS Touch:
Mixed Data Sampling Regression Models, CIRANO Working Pa-
per 2004s-20, Montreal, Canada, Center for Interuniversity
Research and Analysis of Organizations (CIRANO).
Ghysels, E., A. Sinko and R. Valkanov (2007), “ midas Regressions:
Further Results and New Directions,” Econometric Reviews,
26(1), pp. 53–90.
Giacomini, R., and H. White (2006), “Tests of Conditional Predic-
tive Ability,” Econometrica, 74(6), pp. 1545–1578.
Guérin, P., and M. Marcellino (2013), “Markov-Switching midas
Models,” Journal of Business & Economic Statistics, 31(1), pp.
45–56.
Judge, G.G. et al. (1985), Theory and Practice of Econometrics. Second
Edition, New York, United States, John Wiley & Sons.
Marcellino, M., and C. Schumacher (2010), “Factor midas for
Nowcasting and forecasting with Ragged-Edge Data: A Model
Comparison for German GDP,” Oxford Bulletin of Economics
and Statistics, 72(4), pp. 518–550.
Marcellino, M., J.H. Stock and M.W. Watson (2006), “A Comparison
of Direct and Iterated Multistep AR Methods for Forecasting
Macroeconomic Time Series,” Journal of Econometrics, 135(1),
pp. 499–526.
Miller, J.I. (2013), “Mixed-Frequency Cointegrating Regressions
with Parsimonious Distributed Lag Structures,” Journal of
Financial Econometrics, 12(3), pp. 584–614.
Monteforte, L., and G. Moretti (2013), “Real-Time forecasts of
Inflation: The Role of Financial Variables,” Journal of Forecast-
ing, 32(1), pp. 51–61.
Nash, J.C., and R. Varadhan (2011), “Unifying Optimization Algo-
rithms to Aid Software System Users: Optimx for R,” Journal
of Statistical Software, 43(9), pp. 1–14.
Schumacher, C. (2016), “A Comparison of midas and Bridge Equa-
tions,” International Journal of Forecasting, 32(2), pp. 257–270.

282
Stock, J.H., and M.W. Watson (2007), “Why Has US Inflation
Become Harder to Forecast?,” Journal of Money, Credit and
Banking, 39(1), pp. 3–33.

midas Modeling for Core Inflation Forecasting 283

Evaluation of Inflation Forecasting
Models in Guatemala

Juan Carlos Castañeda-Fuentes

Carlos Eduardo Castillo-Maldonado
Héctor Augusto Valle-Samayoa
Douglas Napoleón Galindo-Gonzáles
Juan Carlos Catalán-Herrera
Guisela Hurtarte-Aguilar
Juan Carlos Arriaza-Herrera
Edson Roger Ortiz-Cardona
Mariano José Gutiérrez-Morales

Abstract
Inflation forecasts are a relevant input for monetary policy decision in central
banks, particularly for those operating under inflation targeting. Therefore,
central banks must continuously evaluate the forecasting accuracy of the mod-
els used to produce inflation forecasts. In this study, we present the results
of an exhaustive evaluation of historical forecast performance for the most
important models used to forecast inflation at Banco de Guatemala. We find
evidence supporting the claim that time series models perform better for short
time horizons while conditional dsge models (both structural and semi-struc-
tural) do better in medium and long time horizons.
Keywords: economic forecasting, forecasting accuracy, forecasting
efficiency.
jel classification: C530.

J. C. Castañeda-Fuentes <[email protected]>, Director, Economic Research De-

partment, Banco de Guatemala. Corresponding Author. This research project was de-
veloped at Banco de Guatemala’s Economic Research Department, within the frame-
work of the cemla’s Joint Research Program 2017, coordinated by the Banco de la
República, Colombia. The authors thank counseling and technical advice provided
by the Financial Stability and Development (fsd), Group of the Inter-American De-
velopment Bank in the process of writing this document. The opinions expressed
herein are those of the authors and do not necessarily reflect the views of cemla,
the fsd group, the Inter-American Development Bank or Banco de Guatemala.

285
1. INTRODUCTION

B
anco de Guatemala adopted a monetary policy framework based
on inf lation targeting (it) in 2005. Because of the forward-
looking nature of that regime, central bank authorities should
base their policy decisions on reliable inflation forecasts. In fact,
Banco de Guatemala employs an array of models to forecast inflation,
which include ols, arima , structural and semi-structural dsge type
of models, as well as forecast combinations of all, or some of these
approaches. Since each of these models provides different informa-
tion about the future path of inflation, a rigorous evaluation of their
performance is required in order to determine their reliability, so that
the central bank staff could give more weight to more reliable models,
and improve the less reliable ones or get rid of them.
This document presents the results of a thorough evaluation of the
most frequently used models by Banco de Guatemala to forecast in-
flation. Our evaluation is divided according to the type of model
employed to produce a forecast. First, we evaluate models that pro-
duce unconditional forecasts, based on four different approach-
es: 1) forecasting accuracy and bias; 2) ability to predict a change
of trend; 3) prediction similarity; and 4) forecast efficiency. Second,
we assess the performance of models that produce conditional fore-
casts, by generating in-sample projections for different scenarios
of exogenous and endogenous variables. Our main findings indicate
that time series models perform better for short time horizons, while
the dsge models are more efficient forecasting longer time horizons.
The remaining of this document is organized as follows. Section
2 presents a description of all unconditional and conditional mod-
els employed by Banco de Guatemala to generate inflation forecasts.
Section 3 describes the data and methodology employed for evalua-
tion purposes. Section 4 shows the results obtained. Finally, Section
5 concludes.

2. FORECASTING EVALUATION AT THE

CENTRAL BANK OF GUATEMALA

The prediction of the inflation rate is very important in the case of an

inflation targeting regime, because it allows the central bank to take

286 J. C. Castañeda et al.

the monetary policy actions to keep inflation on target and keep
the credibility of the regime. Therefore, Banco de Guatemala uses
an array of models to forecast the inflation rate. The main forecast
models are divided between those that produce unconditional fore-
casts and those producing conditional forecasts.

2.1 Unconditional-forecasts Models

In this section, we describe the main models used in this paper
to evaluate unconditional inflation forecasts. We start by explaining
the three main models used to explain the inflation rate. The first
one is the indicator variable (iv), which is the inflation forecast em-
ployed at Banco de Guatemala as the main short-term forecast in the
conduction of its monetary policy, and it is estimated by the De-
partment of Macroeconomic Analysis and Forecasts. The forecast
is based on a set of time series models plus the expert knowledge that
the economic analysts have about the inflation series. In particular,
they complement the inflation forecasts generated by the models
with considerations about trend, seasonality, and temporary shocks,
in addition to the overall domestic and foreign economic conditions.
The second one is the forecast combination through individual time-
varying efficient weights (efp). This model is based on assessing past
forecast performance efficiency at each of eight quarters ahead, ac-
cording to an algorithm called the efficient forecast path (efp), de-
scribed in Castillo y Ortiz (2017). The model is explained in detail
in Annex 3, which is delivered upon request. The third one is the av-
erage macroeconomic models (amm), used by the Economic Re-
search Department (die1). The die uses two macroeconomic models
to make forecasts: the semi-structural macroeconomic model 4.0.1
(mms) and the macroeconomic structural model (mme).
Furthermore, we evaluate inflation expectations with two mea-
sures available at Banco de Guatemala. Both are measured monthly.
The first one is from an Economic Expert Panel (eep). Banco de Gua-
temala surveys an independent panel of experts from the private sec-
tor every month on economics, finance, and business in Guatemala.
The objective of the survey is to assess their perception of the future
trend of inflation, economic activity, and public confidence in the
economy. The second one is from the die, which also carries out an
inflation expectations survey among its staff.

Evaluation of Inflation Forecasting Models in Guatemala 287

2.2. Conditional-forecasts Models
In this section, we evaluate the performance of three conditional
models to predict the inflation rate. The first model is the mms 4.0.1
which is a reduced form model, characterized by a difference-equa-
tions system, representing the transmission mechanisms of mone-
tary policy for quarterly data. The current version (mms 4.0.1) is part
of the set of non-micro funded general equilibrium macroeconomic
models used at Banco de Guatemala that have evolved from the first
version launched in 2006. It was built on the basis proposed by Berg,
et al. (2006a and 2006b), who provided a practical guide to non-micro
funded dsge models and their implementations for central banks.
In this regard, the mms 4.0.1 is a semi-structural model (non-micro
funded) for a small, open economy, where monetary authorities op-
erate policy within an inflation-targeting framework and implement
monetary policy through a Taylor-type rule. All variables in the model
are specified in annual growth rates. The mms 4.0.1 has 40 equations
(and 40 variables), of which 28 (70%) are endogenous and 12 (30%)
are exogenous variables. The model delivers forecasts for both core
inflation and headline inflation, and it is currently used for produc-
ing inflation and monetary policy interest rate forecasts that are in-
puts for Banco de Guatemala’s monetary policymaking process.
Those variables that display high volatility are transformed through
a moving sum (or average) scheme in order to reduce that volatility
and avoid possible outliers. At that respect, we get smoothed series.
The second model is a macroeconomic model of inflation fore-
cast for Guatemala (pigu). It is also a semi-structural macroeconom-
ic model, very similar to the mms 4.0.1. Variables in pigu are also
expressed as annual rates of change. There are three main differ-
ences between pigu and mms 4.0.1: the set of exogenous variables,
the exogenous variables’ volatility, and the type of inflation. First,
the set of exogenous variables: Even though some exogenous vari-
ables are common to both models, others are not. For example,
foreign inflation in mms 4.0.1 is the us core-pce inflation, while
in pigu is us Headline cpi inflation. Second, the exogenous vari-
ables’ volatility: many mms 4.0.1’s exogenous variables are smoothed
(four-quarter averages), while pigu uses quarterly variables. Final-
ly, the type of inflation: mms 4.0.1 forecasts both core and headline
inflation, while pigu forecasts headline inflation only. The model

288 J. C. Castañeda et al.

is currently available to all the central bank’s staff, through a cus-
tom-made interface.
The third model is the macroeconomic structural model (mme),
which is a medium scale dsge model, built within the new-Keynes-
ian framework. It features a financial accelerator à la Bernanke,
Gertler and Gilchrist (1999) and other frictions relevant for emerg-
ing or developing economies, such as deviations from the law of one
price and the uip. It is a model of heterogeneous agents; households
supply labor services to entrepreneurs. They consume domestic
and foreign goods, constitute deposits in domestic currency, take
foreign debt and collect remittances from abroad. Firms, operating
in a perfectly competitive market, assemble differentiated varieties
to produce the home (or domestic) homogeneous final good. There
are other firms producing the intermediate good, operating in a mo-
nopolistic competitive market; they buy a homogeneous wholesale
good from entrepreneurs to differentiate it and produce a particu-
lar variety. When these firms decide to change their prices, they face
adjustment costs, à la Rotenberg (1982), introducing nominal price
rigidities into the model. Entrepreneurs use three inputs to produce
the wholesale good: capital, labor, and imported raw materials. They
buy capital from capital producing firms using their own wealth
and loans granted by banks since they are not able to self-finance
their entire capital purchases. The financial sector is comprised
of private banks divided into two activities: narrow banks that carry
out passive operations gathering deposits from households and retail
banks using those deposits to grant loans to entrepreneurs. There
is also a central bank setting the short-term interest rate–the policy
rate–according to a Taylor-type rule and a central government car-
rying out unproductive spending.

3. DATA AND FORECAST EVALUATION

METHODOLOGY

In this section, we describe the data and explain the methodology

chosen in order to examine the forecasting accuracy of both the un-
conditional and conditional models. In the case of the forecast evalu-
ation of unconditional models, the statistical tests are not included
in this paper; however, they can deliver upon request (see Annex 3).

Evaluation of Inflation Forecasting Models in Guatemala 289

3.1 Data
First, we begin describing the dataset used for the unconditional
models. First, we use quarterly data to evaluate the forecasting ac-
curacy of the unconditional models. Each quarter, the iv and the
amm model forecast inflation for the next eight quarters, starting
at 2011Q1 and finishing at 2017Q2. The efp model starts forecast-
ing inflation every quarter for the next eight quarters only from
2014Q2 to 2017Q2. Then, we classify the forecasts of each quan-
titative model into different time-horizons (one, two, three, four,
and eight quarter) to evaluate the forecasting performance of each
time horizon, in order to find which model is best to forecast the in-
flation patterns in every one of them. The evaluation sample is rath-
er short, especially in the case of the efp’s forecasts, for which there
are only 13 quarters. Also, we evaluate how well the quantitative
models predict the inflation rate in December the current and the
next year. Second, we use the monthly data on inflation expectations
from both an economic experts’ panel (eep) and the die to exam-
ine the accuracy of the inflation expectations in prediction the in-
flation of December over a one and two-year horizon. The sample
of forecasting errors is from 2015M07 to 2017M06 in the case of the
one-year horizon and from 2016M07 to 2017M06 in the case of two-
year horizon predictions.
Second, we describe the data used in the case of the conditional mod-
els. For each of the three evaluated models, we generate quarterly
headline inflation forecasts with a sample from 2011Q1 to 2017Q2. 2
In addition, we consider five forecasting horizons: One quarter,
two quarters, four quarters, six quarters, and eight quarters.

3.2. Forecast Evaluation Methodology

First, we explain the methodology to evaluate the forecasting accura-
cy of the unconditional models. We evaluate the key properties of the
forecasting errors; i.e., we perform precision, accuracy, directional

2
A first evaluation was conducted considering a wider sample (2006Q1-
2017Q2), but results from this exercise were not as expected, in par-
ticular for headline inflation forecasts. This could be due to some
periods of high volatility in headline inflation. For example, inflation
went from 14.16% in the third quarter of 2008 towards a negative value
(–0.73%) one year later (in August 2009). Therefore, in order to get
robust results, we began our evaluation from 2011Q1.

290 J. C. Castañeda et al.

change, and efficiency tests to evaluate which model is best to predict
the future path of inflation. We start examining the residuals distri-
bution of the forecast, checking for normality and skewness. Then,
we compare the root mean square error (rmse) values to find which
model predicts the inflation rate best. After that, we use the Diebold-
Mariano (dm) test to examine if the difference between the mse of the
two competing models is statistically significant at least at the 10%
level. Also, we use the Giacomini-Rossi fluctuation (gr) test to ex-
amine the forecasting accuracy between the two competing models
over forecasting horizons with rolling windows of four. With this
test, we examine if the forecasts of one model are better than an-
other in every rolling window or if there is a change (fluctuation)
in the accuracy. In addition, we use the Pesaran-Timmerman (pt)
test to determine if the forecasts of the models can correctly predict
the directional change of inflation. Finally, we test the efficiency
of the forecasts by calculating the weak and strong efficiency tests.
Second, we explain the methodology to evaluate the performance
of the conditional models to predict the inflation rate. The quality
of any variable’s conditional forecasts depends on two elements:
The performance of the forecasting model (as such) and the qual-
ity of the forecasting model’s inputs on which the forecasts are con-
ditioned (e.g., the quality of the exogenous variables’ forecasts).
We evaluate the forecasting model’s performance by generating
in-sample forecasts in hindsight for different scenarios for the ex-
ogenous variables and for some endogenous variables as well. Some
of these scenarios involve historically observed values for the exog-
enous and some endogenous variables, to evaluate forecasts as if
we had the best possible forecast for these variables and thus, elimi-
nate one source of error. In the case of the semi-structural models
(mms and pigu), we plug, for each forecasted period, the historically
observed values of exogenous and some endogenous variables. In the
case of the structural model (mme), exogenous variables are repre-
sented by stochastic processes, typically of autoregressive nature.
Therefore, alternative scenarios are only conditioned by historically
observed values of two endogenous variables: inflation and output.
First, the mms 4.0.1 considers the scenarios: free, anchor 1, anchor
2, and anchor 3. In the free scenario, the exogenous variables’ fore-
casts are generated by the model’s laws of motion and all endogenous’
forecasts are generated by the model. In the anchor 1 scenario, the ex-
ogenous variables’ forecasts are generated by the corresponding

Evaluation of Inflation Forecasting Models in Guatemala 291

historically observed data, and some endogenous variables’ fore-
casts generated by the corresponding historically observed data:
monetary aggregates and economic output. The anchor 2 scenario
considers that the inflation forecast for the first quarter in the fore-
casting horizon is anchored by the corresponding historically ob-
served data, besides the characteristics of the anchor 1 scenario.
The last scenario (anchor 3) considers that the monetary policy in-
terest rate is anchored by the corresponding historically observed
data, as well as the characteristics of the anchor 2 scenario.
Second, pigu considers the scenarios: free, anchor 1, anchor 2,
and anchor 3. The free scenario contains the same characteristics
than in the case of the mms 4.0.1. In the anchor 1 scenario, the ex-
ogenous variables’ forecasts are generated by the corresponding
historically observed data, and all endogenous variables’ forecasts
are generated by the model. In the anchor 2 scenario, the exogenous
variables’ forecasts are generated by the corresponding historically
observed data, the inflation forecasts for the first two quarters in the
forecasting horizon are anchored by the corresponding historically
observed data, while all other endogenous forecasts are generated
by the model. In the anchor 3 scenario, the exogenous variables’ fore-
casts are generated by the corresponding historically observed data,
the inflation forecasts for the first two quarters in the forecasting ho-
rizon are anchored by the corresponding historically observed data,
and all other endogenous variables’ forecasts are anchored by the
corresponding historically observed data.
Third, the mme considers two scenarios: free and anchor 1. In the
free scenario, the exogenous variables forecasts are generated by the
model’s law of motion. In the anchor 1 scenario, the exogenous vari-
ables are generated by the model’s laws of motion; and the inflation
and output forecasts for the first quarter in the forecasting horizon
are anchored by the corresponding historically observed data. 3
For each model’s horizon-scenario combination, we compute
the mean error and the root mean squared error. The quantitative
results allow us to compare the models’ forecasting performances
(provided that they are fed with the best possible inputs; i.e., they

3
Anchored values of inflation are slightly different from the correspond-
ing observed values because the inflation series generated by the model
has a quarterly frequency; hence, its annualized inflation rate is the
sum of four quarterly values rather than a 12-month variation rate.

292 J. C. Castañeda et al.

are fed with historically observed data for the relevant variables)
and to assess the informative contribution of exogenous and endog-
enous variables for forecasting headline inflation.

4. RESULTS

In this section, we present the main results of the forecasting accu-

racy of both the unconditional and the conditional models. Most
of the tables and figures are presented in Annex 5, which do not ap-
pear in this paper. However, they are delivered upon request.

4.1. Unconditional Forecast Evaluation

We compare the forecasting performance to predict the inflation
patterns between the amm, the iv, and the efp model. Also, we evalu-
ate the forecasting performance of the inflation expectations gener-
ated by both the eep and the die. First, we compare the performance
of the forecasts of the models to predict inflation one, two, three,
four, and eight quarters ahead. Second, we analyze the accuracy
of the forecasts to predict the inflation rate in December in either
the current or the following year. The December inflation forecast is a
monetary policy indicator variable at Banco de Guatemala; hence,
its evaluation is very important.

4.1.1. Skewness and Normality

We start by evaluating the key properties of the forecasting error
distribution: normality and bias. To examine normality, we use
the jb test developed by Jarque and Bera (1980). The tables are in-
cluded in Annex 5, which is delivered upon request. First, we eval-
uate the properties of the forecasts through different forecasting
horizons. The forecast errors of the three models follow a normal
distribution according to the Jarque-Bera test, at the conventional
levels of significance. Also, the iv’s forecast shows a negative skew-
ness while the amm’s and efp’s forecasts show a positive skewness.
However, the skewness is low in all cases. Also, the forecast errors
of the inflation expectation predictions (both the eep and the die)
also follow a normal distribution. There is a positive bias in the in-
flation expectations predictions in the case of the die in both one-
and the two-year horizons.

Evaluation of Inflation Forecasting Models in Guatemala 293

 Second, we evaluate the properties of the forecasts in the case of De-
cember evaluation. The forecast errors of the three models also follow
a normal distribution in all forecasting horizons. In addition, there
is a positive bias in the efp’s forecast in the first three quarters while
there is no skewness in the remaining ones. iv’s and amm’s forecasts
both tend to have a negative bias.

4.1.2. rmse and mpe

We compute the rmse and mpe to determine which forecasting model
performs best, in the case of both the quantitative and the inflation
expectations. The tables are included in Annex 5, which is delivered
upon request. In the case of the quantitative models, the forecasts of the
iv model are better in the short run–one and two quarters–based on the
rmse. In the middle run, the forecasts of the amm model are more ac-
curate. However, in the long run–eight quarters–, the forecasts of the
efp model outperform the others. Also, we also analyze the inflation
expectations predictions. Based on the rmse, the eep’s inflation ex-
pectations are more accurate than those of the die’s in both the one-
and two-years horizons
Second, we proceed to analyze the forecasting accuracy of the quan-
titative models in their ability to predict the inflation rate in December
for the current and the following year, based on the rmse. We observe
that the forecasts of the amm model are better than the others in the
first five forecasting horizons, while the iv’s forecasts are best for the
last three horizons.

4.1.3. Diebold-Mariano Test

First, we use the dm test developed by Diebold and Mariano (1999)
to compare the predictive accuracy between two competing models,
of both the quantitative and the inflation expectations predictions.
The null hypothesis is that the two models have equal accuracy. The re-
sults of the dm test in the case of the quantitative models are presented
in Table 1 (the p-values of the test are shown in parenthesis). In Column
2, it is shown the test between the amm and the iv model. Only in the
case of four- and eight-quarter forward forecasting horizons, the dm-
statistic is negative and statistically significant at 5% level; therefore,
we reject the null hypothesis, and conclude that the forecasting accu-
racy of the amm model is best for both the intermediate and long time
horizons. Then, the dm-statist between the efp and the iv model is pre-
sented in Column 3. The statistic is positive and statistically significant

294 J. C. Castañeda et al.

at 5% level in all forecasting horizons, which means that all mse of the
iv model are lower than those of the efp model; therefore, the fore-
casting accuracy of the iv model is best to predict inflation.
After that, the dm-statistic between the efp and the amm mod-
el is presented in Column 4. The statistic is only positive and sta-
tistically significant for the one-, two-, and three-quarter forward
forecasting horizons. This means that for those horizons, the mse
of the amm model are lower than those of the efp model; therefore,
the amm’s forecasts are best to predict inflation in the short run.
Also, we evaluate the predictive performance of the inflation expec-
tations of both the eep and the die. The dm-statistic is only statisti-
cally significant for the two-year horizon with a sample of 12 months.
This means that the mse of the eep is lower than the mse of the die.
Thus, we conclude that the inflation expectation predictions of the
eep are more accurate than those of the die, only on this horizon.
Second, we compare the forecasting accuracy of the quantitative
models to predict the December inflation rate, from different hori-
zons. The results of the dm test are presented in Table 2. In Column
2, it is shown the test between the amm and the iv model. The dm-sta-
tistics are negative and statistically significant starting from three-
quarter forward forecasting horizon, so the mse of the amm model
are lower than those of the iv model. Therefore, the forecasts of the
amm model are best to predict inflation.

Table 1
DM TEST, QUANTITATIVE MODELS

Forecasting
horizons in dm statistic dm Statistic dm statistic
quarters (amm-iv) (efp-iv) (efp-amm)
1 1.44 (0.15) 1.71 (0.087) 1.65 (0.09)
2 1.30 (0.19) 1.97 (0.049) 2.03 (0.04)
3 0.21 (0.84) 1.79 (0.074) 1.70 (0.09)
4 –2.95 (0.00) 1.76 (0.079) 1.61 (0.11)
8 –3.35 (0.02) 2.91 (0.004) –0.87 (0.38)

Sources: author’s elaboration, central bank’s forecasts.

Evaluation of Inflation Forecasting Models in Guatemala 295

 Table 2
DM TEST, QUANTITATIVE MODELS

Forecasting dm statistic dm statistic dm statistic

horizons in quarters (amm-iv) (efp-iv) (efp-amm)
1 1.44 (0.15) 2.10 (0.036) 1.65 (0.10)
2 –0.95 (0.34) 2.70 (0.007) 2.55 (0.01)
3 –4.60 (0.00) 2.62 (0.009) 2.58 (0.00)
4 –2.33 (0.01) 4.75 (0.000) 7.32 (0.00)
5 –3.20 (0.00) 2.09 (0.036) 3.16 (0.00)
6 –2.93 (0.00) –5.61 (0.000) –22.50 (0.00)
7 –2.98 (0.00) –62.39 (0.000) 2.58 (0.01)
8 –1.95 (0.05) – –

Sources: author´s elaboration, central bank’s forecasts.

Then, the dm-statistic between the efp and the iv model is pre-
sented in Column 3. The statistic is statistically significant in all
forecasting horizons, which means that the mse s of the iv model
are lower than those of the efp model. Hence, we reject the null hy-
pothesis of equal accuracy. Also, the statistic is positive for the one-
to five-quarter horizons, which means that the mse s of the iv model
are lower than those of the efp. Hence, the iv model is more accurate
in its prediction of December inflation rate in the short and interme-
diate time horizons. On the other hand, the statistic is negative from
six to seven quarters ahead; therefore, the efp model is best in the
long run to predict the inflation rate. After that, the dm-statistic be-
tween the efp and the amm model is presented in Column 4. This
is statistically significant in all forecasting horizons, which means
that we reject the null hypothesis of equal accuracy. Also, in almost
all forecasting horizons, the mse s of the amm model are lower than
those of the efp model. Therefore, the amm model is best to predict
inflation rate in December.

296 J. C. Castañeda et al.

4.1.4. Pesaran-Timmerman Test
We use the pt test developed by Pesaran and Timmerman (1992)
to evaluate the directional forecasting of both the quantitative models
and the inflation expectations predictions. The critical values to reject
the null hypothesis of independence are ± 1.645 for 10% level of signif-
icance. First, we examine the directional forecasting accuracy in the
case of the iv model (see Annex 1, Table A.1.1). The Sn statistic is only
higher than its critical value in the case of one-, two- and three-quarter
horizons, so we can reject the null hypothesis of independence and con-
clude that the forecasts of the iv model can predict successfully the di-
rection of inflation in the short run. Now, we evaluate the directional
accuracy in the case of the amm model (see Column 3). We observe that
the Sn statistic is higher than its critical value only in the case of one-
and two-quarter horizons, so we can reject the null hypothesis of in-
dependence only for those two horizons and conclude that the model
can successfully predict the direction of the inflation in the short run.
We proceed to analyze the directional accuracy of the forecast in the
case of the efp mode (see Column 4). The Sn statistic is higher than
the critical value only in the case of one-quarter horizon; therefore,
we can only reject the null hypothesis of independence for this hori-
zon and conclude that the forecast of the efp model can predict suc-
cessfully the direction of the inflation in the case of that particular
horizon. Also, we analyze the directional forecasting accuracy of the
inflation expectations predictions of both the eep and the die. We re-
ject the null hypothesis of independence only in the case of the eep’s
forecasts in the case of a two-year horizon. Hence, we can conclude
that the panel can predict successfully the direction of inflation.
Second, we examine the directional forecasting accuracy of the
inflation rate for December (see Table Annex 2, A.2, which is deliv-
ered upon request) only for the case of the iv and amm models, since
we do not have enough data for the case of the efp model. We start with
the iv model (see the first column). We can reject the null hypothesis
of independence in the case of one-, three-, four-, five-, and six-quar-
ter horizons, so the model can predict successfully the directional
change of inflation in the short and middle run. Then, evaluate the per-
formance of the amm model (see the second column). We can reject
the null hypothesis of independence in the case of one-, two-, three-,
six-, seven-, and eight-quarter horizons, which implies that the model
can predict successfully the directional change of inflation in both
the short and the long run.

Evaluation of Inflation Forecasting Models in Guatemala 297

4.1.5. Giacomini-Rossi Fluctuation Test
We use the Giacomini and Rossi fluctuation test developed by Giaco-
mini and Rossi (2010) to examine the performance of two competing
models in the presence of possible instabilities. We use the iv model
as the benchmark model in the case of the quantitative model, and the
inflation expectations’ predictions of the eep in the case of expecta-
tions’ forecasts. The test is only used in some of the forecasting hori-
zons due to data availability. We set the rolling windows equal to four
quarters to make the forecasting analysis. Also, we use graphical anal-
ysis to examine the performance of the forecasts of the two compet-
ing models in the different rolling windows to see whether there is a
fluctuation in the forecasting accuracy. This is available in Annex 4,
which is delivered upon request.
First, we start with the forecasting accuracy evaluation of the quan-
titative models (see Annex 1, Table A1.3). We define the loss function
between the amm and the iv model in Equation 1. If the loss func-
tion turns out to be negative, we conclude that the forecasts of the
amm model are more accurate than those of the iv model. On the
other hand, if the loss function turns out to be positive, the forecasts
of the iv model are better at predicting inflation than those of the
amm model. We observe that we reject the null hypothesis of equal
forecasting accuracy over every forecasting horizon since the gr-sta-
tistic is higher than its critical value (see Table A1.3, Column 2). This
means that one model displays better predictive ability to forecast in-
flation in at least one period of time. Also, the graphical analysis re-
veals that the forecasts of the iv model are more accurate than those
of the amm one step ahead. However, it seems that the forecasts of the
amm model predict better the inflation patterns in four- and eight-
quarter horizons.

1 ( )
Lt θˆj −h ,R ,γˆ j −h ,R = MSE AMM ,t − MSE IV ,t

Then, we compare the forecasting accuracy between the efp and the
iv model with the use of the gr test (see Column 3). The loss func-
tion between the two models is defined by Equation 2. In this case,
the null hypothesis of equal accuracy is rejected in every forecasting
horizon since the gr-statistic is higher than the critical value. This
means that, at least in one period, one model generates more accurate
forecasts of inflation. The graphical analysis shows that the forecasts

298 J. C. Castañeda et al.

of the iv model are more accurate in almost all the evaluation sample
in each forecasting horizon. Therefore, the forecasts of the iv model
seem to be more accurate than the efp model in all forecasting hori-
zons (see Annex 5, delivered upon request).

2 ( )
Lt θˆj −h ,R ,γˆ j −h ,R = MSE EFP ,t − MSE IV ,t

Second, we use the gr test to examine the performance of the infla-

tion expectations predictions from the die and the eep. We consider
the eep data as a benchmark model. The loss function is set up in Equa-
tion 3. The graphical analysis shows that there is a fluctuation of the
forecasting accuracy of the inflation expectations between the two
models in the case of one-year horizon. However, the inflation expecta-
tions of the eee predict better the inflation patterns in the case of the
two-year horizon (see Annex 4, delivered upon request).

3 ( )
Lt θˆj −h ,R ,γˆ j −h ,R = MSE EEP ,t − MSE DIE ,t

4.1.6. Weak Efficiency Test

We examine the efficiency of the unconditional forecasts of both
the quantitative and the qualitative models with a variant of the weak
efficiency test developed by Mincer and Zarnowitz (1969). First, we start
with the quantitative models (see Annex 1, Table A1.4). From the sec-
ond column, we observe that amm’s forecasts satisfy the weak efficiency
hypothesis only in the case of one quarter ahead. From the third col-
umn, we analyze the weak efficiency of the iv forecasts (see the third
column) We observe that forecasts of the model satisfy the weak ef-
ficiency only in the case of one and two forecasting horizons. From
the fourth column, we evaluate the weak efficiency of the efp fore-
casts (see the fourth column). We observe that the forecasts of the
model satisfy the weak efficiency in almost all forecasting horizons
with the exception of four quarters ahead. In sum, the forecast of the
efp is more efficient than those of the other models based on the re-
sults of the weak efficiency test. Also, the forecast of the amm and
the iv are weakly efficient in the short run. In addition, the inflation

Evaluation of Inflation Forecasting Models in Guatemala 299

expectations predictions of both the eep and the die model do not
satisfy de weak efficiency test at 5% level in all forecasting horizons.
Second, we test for the weak efficiency only in the case of the
amm and the iv models, in the prediction of the inflation rate of De-
cember, because of data availability (see Annex 5, which is delivered
upon request). In the case of the amm’s forecasts, we cannot reject
the null hypothesis of weak efficiency only in the case of two and three
quarters ahead. Also, the forecasts of the iv model satisfy the weak ef-
ficiency tests in five out of eight forecasting horizons. In sum, the fore-
casts of the iv model are more efficient than those of the amm model
in evaluating the December predictability of inflation.

4.1.7. Strong Efficiency Test

We perform the strong efficiency test for the two econometric models:
iv and efp. The null hypothesis establishes that a new variable (which
is not included in the econometric models) does not explain the fore-
casting error. Therefore, the rejection of the null hypothesis means
that the errors are strongly efficient. Otherwise, if the null hypothesis
is not rejected, then the inclusion of a new variable can add informa-
tion to improve the forecasts. We consider five variables in logs of the
structural model of the Banco de Guatemala to make the test: con-
sumption, index of raw materials, investment, government spend-
ing, and credit.
First, we start with the iv model; the tests are shown in Annex 5, Ta-
ble A5.7, which is delivered upon request. In the second column, we list
the coefficient of consumption. We cannot reject the null hypothe-
sis at the 5% level of significance in the case of one and two quarters
ahead. Therefore, the forecasts are strongly efficient for those hori-
zons. However, for three to eight quarters ahead, consumption does
explain the forecasting error, which means that they are not strongly
efficient for these horizons. Similarly, in the third column, the null
hypothesis is not rejected at the 5% significance level. Therefore,
the forecasts are strongly efficient in those horizons. However, from
three to eight quarters ahead, the inclusion of the raw material index
can improve the forecasts, which mean that they are not strongly effi-
cient. Then, in the fourth column, we observe that the null hypothe-
sis is not rejected in one, two and three quarters ahead, which means
that the forecasts are strongly efficient in those horizons. However,
from four to eight quarters ahead, investment explains the forecast-
ing errors, therefore; the forecasts are not strongly efficient. After

300 J. C. Castañeda et al.

that, in the fifth column, we observe that the null hypothesis is not
rejected in all forecasting horizons, which means that the forecasts
are strongly efficient, and the inclusion of government spending
will not improve them. Finally, in the sixth column, we observe that
the forecasts are strongly efficient from one to three quarters ahead.
However, from four to eight quarters ahead, the inclusion of credit
can improve the forecasts, which implies that they are not strongly
efficient in those horizons.
We continue with the efp model; the tests are shown in Annex 5,
Table A5.8, which is delivered upon request. We observe that we re-
ject the null hypothesis for one-quarter predictions for the five vari-
ables, which means that the forecasts of the iv model are not strongly
efficient and the inclusion of the consumption, raw material index,
investment, government spending, and credit can improve the fore-
casts for this forecasting horizon. However, the forecasts are strongly
efficient in the case of the remaining forecasting horizons for the five
variables, because we cannot reject the null hypothesis.
Second, we perform the strong efficiency tests in the case of the
evaluation of December, only for the iv model due to data availability
(see Annex 5, Table A5.9). We observe that we cannot reject the null
hypothesis for all forecasting horizons in the case of the raw mate-
rial index, investment, government spending, and credit, at the
5% level of significance, which means that the forecast are strong-
ly efficient. However, in the case of consumption, we cannot reject
the null hypothesis in all forecasting horizons except for the three
quarters ahead, which means that the forecast is strongly efficient
for most horizons.

4.2 Conditional Forecast Evaluation

We make a headline inflation forecasting exercise in hindsight for the
three models. Also, we consider four scenarios for both the mms 4.01.1
and pigu and two scenarios for mme. The forecasting horizon begins
on 2011Q1. First, we show the inflation patterns and the forecasts
of each model (see Annex 5, Figures A5.1, A5.2, and A5.3, which
are delivered upon request). Second, we calculate the me and the
rmse (see Annex 1, Tables A1.6, A1.7 y A1.8).
In the case of the mms 4.0.1, the model generates core inflation
forecasts, and therefore headline inflation is constructed based
on those projections. This explains that, in the case of anchor 2 and
anchor 3, we have values different from zero in 1 and 2 quarters

Evaluation of Inflation Forecasting Models in Guatemala 301

ahead for the me and rmse (see Annex 1, Table A1.6). pigu model
minimizes the rmse in the fourth scenario (anchoring exogenous
variables, all other endogenous variables and two quarters of infla-
tion) for all forecasting horizons (see Annex 1, Table A1.7). In this
case, the model’s forecasts are negatively biased for all relevant ho-
rizons (the first two horizons are trivially unbiased since the histori-
cally observed inflation values are imposed as the model’s forecasts).
In order to compare the two models’ forecasting performances,
we pick the best scenario for each model. In particular, we compare
the mms 4.0.1’s performance in the third scenario with the pigu’s per-
formance in the fourth scenario. We focus on the last three forecast-
ing horizons since pigu’s rmse for the first two horizons is trivially
equal to zero. The results show that pigu’s rmse for the three relevant
horizons are less than the corresponding values for mms 4.0.1 and,
hence, pigu is preferred in this evaluation exercise, even though
its forecasts tend to underestimate inflation (i.e., its forecasts are neg-
atively biased). See Table 3.
For the mme, the me suggests that there is a positive inflation bias
(see Annex 1, Table A1.8). Results also suggest that forecasts gener-
ated by the model can benefit from anchoring inflation and output
one quarter ahead since doing so reduces the rmse (or its mean across
different forecasting horizons). This improvement will require that
better short-term projections (from outside the model) are available.

Table 3
COMPARISON OF THE BEST SCENARIOS BETWEEN MMS 4.0.1 AND PIGU

Forecasting mms 4.0.1, pigu, anchoring exogenous and endogenous

horizons in years anchor 2 variables, plus two periods of inflation
4 1.37 0.61
6 1.36 0.62
8 1.57 0.65

Source: author’s elaboration, central bank’s forecasts.

302 J. C. Castañeda et al.

 5. CONCLUSIONS

In this paper, we evaluated Banco de Guatemala’s most important

models used to forecast inflation. Forecast accuracy for uncondition-
al models (i.e., iv, amm, and forecast combinations of ols and time
series models) was evaluated for end of the year forecasts, and for
a two-year forecast horizon, using a variety of measurements and tests
(i.e., normality, rmse, dm, pt, gr, and weak and strong efficiency
tests). In the case of a conditional forecast, we evaluated the forecast-
ing accuracy of three models: mms 4.0.1, pigu, and mme.
We found empirical evidence supporting a higher degree of accu-
racy for time series models for the short forecast-horizons, and better
performance for models generating conditional-forecasts in lon-
ger forecast-horizons. The main purpose of this study was to assess
the accuracy and precision of the main inflation forecasts generated
at Banco de Guatemala. The next step is to take advantage of the ob-
tained results in order to improve the quality of the inflation fore-
casting models in use at the central bank. In particular, we should
continuously reevaluate model specifications, the quality of the
data sets, and the variable-transformation procedures. In addition,
we should perform a complete evaluation of the inflation forecasts
at least once a year, as some central banks already do.

ANNEX

Annex 1. Tables of the Unconditional Forecast Evaluation

Table A1.1
PT TEST, QUANTITATIVE MODELS

Forecasting
horizons in quarters Sn statistic (iv) Sn Statistic (amm) Sn statistic (efp)
1 4.28 3.98 2.41
2 3.77 2.93 1.62
3 2.57 1.54 0.73
4 0.00 0.88 –1.01
8 0.00 –1.49 –1.53
Sources: author’s elaboration, central bank’s forecasts.

Evaluation of Inflation Forecasting Models in Guatemala 303

 Table A1.2
PT TEST, QUANTITATIVE MODELS, DECEMBER EVALUATION

Forecasting horizons in quarters Sn statistic (iv) Sn statistic (amm)

1 1.67 1.67
2 1.02 1.67
3 –1.67 1.67
4 1.67 –1.46
5 –2.31 –1.33
6 –2.31 –2.31
7 – –2.31
8 – –2.31

Sources: author’s elaboration, central bank’s forecasts.

Table A1.3
gr TEST, QUANTITATIVE MODELS

Forecasting horizons in quarters gr statistic ( amm-iv ) gr statistic (efp-iv)

1 4.77 5.68
2 15.28 5.93
3 9.93 11.29
4 9.07 7.39
8 11.28 –

Sources: author’s elaboration, central bank’s forecasts.

Table A1.4
WEAK EFFICIENCY TEST, QUANTITATIVE MODELS

Forecasting horizons Weak efficiency Weak efficiency Weak efficiency

in quarters test (amm) test (iv) test (efp)
1 0.11 (0.89) 6.33 (0.24) 3.58 (0.08)
2 4.41 (0.02) 3.29 (0.12) 0.22 (0.89)
3 6.18 (0.00) 11.57 (0.01) 12.08 (0.97)
4 5.39 (0.01) 21.81 (0.00) –
8 104.62 (0.00) 62.16 (0.00) 0.20 (0.83)
Sources: author’s elaboration, central bank’s forecasts.

304 J. C. Castañeda et al.

 Table A1.5
WEAK EFFICIENCY TEST, QUANTITATIVE
MODELS, DECEMBER EVALUATION

Weak efficiency test Weak efficiency test

Forecasting horizons in quarters (amm) (iv)
1 83.48 (0.00) 1.17E+12 (0.00)

2 1.36 (0.35) 1.5268 (0.32)

3 1.45 (0.34) 1.2242 (0.38)

4 8.87 (0.034) 9.5156 (0.03)

5 1.71E+11 (0.00) 14.1267 (0.03)

6 1.85E+11 (0.00) 1.6197 (0.33)

7 2.03E+10 (0.00) 0.9950 (0.47)

8 1.66E+10 (0.00) 1.8451 (0.30)

Sources: author’s elaboration, central bank’s forecasts.

Tables of the Conditional Forecast Evaluation

Table A1.6
ME AND RMSE, MMS 4.01, 2011Q1-2017Q2

Free model Anchor 1 Anchor 2 Anchor 3

Forecasting
horizons in
quarters me rmse me rmse me rmse me rmse

1 –0.11 0.73 –0.03 0.71 0.01 0.33 0.01 0.33

2 –0.13 1.21 –0.02 1.27 0.01 0.87 0.01 0.87

4 0.22 1.43 0.29 1.58 0.29 1.37 0.29 1.37

6 0.55 1.47 0.54 1.4 0.52 1.36 0.52 1.36

8 0.27 1.72 0.55 1.63 0.54 1.57 0.54 1.57

Mean 0.16 1.31 0.26 1.32 0.27 1.1 0.27 1.1

Sources: author’s elaboration, central bank’s forecasts.

Evaluation of Inflation Forecasting Models in Guatemala 305

 Table A1.7
ME AND RMSE, PIGU, 2011Q1-2017Q2

Anchoring
Anchoring exogenous and
exogenous endogenous
Anchoring variables and variables, plus
Forecasting exogenous two periods of two periods of
horizons in Free model variables inflation inflation
quarters me rmse me rmse me rmse me rmse

1 –0.22 0.83 –0.25 0.72 0 0 0 0

2 –0.3 1.26 –0.38 0.9 0 0 0 0

4 0 1.44 –0.47 0.88 –0.39 0.82 –0.27 0.61

6 0.34 1.11 –0.58 1.12 –0.56 1.13 –0.32 0.62

8 0.41 0.89 –0.79 1.29 –0.79 1.29 –0.38 0.65

Mean 0.05 1.11 –0.49 0.98 –0.35 0.65 –0.19 0.38

Sources: author’s elaboration, central bank’s forecasts.

Table A1.8
ME AND RMSE, MME, 2011Q1-2017Q2

Forecasting Free model Anchor 1

horizons in
quarters me rmse me rmse

1 0.3 0.62 –0.09 0.1

2 0.89 1.28 0.36 0.61

4 2.37 2.72 1.81 2.09

6 2.82 2.98 2.87 3.04

8 2.82 2.93 2.86 2.96

Mean 1.84 2.11 1.56 1.76

Sources: author’s elaboration, central bank’s forecasts.

306 J. C. Castañeda et al.

References
Berg, Andrew, Philippe D. Karam, and Douglas Laxton (2006a),
A Practical Model-based Approach to Monetary Policy Analysis-
Overview.
Berg, Andrew, Philippe D. Karam, and Douglas Laxton (2006b),
Practical Model-based Monetary Policy Analysis: A How-to Guide,
No. 6-81, International Monetary Fund.
Bernanke, Ben S., Mark Gertler, and Simon Gilchrist (1999),
“The Financial Accelerator in a Quantitative Business Cycle
Framework,” Handbook of Macroeconomics 1, pp. 1341-1393.
Castillo-Maldonado, Carlos, and Edson R. Ortiz-Cardona (2018),
“Evaluación de combinaciones de inflación en Nicaragua
(nica): Un método eficiente para combinar pronósticos,”
Revista de Economía y Finanzas, Vol. 5, Banco Central de Ni-
caragua, octubre, pp. 1-34.
Diebold, F. X., and R. S. Mariano (1995), “Comparing Predictive
Accuracy,” Journal of Business and Economic Statistics, Vol. 13,
No. 3, July, pp. 253-263.
Giacomini, Raffaella, and Barbara Rossi (2010), “Forecast Com-
parisons in Unstable Environments,” Journal of Applied Econo-
metrics, Vol. 25, No. 4, pp. 595-620.
Jarque, Carlos M., and Anil K. Bera (1980), “Efficient Tests for
Normality, Homoscedasticity and Serial Independence of
Regression Residuals,” Economics Letters, Vol. 6, No. 3, pp.
255-259.
Mincer, Jacob A., and Victor Zarnowitz (1969), “The Evaluation of
Economic Forecasts,” in Economic Forecasts and Expectations:
Analysis of Forecasting Behavior and Performance, nber, pp. 3-46.
Pesaran, M. Hashem, and Allan Timmermann (1992), “A Simple
Nonparametric Test of Predictive Performance,” Journal of
Business & Economic Statistics, Vol. 10, No. 4, pp. 461-465.
Rotemberg, Julio J. (1982), “Sticky Prices in the United States,”
Journal of Political Economy, Vol. 90, No. 6, pp. 1187-1211.

Evaluation of Inflation Forecasting Models in Guatemala 307

Forecasting Inflation Expectations from
the cesifo World Economic
Survey: An Empirical Application
in Inflation Targeting Countries
Héctor M. Zárate Solano
Daniel R. Zapata Sanabria

Abstract
This paper has two purposes. First, it evaluates the responses to the ques-
tions on inflation expectations in the World Economic Survey (wes) for six-
teen inflation targeting countries. Second, it compares inflation expectation
forecasts across countries by using a two-step approach that selects the most
accurate linear or non-linear forecasting method for each country. Then,
Self-Organizing Maps are used to cluster inflation expectations, setting as a
benchmark June 2014, when there was a sharp decline in oil prices. Analyz-
ing inflation expectations in the context of this price change makes it pos-
sible to distinguish between countries that anticipated the oil shock smoothly
and those that had to adjust their expectations significantly. The main find-
ings from the wes in-sample comparison suggest that expert forecasts of in-
flation expectations are systematically distorted in 83 percent of the countries
in the sample. On the other hand, the out of sample forecast analysis indicates
that Non-linear Artificial Neural Networks combined with Bayesian regular-
ization outperform arima linear models for longer forecasting horizons.
This holds true for countries with both soft and brisk changes of expectations.
However, when forecasting one step ahead, the performance between the two
methods is similar.

Héctor M. Zárate-Solano <[email protected]>, Senior Econometrician, Banco

de la República Colombia (Central Bank of Colombia) and Daniel R. Zapata-Sanabria
<[email protected]>, Research Assistant, Banco de la República Colom-
bia (Central Bank of Colombia). The authors would like to thank Johanna Garnitz
at the ifo Institute for Economic Research for providing the data used in this paper.
Also, we have greatly benefited from discussions with professor Marcellino Massimil-
liano. The opinions expressed in this publication are those of the authors and do
not reflect the views of cemla, the fsd group, the Inter-American Development Bank
or the Banco de la República or its board of Directors.

309
 jel classifications: C02, C222, C45, C63, E27.
Keywords: Inflation expectations, Machine learning, Self-organizing
maps, Nonlinear auto-regressive neural network, Expectation surveys, Time
series models.

1. INTRODUCTION

C
ross-country data from economic expectations surveys have
recently highlighted the importance of analyzing and forecast-
ing public expectations to gain insight into crucial empirical is-
sues in macroeconomics. Expectations can influence the future path
of real economic variables and help guide policy decision-makers,
and inflation expectations are particularly important for countries
that utilize inflation targeting as their primary monetary policy
framework. The usefulness of inflation expectations is manifested
in various realms of economic analysis. They are critical for i) testing
theories of informational inflation rigidity (Coibion et al., 2012); ii)
estimating key structural parameters, such as the intertemporal
substitution elasticity (Crump et al., 2015); iii) testing public un-
derstanding of monetary policy, such as the Taylor rule (Carvalho
and Nechio, 2014); and iv) assessing how well inflation expectations
may be anchored among economic agents, which is key in assessing
the effectiveness of central bank communication. Lastly, New Keynes-
ian macroeconomic models have successfully used inflation expecta-
tions to predict real inflation (Henzel and Wollmershäuserab, 2008).
Expectation surveys have featured a wide range of respondents,
including economic experts, central bankers, financial agents, con-
sumers, and firms. Those surveyed often have to make important
decisions that take into account inflation and survey data, and their
responses provide information on the effectiveness of economic poli-
cies and institutional confidence. The World Economic Survey (wes)
collects data on inflation expectations across countries and surveys
more than 1,000 economic experts in approximately 120 countries.
The respondents evaluate present economic conditions and predict
the economic outlook of the country in which they reside, giving
special attention to price trends in their answers to both qualitative
and quantitative questions.
Thus, we must assess the suitability of wes data surveys and select
the appropriate methods to accurately forecast inflation expectations.

310 H. M. Zárate-Solano, D. R. Zapata-Sanabria

In regard to suitability, we can use simple exploratory data analysis
based on time plots and correlations, and we can calculate the in-
sample forecast errors within a sample of 16 inflation-targeting coun-
tries. To find the appropriate forecasting method, we use a two-step
approach centered on both clustering and forecasting techniques.
Specifically, we analyze the June 2014 oil price shock and its effect
on inflation expectations and other macroeconomic indicators.
We consider this oil shock relevant because the decline in oil prices
was significantly larger than in any previous episode during the past
30 years. The decline weakened fiscal policy and reduced the eco-
nomic activity of oil exporters, but for oil importers, inflationary
and fiscal pressures were alleviated. The oil price shock is also sig-
nificant because it affected growth and inflation through two chan-
nels: input costs and real income shifts. Changes through either
of these channels then led to changes in inflation expectations. Thus,
we evaluate different forecasting methods in the period after the oil
shock from Q3 2014 to Q2 2016. To obtain optimal forecasts, a com-
bination of clustering and forecasting analysis can be used. Data
visualization techniques are useful for discovering important char-
acteristics and potential clusters of economic agents. In addition,
we use machine learning and statistical methodologies to improve
inflation expectation forecasts based on qualitative and quantita-
tive questions from the wes.
This paper examines the data on inflation expectations from
the wes for 16 inflation- targeting countries. Then, by making use of
Self-Organizing Maps (som) we cluster agents’ expectations for these
countries to classify them either as “soft” or “brisk” based on the speed
of their expectations change after the oil shock of 2014 (Claveria,
Monte and Torra, 2016). After that, we combine the som representa-
tions with different forecasting methods to select models for infla-
tion expectation forecasting. The arima model reflects the linear
class of models and the Non-linear Auto-regressive Neural network
(nar-nn) reflects the non-linear class of models.
Our main findings are the following. First, we present evidence
of heterogeneity in the correlation patterns between inflation expec-
tations and observed inflation. There are increasing, descending,
and inverted U-shaped correlations over time. Regarding frequen-
cy domain analysis, the highest coherence values were often found
in periods of higher frequencies in most countries, implying that
there is a strong relationship between cycles of short periods.

311
 According to the wes forecast error analysis, we observe that even
though the forecasts meet at least the minimum standard when com-
pared to a random walk, economic experts have made systematic
errors in their predictions. That is, inflation was under- predicted
while increasing and over-predicted while declining in most of the
countries. Moreover, the mean squared error decomposition illus-
trated that there were systematic distortions in the inflation forecasts
in around 83 percent of the countries. The evidence suggests that al-
though the accuracy of the forecasts increases as the forecasting hori-
zon decreases, this relationship is not monotonic. This finding does
not support the hypothesis that forecasts have improved over time,
which may signal that there is a non-linear data- generating process.
Second, turning to a much more complex analysis, the som rep-
resentation allows us to cluster countries based on the evolution
of inflation expectations before the oil price shock. It is important
to note that the low inflation expectations cluster is relatively small
compared to the high and neutral clusters for inflation-targeting
countries. We find that in the one step- forward forecasts, the neural
network only slightly improves on forecasts of the arima , but that
it outperforms the arima model in the two step-forward forecasts
for Canada, Colombia, Chile, Poland, Hungary, and Sweden. There-
fore, using a non-linear neural network along with Bayesian regular-
ization leads to an improvement in expectations forecasts.
This paper contains five sections apart from this introduction
and proceeds as follows. In Section 2 we describe the wes data
and evaluate the responses to both qualitative and quantitative
inflation questions. In Section 3, we provide the methodologies
for clustering and forecasting, emphasizing the merits of the artifi-
cial neural network approach. In Section 4, we summarize the main
results, including the cluster analysis and forecasting accuracy. Fi-
nally, in Section 5 we present our conclusions and propose future
lines of research.

2. WORLD ECONOMIC SURVEY DATA AND THEIR

SUITABILITY FOR FORECASTING INFLATION

Surveying economic experts across different countries, the cesifo

World Economic Survey (wes) carried out by the ifo Institute for Eco-
nomic Research collects data on how experts view their country’s

312 H. M. Zárate-Solano, D. R. Zapata-Sanabria

economic outlook. In this paper, we use the term economic experts
to include representatives of multinational enterprises, banks, cham-
bers of commerce, academic institutions, and individual economists.
The questionnaire is distributed every quarter (January, April,
July, and October) with qualitative and quantitative questions relat-
ed to the general economic situation and expectations regarding
key macroeconomic indicators: economic growth, interest rates,
consumption, capital, exchange rates, and inflation, among oth-
ers.1 The questions on the expected inflation rate, which are the
main focus of this paper, reveal qualitative and quantitative infor-
mation on the economic experts of each country. Thus, the partici-
pants are asked to give their expectations of what the inflation rate
will be by the end of the next six months. They indicate “HIGHER”
for an expected rise in the inflation rate, “ABOUT THE SAME”
for no change in the inflation rate, and “LOWER” for an expected
fall in the expected inflation rate by the end of the next six months.
We transformed these responses into a cardinal time series of ex-
pected inflation by applying the following standard approach: where
the response is considered high, a numerical value of 9 is coded; where
the response is considered neutral, a value of 5 is coded; and where
the response is considered low, a value of 1 is recorded. Next, we cal-
culate the average rating for each question for each country. Tradi-
tionally, analysts have categorized these country ratings by terming
an average greater than 5 a positive zone and an average below 5 a
negative zone. The neutral zone depends simply on the analyst’s
subjective decision. One of the results of this paper is to establish
the limitations that come with this three-zone categorization and in-
stead, we let the data speak for itself.
In the quantitative question the experts of each country are asked
to predict the future inflation rate: “the rate of inflation on average
this year will be: % p.a.” We analyze the responses to this question
through an in-sample statistical analysis of forecasting error. Further
information on the wes can be found in Stangl (2007a and 2007b).
We analyze expectations for 16 inflation-targeting countries from
Q3 1991 to Q2 2016. The countries included in our analysis are Bra-
zil, Canada, Switzerland, Chile, Colombia, Czech Republic, United
Kingdom, Hungary, Korea Republic, Mexico, Norway, Philippines,

1
A survey form of the World Economic Survey, the wes questionnaire,
is included in Appendix A, see Figure 14.

313
Poland, Sweden, Thailand, and South Africa. 2 The relationship be-
tween the indicator of wes inflation expectations and the observed
annual inflation rate is illustrated through a simple exploratory analy-
sis that uses time plots and correlation statistics.3 The observed infla-
tion rate and the corresponding inflation expectations are depicted
in Figure 1 for some selected countries. For each country, inflation
was measured by annual changes in the Consumer Price Index. Ac-
cording to Figure 1, wes expectations move in tandem with actual
inflation for most of the period under study except during idiosyn-
cratic and global shocks that affected specific national economies.4
Figure 2 displays the correlation coefficient over time and the
coherence as a function of the frequency between the wes inflation
expectations and real annual inflation. The plot of the correlation
coefficient shows the existence of different patterns of linear asso-
ciation. For example, while the correlation in Mexico has increased
over time, it has decreased in Canada. On the other hand, Colom-
bia has experienced an inverted u-shaped correlation pattern that
peaks in the middle of 2002. According to frequency domain analy-
sis, higher coherence was found in higher frequencies of the spectral
distribution in most of the countries, which suggests that the relation-
ship between inflation expectations and observed inflation is strong
predominantly during short cycles. It is important to note that Asian
countries have higher coherence in lower frequencies, which points
to a different trend between expectation and observed inflation. 5

2
Figure 11 in Appendix A contains the full-time series length.
3
To see the other countries’ inflation expectations, see Figure 15 in
the Appendix.
4
In addition, we include a summary of the data, their histograms and cor-
relations which are relevant to the som analysis: Figure 12 in the Ap-
pendix reveals the heterogeneity of the variables, and Figure 13 displays
the correlation between them. Table 9 in the Appendix shows a brief
summary of the wes expectations data.
5
To see the spectral decomposition of the other countries, see Figure
16 in the Appendix.

314 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 −2
−1
0
1
2
3
4
5

0
2
4
6
8
10
12
Mar 2000 Dec 1991
Jun 1992
Sep 2000 Dec 1992
Mar 2001 Jun 1993
Sep 2001 Dec 1993
Jun 1994
Mar 2002 Dec 1994
Sep 2002 Jun 1995
Dec 1995
Mar 2003 Jun 1996
Sep 2003 Dec 1996
Jun 1997
Mar 2004 Dec 1997
Sep 2004 Jun 1998
Dec 1998
Mar 2005 Jun 1999
Sep 2005 Dec 1999
Jun 2000

Annual Inflation
Mar 2006 Dec 2000
Sep 2006 Jun 2001
Dec 2001
Mar 2007 Jun 2002
Sep 2007 Dec 2002
Jun 2003
Mar 2008 Dec 2003
Sep 2008 Jun 2004
Dec 2004
Mar 2009 Jun 2005
Figure 1

() 

Sep 2009 Dec 2005

() 
Jun 2006
Mar 2010 Dec 2006
Sep 2010 Jun 2007
Dec 2007
Mar 2011 Jun 2008
Sep 2011 Dec 2008
Jun 2009
Mar 2012 Dec 2009
Sep 2012 Jun 2010
Dec 2010

Source:  survey and  statistics. Some selected countries.

Mar 2013 Jun 2011
Sep 2013 Dec 2011
WITH THE OBSERVED ANNUAL INFLATION

Jun 2012
Mar 2014 Dec 2012
Jun 2013
COMPARISON OF THE INFLATION EXPECTATIONS

Sep 2014

Inflation rate next 6 months

Dec 2013
Mar 2015 Jun 2014
Sep 2015 Dec 2014
Jun 2015
Mar 2016 Dec 2015
Sep 2016 Jun 2016
0
1
2
3
4
5
6
7
8
9

0
1
2
3
4
5
6
7
8
9

315
 −1
0
1
2
3
4
5
6
4
8
−1
0
1
2
3
4
5
6
Dec 1991 Dec 1991
Jun 1992 Jun 1992
Dec 1992 Dec 1992
Jun 1993 Jun 1993
Dec 1993 Dec 1993
Jun 1994 Jun 1994
Dec 1994 Dec 1994
Jun 1995 Jun 1995
Dec 1995 Dec 1995
Jun 1996 Jun 1996
Dec 1996 Dec 1996
Jun 1997 Jun 1997
Dec 1997 Dec 1997
Jun 1998 Jun 1998
Dec 1998 Dec 1998
Jun 1999 Jun 1999
Dec 1999 Dec 1999
Jun 2000 Jun 2000

Annual Inflation
Dec 2000 Dec 2000
Jun 2001 Jun 2001
Dec 2001 Dec 2001
Jun 2002 Jun 2002
Dec 2002 Dec 2002
Jun 2003 Jun 2003
Dec 2003 Dec 2003
Jun 2004 Jun 2004
Dec 2004 Dec 2004
Jun 2005 Jun 2005
() 

Dec 2005 Dec 2005

Jun 2006 Jun 2006
Figure 1 (cont.)

Dec 2006 Dec 2006

Jun 2007 Jun 2007

()  

316 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Dec 2007 Dec 2007
Jun 2008 Jun 2008
Dec 2008 Dec 2008
Jun 2009 Jun 2009
Dec 2009 Dec 2009
Jun 2010 Jun 2010
Dec 2010 Dec 2010

Source:  survey and  statistics. Some selected countries.

Jun 2011 Jun 2011
Dec 2011 Dec 2011
WITH THE OBSERVED ANNUAL INFLATION

Jun 2012 Jun 2012

Dec 2012 Dec 2012
Jun 2013 Jun 2013
COMPARISON OF THE INFLATION EXPECTATIONS

Inflation rate next 6 months

Dec 2013 Dec 2013
Jun 2014 Jun 2014
Dec 2014 Dec 2014
Jun 2015 Jun 2015
Dec 2015 Dec 2015
Jun 2016 Jun 2016

0
1
2
3
4
5
6
7
8
9
0
1
2
3
4
5
6
7
8
9
 Figure 2
CORRELATION AND COHERENCE COEFFICIENTS OF QUALITATIVE WES
INFLATION EXPECTATION WITH OBSERVED ANNUAL INFLATION

() :    () : 

1.0
0.5
Correlation coefficient

0.8
0.4
0.6

Coherence
0.3
0.4
0.2 0.2

0.1 0.0
1990 1995 2000 2005 2010 2015 2020 0.00 0.62 1.24 1.86 2.48 3.10
Date Frequency from 0 to 

() :    () : 

 
0.4
0.8
Correlation coefficient

0.2
0.6
Coherence

0.0
0.4
−0.2
0.2
−0.4
0.0
1990 1995 2000 2005 2010 2015 2020 0.00 0.62 1.24 1.86 2.48 3.10
Date Frequency from 0 to 

() :    () : 

0.4 1.0
Correlation coefficient

0.8
0.2
0.6
Coherence

0.0 0.4

0.2
−0.2
0.0
1990 1995 2000 2005 2010 2015 2020 0.00 0.62 1.24 1.86 2.48 3.10
Date Frequency from 0 to 

Source:  survey and  statistics and  data.

317
 Figure 2 (cont.)
CORRELATION AND COHERENCE COEFFICIENTS OF QUALITATIVE WES
INFLATION EXPECTATION WITH OBSERVED ANNUAL INFLATION

()  : ()  :

   

0.2 0.8
Correlation coefficient

0.0 0.6

Coherence
−0.2 0.4
−0.4
0.2
−0.6
0.0
1990 1995 2000 2005 2010 2015 2020 0.00 0.62 1.24 1.86 2.48 3.10
Date Frequency from 0 to 

Source:  survey and  statistics and  data.

2.1 Quantitative Forecasting Inflation Expectations

In this section, we perform an in-sample forecasting analysis based
on the forecasting error. We compute the forecasting error as the
difference between annual average inflation based on the CPI and
the corresponding quantitative wes inf lation assessment from
the survey question “the rate of inflation on average this year will
be: % p.a.”. We follow previous work by Fildes and Stekler (2002)
and Hammella and Haupt (2007) to quantify and examine the ac-
curacy of wes forecasts at different horizons. It is important to note
that the experts receive more information from quarter to quarter
during the year as data on the observed inflation rate is released.

2.1.1 Statistical Analysis of the Forecasting Error

The forecasting error is calculated in the following way:

1 e ( L,Q (h ),t ) = p ( L,t ) − q ( L,Q (h ),t )

318 H. M. Zárate-Solano, D. R. Zapata-Sanabria

where L = countries, h = I, II, III, IV, and t = 1991,. . . , 2016. First,
we compute some standard error statistics for each quarter includ-
ing the rmsfe (root mean squared forecast error), mae (mean ab-
solute error), and Theil U-statistic. See Hamella and Haupt (2007).6
Second, we used the additive mean squared error decomposition
proposed by Theil in 1966 (see Theil et al., 1975) to obtain insight
into the structure of the forecast error. The decomposition is meant
to illustrate how the error changes conditional on the different
forecasting horizons through three components: the bias share Vh ,
the spread share Sh , and the covariance share K h . The Vh  bias com-
ponent measures systematic distortions in the forecast, where bias
should decrease through forecast horizons only if the expectations
are anchored. Sh measures the dispersion between observed infla-
tion and the wes forecast. Finally, K h assesses the linear association
between average inflation and the wes forecast; if the correlation
is perfect then K =0. Notice that the components should sum up to one.

2.1.2 Quantitative Inflation Expectation Results

Tables 1 and 2 summarize the rmsfe and its decomposition for the
sample of countries at different time horizons. The results illustrate
that the rmsfe decreases throughout the year for countries such
as Switzerland, Colombia, Korea, and Norway. Nevertheless, there
are some countries which exhibit a different pattern in which the last
forecast is more uncertain. The countries in this group include Bra-
zil, Canada, Chile, Czech Republic, and United Kingdom. The het-
erogeneity among rmsfe values across countries can be explained
by the fact that the rmsfe relies on the restricted assumption that
survey forecasters have a symmetric loss function. The rmsfe also
depends on the unit of measurement and the inflation rate in each
country. These diagnoses remain by observing the mae and U- sta-
tistics. Figure 3 compares the respective observed annual inflation
(bar line) and the wes expectation for each quarter for some select-
ed countries.7,8

6
The respective statistics equations are presented in Appendix A.3,
and mae and U-statistic results are in Tables 10 and 11, respectively.
See Appendix.
7
To see the other countries quantitative inflation expectation, see Figure
17 in Appendix A.3.
8
The quarter-specific forecasting error by country is plotted in Figure
18, Appendix A.3.

319
 The evidence for Colombia suggests that actual annual inflation
was overestimated during the period from 2000 to 2003, and from
2003 to 2007 the expectations were close to the observed inflation
rate. The 2008 financial crises led expectations to undershoot ob-
served inflation for a short period of time, but soon after, expecta-
tions began to overshoot observed inflation until 2014. Eventually,
the 2014 oil shock induced a period of undershooting. There are dif-
ferent patterns across the countries. For example, in Mexico expec-
tations were close to actual inflation until the oil shock, but after
the shock, they overestimated observed inflation rates. In Tables
3 and 4 we count the number of years in which inflation was over-
estimated and underestimated respectively by respondents, to the
quarterly wes survey. For instance, the results indicate that annual
inflation in Colombia was overestimated, on average, in 14 of 25 years
and for Mexico in 17 of 26 years. There is evidence that systematic
overestimation was greater than underestimation. The exception
occurs in the case of Brazil in which, on average, in 15 of 26 years in-
flation was underestimated by economic experts.
Finally, a cross-country comparison using the U-statistic confirms
that the wes -forecasts in every country at least meet the minimum
standard when compared with the random walk alternative.

320 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Table 1
ROOT MEAN SQUARED FORECAST ERRORS OF WES SURVEY
QUANTITATIVE INFLATION QUESTION Q1 1991 TO Q3 2016

4-step forecast 3-step forecast 2-step forecast 1-step forecast

Countries (QI) (QII) (QIII) (QIV)

Brazil 182.71 321.48 354.44 431.01

Canada 0.70 0.57 0.42 0.58

Switzerland 0.75 0.50 0.41 0.38

Chile 1.23 1.46 1.36 1.66

Colombia 1.80 1.67 1.43 1.00

Czech 4.97 4.81 6.87 3.08

Republic

United 0.89 0.88 0.90 0.99

Kingdom

Korea 1.61 1.41 1.16 1.09

Mexico 3.37 2.03 4.48 3.62

Norway 0.78 0.65 0.52 0.39

Hungary 2.12 1.32 1.12 1.54

Philippines 2.29 1.77 1.29 1.22

Poland 5.48 2.07 10.48 11.47

Sweden 1.05 0.80 0.99 1.19

Thailand 2.05 1.56 1.51 1.04

South Africa 1.77 1.57 1.49 1.27

321
 Table 2
THEIL ERROR DECOMPOSITION OF THE WES
FORECAST ERRORS Q1 1991 TO Q2 2016

4-step 3-step 2-step 1-step

Error forecast forecast forecast forecast
Countries decomposition (QI) (QII) (QIII) (QIV)
V 0.13 0.06 0.07 0.01

Brazil S 0.84 0.81 0.53 0.10

K 0.06 0.14 0.45 0.92

V 0.16 0.20 0.31 0.16

Canada S 0.05 0.14 0.26 0.26

K 0.83 0.70 0.46 0.61

V 0.22 0.32 0.30 0.19

Switzerland S 0.22 0.28 0.37 0.54

K 0.60 0.55 0.35 0.31

V 0.00 0.02 0.05 0.02

Chile S 0.02 0.20 0.74 0.75

K 1.02 0.84 0.25 0.27

V 0.003 0.06 0.04 0.01

Colombia S 0.08 0.02 0.05 0.33

K 0.96 0.99 0.95 0.71

V 0.10 0.14 0.06 0.02

Czech R. S 0.17 0.21 0.33 0.002

K 0.77 0.77 0.65 1.02

V 0.23 0.26 0.18 0.14

United K. S 0.16 0.28 0.43 0.30

K 0.64 0.56 0.43 0.60

V 0.37 0.44 0.52 0.39

Korea S 0.03 0.002 0.0003 0.02

K 0.62 0.45 0.50 0.62

322 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 4-step 3-step 2-step 1-step
Error forecast forecast forecast forecast
Countries decomposition (QI) (QII) (QIII) (QIV)

V 0.03 0.13 0.01 0.01

Mexico S 0.43 0.002 0.11 0.03

K 0.57 1.04 0.92 1.01

V 0.06 0.02 0.08 0.02

Norway S 0.19 0.14 0.13 0.18

K 0.79 0.79 0.83 0.84

V 0.04 0.15 0.01 0.01

Hungary S 0.00 0.08 0.11 0.27

K 1.00 0.96 0.92 0.76

V 0.20 0.15 0.18 0.35

Philippines S 0.01 0.15 0.27 0.07

K 0.81 0.76 0.59 0.61

V 0.06 1.20 0.07 0.05

Poland S 0.44 0.05 0.58 0.36

K 0.54 0.96 0.39 0.62

V 0.35 0.28 0.16 0.03

Sweden S 0.07 0.39 0.49 0.74

K 0.61 0.39 0.39 0.27

V 0.18 0.47 0.40 0.56

Thailand S 0.005 0.001 0.001 0.002

K 0.84 0.77 0.62 0.45

V 0.14 0.25 0.27 0.32

South A. S 0.17 0.21 0.25 0.18

K 0.72 0.65 0.51 0.53

323
 Table 3
OVERESTIMATION OF WES FORECASTS QI 1991 TO Q2 2016

3-step 2-step 1-step

4-step forecast forecast forecast
Countries forecast (QI) (QII) (QIII) (QIV)
Brazil 10 casesof 10 cases 10 cases 13 cases
( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.95 −5.09 −4.54 −70.26
Std. Deviation 1 11.83 11.75 201.59
Canada 16 cases 16 cases 17 cases 20 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.64 −0.55 −0.42 −0.4
Std. Deviation 0.58 0.42 0.27 0.42
Switzerland 20 cases 18 cases 19 cases 19 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.61 −0.45 −0.36 −0.3
Std. Deviation 0.44 0.27 0.27 0.24
Chile 15 cases 15 cases 12 cases 15 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.9 −0.91 −0.61 −0.56
Std. Deviation 0.72 0.7 0.54 0.45
Colombia 13 cases 15 cases 13 cases 13 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.23 −1.23 −1.15 −0.54
Std. Deviation 1.41 1.53 1.32 0.36
Czech Republic 21 cases 18 cases 19 cases 20 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −2.36 −2.05 −2.48 −1.15
Std. Deviation 4.94 5.48 7.6 2.53
United 20 cases 19 cases 17 cases 20 cases
Kingdom of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.78 −0.76 −0.78 −0.71
Std. Deviation 0.5 0.49 0.52 0.47
Korea 20 cases 24 cases 22 cases 20 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.46 −1.18 −0.99 −0.91
Std. Deviation 1.06 0.89 0.74 0.82

324 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 3−step 2−step 1−step
4−step forecast forecast forecast
Countries forecast (QI) (QII) (QIII) (QIV)
Mexico 17 cases 18 cases 15 cases 17 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.8 −0.96 −2.15 −1.12
Std. Deviation 0.71 1.73 4.94 3.28
Norway 16 cases 18 cases 15 cases 14 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.64 −0.52 −0.48 −0.33
Std. Deviation 0.51 0.43 0.32 0.27
Hungary 17 cases 13 cases 15 cases 13 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.5 −1.02 −0.77 −0.76
Std. Deviation 1.3 0.67 0.74 0.77
Philippines 21 cases 19 cases 19 cases 20 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.77 −1.41 −1 −1.1
Std. Deviation 1.42 0.9 0.73 0.64
Poland 17 cases 19 cases 14 cases 13 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −3.19 −1.21 −0.65 −0.45
Std. Deviation 5.62 1.28 0.39 0.28
Sweden 21 cases 21 cases 21 cases 20 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −0.89 −0.66 −0.66 −0.59
Std. Deviation 0.71 0.41 0.44 0.3
Thailand 17 cases 20 cases 21 cases 22 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.69 −1.27 −1.26 −0.91
Std. Deviation 1.83 1.2 1.04 0.65
South Africa 17 cases 18 cases 20 cases 21 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean −1.51 −1.27 −1.12 −0.93
Std. Deviation 1.15 0.41 0.62 0.23

325
 Table 4
UNDERESTIMATION OF WES FORECASTS Q1 1991 TO Q2 2016

4-step 3-step 2-step 1-step

forecast forecast forecast forecast
Countries (QI) (QII) (QIII) (QIV)
Brazil 16 cases 16 cases 15 cases 12 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 109.13 157.78 153.64 178.45
Std. Deviation 212.5 390.49 446.02 580.77
Canada 10 cases 10 cases 8 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.29 0.23 0.16 0.45
Std. Deviation 0.25 0.19 0.1 0.46
Switzerland 6 cases 8 cases 6 cases 6 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.53 0.29 0.2 0.27
Std. Deviation 0.6 0.37 0.21 0.27
Chile 11 cases 11 cases 13 cases 10 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 1.07 1.22 1.15 1.48
Std. Deviation 0.87 1.42 1.33 2.09
Colombia 13 cases 11 cases 12 cases 12 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 1.43 1.01 0.67 0.78
Std. Deviation 1.1 0.73 0.83 1.06
Czech Republic 5 cases 8 cases 6 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 1.76 0.79 1.03 2.33
Std. Deviation 2.32 1.15 1.74 3.93
United Kingdom 6 cases 7 cases 8 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.72 0.62 0.47 1
Std. Deviation 0.36 0.57 0.75 1.13
Korea 6 cases 2 cases 3 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.61 0.34 0.29 0.24
Std. Deviation 0.46 0.46 0.33 0.15

326 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 4-step 3-step 2-step 1-step
forecast forecast forecast forecast
Countries (QI) (QII) (QIII) (QIV)
Mexico 9 cases 8 cases 10 cases 8 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 3.2 1.8 2.12 3.22
Std. Deviation 4.8 1.38 2.25 2.69
Norway 10 cases 8 cases 10 cases 11 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.55 0.51 0.35 0.29
Std. Deviation 0.5 0.39 0.26 0.23
Hungary 9 cases 13 cases 10 cases 12 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 1.58 0.94 0.9 1.21
Std. Deviation 1.92 1.1 0.86 1.57
Philippines 5 cases 7 cases 6 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 2.04 1.56 0.87 0.76
Std. Deviation 1.52 1.41 1.32 0.82
Poland 9 cases 7 cases 11 cases 12 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 2.04 2.04 7.17 6.03
Std. Deviation 2.87 2.03 14.73 16.1
Sweden 5 cases 5 cases 4 cases 5 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.52 0.68 0.97 1.36
Std. Deviation 0.3 0.67 1.61 2.09
Thailand 9 cases 6 cases 4 cases 3 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.66 0.76 0.63 0.14
Std. Deviation 0.56 0.34 0.39 0.07
South Africa 9 cases 8 cases 5 cases 4 cases
of ( 26 ) of ( 26 ) of ( 25 ) of ( 25 )
Mean 0.96 0.72 0.6 0.4
Std. Deviation 1.15 0.41 0.62 0.23

327
 0.0
2.0
4.0
6.0
8.0
10.0
12.0
14.0
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
Ene 2000 Ene 2000
Mar 2000 Mar 2000
Ene 2001 Ene 2001
Mar 2001 Mar 2001
Ene 2002 Ene 2002
Mar 2002 Mar 2002
Ene 2003 Ene 2003
Mar 2003 Mar 2003
Ene 2004 Ene 2004
Mar 2004 Mar 2004
Ene 2005 Ene 2005
Mar 2005 Mar 2005
Ene 2006 Ene 2006

wes expectations
Mar 2006 Mar 2006
Ene 2007 Ene 2007
Mar 2007 Mar 2007
Ene 2008 Ene 2008
Figure 3

() 

Mar 2008 Mar 2008

() 
Ene 2009 Ene 2009

Source:  survey and  statistics and  data.

Mar 2009 Mar 2009

328 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Ene 2010 Ene 2010
Mar 2010 Mar 2010
Ene 2011 Ene 2011
Mar 2011 Mar 2011
Ene 2012 Ene 2012
Mar 2012 Mar 2012
ANNUAL INFLATION, AND INFLATION TARGETS

Ene 2013 Ene 2013

Mar 2013 Mar 2013

Annual inflation rates

Ene 2014 Ene 2014
Mar 2014 Mar 2014
COUNTRIES WES QUANTITATIVE INFLATION EXPECTATIONS,

Ene 2015 Ene 2015

Mar 2015 Mar 2015
Ene 2016 Ene 2016
 0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
4.50
5.00
0.00
0.50
1.00
1.50
2.00
2.50
3.00
3.50
4.00
Ene 2000 Ene 2000
Mar 2000 Mar 2000
Ene 2001 Ene 2001
Mar 2001 Mar 2001
Ene 2002 Ene 2002
Mar 2002 Mar 2002
Ene 2003 Ene 2003
Mar 2003 Mar 2003
Ene 2004 Ene 2004
Mar 2004 Mar 2004
Ene 2005 Ene 2005
Mar 2005 Mar 2005
Ene 2006 Ene 2006

wes expectations
Mar 2006 Mar 2006
Ene 2007 Ene 2007
Mar 2007 Mar 2007
Ene 2008 Ene 2008
Figure 3

Mar 2008 Mar 2008

() 

Ene 2009 Ene 2009

Source:  survey and  statistics and  data.

Mar 2009 Mar 2009

()  

Ene 2010 Ene 2010
Mar 2010 Mar 2010
Ene 2011 Ene 2011
Mar 2011 Mar 2011
Ene 2012 Ene 2012
Mar 2012 Mar 2012
ANNUAL INFLATION, AND INFLATION TARGETS

Ene 2013 Ene 2013

Mar 2013 Mar 2013

Annual inflation rates

Ene 2014 Ene 2014
Mar 2014 Mar 2014
COUNTRIES WES QUANTITATIVE INFLATION EXPECTATIONS,

Ene 2015 Ene 2015

Mar 2015 Mar 2015
Ene 2016 Ene 2016

329
3. METHODOLOGY

In this section, we describe the Artificial Neural Networks (anns)

models applied to cluster and forecast inflation expectations from
the wes surveys. To cluster we relied on Kohonen self-organizing
maps (som s), and to forecast we employed the multilayer percep-
tron from which the Non-linear autoregressive neuronal network,
nar-nn, is a subclass. The learning procedures to train anns is a sta-
tistical technique from which the weights are the relevant statistics
that could be found through an optimal solution, White (1989). Pre-
vious work that employed anns to forecast inflation include Stock
and Watson (1998) and Marcellino (2004) who conducted an exten-
sive successful forecasting study on emu macroeconomic variables.
On the other hand, Kock and Teräsvirta (2016) considered macro-
economic forecasting with a flexible single-hidden layer fed-forward
neural network.

3.1 Artificial Neural Networks

In order to explain the anns framework, we start looking at the
key points of the simple neural network model that form the base
of the som and nar-nn models.
anns are a type of parallel computing system consisting of several
simple interconnected processors called neurons or nodes, through
which there is a learning process that adjusts the system parameters
to approximate non-linear functions between a set of inputs (vari-
ables) and the output (results). For more information, see Jain,
Mao and Mohiuddin (1996).
Following Hagan et al. (2014), the simplest neuron model is com-
posed of a scalar input p, called a single variable, which is multiplied
by a scalar weight w. Then, w p plus the bias b form the called net in-
put n, which is sent to the activation function f, to produce the sca-
lar neuron output a. However, the ann’s architecture may be more
complex; they can have multiple inputs, layers, and neurons as shown
in Figure 4.
The parameters are constrained by weights and biases and are ad-
justed with some learning rule (e.g., Kohonen’s learning rule), while
the activation function is chosen according to the task at hand. For ex-
ample, in the som, the competitive function is applied. These net-
works are fed forward, which means that there are no loops between

330 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 4
A THREE-LAYER NEURAL NETWORK

Source: Hagan et al. (2014).

the outputs and inputs.9 To see more details about anns see Hagan
et al. (2014).

3.2 Self-Organizing Maps

In this paper, Self-Organized Maps, proposed by Kohonen in 1982
(see Kohonen, 2001), were used to cluster economic agents’ expecta-
tions before the oil shock. Furthermore, mapping those expectations
after the shock in the resulting cluster map, we divide the observa-
tions into two groups based on whether the expectations adjusted
briskly or softly. It is important to note that som s are competitive
feed-forward networks based on unsupervised training and have
the topology preservation property. This means that nearby input

9
In the nar-nn Model, to perform multi-step forecasts, the network
is transformed into a recurrent network after their parameters were
trained as a feed-forward network.

331
patterns should be represented on the map by nearby output units;
see Kohonen (2001).
The som architecture consists of a two-layer network: in the first lay-
er the inputs are multiplied with weights that were initialized as small
numbers. Then the results are evaluated by a competitive function
that produces a wining neuron (Best Matching unit). The weights
are updated according to the learning rule, equation (2), and the
neuron’s neighborhood is updated as well. See Figure 5 below.

2 wi (q ) = (1 − α )w i (q − 1) + α ( p (q ) )

The training stage for each iteration consists of weight adjustments

for the winning neuron and its neighbors and these adjustments
are undertaken using the learning rule. This process guarantees
similarity between the inputs and the neurons represented on the
feature map (the second layer of the map). At the end of the process,
the resulting learned weights capture the data characteristics on the
two-dimensional feature map (Hagan et al., 2014).
Kohonen suggested using rectangular and hexagonal neighbor-
hoods. Furthermore, to improve the som’s performance, we con-
sidered gradually decreasing the neighbor size during the training
so that it only includes the winning neuron. Moreover, to consider
the trade-off between fast learning and stability, the learning rate
can be also decreased in this phase. This is because a high learning
rate at the beginning of the training phase allows for quick but un-
stable learning. On the other hand, with a low rate, learning becomes
slow but more stable.

3.3 Nonlinear Auto-Regressive Neural Network

In this subsection, we describe the main issues of the nar-nn meth-
odology, including the selection of the training algorithm. The mod-
el assumes the current observation is explained by the compromise
of two components: signal and noise. The first is an unknown func-
tion that is approximated by the neural network to the inflation ex-
pectation time series with an autoregressive structure. The second
component is noise, which is assumed to be independent with zero
mean. The model equation is stated below:

332 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 5
A SELF-ORGANIZING MAP OF 5X5 DIMENSION

Source: Hagan et al. (2014).

Figure 6
WEIGHT SOM VECTORS OF WES EXPECTATIONS FOR THE NEXT 6 MONTHS

333
 3 ( )
Yt = g Yt −1 + Yt −2 + ... + Yt − p + et

4 ( ( ) )
Yt +1 = f 2 W 2 f 1 W 1 Yt ,Yt −1 ,...,Yt − p  + b1 + b2 + et +1

In order to obtain the best approximation for g, the neural network

architecture should meet the following three standard conditions:
it has to avoid overfitting,10 the predicted error should be uncorre-
lated over time, and the cross-correlation function between the pre-
dicted errors and the observed time series should be close to zero.
In this paper, we rely on the Bayesian regularization framework to ap-
proximate g  in a parsimonious manner (Titterington, 2004). The ob-
jective function for the Bayesian regularization setup is given by:

 
( ) (Yt − Yt ) +α ∑ni =1xi2
T
F (x ) = β ∑ t =1 Yt − Yt
T
5

This is the weighted combination between the model fit and

the smoothness. The parameter α penalizes model complexity and β
reflects the goodness of fit. The term xi2 is the sum of the squared
parameters values of the network, weights and biases.
Using the Bayes theorem sequentially, the joint posterior distri-
bution of the parameters α and β, given the data D and the neural
network model chosen M, is computed by multiplying the likelihood
times the joint a priori distribution of α and β divided by the evidence:

P ( D |α , β , M ) P (α , β | M )
6 P (α , β | D , M ) =
P ( D |M )

The prior joint density for α and β is assumed from the uniform
distribution. Consequently, the posterior can be obtained by com-
puting the following probabilities:

10
Overfitting is a characteristic that should be avoided and occurs when
the neural network fit the data closely in the training set, but in the test-
ing set and out of sample, the fitting is poor.

334 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 P ( D |X , β , M ) P ( X |α , M )
7 P ( D |α , β , M ) =
P ( X |D ,α , β , M )

P ( D |X , β , M ) P ( X |α , M )
8 P ( X |D ,α , β , M ) =
P ( D |α , β , M )

For more technical details and the full training algorithm

see Hagan et al. (2014).
The adaptation of the algorithm requires a neural network ar-
chitecture, M, which means we have to pick the number of neurons
in the input layer, the number of hidden layers, the number of neu-
rons per hidden layer, and the number of neurons in the output lay-
er. For more details see Zhang, Patuwo and Hu (1998).
Bayesian regularization guarantees that the parameter sum is
the optimal given data. In order to optimize the regularization pa-
rameters, the objective function F(x) should be minimized following
the Levenberg-Marquardt Back propagation algorithm.
The Bayesian regularization results exhibit flexibility to model
the network architecture. Thus, for the hidden layer, we set a fixed
number of nodes and we used just one hidden layer due to the length
of the time series. However, we observed that an extra layer did not
significantly change the results. With respect to the output layer,
one node is used because the forecast is one-step-ahead. The selec-
tion of the adequate number of input nodes or lags will be explained
in the nar-nn results section. In order to improve the generaliza-
tion of the network, the methodology usually requires one to divide
the data into three sets: training, validation, and testing. However,
Bayesian regularization avoids the validation stage because the so-
lution is based on the optimization of equation (3).
Moreover, we employed the hyperbolic Tangent Sigmoid as an ac-
tivation function for the nodes in the hidden layer as shown below.
This function is frequently used in forecasting.

9 e n − e −n
a=
e n + e −n

335
 10 a = n

For the output layer the linear function is used.11 The final archi-
tecture in matrix notations and scalar is:

11
( ( )
Yt +1 = f 2 W 2 f 1 W 1 Yt ,Yt −1 ,...,Yt − p  + b1 + b2 )
10  p 
yt +1 = ∑w 2j fj1  ∑wi1+ pYt1+i − p + b 1  + b 2
 
j =1  i =0 

n
p =
(
2 p − p min ) −1
(p max
−p min
)

where wi1+ p , i =1, ... , p, w 2j , i =1, ... , p  are the weights of the output layer,
b 1 is the biases of the first layer, and b 2 the biases of the second layer.
Figure 7 displays the observed data (black line), the fit in the train-
ing set (blue line), the forecasts in horizons 1 and 2 (green and orange
lines, respectively), and the out- of-sample forecasts eight steps ahead
(yellow line). Also, the figure is divided into three blocks. The block
on the left corresponds to the training set from Q31991 to Q2 2014;
the center block corresponds to the testing set from Q3 2014 to Q2
2016, which occurs after the oil shock period, and the right block
is the forecasting period.

3.4 arima
Box and Jenkins proposed the arima model in 1970 (Box et al., 2016).
The general expression of an arima model is the following:

Θs ( Ls )θ ( L )
Yt = εt
12
( )
Φs Ls ϕ ( L ) ∆sD ∆d

11
Notice that before training the network, data normalization, which
transforms the data in the interval between [-1, 1], is required to make
the training algorithm faster.

336 H. M. Zárate-Solano, D. R. Zapata-Sanabria

337
1
2
3
4
5
6
7
8
9

Jul 1991
Mar 1992
Nov 1992
Jul 1993
Mar 1994
Observed

Nov 1994
Jul 1995
Mar 1996
Nov 1996
Jul 1997
Mar 1998
Nov 1998
Jul 1999
Mar 2000
One-step-ahead Training Nov 2000
Jul 2001
Training set
Mar 2002
Nov 2002
Jul 2003
Mar 2004
Nov 2004
Figure 7

Testing set
Jul 2005
Mar 2006
Nov 2006

One-step-ahead Testing
Jul 2007
Mar 2008
Nov 2008

Out of sample
Jul 2009
DATA BLOCK DIVISION AND OUT OF SAMPLE SETS

Mar 2010
Nov 2010
Jul 2011
Mar 2012

Two-step-ahead
Nov 2012
Jul 2013
Mar 2014
Nov 2014
Jul 2015
Mar 2016

Out of sample
Nov 2016
Jul 2017
 ( ) ( )
where Θs Ls = 1 − Θs Ls − Θ2s L2s − Θ3s L3s − ... − ΘQs LQs is a seasonal

( ) (
moving average polynomial, Φs Ls = 1 − Φs Ls − Φ2s L2s − ... − Φ 3s L3s )
is the sea sona l auto -reg ressive poly nomia l, θ ( L ) = (1 − θ L
1 1 − θ2 L2 − ... −

( )
θ ( L ) = 1 − θ1L1 − θ2 L2 − ... − θq Lq is the regular moving average polynomi-
( )
al, and ϕ ( L ) = 1 − ϕ1L − ϕ2 L2 − ... − ϕ p Lp is a regular auto-regressive
1

polynomial, ∆sD is the seasonal difference operator, ∆d is the dif-

ference operator, s is the periodicity of the considered series (s =4
for quarterly data), and εt is the innovation which is assumed to rep-
resent white noise.12

4. RESULTS

In this section, we present the main results of the clustering and fore-
casting for inflation expectations across countries. First, we present
the som analysis that includes three sequential steps: the choice of the
map topology based on data, the training and validation stages of the
som neural network, and the elaboration of the clustering map of
agent expectations (in Appendix B we include a detailed explana-
tion of these steps). Then we overlap agents’ inflation expectations
on the resulting som map. Finally, the nar-nn results are provided.

4.1 Self-Organizing Maps of Agents’ Expectations

In this subsection, we briefly describe technical details on the imple-
mentation of the som analysis. We set a 10x10 hexagonal map with
a learning rate varying from 0.05 to 0.0001, and we used 1,000 itera-
tions. The computation was accomplished by the Kohonen package
in R developed by Wehrens and Buydens (2007). The training step
used observations before the oil shock identified on Q2 2014 and it
covers a sample of 84 observations per country for the expected situ-
ation by the end of the next six months of the overall economy, capi-
tal expenditures, private consumption, and inflation.13

12
The arima models chosen are described in Appendix D.
13
Appendix B explains the choice of topology as well as the post-training
analysis of the results.

338 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 8
SOMs OF COUNTRIES’ ECONOMY SITUATION EXPECTATIONS
FOR THE NEXT 6 MONTS (Q3 1991 TO Q4 2014)

()   ()  

()   ()  

A key tool in this analysis is the feature map or heat map that is the
representation of a single variable across the map (Figure 6). In this
application, the colors identify the intensity of the indicator. For ex-
ample: while the blue color is associated with low expectations, the red
is associated with high expectations. Clustering can be performed
by using hierarchical clustering on the weight learned vectors of the
variable. This procedure requires one to set the number of clusters.

339
Thus, given the nature of the expectations, we choose three clusters
to represent low, neutral, and high expectations.

4.2 Overlapping Agents’ Inflation Expectations by Country

In order to categorize agents’ inflation expectation patterns after
the oil price shock that took place on June 2014, we overlap those ex-
pectations from the third quarter of 2014 with the second quarter
of 2016 on the resulting heatmap. Next, we classified the expecta-
tions patterns by country into two categories: smooth and brisk ex-
pectation trajectories. For smooth transitions, we expected to find
a path that moves through a single cluster. Otherwise, we identify
a brisk trajectory by observing a changing path among several clus-
ters. In Figure 9, the black arrow represents the trajectory of the infla-
tion expectation with the initial node marked by a black start symbol.
For instance, in the case of Colombia, Figure 9(b), the observed
inflation expectations for July 2014 are in the higher expectation
cluster, then move through the heatmap ending in the lower expec-
tation cluster. We classified this pattern as one of brisk expectations.
Conversely, for the United Kingdom in Figure 9(d), inflation expec-
tations vary only between two clusters. Thus, it can be categorized
into the group with a smooth pattern. Table 5 summarizes the classi-
fication results for our sample of countries. From this table it is plau-
sible that changes in expectations in countries heavily dependent
on oil revenues were brisk, as exemplified by Colombia and Canada.
However, in countries such as Mexico, the change in expectations
is smooth because this economy is much more diversified. However,
we should consider that each country faces global and idiosyncratic
shocks that could have produced this heterogeneity as well.

340 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 9
COUNTRIES’ INFLATION RATE NEXT SIX MONTHS (Q3 2014 TO Q2 2016)
ON THE EXPECTED INFLATION RATE SOM MAP

()  () 

()  ()  

341
 Table 5
CLASSIFICATION OF INFLATION EXPECTATIONS
AND LAG SELECTED IN THE nar-nn MODEL

Country Inflation expectation Lag selected

Brazil Brisk 1

Canada Brisk 8

Chile Smooth 4

Colombia Brisk 5

Czech R. Smooth 6

Korea R. Smooth 2

Mexico Smooth 6

Norway Smooth 1

Switzerland Brisk 8

United K. Smooth 6

Hungary Smooth 10

Philippines Brisk 1

Poland Smooth 7

Sweden Smooth 1

Thailand Brisk 4

South A. Brisk 1

342 H. M. Zárate-Solano, D. R. Zapata-Sanabria

4.3 Non-Linear Auto-Regressive Neural Network Results
We have to select a model M to apply the Bayesian regulation frame-
work to the nar-nn in order to improve its generalization ability.
For each country, the sum of the parameters is conditional on the com-
plexity of the data. In this context, we chose a flexible network where
regularization guarantees the minimum sum of parameters. Thus,
we set an architecture with one hidden layer of 10 neurons. More-
over, at the input layer we have to specify the number of neurons that
correspond to the lag order used to forecast one step ahead. We used
the Neural Network Toolbox (Hagan, Demuth and Beale, 2002).
The lag order selection was based on different criteria: the mean
squared error resulting from the testing data, the error auto-correla-
tion function, and the cross-correlation between the errors and the
observed data. In this way, from lags 1 to 10 we generated 30 neural
networks per lag and obtain the MSE for the training, testing, and the
complete sample. Then, we select the lag that reports the smallest
median from the testing data sample, considering the auto-correla-
tion diagnostics.14 The lags chosen for each country are presented
in Table 5, and the overall results from lags 1 to 10 are shown in Table
6.15 A similar procedure was developed by Ruiz et al. (2016). Next,
we present the forecast results for some selected countries.16,17,18

14
In most of the cases mean and median, of the lag chosen, are both
the smallest. However, in Colombia, Czech Republic and Switzerland
this is not the case, even though the lag’s mean is closer to the small-
est mean.
15
These results for all datasets and training sets are presented in Tables
12 and 13, respectively, in Appendix 3.
16
To see the other countries, see Figure 24 in Appendix C.
17
A summary of results of the neural networks parameters is presented
in Table 14 in Appendix 3.
18
A simulation of 1000 networks was performed to ensure that the mse
presented belongs to the average neural network find after specifying
the model previously described. See Table 15 and Appendix 3.

343
 Table 6
LAG STATISTICS TEST DATA - ONE-STEP AHEAD FORECAST. SAMPLE = 30

Countries Lags 1 2 3 4 5 6 7 8 9 10
Brazil mean 1.65 2.08 1.85 1.86 1.89 1.85 1.83 1.84 1.92 2.04

median 1.57 2.07 1.85 1.86 1.83 1.85 1.83 1.84 1.92 2.04
Canada mean 2.04 1.84 1.59 1.63 1.62 1.54 1.56 1.52 1.75 1.73
median 2.04 1.74 1.59 1.63 1.62 1.54 1.56 1.52 1.75 1.73
Switzerland mean 1.32 1.21 1.04 1.04 1.02 0.98 1.06 0.79 0.93 0.78
median 1.30 1.22 1.04 1.04 1.02 0.98 1.06 0.78 0.94 0.83
Chile mean 4.34 2.76 2.77 2.70 2.74 2.80 2.94 3.13 3.13 2.93
median 4.38 2.76 2.79 2.68 2.76 2.81 3.00 3.06 3.13 2.91
Colombia mean 4.82 2.88 2.91 2.90 2.83 2.88 2.86 3.27 3.21 3.27

344 H. M. Zárate-Solano, D. R. Zapata-Sanabria

median 4.94 2.88 2.92 2.88 2.78 2.83 2.81 3.23 3.15 3.20
Czech R. mean 0.72 0.74 0.73 0.70 0.93 0.68 1.20 1.24 2.18 1.46
median 0.73 0.74 0.73 0.70 0.93 0.67 1.20 1.24 2.10 1.15
United K. mean 0.87 0.87 0.95 0.95 0.88 0.82 0.85 1.14 0.87 0.86
median 0.87 0.86 0.95 0.95 0.88 0.82 0.83 0.84 0.87 0.83
Korea R. mean 2.24 1.87 1.99 2.02 2.03 2.21 2.40 2.06 2.11 2.09
median 2.21 1.86 1.99 2.02 2.03 2.21 2.40 2.06 2.06 2.05
 Countries Lags 1 2 3 4 5 6 7 8 9 10
Mexico mean 0.38 0.42 0.52 0.48 0.48 0.31 0.50 0.36 0.45 1.07
median 0.38 0.42 0.52 0.49 0.48 0.30 0.57 0.37 0.31 0.53
Norway mean 1.41 1.44 1.61 1.67 1.59 2.01 2.04 2.10 1.88 1.80
median 1.41 1.44 1.61 1.67 1.59 2.01 2.04 2.10 1.88 1.80
Hungary mean 3.49 2.92 2.94 3.47 3.75 3.54 3.73 3.46 2.87 2.77
median 3.52 2.91 2.94 3.47 3.71 3.54 3.73 3.46 2.86 2.75
Philippines mean 3.41 3.99 3.86 3.78 3.78 3.83 3.47 3.60 4.14 3.65
median 3.41 3.99 3.86 3.78 3.78 3.83 3.47 3.60 4.05 3.45
Poland mean 1.12 1.02 1.04 1.07 1.37 1.07 0.72 0.87 3.52 7.12
median 1.12 1.04 1.01 1.06 1.37 1.07 0.72 0.86 2.99 7.12
Sweden mean 1.09 1.59 1.52 1.58 1.68 1.74 1.72 1.73 1.74 1.67
median 1.10 1.59 1.52 1.58 1.68 1.74 1.72 1.72 1.74 1.67
Thailand mean 1.67 1.01 1.03 0.90 0.95 1.00 1.06 1.10 1.05 1.02
median 1.68 1.01 1.03 0.91 0.95 1.00 1.06 1.10 1.05 1.01
South A. mean 2.63 3.31 3.48 3.56 3.97 3.85 4.27 4.51 4.59 4.64
median 2.63 3.31 3.48 3.56 3.97 3.85 4.11 4.51 4.59 4.97

345
 1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
Jul 1991 Jul 1991
Mar 1992 Mar 1992
Nov 1992 Nov 1992
Jul 1993 Jul 1993
Mar 1994 Mar 1994
Nov 1994 Nov 1994
Jul 1995 Jul 1995
Mar 1996 Mar 1996
Nov 1996 Nov 1996
Jul 1997 Jul 1997
Mar 1998 Mar 1998
Nov 1998 Nov 1998
Jul 1999 Jul 1999
Mar 2000 Mar 2000
Nov 2000 Nov 2000

Observed
One-step-ahead Testing
Jul 2001 Jul 2001
Mar 2002 Mar 2002
Nov 2002 Nov 2002
Jul 2003 Jul 2003
Mar 2004 Mar 2004
Nov 2004 Nov 2004
Jul 2005 Jul 2005
() 
Figure 10

Mar 2006

() 
Mar 2006
Nov 2006 Nov 2006
Jul 2007 Jul 2007

346 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Mar 2008 Mar 2008
Nov 2008

Two-step-ahead
Nov 2008
Jul 2009 Jul 2009
Mar 2010 Mar 2010
Nov 2010 Nov 2010

One-step-ahead Training
Jul 2011 Jul 2011
Mar 2012 Mar 2012
Nov 2012 Nov 2012
Jul 2013 Jul 2013
Mar 2014 Mar 2014
Nov 2014 Nov 2014
Jul 2015 Jul 2015

Out of sample
Mar 2016 Mar 2016
Nov 2016 Nov 2016
FORECASTS OF INFLATION EXPECTATIONS USING THE NAR-NN MODEL

Jul 2017 Jul 2017

 1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
Jul 1991 Jul 1991
Mar 1992 Mar 1992
Nov 1992 Nov 1992
Jul 1993 Jul 1993
Mar 1994 Mar 1994
Nov 1994 Nov 1994
Jul 1995 Jul 1995
Mar 1996 Mar 1996
Nov 1996 Nov 1996
Jul 1997 Jul 1997
Mar 1998 Mar 1998
Nov 1998 Nov 1998
Jul 1999 Jul 1999
Mar 2000 Mar 2000
Nov 2000 Nov 2000

Observed
One-step-ahead Testing
Jul 2001 Jul 2001
Mar 2002 Mar 2002
Nov 2002 Nov 2002
Jul 2003 Jul 2003
Mar 2004 Mar 2004
Nov 2004 Nov 2004
Jul 2005 Jul 2005
() 

Mar 2006 Mar 2006

Nov 2006 Nov 2006
Figure 10 (cont.)

()  

Jul 2007 Jul 2007
Mar 2008 Mar 2008

Two-step-ahead
Nov 2008 Nov 2008
Jul 2009 Jul 2009
Mar 2010 Mar 2010
Nov 2010 Nov 2010

One-step-ahead Training
Jul 2011 Jul 2011
Mar 2012 Mar 2012
Nov 2012 Nov 2012
Jul 2013 Jul 2013
Mar 2014 Mar 2014
Nov 2014 Nov 2014
Jul 2015 Jul 2015

Out of sample
Mar 2016 Mar 2016
Nov 2016 Nov 2016
FORECASTS OF INFLATION EXPECTATIONS USING THE NAR-NN MODEL

Jul 2017 Jul 2017

347
 4.4 Forecast Accuracy

Table 7
MSE COMPARISON AT TESTING DATA SETS FOR COUNTRIES
WITH BRISK INFLATION EXPECTATIONS

Arima NAR Diebold Diebold

Testing Testing
Testing set Testing set set set
One-step Two-step One-step Two-step One-step Two-step
Countries ahead ahead ahead ahead ahead ahead
Brazil 1.909 3.408 1.470 2.616 −0.988 −1.252
Canada 1.732 2.173 1.519 1.834 −1.402 −2.097
Colombia 2.913 2.926 2.776 2.648 −0.467 −1.763
Philippines 3.052 3.223 3.435 4.291 0.751 2.426
South A. 3.892 6.929 2.580 6.045 −1.571 −0.448
Switzerland 0.894 1.136 0.781 1.414 −0.343 1.041
Thailand 0.797 0.885 0.914 1.041 0.519 0.555
Brisk 2.018 2.961 1.734 2.632 −0.693 −0.702

Table 8
MSE COMPARISON AT TESTING DATA SETS FOR COUNTRIES
WITH SOFT INFLATION EXPECTATIONS

Arima NAR Diebold Diebold

Testing Testing
Testing set Testing set set set
Countries One-step Two-step One-step Two-step One-step Two-step
ahead ahead ahead ahead ahead ahead
Chile 3.577 4.181 2.680 2.429 −1.349 −2.539
Czech R 0.918 2.230 0.665 1.464 −0.763 −1.080
Hungary 3.485 6.850 2.746 4.734 −1.380 −1.610
Korea 1.764 2.812 1.857 3.028 2.870 8.936
Mexico 0.279 0.474 0.299 0.341 0.215 −0.945
Norway 1.484 2.019 1.419 1.221 −0.248 −1.043
Poland 1.028 2.263 0.716 0.925 −1.296 −3.950
Sweden 1.822 2.467 0.905 0.913 −2.087 −2.183
United K. 0.947 2.101 0.820 1.465 −0.945 −1.510
Soft 1.205 2.12 1.043 1.544 −0.033 0.33

348 H. M. Zárate-Solano, D. R. Zapata-Sanabria

5. CONCLUSIONS

Evaluating and forecasting inflation expectations from interna-

tional surveys of economics experts can be valuable for monetary
macroeconomic modeling. In this research, we set two goals. First,
we analyzed wes inflation expectations data for 16 countries that
adopted inflation targeting regimes as the basis of their monetary
policy. Given that the quarterly questions on the evolution of prices
in these surveys consider both qualitative and quantitative scales,
we used a descriptive analysis for the relationship between inflation
expectations and observed inflation, and we study the structure
of the in-sample forecasting errors.
Second, we generated-out-of-sample forecasts for the inflation
expectations of the countries by relying on a two-step approach
to sequentially cluster and forecast inflation expectations. Thus,
the clustering technique known as Self-Organizing Maps and a
predictive model based on artificial neural networks allow us to vi-
sualize and predict different patterns of inflation expectations ac-
cording to their perceptions before the oil shock that took place
in the middle of 2014.
We cluster the countries according to the evolution of their infla-
tion expectations during the transition period to the recent mini-
mum oil price mark. Then, we obtain forecasts of survey expectations
by using linear and non-linear nar-nn methods. For the som analy-
sis, we find that some countries exhibited brisk behavior that is as-
sociated with signs that inflation expectations were de-anchoring.
At the same time, there were countries with a soft evolution of infla-
tion expectations.
The correlation analysis from the time and frequency domain
indicates the existence of different patterns of linear associations
over time and frequency: increasing, descending, and inverted U-
shaped. Moreover, the highest coherence between inflation and ex-
pectations was found mainly in higher frequencies, which suggests
that the relationship between inflation expectations and observed
inflation is present in short duration cycles.
Concerning the statistical evaluation based on the forecasting
errors of the quantitative inflation expectation, we detected uncer-
tainty in the predictions of average annual inflation across coun-
tries that could be classified into two groups. In the first group,

349
the closer the expert is to the end of the year, the smaller the predic-
tion bias. This group includes Colombia and Switzerland among
others. The other group of countries exhibit increasing bias in the
last quarter of the prediction period and include Brazil, Canada,
and Chile.
Additionally, the quality of the quantitative question is judged
by standard measures of forecast evaluation at different horizons:
rmse, mae, and U-Theil. Thus, we concluded that the forecasts
meet a minimum standard compared to the random walk reference
and that economic experts have made systematic errors in their pre-
dictions. Inflation was under- predicted when it was rising and over-
predicted when it was declining in most of the countries. The Theil
decomposition of the mae illustrated that 83 percent of the countries
experienced systematic distortion in their forecasts, which means
that the increase in accuracy with shorter forecast horizons is not
monotonic. The evidence does not support the claim that forecasts
have improved over time due to a non-linear generating data process.
The evidence also suggests that turning points of observed average
inflation were mostly anticipated in most cases. This issue may be
an interesting area for further research.
On the other hand, a Self-Organizing Map analysis of surveys ex-
pectations before the impending oil shock allows us to classify infla-
tion expectations as either brisk or soft based on the speed with which
expectations shift. Using this classification, we can select the most
appropriate forecasting method. We notice that the low-inflation ex-
pectations cluster is relatively small compared to high and neutral
clusters for inflation targeting countries. The Nonlinear auto-re-
gressive neural network and arima methods were used as competing
candidates to forecast inflation expectations. The results indicate
that in the one step ahead forecasts the neural network is slightly bet-
ter, but in two step-ahead forecasts, it outperforms the arima mod-
el significantly. For Canada, Colombia, Chile, Poland, Hungary,
and Sweden in particular, the neural network produces significant
improvement in the two-step ahead forecasts.
Further research is required to provide theoretical economic ex-
planations for the results of each country. Moreover, this combina-
tion between machine learning and statistics can be implemented
in a follow-up paper to forecast actual inflation.

350 H. M. Zárate-Solano, D. R. Zapata-Sanabria

ANNEX A. DATA

A.1 Qualitative Series

Figure 11
EXPECTED INFLATION RATE FOR THE NEXT SIX MONTHS
Wes Qualitative Question


2 3 4 5 6 7 8
Inflation exp

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


7 8
Inflation exp
2 3 4 5 6

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01

351
 Figure 11 (cont.)
EXPECTED INFLATION RATE FOR THE NEXT SIX MONTHS
Wes Qualitative Question

8 6 
Inflation exp
4 2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


8
Inflation exp
4 2 6

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01

 
8
Inflation exp
4 2 6

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01

352 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 11 (cont.)
EXPECTED INFLATION RATE FOR THE NEXT SIX MONTHS
Wes Qualitative Question

 

6
Inflation exp
4
2

1995 2000 2005 2010 2015

From 1991-07-01 to 2016-04-01

8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−

01


8
6
Inflation exp
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01

353
 Figure 11 (cont.)
EXPECTED INFLATION RATE FOR THE NEXT SIX MONTHS
Wes Qualitative Question



Inflation exp
2 3 4 5 6 7

1995 2000 2005 2010 2015


8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01


8
Inflation exp
6
4
2

1995 2000 2005 2010 2015

From 1991−07−01 to 2016−04−01

354 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 12
HISTOGRAMS OF AGENTS’ EXPECTATIONS OF ECONOMIC SITUATION
FOR NEXT SIX MONTHS IN MACROECONOMIC VARIABLES

 

 

 

 

355
 Table 9
DATA SUMMARY OF wes EXPECTATIONS FROM Q3 1991 TO Q2 2016
Selected countries

Capital Private
Overall economy expenditures consumption Inflation rate
Min 1 1 1 1
1stQ 4.8 4.7 4.57 4
Median 5.8 5.7 5.5 5.5
Mean 5.79 5.59 5.44 5.32
3rdQ 6.8 6.6 6.5 6.8
Max 9 9 9 9

Figure 13
SCATTER PLOT OF AGENTS’ EXPECTATIONS OF ECONOMIC SITUATION
FOR NEXT SIX MONTHS

356 H. M. Zárate-Solano, D. R. Zapata-Sanabria

A.2 wes Survey Questionnaire

Figure 14
EXAMPLE OF WORLD ECONOMIC SURVEY (WES) QUESTIONNAIRE

357
 0
2
4
6
8

-2

0
1
2
3
4
5
6
7
8

-1
10
12
0
5
10
15
20
25
Mar-2000 Mar-2000 Dec-1995
Sep-2000 Sep-2000 Jun-1996
Mar-2001 Mar-2001 Dec-1996
Jun-1997
Sep-2001 Sep-2001
Dec-1997
Mar-2002 Mar-2002 Jun-1998
Sep-2002 Sep-2002 Dec-1998
Mar-2003 Mar-2003 Jun-1999
Dec-1999
Sep-2003 Sep-2003 Jun-2000
Mar-2004 Mar-2004 Dec-2000
Sep-2004 Sep-2004 Jun-2001
Mar-2005 Mar-2005 Dec-2001
Jun-2002
Sep-2005 Sep-2005 Dec-2002

Source:  and  statistics.

Mar-2006 Mar-2006 Jun-2003
Sep-2006 Sep-2006 Dec-2003
Jun-2004
Mar-2007 Mar-2007
Dec-2004

Annual Inflation
Sep-2007 Sep-2007 Jun-2005
Mar-2008 Mar-2008 Dec-2005
Sep-2008 Sep-2008 Jun-2006
Dec-2006




Mar-2009 Mar-2009 Jun-2007

Figure 15

Sep-2009 Sep-2009 Dec-2007

 
Mar-2010 Mar-2010 Jun-2008
Sep-2010 Dec-2008
Sep-2010
Jun-2009

358 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Mar-2011 Mar-2011 Dec-2009
Sep-2011 Sep-2011 Jun-2010
Mar-2012 Mar-2012 Dec-2010
Jun-2011
Sep-2012 Sep-2012
Dec-2011

Inflation rate next 6 months

Mar-2013 Mar-2013 Jun-2012
Sep-2013 Sep-2013 Dec-2012
Mar-2014 Jun-2013
Mar-2014
Dec-2013
Sep-2014 Sep-2014 Jun-2014
Mar-2015 Mar-2015 Dec-2014
Sep-2015 Sep-2015 Jun-2015
Mar-2016 Dec-2015
COUNTRIES’ INFLATION EXPECTATIONS AND ANNUAL INFLATION

Mar-2016 Jun-2016
Sep-2016 Sep-2016
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
 0
1
2
3
4
5
6

0
2
4
6
8

0
2
4
6
8
10
12

10
12
Mar-2000 Mar-2000 Mar-2000
Sep-2000 Sep-2000 Sep-2000
Mar-2001 Mar-2001 Mar-2001
Sep-2001 Sep-2001 Sep-2001
Mar-2002 Mar-2002 Mar-2002
Sep-2002 Sep-2002 Sep-2002
Mar-2003 Mar-2003 Mar-2003
Sep-2003 Sep-2003 Sep-2003
Mar-2004 Mar-2004 Mar-2004
Sep-2004 Sep-2004 Sep-2004
Mar-2005 Mar-2005 Mar-2005
Sep-2005 Sep-2005 Sep-2005

Source:  and  statistics.

Mar-2006 Mar-2006 Mar-2006
Sep-2006 Sep-2006 Sep-2006
Mar-2007 Mar-2007 Mar-2007

Annual Inflation
Sep-2007 Sep-2007 Sep-2007
Mar-2008 Mar-2008 Mar-2008
Sep-2008 Sep-2008 Sep-2008



Mar-2009 Mar-2009 Mar-2009


Sep-2009 Sep-2009 Sep-2009
Mar-2010 Mar-2010 Mar-2010
Figure 15 (cont.)

Sep-2010 Sep-2010 Sep-2010

Mar-2011 Mar-2011 Mar-2011
Sep-2011 Sep-2011 Sep-2011
Mar-2012 Mar-2012 Mar-2012
Sep-2012 Sep-2012 Sep-2012

Inflation rate next 6 months

Mar-2013 Mar-2013 Mar-2013
Sep-2013 Sep-2013 Sep-2013
Mar-2014 Mar-2014 Mar-2014
Sep-2014 Sep-2014 Sep-2014
Mar-2015 Mar-2015 Mar-2015
Sep-2015 Sep-2015 Sep-2015
Mar-2016
COUNTRIES’ INFLATION EXPECTATIONS AND ANNUAL INFLATION

Mar-2016 Mar-2016
Sep-2016 Sep-2016 Sep-2016

359
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

10.00
 0
1
2
3
4
5

-3
-2
-1
0
2
4
6
8

0
2
4
6
8

-4
-2
10
12

10
12
14
16
Mar-2000 Mar-2000 Mar-2000
Sep-2000 Sep-2000 Sep-2000
Mar-2001 Mar-2001 Mar-2001
Sep-2001 Sep-2001 Sep-2001
Mar-2002 Mar-2002 Mar-2002
Sep-2002 Sep-2002 Sep-2002
Mar-2003 Mar-2003 Mar-2003
Sep-2003 Sep-2003 Sep-2003
Mar-2004 Mar-2004 Mar-2004
Sep-2004 Sep-2004 Sep-2004
Mar-2005 Mar-2005 Mar-2005
Sep-2005 Sep-2005 Sep-2005

Source:  and  statistics.

Mar-2006 Mar-2006 Mar-2006
Sep-2006 Sep-2006 Sep-2006
Mar-2007 Mar-2007 Mar-2007

Annual Inflation
Sep-2007 Sep-2007 Sep-2007
Mar-2008 Mar-2008 Mar-2008
Sep-2008 Sep-2008 Sep-2008
Mar-2009



Mar-2009 Mar-2009
Sep-2009 Sep-2009

 
Sep-2009
Mar-2010 Mar-2010 Mar-2010
Figure 15 (cont.)

Sep-2010

360 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Sep-2010 Sep-2010
Mar-2011 Mar-2011 Mar-2011
Sep-2011 Sep-2011
Sep-2011
Mar-2012 Mar-2012
Mar-2012
Sep-2012 Sep-2012
Sep-2012

Inflation rate next 6 months

Mar-2013 Mar-2013
Mar-2013
Sep-2013 Sep-2013
Sep-2013
Mar-2014 Mar-2014
Mar-2014
Sep-2014 Sep-2014
Sep-2014
Mar-2015 Mar-2015
Mar-2015
Sep-2015 Sep-2015
Sep-2015
Mar-2016
COUNTRIES’ INFLATION EXPECTATIONS AND ANNUAL INFLATION

Mar-2016 Mar-2016
Sep-2016 Sep-2016
Sep-2016
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
10.00

10.00

0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
 -2
0
2
4
6
8
10
12
-6
-4
-2
0
2
4
6
8
10
-2
-1.5
-1
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
Mar-2000 Mar-2000 Mar-1995
Sep-2000 Sep-2000 Sep-1995
Mar-2001 Mar-2001 Mar-1996
Sep-1996
Sep-2001 Sep-2001 Mar-1997
Mar-2002 Mar-2002 Sep-1997
Sep-2002 Sep-2002 Mar-1998
Mar-2003 Sep-1998
Mar-2003 Mar-1999
Sep-2003 Sep-2003 Sep-1999
Mar-2004 Mar-2004 Mar-2000
Sep-2004 Sep-2000
Sep-2004 Mar-2001
Mar-2005 Mar-2005 Sep-2001
Sep-2005 Sep-2005 Mar-2002

Source:  and  statistics.

Mar-2006 Mar-2006 Sep-2002
Mar-2003
Sep-2006 Sep-2006 Sep-2003
Mar-2007 Mar-2007 Mar-2004
Sep-2004

Annual Inflation
Sep-2007 Sep-2007
Mar-2008 Mar-2005
Mar-2008 Sep-2005
Sep-2008 Sep-2008 Mar-2006
Mar-2009 Mar-2009 Sep-2006



Sep-2009 Mar-2007


Sep-2009 Sep-2007
Mar-2010 Mar-2010 Mar-2008
Figure 15 (cont.)

Sep-2010 Sep-2010 Sep-2008

Mar-2011 Mar-2011 Mar-2009
Sep-2009
Sep-2011 Sep-2011 Mar-2010
Mar-2012 Mar-2012 Sep-2010
Sep-2012 Sep-2012 Mar-2011
Sep-2011

Inflation rate next 6 months

Mar-2013 Mar-2013 Mar-2012
Sep-2013 Sep-2013 Sep-2012
Mar-2014 Mar-2014 Mar-2013
Sep-2014 Sep-2013
Sep-2014 Mar-2014
Mar-2015 Mar-2015 Sep-2014
Sep-2015 Sep-2015 Mar-2015
Mar-2016 Mar-2016 Sep-2015
COUNTRIES’ INFLATION EXPECTATIONS AND ANNUAL INFLATION

Mar-2016
Sep-2016 Sep-2016 Sep-2016

361
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00
0.00
1.00
2.00
3.00
4.00
5.00
6.00
7.00
8.00
9.00

10.00
10.00
 Figure 16
CORRELATION COEFFICIENTS BETWEEN WES QUALITATIVE
INFLATION EXPECTATION AND ANNUAL INFLATION

. :    . :   

. :    . :   

.  .:    .  .:   

. :    . :   

Source:  survey,  statistics and  data.

362 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 16 (cont.)
CORRELATION COEFFICIENTS BETWEEN WES QUALITATIVE
INFLATION EXPECTATION AND ANNUAL INFLATION

. :    . :   

. :    . :   

. :    . :   

. :    . :   

Source:  survey,  statistics and  data.

363
 Figure 16 (cont.)
CORRELATION COEFFICIENTS BETWEEN WES QUALITATIVE
INFLATION EXPECTATION AND ANNUAL INFLATION

.  :    .  :   

. :    . :   

. :    . :   

. :    . :   

Source:  survey,  statistics and  data.

364 H. M. Zárate-Solano, D. R. Zapata-Sanabria

A.3 Quantitative Forecasting Inflation Expectations
A.3.1 Equations of the Statistical Analysis Forecasting Error
Root mean squared forecast error (rmsfe):

1
∑ e ( L,Q (h ),t )
2016
13
26 1991

Mean absolute error (mae):

1 2
∑ e ( L,Q (h ),t )
2016
14 1991
26

Theil U.statistic:
1 2
∑ e ( L,Q (h ),t )
2016
26 1991
15
1 2 1 2
∑ q ( L,Q (h ),t ) ∑
2016 2016
1991 1991
p ( L,t )
26 26
Bias share:
1 2 1 2
∑
2016
 26 1991 q ( L ,Q ( h ) ,t ) − ∑
2016
26 1991
p ( L ,t ) 
16 V(h)=
1 2
∑ e ( L,Q (h ),t )
2016
26 1991
The spread share:
2
Sq (h ) − Sq (h ) 
S (h ) =  
17 1 2
∑ e ( L,Q (h ),t )
2016
26 1991

where Sq (h ) and Sq (h ) are the standard deviations of the respec-

tive quarter. The covariance share:

K (h ) =
( )
2 1 − rq ,p (h ) Sq (h ) − S (h )
18 1 2
∑ e ( L,Q (h ),t )
2016
26 1991

where rq ,p (h ) is the correlation coefficient between q  and p. Thus

V(h)+S(h)+K(h)=1.

365
 Table 10
MAE OF WES SURVEY QUANTITATIVE INFLATION QUESTION

3-step 2-step 1-step

forecast forecast forecast
4-step forecast (QI) (QII) (QIII) (QIV)

Brazil 67.52 99.05 94.00 122.19

Canada 0.51 0.43 0.34 0.41

Switzerland 0.59 0.41 0.32 0.30

Chile 0.97 1.04 0.89 0.93

Colombia 1.33 1.14 0.92 0.65

Czech Republic 2.25 1.66 2.14 1.39

United Kingdom 0.76 0.72 0.68 0.77

Korea 1.26 1.11 0.91 0.78

Mexico 1.63 1.22 2.14 1.79

Norway 0.61 0.51 0.43 0.31

Hungary 1.53 0.98 0.82 0.98

Philippines 1.82 1.45 0.97 1.03

Poland 2.79 1.44 3.52 3.13

Sweden 0.82 0.67 0.71 0.75

Thailand 1.34 1.15 1.16 0.81

South Africa 1.32 1.10 1.02 0.84

366 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Table 11
U-STATISTIC OF WES SURVEY QUANTITATIVE INFLATION QUESTION

3-step 2-step 1-step

forecast forecast forecast
4-step forecast (QI) (QII) (QIII) (QIV)

Brazil 0.001 0.001 0.001 0.001

Canada 0.138 0.113 0.083 0.115

Switzerland 0.237 0.162 0.126 0.120

Chile 0.022 0.028 0.027 0.033

Colombia 0.010 0.009 0.007 0.005

Czech Republic 0.075 0.074 0.087 0.057

United Kingdom 0.110 0.111 0.112 0.122

Korea 0.075 0.067 0.055 0.054

Mexico 0.022 0.011 0.022 0.019

Norway 0.143 0.118 0.101 0.079

Hungary 0.011 0.007 0.006 0.008

Philippines 0.046 0.039 0.027 0.025

Poland 0.010 0.004 0.032 0.033

Sweden 0.141 0.121 0.151 0.210

Thailand 0.118 0.091 0.081 0.058

South Africa 0.030 0.026 0.024 0.021

367
 0
1
2
3
4
5
6
7
0
1
2
3
4
5
6
7
8
9
10
0
2
4
6
8
10
12
14
16

Ene 2000 Ene 2000 Ene 2000

Mar 2000 Mar 2000 Mar 2000
Ene 2001 Ene 2001 Ene 2001
Mar 2001 Mar 2001 Mar 2001
Ene 2002 Ene 2002 Ene 2002
Mar 2002 Mar 2002 Mar 2002
Ene 2003 Ene 2003 Ene 2003
Mar 2003 Mar 2003 Mar 2003
Ene 2004 Ene 2004 Ene 2004
Mar 2004 Mar 2004 Mar 2004
Ene 2005 Ene 2005 Ene 2005
Mar 2005 Mar 2005 Mar 2005
Ene 2006 Ene 2006 Ene 2006

wes expectations
Mar 2006 Mar 2006 Mar 2006
Ene 2007 Ene 2007 Ene 2007
Mar 2007 Mar 2007 Mar 2007
Ene 2008 Ene 2008 Ene 2008

() 
() 

Source:  survey and  statistics and  data.

Figure 17

Mar 2008 Mar 2008 Mar 2008

Ene 2009 Ene 2009 Ene 2009

()  

Mar 2009 Mar 2009 Mar 2009

368 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Ene 2010 Ene 2010 Ene 2010
Mar 2010 Mar 2010 Mar 2010
Ene 2011 Ene 2011 Ene 2011
Mar 2011 Mar 2011 Mar 2011
Ene 2012 Ene 2012 Ene 2012
Mar 2012 Mar 2012 Mar 2012
Ene 2013 Ene 2013 Ene 2013
Mar 2013 Mar 2013 Mar 2013

Annual inflation rates

Ene 2014 Ene 2014 Ene 2014
Mar 2014 Mar 2014 Mar 2014
Ene 2015 Ene 2015 Ene 2015
Mar 2015 Mar 2015 Mar 2015
Ene 2016 Ene 2016 Ene 2016
COUNTRIES’ QUANTITATIVE INFLATION EXPECTATIONS AND ANNUAL INFLATION
 0
1
2
3
4
5
6

0
1
2
3
4
5
6
7
8
9
10
0
2
4
6
8
10
12
14
Ene 2000 Ene 2000 Ene 2000
Mar 2000 Mar 2000 Mar 2000
Ene 2001 Ene 2001 Ene 2001
Mar 2001 Mar 2001 Mar 2001
Ene 2002 Ene 2002 Ene 2002
Mar 2002 Mar 2002 Mar 2002
Ene 2003 Ene 2003 Ene 2003
Mar 2003 Mar 2003 Mar 2003
Ene 2004 Ene 2004 Ene 2004
Mar 2004 Mar 2004 Mar 2004
Ene 2005 Ene 2005 Ene 2005
Mar 2005 Mar 2005 Mar 2005
Ene 2006 Ene 2006 Ene 2006
Mar 2006 Mar 2006 Mar 2006
Ene 2007 Ene 2007 Ene 2007
Mar 2007 Mar 2007 Mar 2007
Ene 2008 Ene 2008 Ene 2008
() 

Mar 2008
() 

Source:  survey and  statistics and  data.

Mar 2008 Mar 2008

() 
Ene 2009 Ene 2009 Ene 2009
Figure 17 (cont.)

Mar 2009 Mar 2009 Mar 2009

Ene 2010 Ene 2010 Ene 2010
Mar 2010 Mar 2010 Mar 2010
Ene 2011 Ene 2011 Ene 2011
Mar 2011 Mar 2011 Mar 2011
Ene 2012 Ene 2012 Ene 2012
Mar 2012 Mar 2012 Mar 2012
Ene 2013 Ene 2013 Ene 2013
Mar 2013 Mar 2013 Mar 2013
Ene 2014 Ene 2014 Ene 2014
Mar 2014 Mar 2014 Mar 2014
Ene 2015 Ene 2015 Ene 2015
Mar 2015 Mar 2015 Mar 2015
Ene 2016 Ene 2016 Ene 2016

369
COUNTRIES’ QUANTITATIVE INFLATION EXPECTATIONS AND ANNUAL INFLATION
 −2
0
2
4
6
8
10
12
14
−2
0
2
4
6
8
10
12

−1.0
−0.5
0.0
0.5
1.0
1.50
2.0
2.5
3.0
3.5
4.0
Ene 2000 Ene 2000
Ene 2000
Mar 2000 Mar 2000
Mar 2000
Ene 2001 Ene 2001
Ene 2001
Mar 2001 Mar 2001
Mar 2001
Ene 2002 Ene 2002
Ene 2002
Mar 2002 Mar 2002
Mar 2002
Ene 2003 Ene 2003
Ene 2003
Mar 2003 Mar 2003
Mar 2003
Ene 2004 Ene 2004
Ene 2004
Mar 2004 Mar 2004
Mar 2004
Ene 2005 Ene 2005
Ene 2005
Mar 2005 Mar 2005
Mar 2005
Ene 2006 Ene 2006
Ene 2006
Mar 2006 Mar 2006
Mar 2006
Ene 2007 Ene 2007
Ene 2007
Mar 2007 Mar 2007
Mar 2007
Ene 2008 Ene 2008
() 

Ene 2008

() 

Source:  survey and  statistics and  data.

Mar 2008 Mar 2008
Mar 2008
()  
Ene 2009 Ene 2009
Ene 2009
Figure 17 (cont.)

Mar 2009 Mar 2009

370 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Mar 2009
Ene 2010 Ene 2010
Ene 2010
Mar 2010 Mar 2010
Mar 2010
Ene 2011 Ene 2011
Ene 2011
Mar 2011 Mar 2011
Mar 2011
Ene 2012 Ene 2012
Ene 2012
Mar 2012 Mar 2012
Mar 2012
Ene 2013 Ene 2013
Ene 2013
Mar 2013 Mar 2013
Mar 2013
Ene 2014 Ene 2014
Ene 2014
Mar 2014 Mar 2014
Mar 2014
Ene 2015 Ene 2015
Ene 2015
Mar 2015 Mar 2015
Mar 2015
Ene 2016 Ene 2016
Ene 2016
COUNTRIES’ QUANTITATIVE INFLATION EXPECTATIONS AND ANNUAL INFLATION
 0.0
0.5
1.0
1.5
2.0
2.5
3.0
3.5
4.0
4.5
5.0
−2
−1
0
1
2
3
4
5
6
7
8
9
−1.5
−1.0
−0.5
0
0.5
1.0
1.5
2.0
2.5
3.0

Ene 2000 Ene 2000

Ene 2000
Mar 2000 Mar 2000
Mar 2000
Ene 2001 Ene 2001
Ene 2001
Mar 2001 Mar 2001
Mar 2001
Ene 2002 Ene 2002
Ene 2002
Mar 2002 Mar 2002
Mar 2002
Ene 2003 Ene 2003
Ene 2003
Mar 2003 Mar 2003
Mar 2003
Ene 2004 Ene 2004
Ene 2004
Mar 2004 Mar 2004
Mar 2004
Ene 2005 Ene 2005
Ene 2005
Mar 2005 Mar 2005
Mar 2005
Ene 2006 Ene 2006
Ene 2006
Mar 2006 Mar 2006
Mar 2006
Ene 2007 Ene 2007
Ene 2007
Mar 2007 Mar 2007
Mar 2007
Ene 2008 Ene 2008
Ene 2008

Source:  survey and  statistics and  data.

() 
Mar 2008 Mar 2008
Mar 2008
Ene 2009 Ene 2009
() 

Ene 2009
Figure 17 (cont.)

Mar 2009 Mar 2009

()  

Mar 2009
Ene 2010 Ene 2010
Ene 2010
Mar 2010 Mar 2010
Mar 2010
Ene 2011 Ene 2011
Ene 2011
Mar 2011 Mar 2011
Mar 2011
Ene 2012 Ene 2012
Ene 2012
Mar 2012 Mar 2012
Mar 2012
Ene 2013 Ene 2013
Ene 2013
Mar 2013 Mar 2013
Mar 2013
Ene 2014 Ene 2014
Ene 2014
Mar 2014 Mar 2014
Mar 2014
Ene 2015 Ene 2015
Ene 2015
Mar 2015 Mar 2015
Mar 2015
Ene 2016 Ene 2016
Ene 2016

371
COUNTRIES’ QUANTITATIVE INFLATION EXPECTATIONS AND ANNUAL INFLATION
 Figure 18
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -   

. :   

372 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -   

. :   

373
 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -    

. -   

. :   

374 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -   

. :   

375
 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -   

. :   

376 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -    

. :   

377
 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -   

. :   

378 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 18 (cont.)
QUARTER-SPECIFIC FORECASTING ERROR BY COUNTRY

. -   

. -    

. :   

379
ANNEX B. SELF-ORGANIZING MAP VALIDATION

B.1 Choice of Topology

In this section, we present the best topology according to avail-
able data. This includes presenting the dimensions of the map and
the form of the neighborhood. In order to have more neighbors
around the winning neuron, we choose the hexagonal topology that
allocates six neurons around the center one. For the dimensions
we found several empirical rules. The first rule is to have the num-
ber of neurons increase with the square root y of the number of data
points. This give us a map of 40 neurons. The second rule is to have
10 samples per neuron, which gives a total of 192 neurons.
We tried different architectures to try to get enough granular-
ity on the map with small topographic error. Unfortunately, there
is not a set criterion by which to judge performance in som networks.
Therefore, to complete our goal of finding the agent’s clusters before
the oil price shock, we divide our data into two sets, before and af-
ter the shock. Thus, the training data will be from the third quarter
of 1991 to the second quarter of 2014.
Using the R software, we analyzed various architectures: the di-
mensions of the map (3x10 vs. 18x10), the storage of their topographic
errors, and their granularity.19 Figure 19 shows us the choice of hex-
agonal topology of 10x10.

B.2 Post-Training Analysis

Following Wehrens (2007) and Lynn (2014) we analyze the results
from the trained map to validate the previous results. The train-
ing progress shows the mean distance between neuron’s weights
to the samples represented through each iteration. When the train-
ing progress reaches a minimum, no more iterations are required.
See Figure 20.

19
The quantization error is not comparable between maps because it is
susceptible to map size. To see more about topographic errors see the
Post-training analysis section.

380 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 19
BEST MATCHING UNIT ERROR, ERROR NODE DISTANCE, QUANTIZATION
ERROR, AND SAMPLE PER NEURON VS. MAP WIDTH NODE SIZE

1.9
Best matching unit error 1.8
1.7
1.6
1.5
1.4
5 10 15
1.2
Quantization error

1.0

0.8

0.6

0.4

5 10 15

1.8
Error node distant

1.6

1.4

1.2

5 10 15
50
Average sample per node

40

30

20

10

5 10 15

Source:  survey and  statistics and  data.

381
 Figure 20
POST-TRAINING ANALYSIS

.  /

.  

382 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 In Figure 20(a), the node or quality distance map is shown. This
map displays an approximation of the distance per node to the sam-
ple that they are representing; this is known as the quantization er-
ror. According to the quantization error, the smaller the distance,
the better the map. When it is large, some input vectors are not ad-
equately represented on the map. However, the error is also sub-
ject to map sizes: if the map is large, it could be close to zero. This
would represent overfitting because the number of neurons on the
map should be significantly smaller that the sample size. The mean
quantization error found is 0.5888693.
In Figure 20(b), one can analyze how many samples are mapped
to each node on the map. Ideally, we want the sample distributions
to be relatively uniform. Our map is relatively uniform, including be-
tween 10 to 15 samples per neuron, and there are non-empty neurons.
Figure 21(b) shows a map that is also named the U-matrix and which
shows the distance between each neuron and its immediate neigh-
bors. Because we choose a hexagonal neighbor, each neuron has six
neurons in it neighborhood. This map also assists in identifying
similar neurons.
The weight vectors plot, Figure 22, shows the weights associated
with each neuron. Each weight vector is similar to the variable that
it represents due to Kohonen’s learning rule. The weight distribu-
tions on the map represent: green for the overall economy, yellow
for capital expenditures, orange for private consumption, and white
for inflation expectations. This allows us to distinguish patterns
of the variables.
Finally, we present three measures of topographic errors. We al-
ready looked at the first one, the quantization error, which is the
average distance between each variable and the closest neuron. To re-
iterate our quantization error is 0.5888693. The best-matching er-
ror is the average distance between the best matching unit and the
following, which is 1.568656. This error is in terms of coordinates
in the map. Similarly, the node distance error is the average distance
between all pairs of most similar codebook vectors, which is 1.387984.

383
 Figure 21
THE U-MATRIX

.    45

.  

384 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Figure 22
WEIGHT VECTORS

385
 B.3 Non-Linear Auto-Regressive Neural Networks Validation and Other Results
B.3.1 Lag Selection
Table 12
LAG STATISTICS ON ALL DATA, ONE STEP-AHEAD FORECASTS, SAMPLE OF 30

Countries Lags 1 2 3 4 5 6 7 8 9 10
Brazil mean 2.01 1.99 1.88 1.90 1.93 1.86 1.87 1.88 1.89 1.88
median 1.99 1.99 1.88 1.90 1.90 1.86 1.87 1.88 1.89 1.88
Canada mean 1.17 1.38 1.30 1.29 1.30 1.30 1.31 1.32 1.29 1.30
median 1.16 1.38 1.31 1.29 1.30 1.30 1.31 1.32 1.29 1.30
Switzerland mean 1.03 1.08 1.04 1.04 1.03 1.03 0.83 0.33 0.50 0.65
median 1.03 1.09 1.04 1.04 1.03 1.03 0.87 0.31 0.50 0.83

386 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Chile mean 2.22 2.09 2.08 2.19 2.10 2.04 2.07 2.10 2.04 2.09
median 2.22 2.09 2.05 2.23 2.17 1.95 2.19 2.15 2.04 2.10
Colombia mean 1.56 1.58 1.58 1.60 1.59 1.61 1.57 1.61 1.57 1.60
median 1.56 1.58 1.58 1.59 1.57 1.59 1.55 1.58 1.54 1.57
Czech R. mean 1.86 2.05 1.99 1.94 1.85 0.89 1.73 1.78 0.39 0.37
median 1.87 2.05 1.99 1.94 1.85 0.89 1.73 1.78 0.32 0.29
United K. mean 1.28 1.40 1.39 1.38 1.26 1.21 1.09 0.99 1.15 0.94
median 1.27 1.41 1.39 1.38 1.26 1.21 1.10 1.00 1.16 0.85
 Korea R. mean 1.36 1.47 1.45 1.46 1.45 1.35 1.36 1.30 1.31 1.32
median 1.36 1.47 1.45 1.46 1.45 1.35 1.36 1.30 1.29 1.31
Mexico mean 1.27 1.37 1.33 1.14 1.12 0.91 1.24 1.31 1.02 0.28
median 1.24 1.37 1.32 1.14 1.12 0.90 1.44 1.35 1.18 0.25
Norway mean 1.86 2.08 1.97 1.98 1.92 1.91 1.83 1.81 1.76 1.75
median 1.86 2.08 1.97 1.98 1.92 1.91 1.83 1.81 1.76 1.75
Hungary mean 1.64 1.86 1.78 1.68 1.72 1.63 1.52 1.51 1.64 1.60
median 1.64 1.86 1.78 1.68 1.73 1.63 1.52 1.51 1.64 1.59
Philippines mean 2.66 2.50 2.39 2.41 2.43 2.45 2.26 2.17 1.51 1.71
median 2.65 2.50 2.39 2.41 2.43 2.45 2.26 2.23 1.94 1.70
Poland mean 1.74 1.82 1.25 0.95 1.35 1.30 0.89 0.87 0.32 0.63
median 1.73 1.84 1.26 0.95 1.35 1.30 0.89 0.87 0.27 0.63
Sweden mean 1.25 1.16 1.11 1.13 1.11 1.12 1.07 1.07 1.09 1.06
median 1.24 1.16 1.11 1.13 1.11 1.12 1.07 1.07 1.09 1.06
Thailand mean 1.67 1.87 1.46 1.68 1.35 1.35 1.27 1.29 1.32 1.35
median 1.67 1.87 1.46 1.63 1.35 1.35 1.27 1.29 1.32 1.36
South A. mean 2.25 2.41 2.25 2.25 2.18 2.19 2.19 2.16 1.74 1.92
median 2.25 2.41 2.25 2.25 2.18 2.19 2.21 2.16 1.69 1.85

387
 Table 13
LAG STATISTICS ON TRAIN DATA, ONE STEP-AHEAD FORECAST, SAMPLE OF 30

Countries Lags 1 2 3 4 5 6 7 8 9 10
Brazil mean 2.04 1.98 1.88 1.90 1.94 1.86 1.87 1.89 1.89 1.86
median 2.03 1.98 1.88 1.90 1.90 1.86 1.87 1.89 1.89 1.86
Canada mean 1.10 1.34 1.28 1.26 1.27 1.27 1.28 1.30 1.24 1.26
median 1.07 1.34 1.28 1.26 1.27 1.27 1.28 1.30 1.24 1.26
Switzerland mean 1.01 1.07 1.05 1.04 1.03 1.03 0.80 0.29 0.46 0.64
median 1.01 1.07 1.05 1.04 1.03 1.03 0.85 0.27 0.46 0.83
Chile mean 2.04 2.03 2.02 2.15 2.04 1.96 1.99 2.00 1.94 2.01
median 2.03 2.03 1.98 2.18 2.11 1.87 2.11 2.07 1.94 2.02

388 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Colombia mean 1.27 1.46 1.46 1.48 1.48 1.49 1.45 1.45 1.41 1.44
median 1.26 1.46 1.46 1.48 1.46 1.47 1.43 1.43 1.38 1.41
Czech R. mean 1.96 2.17 2.11 2.05 1.94 0.91 1.78 1.83 0.22 0.26
median 1.97 2.17 2.10 2.05 1.94 0.91 1.78 1.83 0.15 0.19
United K. mean 1.31 1.45 1.43 1.42 1.29 1.25 1.12 0.98 1.18 0.95
median 1.31 1.46 1.43 1.42 1.29 1.25 1.13 1.01 1.18 0.85
Korea R. mean 1.28 1.44 1.40 1.41 1.39 1.27 1.26 1.23 1.23 1.24
median 1.28 1.44 1.40 1.41 1.39 1.27 1.26 1.23 1.22 1.23
 Mexico mean 1.34 1.45 1.40 1.20 1.18 0.97 1.31 1.40 1.07 0.20
median 1.32 1.45 1.39 1.20 1.18 0.95 1.53 1.44 1.27 0.22
Norway mean 1.91 2.14 2.00 2.01 1.95 1.90 1.81 1.79 1.75 1.75
median 1.90 2.14 2.00 2.01 1.95 1.90 1.81 1.79 1.75 1.75
Hungary mean 1.48 1.76 1.67 1.52 1.53 1.45 1.31 1.33 1.52 1.48
median 1.48 1.77 1.67 1.52 1.55 1.45 1.31 1.33 1.52 1.48
Philippines mean 2.59 2.37 2.26 2.29 2.31 2.32 2.15 2.03 1.26 1.52
median 2.58 2.37 2.26 2.29 2.31 2.32 2.15 2.10 1.74 1.53
Poland mean 1.79 1.90 1.27 0.94 1.34 1.32 0.90 0.87 0.01 0.00
median 1.78 1.91 1.29 0.94 1.34 1.32 0.90 0.87 0.01 0.00
Sweden mean 1.27 1.12 1.08 1.09 1.05 1.06 1.00 1.01 1.03 1.00
median 1.26 1.13 1.08 1.09 1.05 1.06 1.00 1.01 1.03 1.00
Thailand mean 1.67 1.95 1.50 1.75 1.39 1.38 1.29 1.31 1.35 1.38
median 1.67 1.95 1.50 1.69 1.39 1.38 1.29 1.31 1.35 1.40
South A. mean 2.22 2.33 2.13 2.13 2.02 2.03 1.99 1.94 1.46 1.66
median 2.21 2.33 2.13 2.13 2.02 2.03 2.03 1.94 1.41 1.54

389
B.3.2 Post-Training Analysis

Table 14
NEURAL NETWORKS RESULTS OF TRAINING PHASE
Total Effective Maximum Sum
number number of sum squared squared of Total
Countries of parameters of parameters parameters epoch

Brazil 31 2.88 2760 1.74 355

Canada 101 7.66 53 1.01 622

Chile 61 4.68 91.3 1.11 228

Colombia 71 5.02 72 0.72 1000

Czech Republic 81 31.17 61.6 20.96 314

Korea 41 2.99 280 1.10 1000

Mexico 81 20.71 61.7 9.91 114

Norway 31 2.96 2760 1.49 70

Switzerland 101 38.81 53.4 22.16 330

United Kingdom 81 10.20 64.7 3.39 245

Hungary 121 14.04 46 2.9722 889

Philippines 31 2.04 2760 1.30 108

Poland 91 19.48 58.2 7.43 156

Sweden 31 2.75 2760 1.53 484

Thailand 61 9.16 91.3 4.09 298

South A. 31 2.64 2760 1.54 502

390 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Best Error Input-error Correlation coefficient
Countries epoch Autocorrelation Correlation Training R Testing R All R

Brazil 2 1 0 0.605 0.877 0.632

Canada 99 1 0 0.570 0.334 0.551

Chile 56 1 0 0.702 -0.049 0.678

Colombia 429 1 0 0.445 0.560 0.463

Czech Republic 253 1 0 0.885 0.607 0.884

Korea 1000 1 0 0.523 -0.464 0.554

Mexico 64 1 0 0.875 0.474 0.879

Norway 4 1 0 0.641 -0.041 0.640

Switzerland 240 1 0 0.935 0.759 0.921

United Kingdom 77 1 0 0.740 0.473 0.743

Hungary 103 1 0 0.820 -0.157 0.826

Philippines 12 0 0 0.678 0.077 0.652

Poland 129 1 0 0.887 0.605 0.895

Sweden 9 1 0 0.741 0.108 0.746

Thailand 151 1 0 0.674 0.181 0.664

South A. 8 0 0 0.744 0.545 0.739

391
 B.3.3 MSE Evaluation

Table 15
NEURAL NETWORK SIMULATIONS STATISTICS BY DATASETS, SAMPLE OF 1,000

Brazil Korea
Training Testing Training Testing
All data set set All data set set
mean 2.01 2.05 1.65 1.47 1.44 1.86
median 2.00 2.04 1.61 1.47 1.44 1.86
std 0.04 0.03 0.18 0.01 0.01 0.03
maximum 2.09 2.09 2.11 1.52 1.44 2.51
minimum 1.92 1.97 1.24 1.36 1.34 1.62

Canada Mexico
Training Testing Training Testing
All data set set All data set set
mean 1.33 1.31 1.52 0.97 1.03 0.34
median 1.32 1.30 1.52 0.90 0.95 0.30
std 0.06 0.06 0.01 0.24 0.24 0.17
maximum 2.09 2.09 2.11 3.91 4.00 2.94
minimum 1.92 1.97 1.24 0.90 0.95 0.30

Chile Norway
Training Testing Training Testing
All data set set All data set set
mean 2.21 2.17 2.69 1.87 1.91 1.41
median 2.23 2.18 2.68 1.86 1.90 1.42
std 0.06 0.07 0.03 0.04 0.04 0.03
maximum 1.33 1.36 1.01 2.06 2.12 1.51
minimum 0.55 0.54 0.67 1.82 1.86 1.25

Colombia Switzerland
Training Testing Training Testing
All data set set All data set set
mean 1.59 1.48 2.83 0.31 0.27 0.78
median 1.57 1.46 2.78 0.31 0.27 0.78
std 0.08 0.07 0.21 0.00 0.00 0.00
maximum 1.93 1.76 3.69 0.31 0.27 0.78
minimum 1.57 1.46 2.77 0.31 0.27 0.78

392 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Czech Republic United Kingdom
Training Testing Training Testing
All data set set All data set set
mean 0.90 0.92 0.69 1.21 1.25 0.82
median 0.89 0.91 0.67 1.21 1.25 0.82
std 0.08 0.08 0.08 0.00 0.00 0.00
maximun 1.21 1.25 0.82 1.21 1.25 0.82
minimum 1.21 1.25 0.82 1.21 1.25 0.82

Hungary Philippines
Training Testing Training Testing
All data set set All data set set
mean 1.59 1.47 2.85 2.66 2.60 3.42
median 1.59 1.48 2.75 2.66 2.59 3.43
std 0.11 0.17 0.54 0.04 0.05 0.07
maximun 1.73 1.58 7.60 2.81 2.74 3.88
minimum 0.68 0.00 2.74 2.59 2.51 2.98

Poland Sweden
Training Testing Training Testing
All data set set All data set set
mean 0.89 0.90 0.72 1.25 1.26 1.14
median 0.89 0.90 0.72 1.24 1.25 1.15
std 0.01 0.01 0.01 0.03 0.04 0.09
maximun 1.02 1.04 0.80 1.35 1.37 1.31
minimum 0.88 0.90 0.66 1.21 1.21 0.86

Thailand South Africa

Training Testing Training Testing
All data set set All data set set
mean 1.69 1.76 0.90 2.27 2.23 2.64
median 1.63 1.69 0.91 2.25 2.22 2.63
std 0.09 0.10 0.02 0.07 0.07 0.10
maximum 1.82 1.91 0.91 2.64 2.58 3.35
minimum 1.63 1.69 0.86 2.22 2.18 2.43

393
 1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
9
01/07/91 01/07/91 01/07/91
01/02/92 01/02/92 01/02/92
01/09/92 01/09/92 01/09/92
01/04/93 01/04/93 01/04/93

Observed
01/11/93 01/11/93 01/11/93
01/06/94 01/06/94 01/06/94
01/01/95 01/01/95 01/01/95
01/08/95 01/08/95 01/08/95
01/03/96 01/03/96 01/03/96
01/10/96 01/10/96 01/10/96
01/05/97 01/05/97 01/05/97
01/12/97 01/12/97 01/12/97
01/07/98 01/07/98 01/07/98
01/02/99 01/02/99 01/02/99
01/09/99 01/09/99 01/09/99
01/04/00 01/04/00 01/04/00
01/11/00 01/11/00 01/11/00

One-step-ahead Training
01/06/01 01/06/01 01/06/01
01/01/02 01/01/02 01/01/02
01/08/02 01/08/02 01/08/02
01/03/03 01/03/03 01/03/03
01/10/03 01/10/03 01/10/03
B.3.4 Results, Other Countries

01/05/04 01/05/04 01/05/04

01/12/04 01/12/04 01/12/04

. 
01/07/05 01/07/05
. 

01/07/05
Figure 23

01/02/06 01/02/06 01/02/06

01/09/06 01/09/06 01/09/06

.  
01/04/07 01/04/07 01/04/07

394 H. M. Zárate-Solano, D. R. Zapata-Sanabria

01/11/07 01/11/07

One-step-ahead Testing
01/11/07
01/06/08 01/06/08 01/06/08
01/01/09 01/01/09 01/01/09
01/08/09 01/08/09 01/08/09
01/03/10 01/03/10 01/03/10
01/10/10 01/10/10 01/10/10
01/05/11 01/05/11 01/05/11
01/12/11 01/12/11 01/12/11
01/07/12 01/07/12 01/07/12
01/02/13 01/02/13

Two-step-ahead
01/02/13
01/09/13 01/09/13 01/09/13
01/04/14 01/04/14 01/04/14
01/11/14 01/11/14 01/11/14
01/06/15 01/06/15 01/06/15
01/01/16 01/01/16 01/01/16
01/08/16 01/08/16 01/08/16
01/03/17 01/03/17 01/03/17
COUNTRIES’ INFLATION RATE FORECAST FOR THE NEXT SIX MONTHS BY NAR-NN

01/10/17 01/10/17

Out of sample
01/10/17
 1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
01/07/91 01/07/91 01/07/91
01/02/92 01/02/92 01/02/92
01/09/92 01/09/92 01/09/92
01/04/93 01/04/93 01/04/93

Observed
01/11/93 01/11/93 01/11/93
01/06/94 01/06/94 01/06/94
01/01/95 01/01/95 01/01/95
01/08/95 01/08/95 01/08/95
01/03/96 01/03/96 01/03/96
01/10/96 01/10/96 01/10/96
01/05/97 01/05/97 01/05/97
01/12/97 01/12/97 01/12/97
01/07/98 01/07/98 01/07/98
01/02/99 01/02/99 01/02/99
01/09/99 01/09/99 01/09/99
01/04/00 01/04/00 01/04/00
01/11/00

One-step-ahead Training
01/11/00 01/11/00
01/06/01 01/06/01 01/06/01
01/01/02 01/01/02 01/01/02
01/08/02 01/08/02 01/08/02
01/03/03 01/03/03 01/03/03
01/10/03 01/10/03 01/10/03
01/05/04 01/05/04 01/05/04
01/12/04 01/12/04 01/12/04

. 
01/07/05 01/07/05

. 
01/07/05
. 

01/02/06 01/02/06 01/02/06

01/09/06 01/09/06 01/09/06
Figure 23 (cont.)

01/04/07 01/04/07 01/04/07

One-step-ahead Testing
01/11/07 01/11/07 01/11/07
01/06/08 01/06/08 01/06/08
01/01/09 01/01/09 01/01/09
01/08/09 01/08/09 01/08/09
01/03/10 01/03/10 01/03/10
01/10/10 01/10/10 01/10/10
01/05/11 01/05/11 01/05/11
01/12/11 01/12/11 01/12/11
01/07/12 01/07/12 01/07/12

Two-step-ahead
01/02/13 01/02/13 01/02/13
01/09/13 01/09/13 01/09/13
01/04/14 01/04/14 01/04/14
01/11/14 01/11/14 01/11/14
01/06/15 01/06/15 01/06/15
01/01/16 01/01/16 01/01/16

395
01/08/16 01/08/16 01/08/16
01/03/17 01/03/17 01/03/17
COUNTRIES’ INFLATION RATE FORECAST FOR THE NEXT SIX MONTHS BY NAR-NN

Out of sample
01/10/17 01/10/17 01/10/17
 1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
9

1
2
3
4
5
6
7
8
9
01/07/91 01/07/91 01/07/91
01/02/92 01/02/92 01/02/92
01/09/92 01/09/92 01/09/92
01/04/93 01/04/93 01/04/93

Observed
01/11/93 01/11/93 01/11/93
01/06/94 01/06/94 01/06/94
01/01/95 01/01/95 01/01/95
01/08/95 01/08/95 01/08/95
01/03/96 01/03/96 01/03/96
01/10/96 01/10/96 01/10/96
01/05/97 01/05/97 01/05/97
01/12/97 01/12/97 01/12/97
01/07/98 01/07/98 01/07/98
01/02/99 01/02/99 01/02/99
01/09/99 01/09/99 01/09/99
01/04/00 01/04/00 01/04/00

One-step-ahead Training
01/11/00 01/11/00 01/11/00
01/06/01 01/06/01 01/06/01
01/01/02 01/01/02 01/01/02
01/08/02 01/08/02 01/08/02
01/03/03 01/03/03 01/03/03
01/10/03 01/10/03 01/10/03
01/05/04 01/05/04 01/05/04
01/12/04 01/12/04 01/12/04
01/07/05 01/07/05 01/07/05

. 
01/02/06 01/02/06 01/02/06
. 

.  
01/09/06 01/09/06 01/09/06
Figure 23 (cont.)

01/04/07 01/04/07 01/04/07

396 H. M. Zárate-Solano, D. R. Zapata-Sanabria

One-step-ahead Testing
01/11/07 01/11/07 01/11/07
01/06/08 01/06/08 01/06/08
01/01/09 01/01/09 01/01/09
01/08/09 01/08/09 01/08/09
01/03/10 01/03/10 01/03/10
01/10/10 01/10/10 01/10/10
01/05/11 01/05/11 01/05/11
01/12/11 01/12/11 01/12/11
01/07/12 01/07/12 01/07/12

Two-step-ahead
01/02/13 01/02/13 01/02/13
01/09/13 01/09/13 01/09/13
01/04/14 01/04/14 01/04/14
01/11/14 01/11/14 01/11/14
01/06/15 01/06/15 01/06/15
01/01/16 01/01/16 01/01/16
01/08/16 01/08/16 01/08/16
01/03/17 01/03/17 01/03/17
COUNTRIES’ INFLATION RATE FORECAST FOR THE NEXT SIX MONTHS BY NAR-NN

Out of sample
01/10/17 01/10/17 01/10/17
 1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
1
2
3
4
5
6
7
8
9
01/07/91 01/07/91 01/07/91
01/02/92 01/02/92 01/02/92
01/09/92 01/09/92 01/09/92
01/04/93 01/04/93 01/04/93

Observed
01/11/93 01/11/93 01/11/93
01/06/94 01/06/94 01/06/94
01/01/95 01/01/95 01/01/95
01/08/95 01/08/95 01/08/95
01/03/96 01/03/96 01/03/96
01/10/96 01/10/96 01/10/96
01/05/97 01/05/97 01/05/97
01/12/97 01/12/97 01/12/97
01/07/98 01/07/98 01/07/98
01/02/99 01/02/99 01/02/99
01/09/99 01/09/99 01/09/99
01/04/00 01/04/00 01/04/00
01/11/00

One-step-ahead Training
01/11/00 01/11/00
01/06/01 01/06/01 01/06/01
01/01/02 01/01/02 01/01/02
01/08/02 01/08/02 01/08/02
01/03/03 01/03/03 01/03/03
01/10/03 01/10/03 01/10/03
01/05/04 01/05/04 01/05/04
01/12/04 01/12/04 01/12/04
01/07/05 01/07/05
. 

01/07/05

. 
01/02/06 01/02/06 .  01/02/06
01/09/06 01/09/06 01/09/06
Figure 23 (cont.)

01/04/07 01/04/07 01/04/07

One-step-ahead Testing
01/11/07 01/11/07 01/11/07
01/06/08 01/06/08 01/06/08
01/01/09 01/01/09 01/01/09
01/08/09 01/08/09 01/08/09
01/03/10 01/03/10 01/03/10
01/10/10 01/10/10 01/10/10
01/05/11 01/05/11 01/05/11
01/12/11 01/12/11 01/12/11
01/07/12 01/07/12 01/07/12

Two-step-ahead
01/02/13 01/02/13 01/02/13
01/09/13 01/09/13 01/09/13
01/04/14 01/04/14 01/04/14
01/11/14 01/11/14 01/11/14
01/06/15 01/06/15 01/06/15
01/01/16 01/01/16 01/01/16

397
01/08/16 01/08/16 01/08/16
01/03/17 01/03/17 01/03/17
COUNTRIES’ INFLATION RATE FORECAST FOR THE NEXT SIX MONTHS BY NAR-NN

Out of sample
01/10/17 01/10/17 01/10/17
B.4 arima
In the arima modeling, various tests were performed before mod-
eling the series in order to understand the generating data process
and find the best (p,d,q)(P,D,Q) order suit to the series. We began
to perform the Augmented Dickey-Fuller (ADF) test (see Dickey
and Fuller, 1981) and the Kwiatkowski, Phillips, Schmidt, and Shin
(KPSS) test (see Kwiatkowski et al., 1992 to find the differentiation
order (Table 16). In the Dickey-Fuller test, we started including
the trend and constant over the regression for which all the series
rejected the null hypothesis of the unit root. For the KPSS test, like
the ADF test, we included the trend and constant terms and almost
all the series did not reject the null hypothesis of stationary except
for Switzerland and Norway, where the Switzerland series became
stationary after the first 8 observations were excluded from the tests.
To find the seasonal difference order, the Canova-Hansen test (see
Canova and Hansen, 1995) was implemented, which has a null hy-
pothesis of no unit roots at seasonal frequencies. This test comple-
ments the HEGGY test of seasonal unit roots.
Once the difference orders were determined and the respective
transformations were applied, such as applying logarithms if nec-
essary, we proceed to explore the autocorrelation function, partial
autocorrelation, extended autocorrelation function, and informa-
tion criterion AIC and BIC. We used these factors to find the autore-
gressive and moving average coefficients. A group of possible models
were tested on each country, for which the most suitable model had to
accomplish five conditions:
• Low BIC, AICc, and rmse
• coefficients statistically different to zero.
• the residuals should be uncorrelated through time.
• the cross-correlation function between the predicted errors
and the observed time series should be close to zero.

• The high order closest model should fail in comparison.

Then, after we found the best arima model possible, we forecast
one step ahead and two step ahead on the testing set and calculate
the respective MSE to compare with the nar-nn Model.

398 H. M. Zárate-Solano, D. R. Zapata-Sanabria

 Table 16
UNIT ROOT, STATIONARITY TESTS AND MODEL IDENTIFICATION

KPSS
ADF t-Stat Stat (p,d,q)(P,D,Q) order
Brazil 5.871 0.089 (1,0,0)

Canada 5.357 0.040 (1,0,0)

Switzerland 4.085 0.188 (2,0,1)

Chile 3.377 0.143 (1,1,1)

Colombia* 4.892 0.059 (1,0,0)

Czech Republic* 4.431 0.086 (1,1,1)

United Kingdom 5.294 0.069 (1,0,0)(1,0,0)

Korea 4.997 0.065 (1,0,1)

Mexico 5.179 0.056 (1,1,1)

Norway 4.846 0.150 (1,1,1)

Hungary* 4.022 0.089 (1,0,0)

Philippines* 6.370 0.077 (1,0,0)

Poland* 3.537 0.122 (0,1,2)

Sweden 5.545 0.065 (2,0,0)

Thailand* 4.928 0.045 (1,0,0)

South Africa 5.515 0.044 (1,0,0)(1,0,0)

Test critical values:

1% level -4.04 0.216

5% level -3.45 0.146

10% level -3.15 0.119

*Log transformation

399
References
Box, G.E.P. et al. 2016. Time Series Analysis: Forecasting and Control.
New York, United States: Wiley.
Canova, F., and B.E. Hansen. 1995. “Are Seasonal Patterns Constant
over Time? A Test for Seasonal Stability.” Journal of Business
Economic Statistics 13(3): 237-252. doi:10.1080/07350015
.1995.10524598.
Carvalho, C., and F. Nechio. 2014. “Do People Understand Monetary
Policy?” Journal of Monetary Economics 66: 108-123.
Claveria, O., E. Monte and S. Torra. 2016. “A Self-Organizing Map
Analysis of Survey-Based Agents’ Expectations before Im-
pending Shocks for Model Selection: The Case of the 2008
Financial Crisis.” International Economics 146: 40-58. doi:
10.1016/j.inteco.2015.11.003.
Coibion, O. 2012. “Are the Effects of Monetary Policy Shocks Big or
Small?” American Economic Journal: Macroeconomics 4(2): 1-32.
Crump, R.K. et al. 2015. “Subjective Intertemporal Substitution.”
Federal Reserve Bank of New York Staff Reports 734. New
York, United States: Federal Reserve Bank of New York.
Dickey, D.A, and W.A. Fuller. 1981. “Likelihood Ratio Statistics for
Autoregressive Time Series with a Unit Root.” Econometrica
49(4): 10-57. doi:10.2307/1912517.
Fildes, R., and H. Stekler. 2002. “The State of Macroeconom-
ic Forecasting.” Journal of Macroeconomics 24(4): 435-468.
doi:10.1016/s0164-0704(02)00055-1.
Hagan, M.T., H.B. Demuth and M.H. Beale. 2002. Neural Network
Toolbox: User’s Guide. Natick, United States: The MathWorks,
Inc.
Hagan, M.T. et al. 2014. Neural Network Design. Second edition.
Boulder, United States: University of Colorado.
Hamella, S., and H. Haupt. 2007. “Suitability of wes Data for
Forecasting Inflation.” In: G. Goldrian, editor. Handbook
of Survey-Based Business Cycle Analysis. Cheltenham, United
Kingdom: Edward Elgar.

400 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Henzel, S., and T. Wollmershäuserab. 2008. “The New Keynesian
Phillips Curve and the Role of Expectations: Evidence from
the CesIfo World Economic Survey.” Economic Modelling 25(5):
811-832. doi:10.1016/j.econmod.2007.11.010.
Jain, A., J. Mao and K. Mohiuddin. 1996. “Artificial Neural Networks:
A Tutorial.” Computer 29(3): 31-44. doi:10.1109/2.485891.
Kock, A.B., and T. Teräsvirta. 2016. “Forecasting Macroeconomic
Variables Using Neural Network Models and Three Auto-
mated Model Selection Techniques.” Econometric Reviews
35(8-10): 1753-1779. doi: 10.1080/07474938.2015.1035163.
Kohonen, T. 2001. Self-Organizing Maps. New York, United States:
Springer.
Kwiatkowski, D. et al. 1992. “Testing the Null Hypothesis of Station-
arity against the Alternative of a Unit Root.” Journal of Econo-
metrics 54(1-3): 159-178. doi: 10.1016/0304-4076(92)90104-y.
Lynn, S. 2014. “Self-Organising Maps for Customer Segmentation
Using R.” https://www.r-bloggers.com/self-organising-
maps-for-customer-segmentation-using-r
Marcellino, M. 2004. “Forecasting emu Macroeconomic Variables.”
International Journal of Forecasting 20: 359-72. doi:10.1016/j.
ijforecast.2003.09.003
Ruiz, L. et al. 2016. “An Application of Non-Linear Autoregressive
Neural Networks to Predict Energy Consumption in Public
Buildings.” Energies 9(9): 1-21. 684. doi:10.3390/en9090684.
Stangl, A. 2007a. “World Economic Survey.” In: G. Goldrian, editor.
Handbook of Survey- Based Business Cycle Analysis. Cheltenham,
United Kingdom: Edward Elgar.
Stangl, A. 2007b. “European Data Watch: Ifo World Economic
Survey Micro Data.” Schmollers Jahrbuch: Journal of Applied
Social Science Studies / Zeitschrift fr Wirtschafts- und Sozialwis-
senschaften 127(3): 487–496. http://EconPapers.repec.org/
RePEc:aeq:aeqsjb:v127_y2007_i3_q3_p487-496
Stock, J.H., and M.W. Watson. 1998. “A Comparison of Linear and
Nonlinear Univariate Models for Forecasting Macroeconomic
Time Series.” NBER Working Paper 6607. Cambridge, United
States: National Bureau of Economic Research
Theil, H. et al. 1975. Applied Economic Forecasting. Amsterdam, The
Netherlands: North- Holland.

401
Titterington, D.M. 2004. “Bayesian Methods for Neu-
ral Networks.” Statistical Science 19(1): 128–139.
doi:10.1214/088342304000000099.
Wehrens, R., and L. Buydens. 2007. “Self- and Super-Organising
Maps in R: The Kohonen Package.” Journal of Statistical Software
21 (5). URL http://www.jstatsoft.org/v21/i05
White, H. 1989. “Learning in Artificial Neural Networks: A Statistical
Perspective.” Neural Computation 1 (4): 425-464. doi:10.1162/
neco.1989.1.4.425.
Yuriy, Y., and C. Olivier. 2015. “Is the Phillips Curve Alive and
Well after All? Inflation Expectations and the Missing Dis-
inflation.” American Economic Journal: Macroeconomics 7(1):
197-232. https://ideas.repec.org/a/aea/aejmac/
v7y2015i1p197- 232.html
Zhang, G., B.E. Patuwo and M.Y. Hu. 1998. “Forecasting with Arti-
ficial Neural Networks: “The State of the Art.” International
Journal of Forecasting 14(1): 35–62. doi: http://doi.org/10.
1016/S0169-2070(97)00044-7. URL http://www.sciencedi-
rect.com/science/article/pii/S0169207097000447.

402 H. M. Zárate-Solano, D. R. Zapata-Sanabria

Inflation Expectations and Its
Relation with Economic Policy
Did the Introduction of Inflation
Targeting Represent a
Regime Switch of Monetary
Policy in Latin America?
Sebastián Cadavid Sánchez
Alberto Ortiz Bolaños

Abstract
In the 1990s, after experiencing high levels of inflation, several countries
in Latin America passed constitutional amendments providing greater
autonomy to their central banks. A few years later, many central banks
increased their exchange rate flexibility and later adopted inflation targeting
frameworks. These institutional changes coincided with sharp reductions
in inflation and its variability. In this paper, we ask if the observed reduction
of inflation is possibly related to changes in monetary policy. To answer this
question, we build and estimate a Markov-Switching Dsge model for an
open economy with monetary factors for Brazil, Chile, Colombia, Mexico,
and Peru, all of whom formally adopted inflation targeting regimes between
1999 and 2002. Regimes are classified according to their relative weights
of inflation in an interest rate reaction function. Although ex-ante these
regimes need not be associated with the introduction of the inflation targeting
framework, the coincidence of a regime switch with a more responsive interest
rate - inflation relationship is striking. Furthermore, the Markov-Switching
Dsge model allows us to generate counterfactuals of what could have
happened if the observed change towards a more aggressive fight against
inflation had not taken place. In general, we observe that if monetary
policy had remained dovish, these countries would have experienced higher
and more variable levels of inflation and more pronounced variations in Gdp
with small gains in average economic growth. Therefore, we conclude that

The authors thank Junior Maih for making his rise toolbox for the solution and esti-
mation of Markov Switching Rational Expectations models available and for patiently
answering all of our questions. The views expressed in this presentation are those
of the author, and not necessarily those of cemla or egade Business School of Tec-
nológico de Monterrey.

405
the introduction of inflation targeting represented a favorable regime switch
in the implementation of monetary policy in Latin America.
Keywords: Monetary policy, inflation, Markov-switching Dsge, Bayesian
Maximum Likelihood methods.
jel: E31, E37, E52, E58, C11.

1. INTRODUCTION

B
eginning in the late 1980s, many countries around the world en-
acted new central banking legislation to grant more autonomy
to their monetary authorities. For example, see Figure 1,
which uses a sample of indexes of central bank independence from
182 countries since 1970, produced by Garriga (2016). Figure 1 shows
a sharp increase in the number of reforms toward increased central
bank independence in the 1990s. This shift came in response to the
traumatic inflationary and hyper-inflationary episodes experienced
in the previous decades, and it was reinforced by evidence showing
that “central bank independence promotes price stability” without
“measurable impact on real economic performance” (e.g., Alesina
and Summers (1993)).
In Latin America, starting with Venezuela in 1974, several coun-
tries had reforms to strengthen the independence of their central
banks1. In some countries, and for different reasons (from depletion
of reserves to the desire to gain greater control of monetary policy),
many central banks increased their exchange rate flexibility. The pro-
cess continued with the adoption of inflation targeting frameworks
to direct monetary policy. These institutional changes coincided with

1
According to Garriga (2016), since 1970, countries that took positive
reforms towards independence were the following: Venezuela in 1974;
Chile in 1975; Haiti in 1979; Mexico in 1985; Brazil in 1988; Chile
in 1989; El Salvador in 1991; Argentina, Colombia, Ecuador, Nica-
ragua, Peru, and Venezuela in 1992; Mexico in 1993; Bolivia, Costa
Rica, Paraguay, and Uruguay in 1995; Honduras in 1996; Cuba in 1997;
Nicaragua and Venezuela in 1999; El Salvador in 2000; Guatemala
and the Dominican Republic in 2002; and Uruguay in 2008 and 2010.
Meanwhile, negative reforms hindering Central Bank independence
include the following: Argentina and El Salvador in 1973, Panama
in 1975, El Salvador in 1982, Uruguay in 1997, Venezuela in 2001,
Argentina in 2003, Ecuador in 2008, Venezuela in 2009, Nicaragua
in 2010, and Argentina in 2012.

406 S. Cadavid, A. Ortiz

 Figure 1
REFORMS TO THE INDEPENDENCE
OF CENTRAL BANKS AROUND THE WORLD

15

10
Number of reforms

−5

1970 1980 1990 2000 2010

Year

More la More Rest of the World

Less la Less Rest of the World

sharp reductions of inflation and its variability. Table 1 summarizes

the average inflation for each decade together with the years when
positive reforms toward central bank independence were enacted,
greater exchange rate flexibility was pursued, and inflation target-
ing was introduced. The selected countries for this analysis are Bra-
zil, Chile, Colombia, Mexico, and Peru, which were early adopters
of inflation targeting in Latin America between 1999 and 2002.
Although common sense provides a reason to believe that there
could be a relation between institutional changes and inflation reduc-
tion, to the best of our knowledge, there is no quantitative evidence
measuring if and how these changes determined inflation. In this
paper, we provide this evidence by analyzing a Markov-Switching
Dynamic Stochastic General Equilibrium (ms-dsge) model for an
open economy with monetary factors estimated for Brazil, Chile, Co-
lombia, Mexico, and Peru. Regimes are classified according to their

Regime Switch of Monetary Policy in Latin America 407

 Table 1
INFLATION AND CENTRAL BANKS CHANGES IN
SELECTED COUNTRIES OF LATIN AMERICA

Year of
Positive Inflation
reforms Exchange Targeting
Average 1980- 1990- 2000- 2010- towards rate introduc-
inflation 1989 1999 2009 2015 independence flexibility tion
Brazil 121.7 147.1 6.6 6.2 1988 1999 1999
Chile 19.9 11.8 3.5 3 1975 and 1999 1999
1989
Colombia 20.8 19.9 6.1 3.1 1992 1999 1999
Mexico 53.1 18.3 5.1 3.6 1985 and 1995 2001
1993
Peru 111 78.5 2.6 3 1992 2002 2002

relative weights of inflation in an interest rate reaction function. Al-

though ex-ante these regimes need not be associated with the intro-
duction of the inflation targeting framework, the coincidence of a
more responsive monetary policy with inflation targeting is strik-
ing. Furthermore, the model allows us to generate counterfactuals
of what could have happened if the observed change toward a more
aggressive fight against inflation would not have taken place. In gen-
eral, we observe that if monetary policy had remained dovish, these
countries would have experienced higher and more variable levels
of inflation and more pronounced variations in gdp with small gains
in average economic growth. Therefore, we conclude that the intro-
duction of inflation targeting represented a favorable regime switch
in the regulation of monetary policy in Latin America.
The rest of the paper is organized as follows. Section 2 pres-
ents a Markov-Switching open-economy dsge model with mone-
tary factors that will serve as the theoretical basis used to perform
our analysis. Section 3 describes the tools used to solve and esti-
mate the Markov-switching dsge model. Section 4 presents results
for the five countries discussed. Specifically, (4.1) displays the prob-
abilities of the high inflation responses and high volatility regimes;

408 S. Cadavid, A. Ortiz

(4.2) reports the parameter estimates; (4.3) shows the model’s impulse
response functions for the high and low inflation response regimes
to analyze the mechanisms; and (4.4) counterfactual simulated vari-
ables under the high and low inflation response regimes to analyze
what could have happened during the sample period if monetary pol-
icy had been conducted differently, together with tables summarizing
the average standard deviation and coefficient of variation of the ob-
served variables and the hypothetical series generated in the counter-
factuals. Section 5 concludes.

2. MODEL

Our model is based on the monetary open economy model presented

by Gali and Monacelli (2005) and later estimated for the Common-
wealth countries by Lubik and Schorfheide (2007) and for a large set of
emerging market countries by Ortiz and Sturzenegger (2007). In es-
sence the economy is summarized by the following three equations:
an open economy Investment-Savings (is) curve, an open economy
Phillips curve and an interest rate rule.
To capture potential regime changes, we specify a Markov-switching
Dsge model where we allow for changes in the parameters associated
with the monetary authority reaction function and the price formation
process, and use a state variable ξ sp to denote the structural parameters
sp regime at time t. To allow for regime changes in the stochastic vola-
tilities we model a second, independent, Markov-Switching process
and use a state variable ξvo to distinguish the volatility vo regime at time t.
In log linearized form, the open economy is -curve is:

2.1

where yt denotes aggregate output, R t nominal interest rate, π t cpi in-

flation, at is the growth rate of a non-stationary technology process At ,
qt terms of trade, defined as the relative price of exports in terms of im-
ports, and yt* world output. E t denotes the conditional expectation
operator. The parameter τ represents the elasticity of inter-temporal

Regime Switch of Monetary Policy in Latin America 409

substitution and α is the import share.2 Technology follows an exog-
enous process: where ρa is
the autoregressive coefficient and σa,ξvo is the standard deviation of the
stochastic volatility of the technology innovations ε a,t , whose ξvo sub-
script denotes that it is allowed to change across regimes at time t.
The same convention in notation follows for the other exogenous
processes as world output yt* that is treated as an unobservable and is
assumed to follow the process
In order to guarantee stationarity of the model, all real variables
are expressed in terms of percentage deviations from At .
The log-linear version of the open economy Phillips curve is:

2.2

where is potential output in the absence of nom-

inal rigidities. β represents the discount factor, χ p is the degree
of lagged price inflation, κ is the structural parameter associated
to the Phillips curve and the ξ sp subscript indicates that these param-
eters are allowed to change across regimes at time t.
The log-linear version of the interest rate rule is given by:

2.3

where et is the nominal effective exchange rate, defined as the price

of domestic currency in terms of foreign currency. The parameter ρR
captures the degree of interest rate smoothing, while ψπ ,ψy and ψΔe
capture the sensitivity of the interest rate with respect to inflation,
output deviation from its steady-state and nominal exchange rate

2
The equation reduces to the closed economy variant when α  = 0.

410 S. Cadavid, A. Ortiz

depreciation, ∆et , respectively. The ξ sp subscript indicates that these
parameters are allowed to change across regimes at time t. σR,ξvo is the
standard deviation of the stochastic volatility of the interest rate
εR ,t ∼ N (0,1), whose ξvo subscript denotes that it is allowed to change
across regimes at time t.
The exchange rate is introduced via Cpi inflation according to:

2.4

where i s a wo r l d i n f l a t io n s h o c k w h i c h i s t r e a t e d
as an unobservable and is assumed to follow an exogenous process:
Terms of trade, in turn, are as-
sumed to follow a law of motion for their growth rate:

2.5

with Equations (2.1) to (2.5), plus the exogenous pro-

cesses for technology, world output and world inflation, constitute
the whole model.

3. SOLUTION AND ESTIMATION OF THE

MARKOV-SWITCHING DSGE MODEL

The dsge system with constant parameters has the following ma-
trix form:

3.1

where Γo , Γ1, Θ and matrices contain the model’s parameters.

xt stands for the vector of endogenous variables,3 Zt is the
vector of exogenous processes and η t corresponds to the

′
3
with X t = ⎡⎢ yt πt Rt Δqt Δet πt* yt*at ⎥⎤ .
⎣ ⎦

Regime Switch of Monetary Policy in Latin America 411

disturbances vector. The conditions for existence and uniqueness
of the solution (3.1) depend on the generalized eigenvalues of the
system’s matrices (Farmer et al., 2008).
Using the solution algorithm proposed by Sims (2002) or Schmitt-
Grohé and Uribe (2003) the unique solution for the system (3.2)
is combined with an observation equation:

3.2

3.3

where stands for the parameters of the model, Ytobs are the ob-
served variables,4 and M provides the policy function for the observ-
ables. Following Bianchi and Ilut (2017), we introduce the possibility
of regime change for the structural parameters and the volatilities
through two Markov chains, ξ sp and ξvo . The former denotes the un-
observed regime associated with the monetary parameters subject
to regime shifts and takes on discrete values 5
and the latter
stands for the shock volatilities, assumes discrete values, 6

and evolves independently of sp.

Both state variables sp and vo are assumed to follow a first-order
Markov chain with the following transition matrices, respectively:

3.4

4
gdp growth, inflation rate, interest rate, change in the terms of trade
and nominal depreciation.
5
Where 1 and 2 are the high and low response to inflation regimes
⎛ ⎞
⎜⎜i.e ψ sp=1 > ψ sp=2 ⎟⎟, respectively.
⎝ π,ξt π,ξt ⎟⎠
6
Where 1 and 2 are the low and high volatility regimes. In order to define
the high volatility regime, we included into the model the following
restriction: σa,ξ vol=1 < σa,ξ vol=2 .

412 S. Cadavid, A. Ortiz

where Hij =p (spt =j|spt−1=i), for i, j=1, 2, and Qij =p(vot =j | vot−1 = i) for i,
j=1, 2. Then, Hij stands for the probability of being in regime j at t giv-
en that one was in regime i. The analysis is symmetric for Qij .
The Markov switching system can be cast in a state-space form
by collecting all the endogenous variables in a vector Xt and all the
exogenous variables in a vector Zt :

3.5

3.6

where the matrices and are

functions of the model parameters. is the covariance matrix

of the shocks,7 which depends on the unobserved state ξvo , controlled

by the transition matrix Q. Therefore, note that, in contrast with
(3.1), (3.5) has a presence of unobserved variables and unobserved
Markov states of the Markov chains.
There are several studies in the ms-dsge literature that an-
alyze the technical aspects of solving this state-space system
(Farmer et al. (2008, 2011); Foerster et al. (2014); Maih (2015) 8
and Cho (2016)), in the sense that solution algorithms developed
for solving dsge models with fixed parameters (e.g. Sims (2002)
and Schmitt-Grohé and Uribe (2003)) are unsuitable. To solve
the system we use the Newton methods developed in Maih (2015),
which expand on the method proposed by Farmer et al. (2011) and con-
centrates on minimum state variable solutions (msv) of the form:

7
( ) ⎛
⎝ q,ξt a,ξt R,ξt y ,ξt π ,ξt ⎟
⎞
Where: Σ ξ vo = diag ⎜⎜σ vo ,σ vo ,σ vo ,σ * vo ,σ * vo ⎟⎟.
⎠
8
The routines used for the computations were implemented using rise,
an object-oriented Matlab toolbox for solving and estimating Markov
switching rational expectation models, developed by Junior Maih.

Regime Switch of Monetary Policy in Latin America 413

 3.7

Where θsp and θvo are the switching parameters controlled by

and respectively.
The complete state form of the model combines (3.7) with the mea-
surement equations (3.8):

3.8

where:

The presence of unobserved dsge states Xt and unobserved pa-

rameters (controlled by the Markov chains), implies that the stan-
dard Kalman filter cannot be used to compute the likelihood. So,
in correspondence with Bianchi and Ilut (2017) we use the Kim et al.
(1999) filter.
We use the Bayesian approach to estimate the model:
• Using Kim et al. (1999) algorithm, we compute the likelihood
introducing non-linearities and unobserved chains employ-
ing the filter with prior distribution of the parameters.

• We construct the posterior kernel with our results from

the Bee_gate9 optimizer routine.

• We use the posterior mode as the initial value for the Metrop-
olis Hasting algorithm, with 100.000 iterations.

• We compute moments utilizing the mean and variance of the

last 50.000 iterations.

9
rise toolbox optimization routine.

414 S. Cadavid, A. Ortiz

3.1 Counterfactuals
To explore the characteristics of the ms-dsge model with multiple
regimes, we generate a counterfactual series based on conditional
forecast simulations. Specifically, this analysis allows us to get an idea
of what would have happened if the monetary policy had not changed,
given the smoothed shocks estimated by the model. The model is re-
solved introducing a law of motion consistent with the fact that no oth-
er regime would have been observed. In this section the algorithm
to generate the simulated series is briefly explained.
Once the model is estimated, we generate forecasts from the ms-
dsge model conditional on the realized path of the five model shocks:
terms of trade, technology, monetary, world output, and world infla-
tion. Our conditional forecasts are generated over the full sample
period for each of the five countries. The data from the first quar-
ter in every sample are used as initial conditions. The parameters
utilized are the estimated posterior distribution of the coefficients
for each regime.
We trace out the counterfactuals’ paths by generating a new data
vector for Zt in (3.7), which includes the smoothed shocks. As differ-
ent paths for the endogenous variables (one for each regime) are ob-
tained for this regime switching model, we utilize the “expected
smoothed series of the shocks, correspond to the weighted average
paths of the exogenous variables.
Once the system is integrated, as in the previous subsection,
the data are filtered and the counterfactual paths for the unobserved
and observable variables are generated.

4. RESULTS

4.1 Regime probabilities

Figures 2 to 6 show the smoothed probabilities for the two Markov-
switching processes. The top panel of each figure shows the prob-
ability that monetary policy is conducted under a high interest rate
response to inflation regime based on the structural parameters
of the interest rate rule. The bottom panel presents the probability
of being on a high volatility regime based on the relative volatility
of the non-stationary technology process. The first thing one must
notice is that high interest rate response regimes have been the most

Regime Switch of Monetary Policy in Latin America 415

prevalent forms of regime during the sample periods. The percent-
age of periods where our estimation assigns a probability higher than
50% of Brazil, Chile, Colombia, Mexico, and Peru being in a high
response regime are 77%, 90%, 77%, 65% and 69%, respectively. Re-
garding the transition matrix, the mean (and 10%-90% confidence
interval in parenthesis) parameter estimates for the probability
of going from a high response to a low response regime,
are 0.1603 (0.039, 0.4719), 0.0808 (0.0141, 0.21), 0.0863 (0.0239,
0.2236), 0.1161 (0.0707, 0.1842) and 0.0721 (0.0276, 0.1129), respec-
tively, while the probability of moving from a low response to a high
response regime are 0.2257 (0.0997, 0.4375), 0.0521 (0.0225,
0.0942), 0.1566 (0.048, 0.3472), 0.2108 (0.097, 0.3049) and 0.0565
(0.0191, 0.101), respectively.

4.1.1 High interest rate response regimes

With the introduction of inflation targeting and greater exchange
rate flexibility, after a 35% real depreciation in 1999, Brazil experi-
enced a regime switch to high response in 1999Q3. Our analysis cap-
tures the 2002 depreciation and the Cardoso-da Silva government
transition as a transitory change of the monetary policy regime from
2002Q4 to 2003Q4. From 2004Q1 onwards, the probability of being
under a high response monetary policy is close to 1.
Chile fully adopted inflation targeting in 1999, but as stated in Cor-
bo et al. (2002) the scheme began to be implemented in the 1990s.
Our estimation captures a high response to inflation from the begin-
ning of the sample in 1996 until 2007Q4.In 2008Q1 and until 2009Q4,
there was a marked shift in policy with smaller weight on inflation
and larger weight on output during a stagflationary period. From
2010Q1 onwards, the interest response of interest rates to inflation
is estimated to be strong with high probability.
Colombia experienced a strong shift in monetary policy during
2000Q1 shortly after the introduction of inflation targeting and great-
er exchange rate flexibility.
Mexico has three periods during which our estimation assigns
a high probability to a high response regime: from 1988Q2 to 1988Q3,
from 1992Q1 to 1994Q4 and from 1997Q2 onwards. The first period
coincides with Pacto de Solidaridad y Estabilidad Económica, signed
in December 1987, which was a heterodox plan committing labor
unions and public and private sectors to limit their price revisions

416 S. Cadavid, A. Ortiz

 Figure 2
SMOOTHED PROBABILITIES FOR BRAZIL

     

100

50

0
1996 1998 2001 2004 2007 2009 2012 2016

     

100

50

0
1996 1998 2001 2004 2007 2009 2012 2016

Figure 3
SMOOTHED PROBABILITIES FOR CHILE

     

100

50

0
1998 2000 2002 2004 2006 2008 2010 2012 2014 2016

     

100

50

0
1998 2000 2002 2004 2006 2008 2010 2012 2014 2016

Regime Switch of Monetary Policy in Latin America 417

 Figure 4
SMOOTHED PROBABILITIES FOR COLOMBIA

     

100

50

0
1995 1998 2001 2004 2007 2009 2012 2015

     

100

50

0
1995 1998 2001 2004 2007 2009 2012 2015

Figure 5
SMOOTHED PROBABILITIES FOR MEXICO

     

100

50

0
1981 1984 1988 1992 1996 2000 2003 2007 2011 2016

     

100

50

0
1981 1984 1988 1992 1996 2000 2003 2007 2011 2016

418 S. Cadavid, A. Ortiz

 Figure 6
SMOOTHED PROBABILITIES FOR PERU

     

100

50

0
1995 1998 2001 2004 2007 2009 2012 2015

     

100

50

0
1995 1998 2001 2004 2007 2009 2012 2015

to anchor inflation expectations. The second period was shortly af-

ter the exchange rate policy changed from fixed exchange rate to a
band system with a floor and a ceiling both adjustable over time.
It includes the 1993 Constitutional reform granting legal autonomy
to the Central Bank and the establishment of the price stability ob-
jective while it recognized that no government authority could force
the Central Bank to grant financing. The December 1994 Tequila
crisis forced the Central Bank to adopt a floating exchange rate re-
gime. The crisis required balancing nominal pressures with an out-
put contraction which required postponing the adoption of a high
response regime until 1997Q2 consolidated in 2001 with the intro-
duction of inflation targeting.
In addition, our analysis estimates Peru had three periods with
a high probability of high response regime: from 1997Q4 to 1998Q1,
in 1998Q4, and from 2002Q1 onwards. Therefore, after brief episodes

Regime Switch of Monetary Policy in Latin America 419

of monetary tightening in 1997/1998, monetary policy switched to-
wards greater responsiveness to inflation in 2002 which coincides
with the adoption of the inflation targeting regime.

4.1.2 High volatility shock regimes

Cogley and Sargent (2005), Sims and Zha (2006) and Bianchi (2012)
highlight the importance of accounting for stochastic volatility of ex-
ogenous shocks when a regime switch in monetary policy is analyzed.
Additionally, Liu and Mumtaz (2011) and Goncalves et al. (2016) show
that the fit of the model is improved when a Markov-Switching process
for regime volatilities is introduced. In our estimation, we classify a re-
gime as one of high volatility if the standard deviation of the stochastic
volatility of the non-stationary technology shock is large. Given that
in order to guarantee stationarity of the model, all real variables must
be expressed in terms of percentage deviations from At , the growth
rate of the non-stationary technology process enters the is -curve. Or-
ganizing countries alphabetically, the percentage of periods where
the estimation assigns a probability higher than 50% of being in a
high volatility regime are 18%, 51%, 22%, 56% and 35%, respec-
tively. Regarding the transition matrix, the mean (and 10%-90%
confidence interval in parenthesis) parameter estimates for the prob-
ability of going from a low volatility to a high volatility regime,
are 0.3071 (0.1241, 0.5589), 0.0307 (0.0107, 0.0589), 0.0607 (0.0089,
0.2931), 0.1922 (0.0958, 0.339) and 0.0849 (0.0103, 0.4463), respec-
tively, while the probability of moving from a low response to a high
response regime are 0.1458 (0.0278, 0.4982), 0.182 (0.1096,
0.2873), 0.1023 (0.0257, 0.2056), 0.109 (0.0577, 0.1836) and 0.1719
(0.0427, 0.4136), respectively. High volatility periods for Brazil
are 1996Q2-1996Q3, 1997Q4-1999Q3, and 2008Q3-2009Q2; while
for Chile they are 1997Q4-2000Q2,2001Q1, and 2003Q1-2010Q3;
for Colombia they are 1995Q4-1996Q3, 1998Q2-2000Q2, 2002Q3-
2003Q1, and 2008Q4-2009Q1; for Mexico they are 1981Q1-1983Q1,
1984Q1-1992Q2, 1994Q1-1998Q3, 2008Q2-2010Q1, 2011Q4-2012Q2,
and 2015Q1-2016Q3; and for Peru it is 1995Q4-2002Q3.

4.2 Estimation results

Table 2, below, reports the mean for the estimated parameters
of the model for each country, while the appendix has individual
tables for each country with the mean, mode, standard deviation

420 S. Cadavid, A. Ortiz

 Table 2
MEAN FOR THE ESTIMATED PARAMETERS FOR
BRAZIL, CHILE, COLOMBIA, MEXICO AND PERU

Country
Parameter Distribution Brazil Chile Colombia Mexico Peru
Beta 0.1738 0.2053 0.7092 0.8564 0.1318

Beta 0.4471 0.5124 0.313 0.6134 0.1471

Gamma 1.1362 0.0765 0.5845 2.1643 0.5011

Gamma 0.6296 0.0631 1.9982 2.3736 0.0565

Beta 0.7629 0.9215 0.7298 0.458 0.697

Beta 0.6113 0.4912 0.7065 0.6279 0.6254

Gamma 3.4901 2.7337 3.2941 1.8458 1.9066

Gamma 1.0417 0.8692 0.9746 0.6154 0.9226

Gamma 0.3013 0.5594 0.3849 0.7265 0.4092

Gamma 0.8799 0.434 0.7379 0.8310 0.5639

Gamma 0.0435 0.0816 0.137 0.1108 0.1725

Gamma 0.0422 0.0662 0.0463 0.3408 0.1506

α Beta 0.076 0.0539 0.1132 0.2689 0.0393

r Gamma 3.6731 2.2813 6.8509 2.1004 8.8041
τ Beta 0.2792 0.16 0.2445 0.3256 0.1306
ρa Beta 0.3014 0.1599 0.1291 0.2007 0.3924

Regime Switch of Monetary Policy in Latin America 421

 Country
Parameter Distribution Brazil Chile Colombia Mexico Peru
ρq Beta 0.424 0.1553 0.1628 0.4305 0.3605
ρy* Beta 0.9818 0.9579 0.9659 0.9042 0.9682
ρπ* Beta 0.3715 0.3129 0.2303 0.7824 0.416
Beta 0.1603 0.0808 0.0863 0.1161 0.0721

Beta 0.2257 0.0521 0.1566 0.2108 0.0565

Inv.Gamma 5.3145 0.5788 0.8134 4.5438 2.4271

Inv.Gamma 3.3642 3.3239 6.8695 5.8216 7.6316

Inv.Gamma 5.791 6.4758 5.5065 3.121 4.1378

Inv.Gamma 4.2554 5.3403 7.2084 4.4066 5.1138

Inv.Gamma 4.6972 3.9563 5.0036 3.2222 2.7075

Inv.Gamma 4.7999 6.1979 6.0725 7.4444 6.0456

Inv.Gamma 3.5522 3.4781 1.6996 6.7571 2.1448

Inv.Gamma 6.9291 5.4652 3.0673 7.3328 3.5942

Inv.Gamma 4.8214 7.2118 5.0864 5.09 5.0435

Inv.Gamma 6.1201 4.6023 2.4292 9.5155 5.0472

Beta 0.3071 0.0307 0.0607 0.1922 0.0849

Beta 0.1458 0.182 0.1023 0.109 0.1719

422 S. Cadavid, A. Ortiz

and confidence intervals. When describing the parameter estimates,
we follow the convention of reporting values of countries ordered
as Brazil, Chile, Colombia, Mexico, and Peru. First, we describe
the values for the high interest rates responses to inflation regimes
and then for the low response regimes, followed by a comparison.
We report the mean for the estimated parameters and, in parenthe-
sis, the estimated values for the 10% and 90% confidence intervals.
Here, we focus on talking about the parameters related to the infla-
tion formation process of the Phillips curve and the interest rate re-
action function.
The persistence of inflation is captured by the parameter χ p in
the Phillips Curve. The parameter estimates for the high interest rate
response regime, χ p,ξcoef =1, are 0.1738 (0.0319, 0.4303), 0.2053 (0.1027,
0.3366), 0.7092 (0.4474, 0.8981), 0.8564 (0.6316, 0.9739) and 0.1318
(0.0321, 0.2885), respectively, while for the low interest rate response
regimes, χ p,ξ coef =2 , they are 0.4471 (0.1352, 0.8285), 0.5124 (0.1913,
0.8204), 0.313 (0.1498, 0.5307), 0.6134 (0.496, 0.7669), and 0.1471
(0.0352, 0.286), respectively. Therefore, average inflation persistence
has been lower for the high interest rate response regimes in Brazil
and Chile, while it has been higher in Colombia and Mexico, and has
remained almost unchanged in Peru. The counterpart to this persis-
tence of inflation is the relative weight that expectations have in the
inflation formation process.
The sensitivity of inflation to the output gap is partially captured
by the parameter κ in the Phillips Curve. The parameter estimates
for the high interest rate response regime, κξcoef =1 , are 1.1362 (0.8484,
1.6328), 0.0765 (0.0368, 0.1346), 0.5845 (0.3863, 0.8068), 2.1643
(1.9357,2.3318) and 0.5011(0.3481,0.6833), respectively, while for the
low interest rate response regimes, κξ coef =2 , they are 0.6296 (0.27,
1.2559), 0.0631 (0.0331, 0.1008), 1.9982 (1.6591,2.3484), 2.3736
(1.7729,3.3246) and 0.0565 (0.0294,0.0863), respectively. There-
fore, average sensitivity of inflation to the output gap has been low-
er for the high interest rate response regime in Colombia, higher
in Brazil and Peru, and it has remained almost unchanged at a fairly
low value in Chile and a high value in Mexico.
Therefore, in the context of the inf lation formation process,
going from a low interest response to a high one, as happened
chronologically in all countries except Chile, Brazil experienced
a drop in inflation inertia and a more responsive trade-off between
output gap and inflation, Colombia has higher inflation inertia

Regime Switch of Monetary Policy in Latin America 423

and a less responsive trade-off, Mexico has higher inflation inertia
and moderate decrease in the responsiveness of the trade-off,
and Peru has the same level of inertia and a more responsive trade-
off. Meanwhile, as stated before, Chile started the sample with a high
interest rate response to inflation and loosened the policy from
2008Q1 to 2009Q4. Then, when moving from a high interest rate
response to a low one, Chile had an increase in inflation inertia
without changes in the slope of its Phillips curve.
Turning to the interest rate reaction function, the persistence
of interest rates is captured by the parameter ρR . The parameter esti-
mates for the high interest rate response regime, ρR,ξ coef =1 ,are 0.7629
(0.6917,0.8144), 0.9215 (0.8525,0.9788), 0.7298 (0.6633, 0.8071),
0.458 (0.3897,0.5541) and 0.697 (0.6211,0.753), respectively, while
for the low interest rate response regime, ρR,ξ coef =2 ,they are 0.6113
(0.2252,0.813),0.4912 (0.4328, 0.5514), 0.7065 (0.6491, 0.7621), 0.6279
(0.3992, 0.7734) and 0.6254 (0.5227, 0.7344), respectively.
Therefore, average persistence of interest rates has been higher
for the high interest rate response regime in Brazil, Chile and Peru,
it has decreased in Mexico and it has remained relatively unchanged
in Colombia.
The sensitivity of interest rates to inflation is captured by the
parameter ψπ. The parameter estimates for the high interest rate
response regime, ψπ,ξ coef =1 , are 3.4901 (2.733, 3.8618), 2.7337
(1.079, 5.4875), 3.2941 (1.8292, 4.9853), 1.8458 (1.7431, 1.9526)
and 1.9066 (1.3059,3.309), respectively, while for the low interest
rate response regime, ψπ,ξcoef=2, are 1.0417 (0.6815,1.4375),0.8692
(0.7058,1.0166),0.9746 (0.7722, 1.1641), 0.6154 (0.4424, 0.823)
and 0.9226 (0.444, 1.7992), respectively.
The sensitivity of interest rates to output deviations is captured
by the parameter ψy. The parameter estimates for the high interest
rate response regime, ψy,ξcoef =1, are 0.3013 (0.075, 0.9818), 0.5594
(0.3015, 0.8963), 0.3849 (0.1969, 0.6058), 0.7265 (0.602, 0.8016)
and 0.4092 (0.1659,0.859), respectively, while for the low interest rate
response regime, ψy,ξcoef =2, are 0.8799 (0.2204, 2.0191), 0.434 (0.2317,
0.7397), 0.7379(0.3355, 1.2305), 0.831 (0.8039, 0.8562) and 0.5639
(0.3263, 1.0481), respectively. Therefore, average sensitivity of in-
terest rates to output deviations has been lower for the high interest
rate response regime in Brazil, Colombia, Mexico and Peru, while
it has been higher in Chile.

424 S. Cadavid, A. Ortiz

 The sensitivity of interest rates to exchange rate deprecia-
tions is captured by the parameter The parameter estimates
for the high interest rate response regime, are 0.0435
(0.0156,0.098), 0.0816 (0.0229,0.2694), 0.137 (0.1068,0.1752), 0.1108
(0.0961,0.1254) and 0.1725 (0.1215,0.2283), respectively, while
for the low interest rate response regimes, are 0.0422
(0.0139,0.1547), 0.0662 (0.026,0.1325), 0.0463 (0.0148,0.0844),
0.3408 (0.0775,0.6386) and 0.1506 (0.1139,0.1925), respectively.
Therefore, average sensitivity of interest rates to exchange rate de-
preciations has been higher for the high interest rate response re-
gime in Colombia, it has decreased in Mexico and it has remained
almost unchanged for Brazil, Chile and Peru.
Therefore, in terms of the interest rate reaction function, going
from a low interest response to a high one as happened chronologi-
cally in all countries except Chile, Brazil exhibited a greater persis-
tence of interest rates, less sensitivity to output deviations, and no
change in the response to exchange rate fluctuations. Colombia ex-
hibited similar persistence of interest rates, decreased sensitivity
to output deviations and larger sensitivity to exchange rate fluctua-
tions. Mexico exhibited less persistence of interest rates, and small-
er sensitivity to output deviations and exchange rate fluctuations.
Peru exhibited larger persistence of interest rates, diminished sen-
sitivity to output deviations, and similar response to exchange rate
fluctuations. Finally, for Chile, when moving from a high interest
rate response to a low one, interest rates exhibited less persistence
and the weight on output deviations was larger, as expected from
the countercyclical stance of their monetary policy.

4.3 Impulse response functions

Figures 7 to 11 show the impulse response functions regarding mon-
etary policy, non-stationary technology, terms of trade, world out-
put, and world inflation shocks, respectively. Each graph compares
the responses under the high and low interest rate response to infla-
tion regimes. Inspecting the different mechanisms prevalent in each
country under each policy stance will allow us to understand the coun-
terfactuals that are presented later where we ask what may have hap-
pened if another regime had been in place for the entire sample.

Regime Switch of Monetary Policy in Latin America 425

 An unexpected expansion of monetary policy appreciates the cur-
rency, while it lowers inflation and output. Under the high policy re-
sponse regime, appreciations are larger in Chile and Peru, where real
interest rates increase by more and inflation drops are larger. Only
in the case of Chile has the observed output contraction been larger
under the high policy response regime, which could be due to the fact
that the low response regime was implemented for countercyclical
motives only once the inflation targeting regime was consolidated.
Technology is assumed to be difference stationary, so innovations
in productivity have permanent effects on output. On average, out-
put increases, inflation is positive, currency depreciates, and real
interest rates decrease. These movements are slightly smaller under
the high policy response regime.
An unexpected improvement in terms of trade raises output,
appreciates the currency, and lowers inflation (except for the high
policy response regime in Peru, where prices increase). On average,
these movements prompt the central banks to loosen policy (except
for the high policy response of Chile). Appreciations are of similar
magnitude under both policy response regimes. Under the high
policy response regime, output expansions are larger in Colombia
and Mexico, the reduction of inflation is smaller in Brazil, Chile
and Mexico, and the real interest rate drops by more in all coun-
tries except Chile.
World demand shocks lower domestic output, increase inflation,
and potentially cause an exchange rate depreciation. These results
arise because, under the estimated elasticities of intertemporal sub-
stitution, world output shocks lower domestic potential output in all
countries. Despite the fact that nominal interest rates increase, real
interest rates decrease. Under high policy response regimes output
contractions are larger, inflation increases less, nominal exchange
rate depreciation is smaller, and the central banks cut real interest
rates by less.
Shocks to import price inflation appreciate the currency, but raise
inflation because, in addition to the inherent foreign price inflation,
the central bank reacts to movements in the exchange rate, and low-
ers real interest rates. Under high policy response regimes output
increases by less, except in the case of Colombia, inflation increas-
es by less, except in the case of Peru and the nominal exchange rate
depreciation is of similar magnitude, except for Mexico where it is
larger under high response.

426 S. Cadavid, A. Ortiz

 Figure 7
MONETARY POLICY SHOCK IRFs

 
2 0.5
Output growth

0 0

−2 −0.5

−4 −1
2 4 6 8 10 12 2 4 6 8 10 12
5 0.5

0 0
Inflation

−5 −0.5

−10 −1
2 4 6 8 10 12 2 4 6 8 10 12
4 0.5
Interest rate

2
0
0

−2 −0.5
2 4 6 8 10 12 2 4 6 8 10 12
10 1
Real interest rate

5
0.5
0

−5 0
2 4 6 8 10 12 2 4 6 8 10 12
5 0.5
∆ Exchange rate

0 0

−5 −0.5

−10 −1
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 427

 Figure 7 (cont.)
MONETARY POLICY SHOCK IRFs

 
0.2 1
Output growth

0 0

−0.2 −1

−0.4 −2
2 4 6 8 10 12 2 4 6 8 10 12
1 5

0 0
Inflation

−1 −5

−2 −10
2 4 6 8 10 12 2 4 6 8 10 12
0.4 2
Interest rate

0.2 1

0 0

−0.2 −1
2 4 6 8 10 12 2 4 6 8 10 12
2 10
Real interest rate

1 5

0 0

−1 −5
2 4 6 8 10 12 2 4 6 8 10 12
1 5
∆ Exchange rate

0 0

−1 −5

−2 −10
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

428 S. Cadavid, A. Ortiz

 Figure 7 (cont.)
MONETARY POLICY SHOCK IRFs


0.5

Output growth
0

−0.5

−1
2 4 6 8 10 12
0
Inflation

−1

−2
2 4 6 8 10 12
2
Interest rate

0
2 4 6 8 10 12
4
Real interest rate

0
2 4 6 8 10 12
0
∆ Exchange rate

−1

−2
2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 429

 Figure 8
TECHNOLOGY SHOCK IRFs

 
1 0.2
Output growth

0.5 0.1

0 0

−0.5 −0.1
2 4 6 8 10 12 2 4 6 8 10 12
2 0.1

1 0.05
Inflation

0 0

−1 −0.05
2 4 6 8 10 12 2 4 6 8 10 12
1 0.1
Interest rate

0.5 0.05

0 0
2 4 6 8 10 12 2 4 6 8 10 12
1 0.05
Real interest rate

0.5
0
0

−0.5 −0.05
2 4 6 8 10 12 2 4 6 8 10 12
2 0.1
∆ Exchange rate

1 0.05

0 0

−1 −0.05
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

430 S. Cadavid, A. Ortiz

 Figure 8 (cont.)
TECHNOLOGY SHOCK IRFs

 
0.2 0.2
Output growth

0.1 0.1

0 0

−0.1 −0.1
2 4 6 8 10 12 2 4 6 8 10 12
1 1

0.5 0.5
Inflation

0 0

−0.5 −0.5
2 4 6 8 10 12 2 4 6 8 10 12
0.4 0.4
Interest rate

0.2 0.2

0 0

−0.2 −0.2
2 4 6 8 10 12 2 4 6 8 10 12
0.5 0.2
Real interest rate

0
0
−0.2

−0.5 −0.4
2 4 6 8 10 12 2 4 6 8 10 12
1 1
∆ Exchange rate

0.5 0.5

0 0

−0.5 −0.5
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 431

 Figure 8 (cont.)
TECHNOLOGY SHOCK IRFs


0.4

Output growth
0.2

−0.2
2 4 6 8 10 12
0.5
Inflation

−0.5
2 4 6 8 10 12
0.3
Interest rate

0.2

0.1

0
2 4 6 8 10 12
0.2
Real interest rate

−0.2
2 4 6 8 10 12
0.5
∆ Exchange rate

−0.5
2 4 6 8 10 12

High response regime Low response regime

432 S. Cadavid, A. Ortiz

 Figure 9
TERMS OF TRADE SHOCK IRFs

 
0.1 0.2
Output growth

0.05 0.1

0 0
2 4 6 8 10 12 2 4 6 8 10 12
0.1 0.2

0 0
Inflation

−0.1 −0.2

−0.2 −0.4
2 4 6 8 10 12 2 4 6 8 10 12
0.1 0.2
Interest rate

0 0

−0.1 −0.2

−0.2 −0.4
2 4 6 8 10 12 2 4 6 8 10 12
0.1 0.2
Real interest rate

0
0
−0.1

−0.2 −0.2
2 4 6 8 10 12 2 4 6 8 10 12
0 5
∆ Exchange rate

−2 0

−4 −5

−6 −10
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 433

 Figure 9 (cont.)
TERMS OF TRADE SHOCK IRFs

 
0.4 0.15
Output growth

0.2 0.1

0 0.05

−0.2 0
2 4 6 8 10 12 2 4 6 8 10 12
0.2 0.4

0.2
Inflation

0
0

−0.2 −0.2
2 4 6 8 10 12 2 4 6 8 10 12
0.2 0
Interest rate

0
−0.2
−0.2

−0.4 −0.4
2 4 6 8 10 12 2 4 6 8 10 12
0.2 0
Real interest rate

0 −0.2

−0.2 −0.4

−0.4 −0.6
2 4 6 8 10 12 2 4 6 8 10 12
5 0
∆ Exchange rate

−1
0
−2

−5 −3
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

434 S. Cadavid, A. Ortiz

 Figure 9 (cont.)
TERMS OF TRADE SHOCK IRFs


0.15

Output growth
0.1

0.05

0
2 4 6 8 10 12
0.2

0.1
Inflation

−0.1
2 4 6 8 10 12
0
Interest rate

−0.1

−0.2

−0.3
2 4 6 8 10 12
0
Real interest rate

−0.1

−0.2

−0.3
2 4 6 8 10 12
5
∆ Exchange rate

−5
2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 435

 Figure 10
WORLD OUTPUT SHOCK IRFs

 
0 0
Output growth

−0.5
−1
−1

−1.5 −2
2 4 6 8 10 12 2 4 6 8 10 12
3 3

2 2
Inflation

1 1

0 0
2 4 6 8 10 12 2 4 6 8 10 12
1.5 2
Interest rate

1
1
0.5

0 0
2 4 6 8 10 12 2 4 6 8 10 12
1 1
Real interest rate

0 0

−1 −1

−2 −2
2 4 6 8 10 12 2 4 6 8 10 12
3 3
∆ Exchange rate

−2 2

1 1

0 0
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

436 S. Cadavid, A. Ortiz

 Figure 10 (cont.)
WORLD OUTPUT SHOCK IRFs

 
0 0
Output growth

−0.5
−5
−1

−1.5 −10
2 4 6 8 10 12 2 4 6 8 10 12
6 30

4 20
Inflation

2 10

0 0
2 4 6 8 10 12 2 4 6 8 10 12
1.5 10
Interest rate

1
5
0.5

0 0
2 4 6 8 10 12 2 4 6 8 10 12
2 10
Real interest rate

0 0

−2 −10

−4 −20
2 4 6 8 10 12 2 4 6 8 10 12
6 30
∆ Exchange rate

4 20

2 10

0 0
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 437

 Figure 10 (cont.)
WORLD OUTPUT SHOCK IRFs


0

Output growth
−0.5

−1

−1.5
2 4 6 8 10 12
1.5

1
Inflation

0.5

0
2 4 6 8 10 12
1
Interest rate

0.5

0
2 4 6 8 10 12
0.5
Real interest rate

−0.5

−1
2 4 6 8 10 12
1.5
∆ Exchange rate

0.5

0
2 4 6 8 10 12

High response regime Low response regime

438 S. Cadavid, A. Ortiz

 Figure 11
WORLD INFLATION SHOCK IRFs

 
0.05 0.1
Output growth

0.05
0
0

−0.05 −0.05
2 4 6 8 10 12 2 4 6 8 10 12
0.15 0.1

0.1
Inflation

0.05
0.05

0 0
2 4 6 8 10 12 2 4 6 8 10 12
0.01 0.1
Interest rate

0 0

−0.01 −0.1

−0.02 −0.2
2 4 6 8 10 12 2 4 6 8 10 12
0 0.1
Real interest rate

−0.05 0

−0.1 −0.1

−0.15 −0.2
2 4 6 8 10 12 2 4 6 8 10 12
0 5
∆ Exchange rate

−2
0
−4

−6 −5
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 439

 Figure 11 (cont.)
WORLD INFLATION SHOCK IRFs

 
0.04 1
Output growth

0.02 0.5

0 0

−0.02 −0.5
2 4 6 8 10 12 2 4 6 8 10 12
0.1 6

0.05 4
Inflation

0 2

−0.05 0
2 4 6 8 10 12 2 4 6 8 10 12
0.02 1.5
Interest rate

0 1

−0.02 0.5

−0.04 0
2 4 6 8 10 12 2 4 6 8 10 12
0.05 2
Real interest rate

0 0

−0.05 −2

−0.1 −4
2 4 6 8 10 12 2 4 6 8 10 12
2 0
∆ Exchange rate

0 −2

−2 −4

−4 −6
2 4 6 8 10 12 2 4 6 8 10 12

High response regime Low response regime

440 S. Cadavid, A. Ortiz

 Figure 11 (cont.)
WORLD INFLATION SHOCK IRFs


0.2

Output growth
0.1

0
2 4 6 8 10 12
0.4
Inflation

0.2

0
2 4 6 8 10 12
0.1
Interest rate

−0.1

−0.2
2 4 6 8 10 12
0
Real interest rate

−0.2

−0.4
2 4 6 8 10 12
0
∆ Exchange rate

−2

−4

−6
2 4 6 8 10 12

High response regime Low response regime

Regime Switch of Monetary Policy in Latin America 441

4.4 Counterfactuals
As shown by the impulse response functions, there are differences
in the magnitudes and even signs of the responses under the different
regimes. Our estimated model allows one to perform counterfactual
analysis of what could have happened if policies had been different.
In Figures 12 to 16, we show the actual behavior of five observables:
gdp growth, inflation, nominal interest rate, ex-post real interest rate,
and nominal depreciation, and compare them with the hypothetical
behavior that may have been observed under a constant high inter-
est rate response regime and a constant low response regime. Table
3 reports the average, standard deviation and coefficient of varia-
tion of the actual observables and their simulated counterfactuals.
Looking at the figures one realizes that the regime switches that
occurred throughout Latin America towards more responsive inter-
est rate reaction functions helped to prevent many inflationary runs,
several large nominal exchange rate depreciations, and large volatil-
ity of the nominal variables. Table 3 confirms that there would have
been less average inflation under the high interest rate response re-
gime than the observed average inflation, which is lower than the av-
erage inflation under the low interest rate response regime. Not only
would average inflation have been lower, but the standard deviation
of inflation would also have been lower under the counterfactual
high response regime than in the observed one, which is lower than
the counterfactual low response regime. The high response regime
does not imply higher average nominal interest rates or higher av-
erage real interest rates, while their variability under that high re-
sponse regime would have been less than the observed ones. Average
nominal depreciation under the high response regime turned out to
be smaller and less volatile. The reduction in the level and volatil-
ity of the nominal variables under the high response regime does
not imply a sacrifice in terms of output growth, or on its volatility.

442 S. Cadavid, A. Ortiz

 Table3
SUMMARY STATISTICS

Variable Series Brazil Chile Colombia Mexico Peru

Average SD CV Average SD CV Average SD CV Average SD CV Average SD CV

Observed 0.64 1.26 1.97 3.85 4.21 1.09 3.44 4.27 1.11 2.26 5.73 2.53 4.65 3.31 0.71

Output High 0.99 3.28 3.30 3.75 2.78 0.74 3.42 4.11 1.22 1.77 4.84 2.73 4.97 2.74 0.55
growth response

Low 1.00 3.63 3.62 3.65 4.75 1.30 3.37 4.47 1.31 3.46 8.85 2.56 5.38 5.71 1.06
response

Observed 6.31 3.72 0.59 3.06 2.51 0.82 9.84 7.43 1.12 20.15 24.78 1.23 3.62 3.18 0.88

Inflation High 3.89 2.73 0.70 2.93 1.90 0.65 7.13 4.93 1.04 11.08 7.65 0.69 3.34 2.12 0.64
response

Low 15.73 5.80 0.37 3.14 3.88 1.23 17.13 12.26 0.89 26.83 23.69 0.88 6.89 8.68 1.26
response

Observed 16.49 7.00 0.42 4.59 2.04 0.44 12.56 10.50 2.33 25.38 26.36 1.04 6.91 6.03 0.87

Regime Switch of Monetary Policy in Latin America 443

 Variable Series Brazil Chile Colombia Mexico Peru
Average SD CV Average SD CV Average SD CV Average SD CV Average SD CV

Interest High 10.69 3.84 0.36 4.76 1.44 0.30 9.01 6.41 1.61 18.90 13.06 0.69 5.43 2.61 0.48
rate response

Low 12.59 4.17 0.33 4.76 3.02 0.63 12.58 8.48 1.57 30.81 27.84 0.90 4.42 5.26 1.19
response

444 S. Cadavid, A. Ortiz

Observed 10.18 7.58 0.74 1.53 2.15 1.40 2.72 5.13 2.43 5.23 9.08 1.74 3.29 5.73 1.74

Real High 6.80 3.49 0.51 1.83 2.04 1.11 1.89 2.80 3.64 7.82 7.94 1.01 2.09 2.61 1.25
interest response
rate

Low −3.14 7.36 2.34 1.63 1.48 0.91 −4.56 8.94 1.07 3.98 15.61 3.92 −2.47 13.33 5.40
response

Observed 1.64 9.13 5.58 0.67 4.87 7.25 1.36 6.60 0.89 −0.59 4.81 8.10 0.45 2.70 6.02

Nominal High −0.60 8.61 14.26 0.72 2.86 4.00 −3.15 8.44 1.34 −5.59 9.92 1.77 0.25 3.29 13.40
depreciation response

Low 11.24 9.23 0.82 1.28 7.50 5.88 6.33 9.15 0.77 5.74 14.88 2.59 3.77 8.15 2.16
response
 Figure 12
COUNTERFACTUAL FOR HIGH AND LOW RESPONSE REGIMES FOR BRAZIL

 

10

−10
2000 2005 2010 2015



10

2000 2005 2010 2015

 

30
20
10
0
2000 2005 2010 2015

  

30
20
10
0
−10

2000 2005 2010 2015

 
30
20
10
0
−10

2000 2005 2010 2015

High response forecast Low response forecast

Actual

Regime Switch of Monetary Policy in Latin America 445

 Figure 13
COUNTERFACTUAL FOR HIGH AND LOW RESPONSE REGIMES FOR CHILE

 
10

−10

2000 2005 2010 2015



10
5
0

2000 2005 2010 2015

 

0.2

0.1

0
2000 2005 2010 2015

  

8
6
4
2
0
−2

2000 2005 2010 2015

 
20
10
0
−10
2000 2005 2010 2015

High response forecast Low response forecast

Actual

446 S. Cadavid, A. Ortiz

 Figure 14
COUNTERFACTUAL FOR HIGH AND LOW RESPONSE
REGIMES FOR COLOMBIA
 

10
5
0
−5

1995 2000 2005 2010



10

1995 2000 2005 2010

 

30
20

10
0
1995 2000 2005 2010

  

20

−20

1995 2000 2005 2010

 

20

−20

1995 2000 2005 2010

High response forecast Low response forecast

Actual

Regime Switch of Monetary Policy in Latin America 447

 Figure 15
COUNTERFACTUAL FOR HIGH AND LOW RESPONSE REGIMES FOR MEXICO

 

20

−20

1985 1990 1995 2000 2005 2010 2015


100

50

0
1985 1990 1995 2000 2005 2010 2015

 

100

50

0
1985 1990 1995 2000 2005 2010 2015

  

60
40
20
0
−20

1985 1990 1995 2000 2005 2010 2015

 
60
40
20
0
−20
−40

1985 1990 1995 2000 2005 2010 2015

High response forecast Low response forecast

Actual

448 S. Cadavid, A. Ortiz

 Figure 16
COUNTERFACTUAL FOR HIGH AND LOW RESPONSE REGIMES FOR PERU

 

15
10
5
0
−5
1990 2005 2010 2015



100

50

1990 2005 2010 2015

 

50

0
1990 2005 2010 2015

  

30
20
10
0

1990 2005 2010 2015

 

5
0
−5

1990 2005 2010 2015

High response forecast Low response forecast

Actual

Regime Switch of Monetary Policy in Latin America 449

5. CONCLUSIONS

In this paper we explore whether the central bank reforms imple-

mented in the 1990s in Brazil, Chile, Colombia, Mexico and Peru,
which lead to an inflation targeting framework, represented a regime
switch in their monetary policies. The estimation of a Markov-switch-
ing dsge open economy monetary model allows us to identify regime
shifts of an interest rate reaction function together with the inflation
determination process of a hybrid New Keynesian open economy
Phillips curve. Our estimation identifies the following periods as hav-
ing high interest rate responses to inflation: from 1999Q3 to 2002Q3
and from 2004Q1 onwards for Brazil; from the beginning of the
sample in 1996Q2 to 2007Q4 and from 2010Q1 onwards for Chile;
from 2000Q1 onwards for Colombia; from 1988Q2 to 1988Q3, from
1992Q1 to 1994Q4, and from 1997Q2 onwards for Mexico; 1997Q4
to 1998Q1, in 1998Q4, and from 2002Q1 onwards for Peru. The in-
troduction of inflation targeting is associated with a marked regime
switch towards a more reactive interest rate policy.
The estimation of the structural parameters associated with the hy-
brid New Keynesian open economy Phillips curve indicates that when
changing from a low interest response to a high one as it happened
chronologically in all countries (except Chile), Brazil experienced
a drop in inflation inertia and a more responsive trade-off between
output gap and inflation, Colombia experienced a higher inflation
inertia and a reduction in the slope of the Phillips curve, Mexico also
experienced higher inflation inertia and a slightly reduction in the
large slope of the Phillips curve, and Peru experienced the same lev-
el of inertia and a more responsive trade-off. Meanwhile, as stated
before, Chile began our sample with a high interest rate response
to inflation and loosened the policy from 2008Q1 to 2009Q4. Then,
when moving from a high interest rate response to a low one, Chile
had an increase in inflation inertia without changes in the small
slope of the Phillips curve.
The estimation of the structural parameters associated with
the interest rate reaction function indicates that when going from
a low interest response to a high one as it happened chronologically
in all countries (except Chile), Brazil exhibited increased persis-
tence of interest rates, decreased sensitivity to output deviations,
and no change in response to exchange rate fluctuations. Colombia

450 S. Cadavid, A. Ortiz

exhibited similar persistence of interest rates, less sensitivity to out-
put deviations, and more sensitivity to exchange rate fluctuations.
Mexico exhibited smaller persistence of interest rates and smaller
sensitivity to output deviations and exchange rate fluctuations. Peru
exhibited higher persistence of interest rates, lower sensitivity to out-
put deviations and similar responses to exchange rate fluctuations.
Finally, for Chile, when moving from a high interest rate response
to a low one, interest rates exhibited less persistence and the weight
on output deviations was larger, as expected from the countercycli-
cal stance of their monetary policy.
When comparing the impulse response functions under the two
regimes, we notice some differences in magnitude and sign. An un-
expected increase in monetary policy, appreciates the currency,
while it lowers inflation and output. Under high policy response re-
gimes appreciations are larger in Chile and Peru, where real inter-
est rates increase by more and inflation drops are larger. Only in the
case of Chile has the observed output contraction been larger un-
der the high policy response regime. This may be explained by the
fact that the Chile’s low response regime was implemented with
countercyclical motives only once the inflation targeting regime
was consolidated.
Our counterfactual analysis allows us to argue that the regime
switches towards more responsive interest rate reaction functions
helped to avoid many inflationary runs, several large nominal ex-
change rate depreciations and large volatility of the nominal vari-
ables. This reduction of nominal volatility did not come at the cost
of smaller output growth or the need of larger output fluctuations.
Therefore, we conclude that the introduction of inflation targeting
represented a favorable regime switch in the conduct of monetary
policy in Latin America.

Regime Switch of Monetary Policy in Latin America 451

 ANNEX

A. Estimated Parameters

Table 4
ESTIMATED PARAMETERS OF BRAZIL

Parameter Distribution Mean Mode Standard dev. 10% 90%

Beta 0.1738 0.0482 0.1299 0.0319 0.4303

Beta 0.4471 0.3213 0.2214 0.1352 0.8285

Gamma 1.1362 0.9582 0.2401 0.8484 1.6328

Gamma 0.6296 0.4708 0.3204 0.27 1.2559

Beta 0.7629 0.7847 0.048 0.6917 0.8144

Beta 0.6113 0.7513 0.1814 0.2252 0.813

Gamma 3.4901 3.6914 0.3406 2.733 3.8618

Gamma 1.0417 0.7656 0.296 0.6815 1.4375

Gamma 0.3013 0.1377 0.3081 0.075 0.9818

Gamma 0.8799 0.378 0.6633 0.2204 2.0191

Gamma 0.0435 0.0323 0.025 0.0156 0.098

Gamma 0.0422 0.0268 0.0391 0.0139 0.1547

α Beta 0.076 0.0436 0.0454 0.0291 0.1778

r Gamma 3.6731 3.0345 0.7512 2.6661 4.8276
τ Beta 0.2792 0.2896 0.1012 0.147 0.4755
ρa Beta 0.424 0.0606 0.2548 0.0349 0.7505

452 S. Cadavid, A. Ortiz

Parameter Distribution Mean Mode Standard dev. 10% 90%
ρq Beta 0.3014 0.1127 0.2563 0.0512 0.8563
ρy* Beta 0.9818 0.999 0.0356 0.9387 0.9992
ρπ* Beta 0.3715 0.3471 0.0892 0.2455 0.535
Beta 0.1603 0.0428 0.1461 0.039 0.4719

Beta 0.2257 0.1262 0.1173 0.0997 0.4375

Inv.Gamma 5.3145 6.0774 0.704 4.0033 6.2124

Inv.Gamma 3.3642 2.3893 1.0409 2.262 5.2165

Inv.Gamma 5.791 6.202 0.342 5.2568 6.2795

Inv.Gamma 4.2554 3.7875 0.9045 3.1056 5.8516

Inv.Gamma 4.6972 4.5965 0.1259 4.5092 4.9424

Inv.Gamma 4.7999 4.6046 0.2162 4.5204 5.188

Inv.Gamma 3.5522 2.4173 0.9944 2.3191 5.2352

Inv.Gamma 6.9291 7.8618 1.0058 5.0694 8.0384

Inv.Gamma 4.8214 4.0789 0.6289 4.002 5.8516

Inv.Gamma 6.1201 6.72 0.7868 4.811 7.1394

Beta 0.3071 0.2284 0.1399 0.1241 0.5589

Beta 0.1458 0.0487 0.1472 0.0278 0.4982

Regime Switch of Monetary Policy in Latin America 453

 Table 5
ESTIMATED PARAMETERS OF CHILE

Parameter Distribution Mean Mode Standard dev. 10% 90%

Beta 0.2053 0.1535 0.0715 0.1027 0.3366

Beta 0.5124 0.7801 0.1992 0.1913 0.8204

Gamma 0.0765 0.0648 0.03 0.0368 0.1346

Gamma 0.0631 0.0363 0.0208 0.0331 0.1008

Beta 0.9215 0.8787 0.0462 0.8525 0.9788

Beta 0.4912 0.4859 0.0359 0.4328 0.5514

Gamma 2.7337 1.2134 1.6982 1.079 5.4875

Gamma 0.8692 0.8601 0.0915 0.7058 1.0166

Gamma 0.5594 0.3792 0.2495 0.3015 0.8963

Gamma 0.434 0.4119 0.1508 0.2317 0.7397

Gamma 0.0816 0.0441 0.0774 0.0229 0.2694

Gamma 0.0662 0.0561 0.0334 0.026 0.1325

α Beta 0.0539 0.0526 0.0144 0.0331 0.0798

r Gamma 2.2813 1.6134 0.9063 0.9927 3.9223
τ Beta 0.16 0.1537 0.0389 0.1068 0.2349
ρa Beta 0.1553 0.1155 0.0403 0.0925 0.224
ρq Beta 0.1599 0.173 0.0632 0.0639 0.2696

454 S. Cadavid, A. Ortiz

 Parameter Distribution Mean Mode Standard dev. 10% 90%
ρ y* Beta 0.9579 0.9601 0.0134 0.9344 0.9784
ρ π* Beta 0.3129 0.2894 0.0601 0.2253 0.4259
Beta 0.0808 0.0234 0.0633 0.0141 0.21

Beta 0.0521 0.0288 0.0222 0.0225 0.0942

Inv.Gamma 0.5788 0.708 0.0994 0.3924 0.7324

Inv.Gamma 3.3239 0.5593 1.72 0.5233 5.5594

Inv.Gamma 6.4758 8.2896 0.9905 5.2872 8.4767

Inv.Gamma 5.3403 2.6141 1.7776 2.2494 7.698

Inv.Gamma 3.9563 3.8918 0.3844 3.4115 4.7018

Inv.Gamma 6.1979 3.9948 1.5389 3.8555 8.1539

Inv.Gamma 3.4781 3.5359 0.5735 2.7111 4.5585

Inv.Gamma 5.4652 3.8204 1.2679 3.5727 7.7182

Inv.Gamma 7.2118 8.1589 0.9146 5.7006 8.754

Inv.Gamma 4.6023 2.8905 1.2114 2.8584 6.702

Beta 0.0307 0.0275 0.0149 0.0107 0.0589

Beta 0.182 0.1298 0.0535 0.1096 0.2873

Regime Switch of Monetary Policy in Latin America 455

 Table 6
ESTIMATED PARAMETERS OF COLOMBIA

Parameter Distribution Mean Mode Standard dev. 10% 90%

Beta 0.7092 0.3151 0.1375 0.4474 0.8981

Beta 0.313 0.592 0.1163 0.1498 0.5307

Gamma 0.5845 0.6924 0.1267 0.3863 0.8068

Gamma 1.9982 1.6185 0.2196 1.6591 2.3484

Beta 0.7298 0.8046 0.0436 0.6633 0.8071

Beta 0.7065 0.6574 0.0349 0.6491 0.7621

Gamma 3.2941 1.8019 1.1685 1.8292 4.9853

Gamma 0.9746 0.8772 0.1204 0.7722 1.1641

Gamma 0.3849 0.2018 0.1315 0.1969 0.6058

Gamma 0.7379 0.5394 0.3263 0.3355 1.2305

Gamma 0.137 0.1147 0.0206 0.1068 0.1752

Gamma 0.0463 0.0296 0.042 0.0148 0.0844

α Beta 0.1132 0.084 0.0292 0.0722 0.1641

r Gamma 6.8509 5.4332 0.5786 5.8481 7.6151
τ Beta 0.2445 0.1537 0.0852 0.1398 0.4253
ρa Beta 0.1628 0.1377 0.0336 0.1122 0.2207
ρq Beta 0.1291 0.1212 0.0685 0.0355 0.2556

456 S. Cadavid, A. Ortiz

Parameter Distribution Mean Mode Standard dev. 10% 90%
ρy* Beta 0.9659 0.9619 0.015 0.9391 0.9864
ρπ* Beta 0.2303 0.2119 0.0575 0.1442 0.3319
Beta 0.0863 0.0163 0.0566 0.0239 0.2236

Beta 0.1566 0.0367 0.0871 0.048 0.3472

Inv.Gamma 0.8134 0.3186 0.4212 0.3368 1.5494

Inv.Gamma 6.8695 5.962 0.4012 6.2046 7.4314

Inv.Gamma 5.5065 5.5265 0.7479 4.3785 6.5215

Inv.Gamma 7.2084 6.3208 0.5752 6.3062 8.2431

Inv.Gamma 5.0036 5.7915 0.6898 3.9735 5.8133

Inv.Gamma 6.0725 8.2629 1.1733 4.3541 7.8001

Inv.Gamma 1.6996 0.6976 0.8043 0.5943 2.9033

Inv.Gamma 3.0673 2.673 0.3084 2.6163 3.5536

Inv.Gamma 5.0864 5.0467 0.3212 4.3974 5.4164

Inv.Gamma 2.4292 2.907 0.4745 1.6841 3.073

Beta 0.0607 0.0553 0.0809 0.0089 0.2931

Beta 0.1023 0.136 0.0601 0.0257 0.2056

Regime Switch of Monetary Policy in Latin America 457

 Table 7
ESTIMATED PARAMETERS OF MEXICO

Parameter Distribution Mean Mode Standard dev. 10% 90%

Beta 0.8564 0.9444 0.1206 0.6316 0.9739

Beta 0.6134 0.7351 0.0816 0.496 0.7669

Gamma 2.1643 2.2281 0.1162 1.9357 2.3318

Gamma 2.3736 2.0484 0.4645 1.7729 3.3246

Beta 0.458 0.4138 0.0551 0.3897 0.5541

Beta 0.6279 0.735 0.142 0.3992 0.7734

Gamma 1.8458 1.7333 0.0627 1.7431 1.9526

Gamma 0.6154 0.8004 0.1313 0.4424 0.823

Gamma 0.7265 0.7031 0.0629 0.602 0.8016

Gamma 0.8310 0.8491 0.1824 0.8039 0.8562

Gamma 0.1108 0.1093 0.0335 0.0961 0.1254

Gamma 0.3408 0.0899 0.2613 0.0775 0.6386

α Beta 0.2689 0.238 0.0362 0.2123 0.3289

r Gamma 2.1004 1.7134 0.2971 1.6491 2.5185
τ Beta 0.3256 0.3347 0.0216 0.2756 0.3478
ρa Beta 0.2007 0.2273 0.0444 0.1302 0.2724
ρq Beta 0.4305 0.3102 0.0879 0.2889 0.5608

458 S. Cadavid, A. Ortiz

Parameter Distribution Mean Mode Standard dev. 10% 90%
ρy* Beta 0.9042 0.9236 0.0217 0.8646 0.9359
ρπ* Beta 0.7824 0.8252 0.0428 0.7059 0.8408
Beta 0.1161 0.1094 0.0361 0.0707 0.1842

Beta 0.2108 0.2528 0.061 0.097 0.3049

Inv.Gamma 4.5438 4.5083 0.1139 4.3641 4.7386

Inv.Gamma 5.8216 5.8513 0.038 5.7552 5.8765

Inv.Gamma 3.121 3.0513 0.0634 3.012 3.2223

Inv.Gamma 4.4066 4.3941 0.0598 4.3069 4.5035

Inv.Gamma 3.2222 3.2116 0.0709 3.101 3.3199

Inv.Gamma 7.4444 7.3618 0.1203 7.2862 7.6952

Inv.Gamma 6.7571 6.7489 0.0777 6.6572 6.9247

Inv.Gamma 7.3328 7.3367 0.0746 7.2085 7.4538

Inv.Gamma 5.09 5.0717 0.0477 5.0186 5.1741

Inv.Gamma 9.5155 9.4522 0.0774 9.3967 9.6475

Beta 0.1922 0.1021 0.0825 0.0958 0.339

Beta 0.109 0.0925 0.0397 0.0577 0.1836

Regime Switch of Monetary Policy in Latin America 459

 Table 8
ESTIMATED PARAMETERS OF PERU

Parameter Distribution Mean Mode Standard dev. 10% 90%

Beta 0.1318 0.0928 0.0787 0.0321 0.2885

Beta 0.1471 0.1609 0.0779 0.0352 0.286

Gamma 0.5011 0.3816 0.1019 0.3481 0.6833

Gamma 0.0565 0.0672 0.0171 0.0294 0.0863

Beta 0.697 0.7132 0.0412 0.6211 0.753

Beta 0.6254 0.6094 0.0656 0.5227 0.7344

Gamma 1.9066 1.4844 0.5911 1.3059 3.309

Gamma 0.9226 0.5921 0.5032 0.444 1.7992

Gamma 0.4092 0.2172 0.2179 0.1659 0.859

Gamma 0.5639 0.4629 0.2286 0.3263 1.0481

Gamma 0.1725 0.1612 0.0326 0.1215 0.2283

Gamma 0.1506 0.1693 0.0247 0.1139 0.1925

α Beta 0.0393 0.0389 0.0166 0.0201 0.0757

r Gamma 8.8041 1.8227 4.3353 1.4432 13.308
τ Beta 0.1306 0.0582 0.0516 0.0522 0.2153
ρa Beta 0.3605 0.3714 0.0521 0.2759 0.4462
ρq Beta 0.3924 0.3134 0.0721 0.2687 0.5028

460 S. Cadavid, A. Ortiz

Parameter Distribution Mean Mode Standard dev. 10% 90%
ρy* Beta 0.9682 0.9756 0.0133 0.9445 0.9877
ρπ* Beta 0.416 0.3717 0.0559 0.3277 0.5085
Beta 0.0721 0.0662 0.0257 0.0276 0.1129

Beta 0.0565 0.0615 0.0265 0.0191 0.101

Inv.Gamma 2.4271 1.0314 1.8471 0.9785 6.2415

Inv.Gamma 7.6316 7.4037 1.2162 5.286 9.3854

Inv.Gamma 4.1378 3.6176 0.7587 3.5144 6.1017

Inv.Gamma 5.1138 3.2364 2.213 2.7457 8.9518

Inv.Gamma 2.7075 2.2299 0.8397 2.0503 4.9969

Inv.Gamma 6.0456 3.837 1.7465 3.4937 8.618

Inv.Gamma 2.1448 0.2633 1.2789 0.2842 4.4775

Inv.Gamma 3.5942 0.2823 2.6484 0.3459 8.0066

Inv.Gamma 5.0435 4.6914 0.5071 4.3391 6.001

Inv.Gamma 5.0472 4.5065 1.3062 3.1803 7.6507

Beta 0.0849 0.0213 0.1287 0.0103 0.4463

Beta 0.1719 0.0582 0.1171 0.0427 0.4136

Regime Switch of Monetary Policy in Latin America 461

References
Alesina, A. and Summers, L. H. (1993). Central Bank Independence
and Macroeconomic Performance: Some Comparative Evi-
dence. Journal of Money Credit and Banking, 25(2):151- 162.
Bianchi, F. (2012). Regime Switches, Agents’ Beliefs, and Post-
World War ii us Macroeconomic Dynamics. Review of Economic
Studies, 80(2):463-490.
Bianchi, F. and Ilut, C. (2017). Monetary/Fiscal Policy Mix and
Agents Beliefs. Review of Economic Dynamics (forthcoming).
Cho, S. (2016). Sufficient Conditions For Determinacy in a Class
of Markov-Switching Rational Expectations Models. Review
of Economic Dynamics, 21:182-200.
Cogley, T. and Sargent, T. J. (2005). Drifts and Volatilities: Monetary
Policies and Outcomes in the Post wwii us. Review of Economic
Dynamics, 8(2):262-302.
Corbo, V., Landerretche, O., Schmidt-Hebbel, K., et al. (2002). Does
Inflation Targeting Make a Difference? Inflation Targeting:
Design Performance Challenges, 5:221-69.
Farmer, R. E., Waggoner, D. F., and Zha, T. (2008). Minimal State
Variable Solutions to Markov-Switching Rational Expectations
models, Working Paper, No. 2008-23a, Federal Reserve Bank
of Atlanta.
Farmer,R. E.,Waggoner,D. F.,and Zha,T. (2011).Minimal State
Variable Solutions to Markov-Switching Rational Expec-
tations models. Journal of Economic Dynamics and Control,
35(12):2150-2166.
Foerster, A., Rubio-Ramirez, J., Waggoner, D. F., and Zha, T. (2014).
Perturbation Methods for Markov-Switching dsge Models. Techni-
cal Report, National Bureau of Economic Research.
Gali, J. and Monacelli, T. (2005). Monetary Policy and Exchange
Rate Volatility in a Small Open Economy. The Review of Eco-
nomic Studies, 72(3):707-734.
Garriga, A. C. (2016). Central Bank Independence in the World:
A New Data Set. International Interactions, 42(5):849-868.
Goncalves, C. C. S., Portugal, M. S., and Aragón, E. K. d. S. B.
(2016). Assessing Brazilian Macroeconomic Dynamics Using
a Markov-Switching dsge Model. EconomiA, 17(1):23-42.

462
Kim, C.-J., Nelson, C. R., et al. (1999). State-Space Models with Regime
Switching: Classical and Gibbs-Sampling Approaches with Applica-
tions. mit Press Books, 1.
Liu, P. and Mumtaz, H. (2011). Evolving Macroeconomic Dynamics
in a Small Open Economy: An estimated Markov Switching
dsge Model for the uk. Journal of Money Credit and Banking,
43(7):1443-1474.
Lubik, T. A. and Schorfheide, F. (2007). Do Central Banks Respond
to Exchange Rate Movements? A Structural Investigation.
Journal of Monetary Economics, 54(4):1069-1087.
Maih, J. (2015). Efficient Perturbation Methods for Solving Regime-Switch-
ing dsge Models. Working Paper 2015/01, Norges Bank.
Ortiz, A. and Sturzenegger, F. (2007). Estimating Sarb’s Policy
Reaction Rule. South African Journal of Economics, 75(4):659-
680.
Schmitt-Grohé, S. and Uribe, M. (2003). Closing Small Open
Economy Models. Journal of International Economics, 61(1):163-
185.
Sims, C. A. (2002). Solving Linear Rational Expectations Models.
Computational Economics, 20(1):1-20.
Sims, C. A. and Zha, T. (2006). Were There Regime Switches in us
Monetary Policy?, The American Economic Review, 96(1):54-81.

Regime Switch of Monetary Policy in Latin America 463

Fiscal Policy and Inflation:
Understanding the Role of
Expectations in Mexico
Bernabe López Martin
Alberto Ramírez de Aguilar
Daniel Sámano

Abstract
We estimate a hidden Markov model where inflation is determined by gov-
ernment deficits financed through money creation and/or by destabilizing
expectations dynamics (expectations can potentially divorce inflation from
fundamentals). The baseline model, proposed by Sargent et al. (2009), is used
to analyze the interaction between fiscal deficits, inflation expectations,
and inflation in Mexico. The model is able to distinguish between causes
and remedies of hyperinflation, such as persistent or transitory shocks to sei-
gniorage-financed fiscal deficits, de-anchoring of inflation expectations from
fiscal fundamentals, and cosmetic (non-fundamental) monetary reforms.
The behavior of monetized deficits provides an adequate account of high in-
flation episodes and stabilizations for the period 1969-1994. We then extend

This research was developed within the framework of cemla’s Joint Research Program
2017 coordinated by the Central Bank of Colombia. We thank counselling and techni-
cal advisory provided by the Financial Stability and Development (FSD) Group of the
Inter-American Development Bank in the process of writing this document. We are
especially grateful to Daniel Chiquiar and our academic advisor Andrea Tambalotti,
for their feedback during this project. We would like to thank, for thoughtful com-
ments and suggestions, Nicolás Amoroso, Oscar Budar Mejía, Julio Carrillo, Alfonso
Cebreros, Gerardo Hernández del Valle, Juan Ramón Hernández, Raul Ibarra, Emil-
iano Luttini, André Martínez-Fritscher, Felipe Meza, Alberto Ortiz-Bolaños, Fernando
Pérez-Cervantes, Manuel Ramos-Francia, Juan Sherwell, our discussant Christian
Velásquez, and the participants of the cemla Research Network and the XXII An-
nual Meeting of the Central Bank Research Network (Bogota, 2017). The opinions
expressed in this publication are those of the authors and do not reflect the views
of cemla, the FSD Group, the Inter-American Development Bank, or those of Banco
de México. E-mails: [email protected], [email protected],
and [email protected].

465
the model to analyze the possibility that fiscal policy can affect inflation expec-
tations in a context of Central Bank independence, as is the case of Mexico
after 1994. We find evidence that the exchange rate and sovereign interest
rate spreads influence the evolution of aggregate prices.
Keywords: inflation, inflation expectations, fiscal policy.
jel: E31, E42, E52, E63.

1.INTRODUCTION

As in other countries in Latin America during the second half of

the twentieth century, Mexico suffered several episodes of annual
inflation rates above one hundred percent. These high inflation epi-
sodes were typically accompanied by elevated levels of public deficit
financed with monetary expansions.1 Until 1994, a regime of fiscal
dominance prevailed, where the Central Bank adjusted its monetary
policy to the financial requirements of the fiscal authority. Thereaf-
ter, the autonomy of Banco de México was established and inflation
started a process of moderation.
To analyze the interaction between inflation, inflation expecta-
tions, and fiscal deficits in Mexico, we utilize the model developed
by Sargent et al. (2009). This model has been used to infer the de-
terminants of hyperinflations and stabilizations in different coun-
tries in Latin America (Argentina, Bolivia, Brazil, Chile, and Peru).
It gives a central role to government deficits financed through money
creation, but also to destabilizing expectations that can, under cer-
tain conditions, divorce inflation from fundamentals. The baseline
framework consists of a non-linear hidden Markov model with the fol-
lowing key components: (i) a standard demand function for real bal-
ances, an adaptive scheme for the expected rate of inflation, 2 (iii)
a government budget constraint that relates fiscal deficits to monetary

1
Fischer et al. (2002), Catao and Terrones (2005), and Lin and Chu
(2013), among others, document international evidence regarding
the relationship between ination rates, scal decits, and money supply.
Rogers and Wang (1994) estimate that between 1977 and 1990, scal
and monetary shocks accounted for 60 percent of the variance of ina-
tion in Mexico.
2
Agents have adaptive expectations or backward-looking expectations
when these are formed by extrapolating past values of the variable be-
ing predicted.

466 B. López Martín, A. Ramírez de Aguilar, D. Sámano

supply, and (iv) a stochastic fiscal deficit that follows a hidden Mar-
kov process. With these components, the model is able to distinguish
between the causes and remedies of hyperinflations, such as per-
sistent or transitory shocks to seigniorage-financed fiscal deficits,
de-anchoring of inflation expectations from fiscal fundamentals,
and cosmetic (non-fundamental) monetary reforms. Sargent et al.
(2009) conclude that the behavior of monetized deficits determined
most hyperinflations and stabilizations for the set of countries they
studied.
We first use the baseline model to account for the evolution of in-
flation in Mexico between 1969 and 2016. The methodology uses a se-
ries for inflation, interpreting the density of the inflation series as a
likelihood function in order to estimate the history of fiscal deficits
and the process of the formation of inflation expectations that better
account for the evolution of inflation. This approach is convenient
given numerous methodological modifications in the construction
of public accounts, the sometimes less-than-ideal transparency in his-
torical series, and the fluctuations in the perception of economic
agents of what constitutes fiscal responsibility for the government
(e.g., bailouts of the financial system or sub-national governments).
These problems plague historical accounts of events in developing
economies. The estimated sequence of fiscal deficits is then compared
to available data for government deficits and a historic narrative of the
events associated with episodes of high inflation and stabilizations.
In line with the results for other countries, the model suggests that
the evolution of fiscal deficits is central in explaining the behavior
of inflation in Mexico. Furthermore, it provides a description of the
formation of inflation expectations. For example, the parameters
of the model suggest that inflation must be high for several consecu-
tive periods in order to de-anchor inflation expectations and gener-
ate an inflation spiral.
For the period of decreasing inflation that started in the second
half of the 1990s, the baseline model suggests that the level of fiscal
deficits financed through monetary expansion is modest. This inter-
pretation, however, is not fully satisfactory as the Central Bank be-
came independent in 1994. Thus, a theory that contemporaneously
links inflation to fiscal deficits through the monetary channel seems
lacking if we aim to understand inflation after 1994. This motivates
the following question; can we find evidence that fiscal policy affects

467
inflation and inflation expectations even in the context of Central
Bank independence?
A strand of the macroeconomic literature proposes that fiscal
policy is relevant to achieving price stability even in an environ-
ment where monetary policy is conducted by an independent Cen-
tral Bank. 3 We extend the baseline model along several dimensions
with the objective of documenting evidence, perhaps indirect, or re-
butting the possibility that fiscal policy is relevant in determining
inflation and inflation expectations in a context of Central Bank in-
dependence. A variable of interest we consider is the spread in the
sovereign interest rate embi. This variable, which can be considered
forward-looking, reflects the fiscal situation of the government.
To the extent that economic agents perceive potential risks in terms
of the ability of the government to make debt payments, it may also
affect the credibility of the Central Bank. The perception of this
type of risk is incorporated in the prices of sovereign debt. The state
of public finances is often considered to affect the exchange rate;
this is the second variable we assess in the model. The results indi-
cate that both variables are relevant in determining inflation expec-
tations and inflation.4
We proceed as follows: Section 2 presents the baseline model
and describes the mechanisms that drive the behavior of the different
variables. Section 3 presents the main results for the baseline model:
(1) the parameter values of the model and their implications in terms

3
There exists a vast literature studying the relevance of fiscal policy
and its interaction with monetary policy for the determination of infla-
tion, a seminal paper is Sargent and Wallace (1981). Though we will
not attempt to provide an exhaustive set of references, some addi-
tional examples are provided by Sims (2016), Leeper (1991), Davig
et al. (2011), Sargent and Zeira (2011), Woodford (2001), and Bianchi
and Ilut (2017). For an introductory treatment of the fiscal theory
of the price level, see Christiano and Fitzgerald (2000). Central Banks
frequently express concern related to how fiscal imbalances may affect
the effectiveness of monetary policy (e.g., Carstens and Jácome (2005)
and Ramos-Francia and Torres-Garcia (2005)).
4
There are different mechanisms through which these variables could
potentially be relevant; we explore the impact through expectations
and the demand for real money balances. We discuss the evidence
of the extent to which these variables are influenced by international
and exogenous factors, with a focus on the case of Mexico, such as prices
of commodities in global markets.

468 B. López Martín, A. Ramírez de Aguilar, D. Sámano

of the behavior of the main variables, (2) a comparison of the infla-
tion series generated by the model and those observed in the data,
with a historical account of the events associated with the different
inflation and stabilization episodes, and (3) a comparison of the se-
ries for fiscal deficits generated by the model with the historical se-
ries. Section 4 presents the extensions of the model and the main
results. Section 5 provides our concluding remarks.

2. THE BASELINE MODEL

The baseline model is the one featured in Sargent et al. (2009), con-
structed to study the relationship between inflation, fiscal deficits,
and inflation expectations. An ad- vantage of this model is its sim-
ple structure, which allows for the estimation of its parameters us-
ing only the historic series of one of the main variables, in our case
the monthly inflation series (the estimation algorithm is described
briefly in the next section and in the Appendix). With these param-
eters, the model accounts for an observed sequence of inflation as a
result of fiscal deficits and a particular process for the formation
of inflation expectations. The framework consists of three main
components: a money demand function, the budget constraint of the
government, a process that models the formation of expectations,
and the (exogenous and stochastic) evolution of deficits. We now de-
scribe each of these components.

2.1 The Money Demand and the Government Budget

Constraint
A standard money demand equation (e.g. Cagan (1956)) establishes
a relationship between the nominal balances as a percentage of out-
put Mt at time t, the price level P t at time t, and the expectations
of agents of the price level for period t + 1:5

5
In a seminal paper, Cagan (1956) specifies a demand for real balances
and backward-looking expectations to explain several European hyper-
inflation episodes.

469
where represents the weight that the expected price level
has on the current price level P t , and is the weight that
the nominal balances relative to output have on the price level at time
t.6 Thus, if the public expects a higher price level in t + 1, their real
balances demand Mt/P t will fall.
The next equation represents the budget constraint of the gov-
ernment, where dt (a stochastic variable) is the part of the real defi-
cit of the government that is monetized (net of debt emissions, so it
must be covered by printing money). Thus, the growth of nominal
balances per unit of output is determined according to the follow-
ing equation:

where parameter adjusts for growth in real output and tax-

es on cash balances.7 This equation implies that larger fiscal deficits
are associated with increases in the level of nominal balances as a
percentage of gdp. 8
We let denote the gross expected inflation rate. Us-
ing (1) and (2) it can be shown that the gross inflation rate at time t is:

Equation (1) can be written as Pt = γMt + λPt +1 . Hence, {λ, γ} represent

6 e
e
the weights that Pt +1 and Mt have on Pt, respectively.
7
Parameter is related to output growth in the model. Let

where are the nominal balances at time t and Yt is output. If Dt

represents the level of real fiscal deficit at time t, then the government
budget constraint is Dividing this equation by Yt

then: Therefore, can be interpreted as the

inverse of the output growth factor. Consequently, this model is assum-

ing a constant output growth rate. Quantitatively, this parameter is not
relevant for our results.
8
We are defining the fiscal deficit as where
gt and τt represent government expenditures and revenues relative
to output, bt is the level of sovereign debt relative to output and rt is the
interest rate on sovereign debt.

470 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 3

This equation suggests that inflation is a function of two variables:

the expected gross inflation rate and the real fiscal deficit. Accord-
ing to (3), if the expected gross inflation rate βt or fiscal deficit dt rise,
current inflation πt will also increase.9 It is worth mentioning that,
equation (3) does not depend on the particular process through
which inflation expectations are formed, or the stochastic process
assumed for fiscal deficits. Nevertheless, these assumptions are cru-
cial to determine a sequence of inflation rates according
to the model. The next two sections will explain the specification
for the evolution of expectations and the dynamics followed by the
real fiscal deficit.

2.2 Inflation Expectations

The baseline specification follows, for example, Marcet and Nicolini
(2003), assuming that the public updates their beliefs on future infla-
tion βt using adaptive expectations. According to Sargent and Wal-
lace (1973), agents have adaptive expectations when they take into
account past information to extrapolate it to form their expecta-
tions. Specifically in this model, the gross expected inflation rate
is a weighted average between the gross inflation rate and the gross
expected inflation lagged one period:

where is the weight that expectations give to past observed

inflation. In related literature, this particular type of adaptive expec-
tations is known as constant-gain expectations, given the constant
weight in the process that determines the formation of expectations.10

9
This is obtained with λ∈(0,1), θ∈(0,1), and γ>0.
10
For example, Branch (2004) develops a micro-founded model where
agents optimally choose not to update their beliefs according to a
rational expectations algorithm because the information it requires
is too costly (rational expectations algorithms usually require a lot
of information). In the type of models we are considering, adaptive

471
 Assuming constant-gain expectations (cge) is key in determining
the dynamics of the model. Panel (a) of Figure 1 shows the change
in gross inflation πt+1−π t as a function of expectations βt , with a con-
stant real fiscal deficit. As shown in the Figure, there are two values
of β that imply a constant inflation equilibrium: β1 and β 2 . In the
adaptive expectations literature, β1 and β 2 are known as self-confirm-
ing equilibria. As implied by the Figure, β1 is a locally stable equi-
librium, thus, if the beliefs of the public regarding future inflation
are not sufficiently high then πt+1−πt will converge to zero and βt+1
to β1. Additionally, equation (4) implies that πt will also converge
to β1. However, if βt > β 2, then πt+1−πt will increase, with unbounded
dynamics. Therefore, βt>β 2 implies that the model will eventually
generate a hyperinflation episode. This phenomenon is called es-
cape dynamics by Sargent et al. (2009).11
Panel (b) of Figure 1 presents another result of cge: assuming
β t induces escape dynamics, a hyperinflation episode can be pre-
vented if the deficit is reduced. This Panel shows two dynamic paths
for πt+1−π t as a function of βt . The only difference between these paths
is the level of fiscal deficit. The dynamics shown in blue correspond
to a high fiscal deficit, while the dynamics in green correspond to a
low fiscal deficit. Assuming a high deficit and βt = βˆ , if the deficit
is not reduced then it will provoke an escape dynamics of inflation
and expectations as shown with blue arrows in Panel (b) of Figure 1.
However, if the government reduces its fiscal deficit to a sufficiently
low level then, even when βt = βˆ , it will be able to prevent an escape
dynamics. Furthermore, πt+1−π t will converge to a low and stable in-
flation equilibrium as shown by the green arrows in the Figure.
Finally, cge implies a non-trivial computational advantage: given
the complexity of the function that will be used to estimate all the pa-
rameters involved in the model, assuming this type of expectations

expectations or other deviations from rational expectations, can be

necessary to generate hyperinflation episodes (e.g. Sallum et al. (2005)).
See Sargent et al. (2009) for a list of references in a growing literature
using calibration or econometric techniques to compare time-series
data with models in which agents use this type of algorithm to form
their beliefs.
11
Williams (2016) characterizes how adaptive expectations can lead to es-
cape dynamics and ex- plains how the likelihood, frequency and direc-
tion of the variables during an escape dynamics can be characterized
by a deterministic control problem.

472 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 1
DYNAMICS INDUCED BY ADAPTIVE EXPECTATIONS

.   

80

60

40
πt+1−πt

20

−20

0 β1 5 βt 10 15 β2

.    

200 High dt

150

100
πt+1−πt

Low dt
50

β2
0
β1

−50
0 β15 βt 10 15 β̂

Note: these figures considers βt−1 =1.02 and the estimated parameters shown in Table 1.

473
allows us to reduce the computational burden.12 We discuss the im-
plications of using rational expectations in the Appendix.

2.3 The Process for Fiscal Deficits

The last key variable that determines inflation rates is the level of real
fiscal deficit relative to output dt . The fact that dt is assumed to be
a random variable is motivated by, among other factors according
to our interpretation, exogenous conditions in global financial mar-
kets, the international price of commodities that are crucial in de-
termining the fiscal situation of many governments in developing
economies, and political processes. With these considerations, in an
admittedly reduced form, it is assumed that dt is a random variable
with the following conditional distribution:

Thus, dt is a random variable with a log-normal distribution that

has a median of and a variance parameter vt . A restriction of as-
suming a log-normal distribution for fiscal deficits relative to output
is that dt cannot be negative (a fiscal surplus is not feasible). Sargent
et al. (2009) explain that even when they allow the distribution of dt
to have negative values, there is not a significant improvement in the
fit of the model. Furthermore, a log-normal distribution captures
the skewness of inflation shown in the data. In the case of Mexico,
we will see that three values for are sufficient to adequately cap-
ture the evolution of deficits during the period we analyze.
Each period, is determined by a discrete Markov process with
D possible states.13 In the same manner, vt follows another Markov
process with V states that is independent of the process that deter-
mines In related literature, the stochastic process followed by dt

12
The next section explains some of the details involved in estimating
the parameters of model.
13
A stochastic process xt is said to be a discrete Markov process if xt takes
values in a set I with and for all the Markov property
is satisfied: This property states
that past realizations of the process do not affect future
values, only the present state xt affects xt+1.

474 B. López Martín, A. Ramírez de Aguilar, D. Sámano

is called a Hidden Markov Process.14 Each Markov process involved
in the model is related to a matrix where the elements represent
the transition probabilities from one state of the process to anoth-
er. We let be the transition matrix associated
to the processes, respectively.15
Another important property of the model is that it generates a non-
linear relation- ship between inflation, its expectations, and fiscal
deficits. The impact that current inflationary expectations βt have
on inflation πt and future expectations βt+1 is a function of the hid-
den Markov state that governs the median fiscal deficit An ex-
ample of the non-linearity generated by the hidden Markov process
of the model can be seen in Panels (a) and (b) of Figure II. Panel
(a) shows that, for the same level of βt , the effect of the fiscal deficit
on inflation is magnified as the median level of fiscal deficit rises
(this Figure considers ). Panel (b) displays a similar ef-
fect of fiscal deficit on the evolution of inflation expectations. This
non-linearity between the inflation rate, its expectations, and fiscal
deficits in the model is consistent with empirical studies. For exam-
ple, Catao and Terrones (2005) and Lin and Chu (2013) provide evi-
dence, utilizing data for more than 100 countries, that fiscal deficits
have a strong and weak impact on the inflation rate in high and low
inflation episodes, respectively. Thus, the data and the model sug-
gest that there is a non-linear impact of fiscal deficits on inflation
and expectations of inflation.

14
Formally, a hidden Markov process is a pair {x t, y t} such that x t is
a (standard) Markov process and there exists a function f such that
for all and:

In these type of processes, yt is known as the observable part of the

process and xt is the hidden component. In the model presented in this
section, yt is the real fiscal deficit relative to output while xt is a vector
that contains the median and variance vt of fiscal deficit at each t.
15
This means, in the case of in its (i, j) component con-
tains the probability of being in a state j in t+1 conditional
on dt = i : Q d (i , j ) = P dt +1 = j |dt = i  .

475
 Figure 2
NON-LINEAR EFFECT OF FISCAL DEFICITS

.   
—
d1
—
d2
60

40

—
πt

d3
20

0 5 10 15 20 25
βt

.   

40
— —
d1 d2
βt+1

20

—
d3

0
0 5 10 15 20 25

βt

Note: these figures consider βt−1 =1.02 and the estimated parameters as described in
the next section.

476 B. López Martín, A. Ramírez de Aguilar, D. Sámano

2.4 Model Restrictions on Expectations
Equation (3) implies that inflation in the model is well defined only
if at each t:1−λβ t−1>0 and 1−λβ t−γdt>0 (otherwise the real balances
demand could become negative). However, there is no restriction
within the model preventing these constraints from being violated.
Furthermore, (3) implies that the gross inflation rate is not bound-
ed.16 Given the numerical problems that this can generate when es-
timating the parameters, it is assumed that there exists a constant
δ>0 such that πt<δ for every t.
The two restrictions that need to be considered such that πt is well
defined and bounded are:

If any of these constraints is violated, then it is assumed that

the gross inflation rate is not determined following (3). Instead,
πt will be determined randomly according to the following log-nor-
mal distribution:

where is the inflation equilibrium determined by (3) in the

model without uncertainty and conditional to a certain fiscal deficit
dt ,17 whereas v π represents the variance of inflation when it is deter-
mined following (7). Additionally, if
Sargent et al. (2009) suggest resetting expected inflation to βt+1=πt ,
otherwise the dynamics between β t+1 and inf lation will provoke
and eventually β,π→∞.
Whenever the current hidden Markov state provokes
dynamics that will eventually make violate (6) or that will

16
17
Certainty in the model implies In equilibrium, Using

(3) it can be shown that:

477
generate an escape dynamics, the government can implement a re-
form to prevent this from happening. Sargent et al. (2009) define
two types of reforms: a reform is said to be cosmetic if the govern-
ment is able to (temporarily) control inflation but the median level
of fiscal deficit is not altered. Following Panel (a) of Figure 1, a cos-
metic reform can fail if the expected inflation rate associated with
inflation βt+1 is such that βt+1>β 2. However, a cosmetic reform can be
successful if 18
A structural reform, on the other hand, oc-
curs when the government is able to control the inflation rate by re-
ducing the median level of fiscal deficit, Panel (b) of Figure 1 is
an example of a structural reform where the government succeeded
in controlling an escape dynamics.
An important contribution of the model is its ability to identi-
fy whether a reform is cosmetic or structural. Previous literature
had only studied structural reforms, even though the notion of a
cosmetic reform was part of academic and economic policy discus-
sions. The inclusion of cosmetic reforms in the model represents
a reduced form approach to consider different episodes in Latin
America, when governments attempted to control inflation without
tackling fiscal deficits. Discussions of eco- nomic events often point
to the role of the exchange rate, which is not explicitly included in the
baseline model, and we explore below through different extensions
of the baseline model.

3. BASELINE MODEL RESULTS

In this section, we present the main results of the baseline mod-

el. We present the fit of the model for real fiscal deficits, inflation,
and its expectations between 1969 and 2016. Then, as a validation
procedure, we compare these model-fitted series with data available
for different variables.

3.1 Baseline Model Estimation

Heuristically, the estimated parameters are obtained as the vector
of values that maximize the likelihood function, which consists of the

18
Sargent et al. (2009) argue that in Peru a cosmetic reform was enough
to control the inflationary crisis this country experienced in 1985.

478 B. López Martín, A. Ramírez de Aguilar, D. Sámano

marginal density of the sequence of inflation.19 The inflation data
corresponds to the Índice Nacional de Precios al Consumidor (inpc) be-
tween 1969 and 2016, at a monthly frequency. The inpc is the Consum-
er Price Index (cpi) computed by the National Institute of Statistics
and Geography, Instituto Nacional de Estadística y Geografía (inegi)
since 2011, and by Banco de México before that year.
We consider a monthly frequency for the model estimation, con-
sistent with the data. Before estimating the parameters, one must
choose the number of states of nature for denoted D and V,
respectively. As D or V become larger, the fit of the model in terms
of approximating the data tends to improve at the expense of in-
creasing the computational burden. Sargent et al. (2009) estimate
two models for each country they study: a model with D =3, V=2 and a
model with D =2, V=3. Then, using the Schwarz information crite-
rion (sic), we select the model that provides a better fit to the data. 20
Table 1 shows the estimation results for a model with three possible
states for and two states for v (V=2). We choose this model
because, after estimating the two models with data for Mexico, the sic
suggests that D = 3, V = 2 provides a better approximation to the data.
The estimated parameters suggest interesting facts about the price
formation process in Mexico: λ=0.7556 implies that the price level
reflects agents’ expectations on the future price level. Hence, if in-
flation expectations are volatile, then the observed inflation will also
have a high variance. This result implies that a necessary condition
to have stable inflation is to anchor expectations. Mexico’s λ is simi-
lar to the estimation by Sargent et al. (2009) for Argentina (λ=0.730)
and Peru (λ=0.740).
The estimated value of ν=0.1147 for Mexico implies that to an-
chor expectations, observed inflation must remain stable for sever-
al months.21 On the other hand, this also implies that the expected

19
In the Appendix we provide further details regarding the estimation
of the model. Ramirez de Aguilar (2017) describes the computational
procedure.
20
The sic is a Bayesian selection criterion between two models, A and B.
Let Lx, Px, nx be the log-likelihood, the number of parameters, and the
sample size in model x ∈ {A, B}, respectively. Then, the Schwarz crite-
rion for model x is computed as SICx=log(nx)Px−2Lx. If SICA<SICB, then
model A is preferred.
21
The estimation of ν=0.1147 implies that the weight agents give to their
past expectations is 0.8853. Hence, if inflation is stable for only

479
 Table 1
PARAMETER ESTIMATION

Parameter Estimation Description

0.7556 (0.0022) weight of expectations on the price level

v 0.1147 (0.0081) weight of past inflation on expectations

0.0075 (0.0001) monthly high median level of fiscal deficits

0.0039 (0.0004) monthly moderate median level of fiscal deficits

0.0023 (0.0002) monthly low median level of fiscal deficits

v1 0.0671 (0.0087) high variance of monthly fiscal deficits

v2 0.0295 (0.0012) low variance of monthly fiscal deficits

0.0753 (0.0010) variance of inflation when it is determined randomly

0.9731 (0.0361) probability of conditional on

0.9787 (0.0390) probability of conditional on

0.9924 (0.0056) probability of conditional on

0.7493 (0.1072) probability of vt+1=v1 conditional on vt=v1

0.7789 (0.0879) probability of vt+1=v2 conditional on vt=v2

Note: the numbers shown in parentheses represent the standard deviation of each parameter,
computed using the Hessian matrix of the maximum likelihood problem (see MacDonald
and Zuccini (2009)).

480 B. López Martín, A. Ramírez de Aguilar, D. Sámano

inflation rate de-anchors only if the observed inflation is high for an
extended period. Sargent et al. (2009)’s estimations for Argentina
(ν=0.023), Chile (ν=0.025), and Peru (ν=0.069) indicate that, in these
countries, observed inflation has a relatively limited effect on infla-
tion expectations, while the estimates for Bolivia (ν=0.232) and Bra-
zil (ν=0.189), suggest that observed inflation has a stronger impact
on expectations.
Regarding fiscal deficits, according to the estimation, when
the government generates a high fiscal deficit for one year (
for twelve consecutive months), fiscal deficit represents approxi-
mately 9.12% of gdp. If the government generates a moderate deficit
for one year, this will amount to approximately 4.76% of gdp. Finally,
if fiscal deficits are low for one year, then it represents 2.78% of the
gdp. These levels of deficit are associated, in steady state, with aver-
age annual inflation rates of 69.41%, 17.53% and 3.54%, respectively.
As it will be shown, these estimates are consistent with fiscal deficit
data between 1977 and 2016.

3.2 Fiscal Deficits, Inflation, and Expectations

Once the parameters are estimated, fiscal deficits relative to output
can be computed in each period exploiting the assumptions made
for and considering that follow a discrete Markov
process. We estimate the conditional density of fiscal deficits given
the sequence of inflation observed in the data πT and the parameter
estimation, Then, we use the median of each density
to construct a sequence that is used to compute
according to the model. Finally, we compare the model implied se-
quence of inflation with the empirical series.
Figure 3 presents the model simulation for fiscal deficits, infla-
tion expectations, observed inflation, and the probability of a re-
gime change in
• Between 1969 and 1972, marked as Region (1) in Figure 3,
a low rate of inflation is associated with the lowest hidden state

one month, this will not be enough to reduce β because past beliefs
have more weight on expectations. Only if the inflation rate is stable
for several consecutive months will β also become stable.

481
 of median deficit This is consistent with the economic his-
tory of Mexico; during the decade of the 1960s, the inflation
rate in Mexico achieved its lowest value during the second half
of the twentieth century: an average of 2.8%, which is repli-
cated by the model.22

• Between 1973 and 1982, marked as Region (2) of the Figure,

the model suggests that fiscal deficits increased from a low to a
moderate median level, accompanied by an increase of the
inflation rate. Since this level of deficit remained constant
for several years, inflation expectations de-anchored. Conse-
quently, the observed inflation rate also presented an increase
between 1973-1982. At the end of 1971, a global recession re-
duced international credit. Fearing a period of stagnation,
the government responded by increasing public expenditures
financed with monetary emission, foreign credit, and reserves
of private financial institutions at the Central Bank. The fiscal
deficit relative to output increased from 2.5% of gdp in 1971
to 4.9% in 1972, while the monetary base grew 14.8% during
1972, the rate of inflation registered an average of 14% during
1973-1976. Meanwhile, government expenditures increased
from 30.9% relative to output to 40.6% in 1981; the fiscal defi-
cit relative to output rose from 6.7% in 1977 to 14.1% in 1982.

• In 1981, the world economy was going through another reces-

sion that once again reduced international credit. In Mexico,
there was not a significant reduction in expenditures and by
1982 the lack of foreign credit led the government to finance
most of its expenditures with monetary emission: between
1981 and 1983, the monetary base was growing at an average
rate of approximately 90% and the inflation rate was 63.1%
on average. During 1983, the model generates an inflation
rate above 80% as a result of an increase in fiscal deficits,
which reached their highest median level. During 1983-1986,
the government raised taxes and renegotiated its foreign debt.

22
In this section we draw from Cardenas (2015), who provides an exhaus-
tive narrative of the economic history of Mexico during the period
of our analysis. Historical series for output and the inflation rate
data presented in this section were obtained in the Historic Statistics
of Mexico published by inegi.

482 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 However, there was not a significant adjustment of expendi-
tures; by 1986 the fiscal deficit reached the same level it regis-
tered in 1982, equal to 14.1% of gdp. In 1985 world oil prices fell
and by 1986 the price of the Mexican oil mix suffered a drop
of 65%, generating a loss equivalent to 6.5% of gdp and a re-
duction of 26% in federal income. By 1987, the annual infla-
tion rate was 159%.23

• Region (3) of Figure 3 presents evidence of a cosmetic reform,

to control inflation: during 1984 the government was able
to reduce inflation from 85% to 56%, according to the mod-
el, due to a temporal reduction of its fiscal deficit. However,
as shown by Panels (a) and (d) the median fiscal deficit between
1985-1987 remained at the highest possible (estimated) value.
As a consequence, inflation began to grow once again in 1985.

• After the 1987 crisis, in 1988 the Mexican government reached

an agreement with representatives of the private sector called
the Economic Solidarity Plan (in Spanish: Pacto de Solidaridad
Económica) in which the government com- mitted to reduc-
ing expenditures and inflation. The fiscal deficit came to his-
toric lows and even achieved surpluses, and the government
was able to restructure its debt. By 1989 the annual inflation
rate was lowered to 20.3%. The model is consistent with this
episode of economic history in Mexico; through the lens of the
model, the government conducted a structural reform: be-
tween 1988 and 1993 (Region (4) of the Figure), fiscal deficits
were reduced from the highest possible median to a mod-
erate level in 1989 and then in 1993 to a lower median
This reduction of the fiscal deficit had an immediate impact
on inflation and its expectations.

• Several factors induced another crisis at the end of 1994

and during 1995. The re-privatization of the banks was fi-
nanced with foreign debt, which left the financial sector ex-
posed to sudden exchange rate movements and increments
in interest rates. Additionally, the government issued bonds
that were paid in pesos but with dollar nominal values (the

23
Cardenas (2015) argues that the crisis presented during 1987 is a
direct consequence of the unwillingness of the government to reduce
its deficit during 1982-1987.

483
 Tesobonos), which required a stable exchange rate in order
to keep this debt sustainable. However, political events led to
a significant depreciation of the domestic currency in 1994
accompanied by capital outflows (Calvo and Mendoza (1996),
Cole and Kehoe (1996) analyze these events). The government
faced a debt crisis, the private financial sector found itself
in bankruptcy, and the inflation rate reached 51% in 1995.
The government negotiated loans with the International
Monetary Fund (imf) and with the United States in order to fi-
nance its debt.

• The model attributes, in Region (5), the escalation in infla-

tion during 1995 to an increase in fiscal deficit between 1994
and 1995. However, this escalation was a consequence, to a
significant extent, of the nominal exchange rate depre- cia-
tion at the end of 1994 and the collapse of the financial sector
in 1995. In this case, there is a discrepancy between the in-sam-
ple predictions of the model concerning fiscal deficit and what
is observed in the data. This discrepancy between the model
and the data motivates the introduction of the nominal ex-
change rate in the model. It will be shown that by introducing
this variable we can better account for the behavior of infla-
tion during 1995 and in general.

• After a constitutional reform in 1993, Banco de México be-

came independent in 1994. The reform established as its
primary mandate to preserve the purchasing power of the
national currency. 24 The average annual inflation rate fell
from 10.95% between 1996-2002 to 3.98% between 2003-
2016, achieving historic minimums during 2015 and 2016.25

24
Some of the policies adopted by the Central Bank after 1994 were: (i)
restoration of the level of international reserves to gain credibility, (ii)
the use of an objective of cumulative current account balances that
private banks held at the Central Bank as the primary monetary policy
instrument, (iii) adoption of an inflation-targeting policy, and (iv)
to improve transparency, the Central Bank began to publish reports
communicating monetary policy decisions as well as quarterly reports
of the economy. For a more detailed description of these policies
see Ramos-Francia and Torres- Garcia (2005).
25
Furthermore, as documented by Chiquiar et al. (2010), the inflation
rate after 2000-2001 became a stationary process and initiated its con-
vergence towards the inflation target.

484 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Meanwhile, fiscal deficits remained relatively low and stable
during 1997-2016.26

• During the last sub-period (Region (6) of Figure 3), the model
predicts that fiscal deficits were at the lowest median and vari-
ance hidden states. The model also shows that the expected
inflation rate has fluctuated within the range of the target
of Banco de México: an inf lation rate of 3% that can vary
between 2% and 4%. The model proposes that a necessary
condition to anchor inflation and its expectations is a low mon-
etization of fiscal deficit. The only year in which the fiscal
deficit had a slight probability of being at a higher median
state was in 2009, in the course of the global financial crisis.
However, since the inflation rate remained low after 2009,
the baseline model predicts that Mexico has remained in a
low fiscal deficit regime.
Considering the inflation history previously described, we ob-
serve that the model predicts a deficit distribution with an elevated
mean and variance during those years in which the inflation rate
was elevated, as in 1987 (a year characterized by the highest infla-
tion rate presented in Mexico during the second half of the twentieth
century). In those years in which the inflation rate was moderately
high, as in 1975, the model predicts a fiscal deficit with a moderate
mean and lower variance than in 1987. Finally, in those years where
the inflation rate is low, the fiscal deficit density is characterized by a
low mean and variance.

3.3 Fiscal Deficits: Data and Model Simulation

The Ministry of Public Finance of Mexico, Secretaría de Hacienda
y Crédito Público (shcp), computes a measure of the fiscal deficit
called Balance Público Tradicional (bpt) since 1977. This measure
represents the difference between current and capital expenditures
and revenue of almost all of the public sector. 27 Since 1990, the shcp

26
In 2008 there was a methodological modification in bpt, it became
a wider measure of fiscal deficits: after 2008 the bpt considers part
of the investments made by two important state- owned firms (pemex
and cfe) that before were considered as long-term debt (this type
of investments are called pidiregas).
27
The bpt does not consider the revenue and expenditures of Banco
de México or the public financial sector. The financial sector of the

485
 Figure 3
DYNAMICS OF THE MODEL

.      

%
15
(1) (2) (4) (5) (6)

(3)

10

0
1970

1973

1976

1979

1982

1985

1988

1991

1994

1997

2000

2003

2006

2009

2012

2015
Year

% .   

200
(1) (2) (3) (4) (5) (6)

150
Data

100

50 Model

0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017

Year
Notes: Panel (a) plots the median real scal deficit relative to output together with the
10th and 90th percentile of the annual deficit distribution. Panel (b) shows the annual
inflation rate predicted by the model given the real scal deficit, and the data. Panel
(c) shows the expected inflation rate according to the  algorithm (4). Panel (d)
plots P [dt = d2|πt, φ] + P[dt = d3|πt, φ] where d2 and d3 are the moderate and low levels
of mean fiscal deficit.
Source:  and model results.

486 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 3 (cont.)
DYNAMICS OF THE MODEL

.   

%
200
(1) (2) (3) (4) (5) (6)

150

100

50

0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017
Year
.      
%
100

80 (1) (3) (4) (5) (6)

60

40

20
(2)
0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017

Year
Notes: Panel (a) plots the median real scal deficit relative to output together with the
10th and 90th percentile of the annual deficit distribution. Panel (b) shows the annual
inflation rate predicted by the model given the real scal deficit, and the data. Panel
(c) shows the expected inflation rate according to the  algorithm (4). Panel (d)
plots P [dt = d2|πt, φ] + P[dt = d3|πt, φ] where d2 and d3 are the moderate and low levels
of mean fiscal deficit.
Source:  and model results.

487
computes an alternative fiscal deficit measure called Requerimien-
tos Financieros del Sector Público (rfsp), which incorporates the fi-
nancial requirements of the government at the federal level. This is a
broader measure of fiscal deficit since it includes the bpt in addition
to all revenues and expenditures of the public financial sector that
provide funds for public policy.28
Panel (a) of Figure 4 displays the estimated sequence of fiscal defi-
cits from the model, as well as the bpt and the rfsp relative to gdp be-
tween 1977 and 2016. As shown in the Figure, there is an adequate
approximation of the model to the bpt data before 1991 and to the rfsp
after 1993. During 1991 and 1992, both series show a fiscal surplus.
The model cannot match this feature of the data given the assumption
of a log-normal distribution, and deficits cannot be negative. Addi-
tion- ally, the model predicts a higher deficit during 1994-1996 rela-
tive to those observed in the data; in 1995 the model predicts a fiscal
deficit relative to output of 6.1% of gdp, while the Rfsp exhibits a fiscal
deficit of 2.5% of gdp. The baseline model can only attribute the spike
in inflation of that year to fiscal deficits. We will see that the exten-
sions of this model can better account for the rates of inflation during
this episode. During 1977-2016, the model’s median deficit variance
is 53.7% of the variance presented in the fiscal deficit data.29
Panel (b) of Figure4 displays the model’s implied monetary base
growth rate compared with Banco de México’s data between 1969
and 1970. 30 The Figure shows that the model approximates the data’s

government includes, among others, trust funds and banks administered

by the federal government.
28
For example, during 1990-1998 the government managed a trust fund
called fobaproa, its objective was to insure private banks against overdue
accounts in case of a financial crisis. If the fund provided resources to a
private bank to cover its overdue accounts, this would be considered in the
Rfsp but not in the bpt. The rfsp are a better approximation of the con-
cept of deficits considered in the model. However, before 1990 the only
official deficit measure available is the bpt. We are grateful to Nicolas
Amoroso, Oscar Budar, and Juan Sherwell for their invaluable guidance
in understanding historical accounts and providing these series.
29
For these results, we considered the Bpt before 1991 and the rfsp after
this year.
30
To compute the monetary base growth according to the model, we con-

sidered equation (1) to show that: Ramirez

de Aguilar (2017) presents further details.

488 B. López Martín, A. Ramírez de Aguilar, D. Sámano

sequence reasonably well, although there are differences in 1990-
1992. The model’s monetary base growth rate variance accounts
for 82% of the variance presented in the data.

4. BEYOND THE BASELINE MODEL

Considering that, since 1994, Banco de México has been an indepen-

dent Central Bank and no longer finances the federal government
through money creation, in this section we present modifications
to the baseline model.31 Before we discuss these extensions, we should
be explicit about the fact that the model by itself does not distin-
guish between periods of monetary or fiscal dominance. Formally,
the estimation of the model will propose a series of deficits that are fi-
nanced with monetary emission, while the classification of differ-
ent periods in terms of the regime rests on the interpretation of the
historical narrative we previously presented. 32 In a similar manner,
Meza (2017) concludes that the change in legislation that granted
independence to Banco de México in 1993 represented a credible
change from fiscal to monetary dominance, and that the transition
to an independent Central Bank has been successful. Furthermore,
Central Bank independence does not imply d≈0 if the target for infla-
tion is, for example, 3%. Through the lens of the model, the Central
Bank would target a long-run level of money growth such that infla-
tion fluctuates around the target of this institution. 33
The extensions we present will allow us to illustrate some of the
channels through which fiscal policy may potentially influence in-
flation even in a context of autonomy of the Central Bank. These

31
As explained by Meza (2017), the Central Bank transfers resources
to the Ministry of Finance (equivalent to the Treasury in the U.s.),
after determining its earnings and following legally specified rules.
This is called the Remanente de Operación de Banco de México. In the
United States, the Federal Reserve transfers to the Treasury most of its
interest earnings from government debt. As further discussed below,
this can be perfectly consistent with a regime of monetary dominance.
32
In this sense, the approach is complementary to models that consider
regime-switching environments, e.g. Chung et al. (2007), Cadavid-
Sanchez et al. (2017), and Bianchi and Ilut (2017).
33
For the period, Meza (2017) estimates seigniorage at an average of 0.66
p.p. of gdp for the period 1995-2016.

489
 Figure 4
DATA AND MODEL COMPARISON

.       1977-2016

%
15
P90
12

6 Model RFSP
3
P10
0

−3
BPT
−6
1977

1980

1983

1986

1989

1992

1995

1998

2001

2004

2007

2010

2013

2016
Year

.    1969-2016

%
150

Data
100

50
Model

−50
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2016

Year

Notes: the series presented are - Panel (a): in blue the estimated scal deficit with the
10th/ 90th percentiles of the estimated deficit distribution. In red/orange the 
or the  relative to . Panel (b): In blue/red the model/data monetary base
annual growth rate, respectively.
Source: Banco de México and .

490 B. López Martín, A. Ramírez de Aguilar, D. Sámano

modifications are inspired by the literature that studies the interac-
tions between fiscal and monetary policy, which suggests that, even
with an independent Central Bank, fiscal policy can still affect infla-
tion. For example, if agents observe an increasing deficit that trans-
lates into higher debt, they may anticipate a regime change to make
the fiscal path sustainable, hence, they may increase their current
inflationary expectations and inflation itself.
First, we present an extension where we consider that the expect-
ed inflation rate may be influenced by fluctuations in the nominal
exchange rate (ner) between the Mexican peso and the u.s. dollar.
An important result of this model is that the effect that the ner has on
inflation (known in the literature as Exchange Rate Pass-Through,
erpt) is a function of the fiscal deficit. According to our estimation,
in a situation with elevated fiscal deficits that generate high inflation
rates, the erpt is considerable. After 1995, the year in which the ner
changed from a fixed to a flexible regime and after Banco de México
became an independent institution, the erpt to inflation and its ex-
pectations has become rather limited.
The second extension considers the sovereign interest rate spread
embi of J.P. Morgan as a variable that reflects the fiscal situation
of governments. We estimate that the embi has a moderate impact
on inflation and its expectations, although its effect is positive and sta-
tistically significant. An increase in the embi spread is associated with
the perception that the government is not in a solid fiscal situation.
Hence, following the example illustrated by Kocherlakota (2012),
agents may incorporate in their inflation expectations the possibility
that the Central Bank may lose independence to the fiscal authority,
and consequently raise their inflation expectations. This, accord-
ing to the model, generates an increase in observed inflation as well.
In the third extension we specify a real-balances demand func-
tion that incorporates the exchange rate, as an alternative channel
through which this variable may influence inflation. 34

34
We have explored additional extensions of the model. For example,
incorporating the cetes interest rate, and another specification that
includes the target for the inflation rate of Banco de México. However,
the fit of these alternative specifications is less favorable (results avail-
able upon request). Further exploration of alternative specifications
would certainly be an interesting topic for future research.

491
 Empirical evidence shows that sovereign interest rate spreads
are, to a large extent, driven by international factors such as risk
appetite, market volatility, terms of trade, global liquidity, conta-
gion from events such as the Russian crisis or the LTCM collapse
in 1998, and even U.s. macroeconomic news. 35 In the same fashion,
exchange rate fluctuations are linked to global financial factors (to
give some recent examples, Gabaix and Maggiori (2015) and Itskhoki
and Mukhin (2017)), and the Mexican peso is sometimes considered
a commodity currency (see Kohlscheen (2010)). The state of public ac-
counts can make the economy vulnerable to these external shocks. 36
Our model allows us to explore empirically the possibility that
fiscal policy can make the evolution of inflation sensitive to events
in international financial markets. The results motivate the need
for further theoretical developments in this area, in particular for de-
veloping economies, where sovereign interest rate spreads and ex-
change rates seem to be of primary relevance. The historical narrative
of events in Mexico for the period 1969-1994 supports this interpre-
tation; events such as significant drops in the price of oil or sudden-
stops make the economy vulnerable when fiscal accounts are in a dire
situation and the government may be forced to turn to the Central
Bank to cover its financial needs. Even in a context of de jure mon-
etary dominance, economic agents may consider that these risks
are still present, and thus we aim to capture this possibility in the
estimation of our model. 37

35
There is an extensive literature that documents these facts, including
Longstaff et al. (2011), González-Rozada and Levy-Yeyati (2008), Bunda
et al. (2009), Ciarlone et al. (2009), Hilscher and Nosbusch (2010),
and Ozatay et al. (2009).
36
The issue of endogeneity is addressed by exploiting alternative method-
ologies in Cortés-Espada (2013) and Lopez-Villavicencio and Mignon
(2016).
37
These channels have been considered by Zoli (2005) in the case of Bra-
zil, by assessing the impact of news concerning fiscal variables and fis-
cal policy on sovereign interest rate spreads and the exchange rate
and discussing the potential implications for monetary policy. Cerisola
and Gelos (2005) find that the stance of fiscal policy (proxied by the
ratio of the consolidated primary surplus to gdp) is important to deter-
mine inflation expectations in the case of Brazil and argue that fiscal
policy is instrumental in anchoring inflation expectations.

492 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 5
ANNUAL INFLATION AND VARIATION OF THE NER

  (%)

200 Fixed Exchange Rate

150 Annual Inflation

100
Floating Exchange Rate
50

−50
1977

1980

1983

1986

1989

1992

1995

1998

2001

2004

2007

2010

2013

2016
Source: Banco de México and .

4.1 The Role of the Exchange Rate

As documented by Rogers and Wang (1994) and Carrasco and Fer-
reiro (2013), an important variable in determining inflation expec-
tations is the nominal exchange rate (ner). Figure 5 presents, as a
motivation for this extension, the annual inflation rate and the an-
nual variation of the ner between 1977 and 2016. This Figure shows
a significant correlation between these variables, particularly dur-
ing episodes of high inflation. An important fact to consider is that
before 1995 Mexico had a fixed exchange rate with bounded depre-
ciations. 38 After 1994, the peso-dollar ner entered a floating regime.
In this extension, we consider that the exchange rate variation
∆NER is a variable that can affect inflation expectations. We assume

38
In the Appendix, we describe the different exchange rate regimes
in Mexico.

493
that this variation has a weight ξ on expectations. Hence, for each pe-
riod t, the expected inflation rate is determined as follows:

Given that Mexico had a fixed ner during 1969-1994 and after 1995
the ner is in a floating regime, we estimate the model allowing ξ to
change during 1969-1994 and 1995-2016. Hence, the model allows agents
to give a weight ξ 1 to the ner variation during a fixed exchange rate re-
gime and a weight ξ 2 when the ner is in a floating regime. To estimate
this model, again we consider the monthly inflation sequence accord-
ing to the inpc between January of 1969 and December of 2016, and the
sequence of the monthly variation in the peso-dollar ner documented
by Banco de México for that period. Table 2 presents the estimated pa-
rameters of this version compared with the baseline model estimation.
Considering the exchange rate as a variable that can influence inflation
expectations (and hence, inflation), the model can account for 75.8%
of the variance observed in the inflation data, while the baseline model
can explain 61.6% of this variance. Also, as suggested by the Diebold-
Mariano test, during 2000-2016 the ner and baseline models produce
different in-sample forecasts of observed inflation (at a 1% significance
level) and the modified model has a higher correlation with the infla-
tion data. 39 This result emphasizes the relevance of the exchange rate
for the determination of the inflation rate in Mexico.40

39
The hypothesis test proposed in Diebold and Mariano (1995) allows to assess

if two forecasts related to a series are statistically different.

Defining ekt=ykt−yt for k ∈ {i, j} and considering a loss-function g(e), the null

hypothesis in the Diebold-Mariano test is that These

authors construct a statistic function that involves the autocorrelations

of the forecasts and show that, if the time series considered are covariance
stationary and short memory, it has a t-Student distribution. Then, they
construct a statistic that, under the same assumptions, is asymptotically
N (0, 1).
40
More formally, according to the sic comparison, the ordering of the models
is the following: the model with the embi spread and the ner in the for-

494 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Table 2
EXTENDED MODEL PARAMETER ESTIMATION

parameter ner model baseline model description

λ 0.7730 0.7556 weight of expectations on the
(0.0013) (0.0022) price level
ν 0.1152 0.1147 weight of past inflation
(0.0049) (0.0081) on expectations

ξ1 0.0215 - weight of ner on expectations

(0.0006) in a fixed regime

ξ2 0.0047 - weight of ner on expectations

(0.0001) in a floating regime
0.0077 0.0075 monthly high median level
(0.0001) (0.0001) of fiscal deficits
0.0039 0.0039 monthly moderate median level
(0.0003) (0.0004) of fiscal deficits
0.0022 0.0023 monthly low median level of fiscal
(0.0003) (0.0002) deficits

Notes: the numbers shown in parentheses represent the standard deviation

of each parameter, computed using the Hessian matrix of the maximum
likelihood problem (see MacDonald and Zuccini (2009)).

The parameters {ξ 1, ξ 2} are statistically different, a result that

can be interpreted as follows: between 1969 and 1994 the erpt to ex-
pectations was 0.0215 p.p. given 1% depreciation of the ner. After
1995 the erpt shows a considerable reduction: a 1% exchange rate
depreciation translates to an increase in the expected inflation rate
of 0.0047 p.p. To assess the erpt into the observed inflation, we must
consider not only the erpt to expectations, but also the fiscal deficit
level relative to gdp. This is because, within the model, both variables
jointly determine the inflation rate. As we detailed in the previous

mation of expectations, the model with the ner in the real balances
demand function (presented in the following section), the model with
only the ner in the formation of expectations and, finally, the base-
line model.

495
 Figure 6
IMPULSE-RESPONSE FUNCTIONS OF INFLATION

1
Difference with Equilibrium Inflation (%)

0.8 0.821 p.p.

0.6

0.4

0.2
−
0.026 p.p. High d (1982-1987)
0
−
Low d (1995-2016)
−0.2

−0.4
0 10 20 30

Time

Source: Banco de México and .

section, a higher fiscal deficit magnifies the effect that βt has on infla-
tion (in fact this effect is nonlinear). Hence, if fiscal deficit increas-
es, the effect that the ner variation has on πt will grow because this
variation affects βt . It can be shown that:

∂πt ∂π ∂βt λξ
9 = t = πt .
∂∆NER ∂βt ∂∆NER 1 − λβt − dt

This equation highlights two important results: (i) the erpt is in-
creasing in dt; (ii) a higher inflation rate implies a higher erpt. Figure
6 shows the impulse-response function of inflation given a 1% depre-
ciation in the ner. As this Figure suggests, when fiscal deficit is high
(e.g., during 1982-1987) the erpt to inflation is 0.821 p.p. However,
if fiscal deficit is low the erpt of a 1% depreciation is 0.026 p.p. Hence,
a low fiscal deficit financed by the Central Bank not only translates
into low inflation, but also into a limited erpt. A low pass-through

496 B. López Martín, A. Ramírez de Aguilar, D. Sámano

contributes to a steady and anchored expected and observed infla-
tion rate.41
Figure 7 shows that the model that considers the ner as a variable
that influences inflation expectations is able to provide a better ac-
count of the behavior of inflation dynamics in general, but espe-
cially during 1982 and 1994-1995, relative to the model that does
not consider the ner, given the depreciation of the ner observed
during those years.

4.2 The Role of the embi Spread

In this section we analyze an extension of the baseline model that
considers the sovereign interest rate spread embi, a variable that
captures the perception of the fiscal situation in Mexico and may in-
fluence inflation expectations. To the extent that this variable is rel-
evant according to the estimation then this would suggest that, even
though Mexico has an independent Central Bank, fiscal policy must
be relevant for monetary policy through its influence on the infla-
tion rate and its expectations.42
As a motivation for this extension, Figure 8 displays, in Panel (a),
the interest rate spread embi and the ner between 1998 and 2016.
This Figure shows that these variables are weakly correlated. Hence,
if we consider the embi and the ner, we will be able to identify the ef-
fect that each variable has on inflation and its expectations. Panel (b)

41
The low level of pass-through is consistent with estimates in the lit-
erature for Mexico, see Albagli et al. (2015), Capistrán et al. (2011),
Cortés-Espada (2013), and Kochen and Samano (2016). Furthermore,
there is evidence of a declining erpt in environments with more stable
inflation and with the adoption of inflation targets (see Baqueiro
et al. (2003), Choudhri and Hakura (2006) and Lopez-Villavicencio
and Mignon (2016)). Capistrán et al. (2011) and Cortés-Espada (2013)
document a lower erpt for Mexico under the inflation targeting regime.
42
The perception of economic agents of the fiscal responsibility of the
government may depend on the particular historical context. For ex-
ample, Sargent and Zeira (2011) describe how the anticipation of a
future government bailout of banks caused a jump in inflation in Israel
in 1983. They argue that the public anticipated that this bailout would
eventually be financed by monetary expansion. Alternatively, Chung
et al. (2007) explore an environment where monetary and fiscal re-
gimes evolve according to a Markov process, this possibility can change
the impact of policy shocks. These authors argue that, to the extent that
there has been a history of changes in policy regimes, private agents
can ascribe a probability distribution over the different regimes.

497
 Figure 7
INFLATION AND EXPECTATIONS IN THE NER MODEL

.   

%
200

Inflation Data
150

Average Inflation
100 2006-2016:
Data: 3.91%
Baseline Model  Model: 3.56%
ner Model
50 Baseline Model: 3.15%

0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017
Year

.  
%
200

150

Average Expected
100 ner Model Inflation
2006-2016:
 Model: 3.37%
50 Baseline Model: 3.15%
Baseline Model

0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017

Year

Source: .

498 B. López Martín, A. Ramírez de Aguilar, D. Sámano

of this Figure shows the relationship between the annual inflation
rate and the variation (in basis points) of the embi spread.
In this extension, we consider two regimes: a fiscal dominance
regime where the fiscal authority can use money creation to finance
its deficit, and Central Bank autonomy, where it cannot. The inter-
pretation we propose is that Mexico had a fiscal dominance regime
between 1969 and 1994. Under fiscal dominance, Mexico had a fixed
ner and under monetary dominance, the peso-dollar ner is under
a floating regime (see the Appendix for a more detailed description
of the exchange rate regimes). We assume that under fiscal domi-
nance, agents determine their expectations according to:

10

After 1994 we allow agents to give some weight σ to the current

fiscal situation (which is reflected in the sovereign embi spread).
Hence, agents determine their inflation expectations according to:

11

We allow the parameters {ν, ξ} to vary because the ner had a change
in its regime.
If parameters ξ and σ are positive and statistically significant,
it would imply that the embi spread and the ner influence inflation.
In fact, these variables can generate the escape dynamics that in the
baseline model could only be ignited by the behavior of fiscal defi-
cits.43 Figure 9 exemplifies how an escape dynamics, that leads to high
or hyperinflation, can occur in this scenario: suppose that initially
βt =β ∗ and that ∆NER t , ∆EMBIt are limited. This implies that infla-
tion and its expectations will converge to a low inflation equilibrium
as the blue arrows show. However, if the fiscal authority starts to con-
siderably increase its deficit (which is no longer financed with money
creation and is therefore translated into debt) this would be reflected

43
In the baseline model, an escape dynamics can only occur if fiscal deficit
increases for a considerable period, because it is the only way to raise
inflation expectations.

499
 Figure 8
EMBI, NER AND INFLATION

.   

Peso-Dollar
EMBI Exchange Rate

20
380 embi ner
18
330
16
280
14
230
12
180

130 10

80 8
1999

2001

2003

2005

2007

2009

2011

2013

2015

2017
Year

.    

EMBIVariation Annual
(Base Points) Inflation (%)
300 20
250 embi Variation
200 Inflation
150 15
100
50
10
0
−50
−100 5
−150
−200
−250 0
1999

2001

2003

2005

2007

2009

2011

2013

2015

2017

Year

Source: Banco de México, Bloomberg and .

500 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 9
ESCAPE DYNAMICS IN THE MODIFIED MODEL

0.8

0.6

0.4
πt+1−πt

∆NER>>0
0.2
∆EMBI>>0
β∗
0

−0.2

β2
−0.4
0 5 10 15

βt

Note: these figures considers βt−1 =1.02 and the estimated parameters of the 
extension.

in the embi spread and influence the ner. In our model, the incre-
ment in these variables will affect inflation expectations. Further-
more, if this effect is large enough, as shown with an orange arrow
in the Figure, it will cause that βt>β 2 , which will lead to high infla-
tion (as shown with red arrows). Consequently, if σ and ξ are signifi-
cant and positive then, even in a context of monetary dominance,
our model suggests the possibility of high inflation caused by the fis-
cal authority via expectations.
To estimate this model, once again we consider the inflation se-
quence according to the INPC during 1969-2016, the ner varia-
tion registered by Banco de México, and the embi spread reported
by Bloomberg after 1994. The main results of this extension are:

501
 • The estimation for σ suggests that, everything else constant,
if the embi spread increases 100 basis points, the rate of infla-
tion rises by 0.24 p.p.44

• On the other hand, the estimation of ξ 2 implies that under

monetary dominance the inflation rate increases 0.011 p.p.
given a 1% depreciation of the ner.

• Finally, with this specification for inf lation expectations,

the model estimates that is almost zero, which is the defi-
cit regime for the period of independence of the Central Bank.
Figure 10 shows that, if we consider the interest rate spread embi
and the ner, then the inflation generated by this model is closer to the
inflation sequence presented in the data. Actually, the incorporation
of these variables allows the model to explain 0.65 p.p. more of the
inflation rate during 2006-2016 compared to the baseline model.
The Diebold-Mariano test also suggests that the in-sample forecast
for the inflation sequence between these years is statistically differ-
ent (at a 1% confidence level) between the embi extension and the
baseline model. Hence, these extensions suggest that the fiscal sit-
uation, to some extent, have caused the inflation rate to be above
Banco de México’s inflation target of 3%.

4.3 The Exchange Rate: An Alternative Channel

A variable such as the exchange rate may affect inflation through
several channels and not only through inf lation expectations.
We now discuss an extension where the ner has an effect on infla-
tion through its direct influence on the price level Pt. We assume that
Pt = γMt + λPte+1 + ψNERt . 45 Hence, the ner has a weight ψ on the price
level, parameter that can be interpreted as the pass-through of the
ner to the price level. This modification implies that the inflation
rate is now given by the following expression:

12

44
To find the impact that the embi spread has on inflation, we again have
to consider an impulse-response function as in Figure 6.
45
Alternatively, this expression can be rewritten as a demand for real
balances that depends on the exchange rate.

502 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 10
EVOLUTION OF INFLATION: MODELS AND DATA DURING 2006-2016

8
Data Average: 3.91%
Data  and  Model Average: 3.80%
Baseline Model Average: 3.15%
Annual Inflation Rate (%)

Baseline Model
2 embi and ner Model
bm Inflation Target

0
2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017
Year

Source: Banco de México, Bloomberg and .

Expectations are given by the cge algorithm β t+1=(1−ν)β t +νπ t ,

when there is fiscal dominance (i.e. before 1994) and by βt+1=(1−ν−σ)
βt +νπt +σ∆EMBIt under Central Bank independence. The main dif-
ference between assuming that the ner affects expectations or P t
is that, in this extension, inflation is a function of the ner dynamics
in two consecutive periods: (NER t−1, NER t). Hence, if the ner depre-
ciates considerably between t− 1 and t, this will have a higher impact
on inflation and on future inflationary expectations.
Figure 11 presents the main results of this extension. As this Fig-
ure shows, the extended model better accounts for the inflation rate
during 1970-2016 than the baseline model. This model performs
particularly better in those periods in which the ner registers a con-
siderable depreciation. For example, during 1982, the peso-dollar
ner suffered a depreciation of over 200% and the model predicts
that inflation at the end of that year was 118.1%. Additionally, during
1995 the ner had a depreciation that surpassed 100%, which implied,
according to the model, an inflation of 49.1% by the end of this year.

503
 Figure 11
INFLATION IN THE EXTENDED MODEL

. 1970-2016
200

Data
150
Annual Inflation (%)

Extension
100

50
Baseline Model

0
1969

1972

1975

1978

1981

1984

1987

1990

1993

1996

1999

2002

2005

2008

2011

2014

2017
Year

. 2006-2016

Data
Data Average: 3.91%
6 Extension Average: 4.01%
Annual Inflation (%)

4
ner in rbd Extension
3%

2
Banco de México’s Extension Target

0
2006

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

Year

Source: Banco de México and .

504 B. López Martín, A. Ramírez de Aguilar, D. Sámano

5. CONCLUDING REMARKS

The baseline model and the extensions that we have presented allow
us to assess the role of fiscal policy in the determination of inflation
and its expectations. Even in a context of Central Bank indepen-
dence, a large literature has explored the role of fiscal policy in de-
termining inflation. We exploit a simple model and provide evidence
of the relevance of fiscal policy in determining the behavior of ag-
gregate prices in Mexico as well as the importance of expectations.
Admittedly, the theoretical framework we utilize is relatively sim-
ple and models with more structure, perhaps in the inter-temporal
dimension, would increase our understanding of the relationship
between fiscal policy and inflation in emerging economies. Further-
more, it is sometimes argued that Central Bank independence acts
as a mechanism that increases fiscal responsibility of the govern-
ment in developing countries (Bodea and Higashijima (2015), Minea
and Tapsoba (2014)). We believe that further research is necessary
to understand the institutional arrangements that govern the rela-
tionship between a central bank and the fiscal authority in the pres-
ence of competing objectives and constraints.

6. APPENDIX

6.1 Parameter Estimation

The following equations, together with transition matrices {Qd , Q v}
define inflation, expected inflation, and fiscal deficits at each t ac-
cording to the baseline model:

505
where Xt is a constant equal to 1 if {πt, βt.dt} satisfy 1−λβt−1>0 and δ(1−
λβt−γdt)>θ(1−λβt−1).46 Assuming β0 =π 0 and given a sequence of fiscal
deficits the model can generate a sequence for the expected
inflation rate and for the actual inflation rate How-
ever, the hidden Markov states among other parameters, must
be estimated to generate a sequence of fiscal deficits. Table 3 shows
the parameters that need to be estimated.

Table 3
MODEL PARAMETERS

parameter restrictions description

λ 0<λ< 1 weight of expectations on the price
level

ν 0<ν<1 weight of past inflation

on expectations

γ γ>0 weight of monetary base on the

price level

θ 0<θ<1 persistence of the monetary base

δ δ>0 constant that bounds inflation

median values of fiscal deficits

variance values of fiscal deficits

vπ vπ > 0 inflation variance when

determined randomly

i, j-component of the transition

matrix Qd

i, j-component of the transition

matrix Qv

46
These constraints guarantee that the model’s inflation rate is bounded
and that the real balances demand is positive.

506 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Let be the vector of all the parameters in the model. Given that
dt is a random variable and because {πt, βt} are a function of fiscal defi-
cits, we can construct a joint density function for a sequence of T pe-
riods of inflation, its expectations and fiscal deficit:
If there was available data on inflation, its expectations and fiscal
deficit for a large T , the estimated parameters can be obtained
using the maximum-likelihood method applied to the joint density
However, data on inflation expectations and fiscal
deficit is hard to find for a large T , or may not be reliable. Further-
more, we find that historical series often go through methodological
modifications. This is particularly true in the case of Mexico, as we
have already discussed.
inpc (consumer price index) data is available since January 1969
at a monthly frequency. Therefore, to estimate the parameters we use
the marginal density of a sequence of inflation πT between January
of 1969 and December of 2016. This marginal density is denoted
The estimated parameters are obtained as the vector
that maximizes given the gross inflation rate sequence
πT (subject to constraints):

13

where Ω is the set of all the vectors φ that satisfy the constraints rel-
evant for each parameter. Because there is no analytical solution
to this maximization problem, has to be approximated numeri-
cally. To do this, we used a constrained optimization algorithm based
on the bfgs (Broyden-Fletcher-Goldfarb-Shanno) method of Nocedal
and Wright (2006) and the block-wise method of Sims et al. (2006).
Given the computational burden of the maximum-likelihood
optimization problem, Sargent et al. (2009) fix three parameters
to reduce the complexity on the estimation. These parameters are:
θ = 0.99, δ = 100, and γ = 1. The value assigned to θ is consistent with
the behavior of nominal balances in the five countries these authors
studied. Fixing δ = 100 implies that, in every period, inflation cannot
surpass 10,000%. Finally, γ was fixed because the maximum-likeli-
hood algorithm cannot identify γ and dt separately. Once dt is esti-
mated for each period, γ is re-normalized so that the mean of fiscal

507
deficits estimated by the model matches the mean observed in the
data (in our case, for Mexico for the period 1977-2016).

6.2 Adaptive vs. Rational Expectations

In this part of the Appendix we discuss some of the implications that
rational expectations have in the baseline model presented in this
paper. Additionally, we compare the main differences induced in the
dynamics of the model between these types of expectations and cge.
One way of modeling that agents are rational when forming their
beliefs on future inflation is to assume:

14

Equation (14) points out one important difference between ratio-

nal expectations and cge in this model. If agents are rational, they
condition their expectations on the median level and the variance
vt of current fiscal deficit since the evolution of the median and vari-
ance of fiscal deficit is known to agents when they are rational. As-
suming cge does not require agents to condition their expectations
on because they update their beliefs according to (4).
Assuming rational expectations also affects the dynamics between
the gross inflation rate of two consecutive periods {πt , πt+1} as a func-
tion of βt . Panel (a) of Figure 12 plots πt+1 −πt as a function of βt assum-
ing βt+1 is determined according to (14) and using the same median
and variance of fiscal deficit in t and t + 1. As this Figure shows, there
is only one value of βt that induce a constant inflation (and expecta-
tions) over time (β1). As the Figure suggests, β1 is a stable equilibrium.
Thus, if fiscal deficit remains with the same median and variance
level, πt+1 −πt will converge to zero and βt to β1.
With rational expectations, contrary to cge, if inflation is high (βt
> β1), agents will not allow their expectations to provoke the escape
dynamics. Their expectations will adjust and converge to β1. How-
ever, the government could prevent expectations from converging
to a high inflation equilibrium by reducing its fiscal deficits as shown
in Panel (b) of Figure 12. This Figure plots πt+1 πt as a function of βt
for two different values (low and high). Assuming and that
the median fiscal deficit level is high, if the government continues
with this deficit level, inflation will con- verge to a high equilibrium

508 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Figure 12
DYNAMICS INDUCED BY RATIONAL EXPECTATIONS

.   

β1
0
πt+1−πt

−5

−10

−15
−5 0 5 10 15
βt

.  ,   

–
10 High dt

β2
0
β1
πt+1−πt

−10

–
Low dt
Š
β
−20
−5 0 5 10 15

Note: These figures considers βt−1 =1.02 and the estimated parameters shown in Table 1.

509
and its expectations to β 2. However, if the government reduces its fis-
cal deficits, it will change the dynamics on inflation and its expecta-
tions inducing a convergence to β1.
Figure 12 points out an important difference between rational
expectations and cge: when agents use the cge algorithm, if the in-
flation rate induces a high βt then this could provoke an escape dy-
namics and eventually a hyperinflation episode, where the dynamics
between inflation and its expectations are unbounded. However,
with rational expectations, even with an extremely high fiscal deficit,
agents always adapt their expectations to prevent a hyperinflation
spiral. If fiscal deficit is high, rational expectations imply a stable
equilibrium with a high inflation rate and no escapes.
Even though cge and rational expectations induce different dy-
namics on the variables involved in the model, the inflation equi-
libria they predict are similar. Sargent et al. (2009) argue that, in the
context of hyperinflation models, “an adaptive expectations version
of the model shares steady states with the rational expectations ver-
sion, but has more plausible out-of-steady state dynamics.” Besides,
rational expectations may induce multiple equilibria that are hard
to compute. Given the computational problem rational expectations
may induce and the fact that some Latin American countries have
experienced hyperinflation episodes with escape dynamics which
a strictly rational expectations model cannot account for, cge are
necessary for the purposes of this study.

6.3 Exchange Rate Regimes

The table in this Annex presents the different regimes that the peso-
dollar ner has had between 1954 and 2016. Before 1994, this ner had
several regimes that can be considered slight variations of a fixed
ner rule. For example: (i) controlled variation, in which the Ban-
co de México established an interval in which the ner was allowed
to vary; (ii) generalized controlled system, in which all credit institu-
tions needed an authorization from the Central Bank to sell or buy
currencies; and (iii) controlled flotation, in which Banco de México
established an interval, changed daily, within which the ner was al-
lowed to fluctuate.

510 B. López Martín, A. Ramírez de Aguilar, D. Sámano

 Table 4
EXCHANGE RATE REGIMES IN MEXICO DURING 1954-2016

Date ner
Begining End Regime Begining End
April 1954 August 1976 Fixed 12.50 12.50
September 1976 August 1982 Controlled 20.50 48.80
Variation
September 1982 December 1982 Generalized 50.00 70.00
Controlled
System
December 1982 August 1985 Controlled 95.00 281.00
System
August 1985 November 1991 Controlled 282.30 3,073.00
Flotation
November 1991 December 1994 Floating 3,074.10 N 3.99
Intervals
with
Controlled
Variation
December 1994 December 2016 Floating N 4.88 N 20.51

Notes: N denotes New Mexican Pesos.

Source: Banco de México.

References
Albagli, E., A. Naudon, and R. Vergara (2015): “Inflation Dynamics
in LATAM: A Comparison with Global Trends and Implica-
tions for Monetary Policy,” Economic Policy Papers Central
Bank of Chile 58, Central Bank of Chile.
Baqueiro, A., A. Díaz De León, and A. Torres-García (2003): “Temor
a La Flotación o a La Inflación? La Importancia del “Traspaso
del Tipo De Cambio A Los Precios,” Working Papers 2003-
02, Banco de México.
Bianchi, F. and C. Ilut (2017): “Monetary/Fiscal Policy Mix and
Agent’s Beliefs,” Review of Economic Dynamics, 26, 113-139.

511
Bodea, C. and M. Higashijima (2015): “Central Bank Independence
and Fiscal Policy: Can the Central Bank Restrain Deficit
Spending?” British Journal of Political Science, 1, 1-24.
Branch, W. A. (2004): “The Theory of Rationally Heterogeneous
Expectations: Evidence from Survey Data on Inflation Ex-
pectations,” Economic Journal, 114, 592-621.
Bunda, L., A. J. Hamann, and S. Lall (2009): “Correlations in
emerging market bonds: The role of local and global fac-
tors,” Emerging Markets Review, 10, 67-96.
Cadavid-Sanchez, S., A. Martinez-Fritscher, and A. Ürtiz-Bolaños
(2017): “Monetary and Fiscal Policies Interactions in Mexico:
1981-2016,” Working paper, cemla.
Cagan, P. (1956): “The Monetary Dynamics of Hyperinflation,”
in Studies in the Quantity Theory of Money, Chicago University
Press, Milton Friedman ed.
Calvo, G. A. and E. G. Mendoza (1996): “Mexico’s balance-of-
payments crisis: a chronicle of death foretold,” Journal of
International Economics, 41, 235-264.
Capistrán, C., R. Ibarra-Ramírez, and M. Ramos-Francia (2011): “Ex-
change Rate Pass-Through to Prices: Evidence from Mexico,”
Working Papers 2011-12, Banco de México.
Cardenas, E. (2015): El Largo Curso de la Economía Mexicana: De
1780 a Nuestros Días, Ciudad de Mexico: Fondo de Cultura
Economica.
Carrasco, C. A. and J. Ferreiro (2013): “Inflation targeting and
inflation expectations in Mexico,” Applied Economics, 45,
3295-3304.
Carstens, A. and L. L. Jácome (2005): “Latin American Central
Bank Reform; Progress and Challenges,” IMF Working Papers
05/114, International Monetary Fund.
Catao, L. A. and M. E. Terrones (2005): “Fiscal deficits and infla-
tion,” Journal of Monetary Economics, 52, 529-554.
Cerisola, M. and G. Gelos (2005): “What Drives Inflation Expecta-
tions in Brazil? An Empirical Analysis,” imf Working Papers,
05, 1.
Chiquiar, D., A. Noriega, and M. Ramos-Francia (2010): “A time-
series approach to test a change in inflation persistence:
the Mexican experience,” Applied Economics, 42, 3067-3075.

512 B. López Martín, A. Ramírez de Aguilar, D. Sámano

Choudhri, E. U. and D. S. Hakura (2006): “Exchange rate pass-
through to domestic prices: Does the inflationary environ-
ment matter?” Journal of International Money and Finance, 25,
614-639.
Christiano, L. J. and T. J. Fitzgerald (2000): “Understanding the
fiscal theory of the price level,” Economic Review, 2-38.
Chung, H., T. Davig, and E. Leeper (2007): “Monetary and Fiscal
Policy Switching,” Journal of Money, Credit and Banking, 39,
809-842.
Ciarlone, A., P. Piselli, and G. Trebeschi (2009): “Emerging markets’
spreads and global financial conditions,” Journal of Interna-
tional Financial Markets, Institutions and Money, 19, 222-239.
Cole, H. L. and T. J. Kehoe (1996): “A self-fulfilling model of
Mexico’s 1994-1995 debt crisis,” Journal of International Eco-
nomics, 41, 309-330.
Cortés-Espada, J. (2013): “Una Estimación del Traspaso de las
Variaciones en el Tipo de Cambio a los Precios en México,”
Working Papers 2013-02, Banco de México.
Davig, T., E. M. Leeper, and T. B. Walker (2011): “Inflation and the
fiscal limit,” European Economic Review, 55, 31-47.
Diebold, F. X. and R. S. Mariano (1995): “Comparing Predictive
Accuracy,” Journal of Business & Economic Statistics, 13, 253-263.
Fischer, S., R. Sahay, and C. A. Vegh (2002): “Modern Hyper- and
High Inflations,” NBER Working Papers 8930, National
Bureau of Economic Research.
Gabaix, X. and M. Maggiori (2015): “International Liquidity and
Exchange Rate Dynamics,” The Quarterly Journal of Economics,
130, 1369-1420.
González-Rozada, M. and E. Levy-Yeyati (2008): “Global Factors and
Emerging Market Spreads,” Economic Journal, 118, 1917-1936.
Hilscher, J. and Y. Nosbusch (2010): “Determinants of Sovereign
Risk: Macroeconomic Fundamentals and the Pricing of Sov-
ereign Debt,” Review of Finance, 14, 235-262.
Itskhoki, Ü. and D. Mukhin (2017): “Exchange Rate Disconnect
in General Equilibrium,” Working Paper 23401, National
Bureau of Economic Research.

513
Kochen, F. and D. Samano (2016): “Price-Setting and Exchange
Rate Pass- Through in the Mexican Economy: Evidence from
CPI Micro Data,” Working Papers 2016-13, Banco de México.
Kocherlakota, N. (2012): “Central Bank Independence and Sover-
eign Default,” Bank du France, Financial Stability Review, 16.
Kohlscheen, E. (2010): “Emerging floaters: Pass-throughs and
(some) new commodity currencies,” Journal of International
Money and Finance, 29, 1580-1595.
Leeper, E. M. (1991): “Equilibria under “active” and “passive”
monetary and fiscal policies,” Journal of Monetary Economics,
27, 129 - 147.
Lin, H.-Y. and H.-P. Chu (2013): “Are fiscal deficits inflationary?”
Journal of International Money and Finance, 32, 214-233.
Longstaff, F. A., J. Pan, L. H. Pedersen, and K. J. Singleton (2011):
“How Sovereign Is Sovereign Credit Risk?” American Economic
Journal: Macroeconomics, 3, 75-103.
Lopez-Villavicencio, A. and V. Mignon (2016): “Exchange Rate
Pass-through in Emerging Countries: Do the Inflation Envi-
ronment, Monetary Policy Regime and Institutional Quality
Matter?” Working Papers 2016-07, CEPII Research Center.
Macdonald, L. and W. Zuccini (2009): Hidden Markov Models for
Time Series, New York: CRC Press, 2nd ed.
Marcet, A. and J. P. Nicolini (2003): “Recurrent Hyperinflations
and Learning,” American Economic Review, 93, 1476-1498.
Meza, F. (2017): “Mexico from the 1960s to the 21st Century: From
Fiscal Dominance to Debt Crisis to Low Inflation,” Working
papers, itam.
Minea, A. and R. Tapsoba (2014): “Does inflation targeting improve
fiscal discipline?” Journal of International Money and Finance,
40, 185-203.
Nocedal, J. and S. J. Wright (2006): Numerical Optimization, New
York: Springer, 2nd ed.
Ozatay, F., E. Ozmen, and G. Sahinbeyoglu (2009): “Emerging
market sovereign spreads, global financial conditions and
U.s. macroeconomic news,” Economic Modelling, 26, 526 - 531.
Ramirez De Aguilar, A. (2017): “Mexico’s Recent Inflation History
as a Result of Fiscal Deficit and its Expectations,” B. A. thesis,
Banco de México and itam.

514 B. López Martín, A. Ramírez de Aguilar, D. Sámano

Ramos-Francia, M. and A. Torres-Garcia (2005): “Reducing Infla-
tion Through Inflation Targeting: The Mexican Experience,”
Working Papers 2005- 01, Banco de México.
Rogers, J. H. and P. Wang (1994): “Output, inflation, and stabi-
lization in a small open economy: evidence from Mexico,”
Journal of Development Economics, 46, 271-293.
Sallum, E.-M., F. Barbosa, and A. Cunha (2005): “Competitive
equilibrium hyperinflation under rational expectations,”
FGV/EPGE Economics Working Papers 578, FGV/EPGE-
Escola Brasileira de Economia e Financas, Getulio Vargas
Foundation (Brazil).
Sargent, T. and N. Wallace (1973): “Rational Expectations and the
Dynamics of Hyperinflation,” International Economic Review,
14, 328-350.
Sargent, T., N. Williams, and T. Zha (2009): “The Conquest of South
American Inflation,” Journal of Political Economy, 117, 211-256.
Sargent, T. and J. Zeira (2011): “Israel 1983: A bout of unpleasant
monetarist arithmetic?” Review of Economic Dynamics, 14,
419-431.
Sargent, T. J. and N. Wallace (1981): “Some Unpleasant Monetarist
Arithmetic,” Quarterly Review.
Sims, C. (2016): “Fiscal Policy, Monetary Policy and Central Bank
Independence,” Working paper, Jackson Hole Economic
Policy Symposium.
Sims, C., D. Waggoner, and T. Zha (2006): “Generalized Methods
for Restricted Markov-Switching Models with Independent
State Variables,” Journal of Econometrics, 146, 255-274.
Williams, N. (2016): “Escape Dynamics in Learning Models,” Work-
ing paper, University of Wisconsin-Madison.
Woodford, M. (2001): “Fiscal Requirements for Price Stability,”
NBER Working Papers 8072, National Bureau of Economic
Research, Inc.
Zoli, E. (2005): “How does Fiscal Policy Affect Monetary Policy in
Emerging Market Countries?” BIS Working Paper, 174, 1.

515
CENTER FOR LATIN AMERICAN MONETARY STUDIES

www.cemla.org