Monetary and Capital Markets Department

TECHNICAL ASSISTANCE HANDBOOK

Liquidity Forecasting—
Part II: The Statistical
Component

Prepared by the Central Bank Operations Division (CO)

Zhuohui Chen, Nikos Kourentzes, Romain Lafarguette,
Anastasios Panagiotelis, and Romain Veyrune

IMF | MCM Technical Assistance Handbook 35

 Liquidity Forecasting—Part II: The Statistical Component

THIS ONLINE HANDBOOK

This handbook aims to distill, document, and make widely available the lessons learned from MCM
technical assistance over a long period, while also incorporating lessons learned globally. It covers a
wide range of central banking topics pertaining to governance and risk management, monetary policy,
monetary and foreign exchange operations, and financial market development and infrastructures,
while highlighting, where relevant, specific issues for low-income, resource-rich countries. The
handbook is intended to document and promote good practices and to support consistency of advice
over time. It is, however, stressed that one-size solutions cannot fit all, and all advice therefore needs
to be tailored to country-specific circumstances. The handbook comprises self-contained, issue-
specific chapters with cross-references on overlapping issues where needed. It is targeted at those
individuals who provide technical assistance (both IMF and non-IMF personnel), and practitioners in
central banks and other relevant institutions.

THIS CHAPTER: LIQUIDITY FORECASTING—PART II: THE STATISTICAL COMPONENT

This chapter elucidates liquidity forecasting within the context of technical assistance. The audience
for this chapter is central bank staff with a strong quantitative background. Liquidity forecasting entails
a process of estimating the near-term path of a bank’s reserves using a centralized framework. Short-
term liquidity forecasts are used to calibrate the volume of central bank monetary operations to align
liquidity with the announced stance of monetary policy, whether expressed as an interest rate or as a
quantity. The best practice would be for the central bank to receive accurate information for
counterparties that have accounts in its books, including its monetary counterparties (banks) or non-
monetary counterparties, such as the government. However, the central bank may not have direct
access to some counterparties (e.g., the public which demands banknotes) or the information could
include significant errors. This chapter presents the statistical methods that have been used in
technical assistance to forecast liquidity factors and the demand for liquidity. It also proposes solutions
to select the best models, measure forecast accuracy, and reconcile forecasts. Some liquidity factors
are relatively easy to forecast due to regular patterns (currency in circulation) while others require
more sophisticated models, such as the government account. There is a tradeoff between the cost of
implementing complex models and the accuracy gains.

CHAPTER AUTHORS

Zhuohui Chen ([email protected]); Nikos Kourentzes ([email protected]);

Romain Lafarguette ([email protected]); Anastasios Panagiotelis
([email protected]); Romain Veyrune ([email protected])

IMF | MCM Technical Assistance Handbook 2

Liquidity Forecasting—Part II: The Statistical Component

Contents Page

GLOSSARY _______________________________________________________________________________________ 4

EXECUTIVE SUMMARY __________________________________________________________________________ 5

I. INTRODUCTION _______________________________________________________________________________ 6

II. LITERATURE REVIEW _________________________________________________________________________ 8

III. CHARACTERISTICS OF THE AUTONOMOUS FACTORS _____________________________________ 9

IV. FORECASTING MODELS ___________________________________________________________________ 11

V. FORECAST EVALUATION AND SELECTION ________________________________________________ 22

VI. TECHNICAL ASSISTANCE APPROACH_____________________________________________________ 30

VII. CONCLUSIONS ____________________________________________________________________________ 32

REFERENCES ___________________________________________________________________________________ 51

FIGURES
1. Equilibrium of Reserve Demand and Supply through OMO ____________________________________ 7
2. Examples of Autonomous Factors Series ______________________________________________________ 10
3. Plots of Seasonal Cycles in CIC ________________________________________________________________ 10
4. Decomposition of an Observed Time Series by Exponential Smoothing _______________________13
5. An Example of Sparse Binary and Trigonometric Seasonality Encoding________________________17
6. Modeling Deterministic Structure as Regressors: Example of CIC ______________________________19
7. Forecasting Models Tiered in Groups of Increasing Complexity _______________________________25
8. The Out-of-sample Cross-validation Scheme __________________________________________________ 26
9. An Example Nemenyi Test Comparison for Three Different Target Forecast Horizons _________27
10. A Schematic of the Process for Liquidity Forecasting _________________________________________ 28
11. The Workflow for Statistical Modeling of Short-term Liquidity _______________________________29

TABLES
1. Stylized Central Bank Balance Sheet ____________________________________________________________ 6
2. Subset of Permitted ETS Models ____________________________________________________________________ 14

APPENDICES
I. Data Request Template for Liquidity Forecasting ______________________________________________ 33
II. Standardized Output _________________________________________________________________________ 37

IMF | MCM Technical Assistance Handbook 3

 Liquidity Forecasting—Part II: The Statistical Component

Glossary
AGG Aggregate
AIC Akaike Information Criterion
ARIMA Autoregressive Integrated Moving Average
CB Central Bank
CIC Currency in Circulation
CO Central Bank Operations Division (IMF)
ETS Exponential Smoothing
GAB Government Account Balance
GARCH Generalized Autoregressive Conditional Heteroskedasticity
MAE Mean Absolute Error
MCM Monetary and Capital Markets Department (IMF)
ME Mean Error
MIS Mean Interval Score
MSE Mean Squared Error
NFA Net Foreign Assets
OMO Open Market Operations
RegARIMA ARIMA with additional regressors
RMSE Root Mean Squared Error
TBATS Trigonometric Seasonality BOX-Cox, ARMA errors, Trend, and Seasonal Components
TA Technical Assistance

IMF | MCM Technical Assistance Handbook 4

Liquidity Forecasting—Part II: The Statistical Component

Executive Summary
In most circumstances, central banks must manage liquidity to implement monetary policy. As an integral
part of liquidity management, liquidity forecasting provides an outlook on the changes in liquidity-
impacting items of central banks' balance sheets. By understanding the forecasted gap between liquidity
supply and demand, central banks can effectively calibrate open market operations (OMO) to control the
cost of refinancing (short-term interest rate), which is a common operational target of monetary policy.
Forecasting is also important for central banks that have reserve money as their operational target.

Statistical forecasts are necessary if the development of liquidity factors is not certain and if accurate
forecasts cannot be obtained otherwise. Some factors impacting liquidity require no "prediction" because
they are predetermined or known in advance by the central banks. However, several factors, such as
changes in Currency in Circulation (CIC) and net flows in the Government Account Balance (GAB), are
not directly controlled by central banks. Fluctuations in Net Foreign Assets (NFA) are also not under the
direct control of the central bank in fixed exchange rate arrangements. These factors require a more
systematic forecasting approach. Forecasts could be obtained directly for the counterparties, e.g., the
government could provide the central bank with its forecast or give prior notice of transactions on its
account. Similarly, foreign exchange transactions could be known exactly at the two-day horizon if it
settles on a T+2. (Considerations regarding the institutional framework for the sharing and centralization
of information are considered in another handbook chapter.) Statistical forecasting is necessary when
“qualitative” information is not available or accurate enough.

Many researchers have developed and experimented with various liquidity forecasting methods.
However, most existing models in the literature rely on sophisticated specifications or focus primarily on
the easiest-to-forecast item, i.e., the CIC. A more comprehensive framework that can forecast not only
autonomous factors but, more importantly, the aggregate liquidity supply is desired.

This handbook proposes a framework for predicting short-term liquidity supply generated by three
autonomous factors: CIC, GAB, and NFA. The framework is designed to cross-validate a family of time
series models and identify the best-performing model for forecasts, while allowing for generic yet flexible
country-specific customizations. Furthermore, reconciliation techniques are employed to enhance the
estimation accuracy of aggregate liquidity supply, which is crucial for the calibration of OMO. While most
central banks use point forecasts of the mean to calibrate their operations, this handbook argues that the
central bank should forecast the predictive distribution and target the percentile of this distribution that
reflects its risk preference. It would likely be different from the mean (point) forecast. If forecasts are
published, the predictive distribution should be published as well, rather than have only the point forecast
and to leave the users to factor the forecast uncertainty into their decisions.

IMF | MCM Technical Assistance Handbook 5

Liquidity Forecasting—Part II: The Statistical Component

I. Introduction
A key component of liquidity management in central banks is accurate forecasting of short-term changes
in liquidity. For commercial banks, liquidity management is a process by which they ensure sufficient cash
or liquid assets to meet their short-term obligations and operating needs. In central banks, it is the
mechanism by which to steer the dynamics of short-term interbank interest rates through controlling the
provision of reserves to commercial banks. Reserves are a highly liquid type of central bank liability held
by commercial banks, deposited to fulfil the required reserve and/or payment settlement. When there is a
shortage of reserves, interbank refinancing rates soar as commercial banks scramble for funding in the
interbank market to avoid penalties for unmet reserve requirements. Conversely, surplus liquidity may
drive short-term rates down, depressing interbank market transactions volumes, lowering banks’ net
interest margin, and flattening the yield curve.

Central banks need to proactively monitor and manage liquidity through open market operations (OMO)
to achieve their mandates of price stability. Hence, central banks require an accurate forecast of the
current and future changes of liquidity impacting items on their balance sheets (Gray 2008). The net of all
changes will represent the gap between liquidity supply and demand and will determine the amount the
central bank should absorb or inject through OMO. Table 1 presents a stylized central bank balance
sheet.

Table 1. Stylized Central Bank Balance Sheet

Assets Liabilities

Net Foreign Assets (NFA) (+) Currency in Circulation (CIC) (-)

OMO: Liquidity Supply (+) Net Government Account Balance (GAB) (-)

Lending Facility (+) OMO: Liquidity Drain (-)

Others Deposit Facility (-)

Commercial Banks’ Required Reserves (+)

Commercial Banks’ Excess Reserves (+)

Capital and Reserves

Note: (+) indicates increase in liquidity; (-) indicates decrease in liquidity

IMF | MCM Technical Assistance Handbook 6

Liquidity Forecasting—Part II: The Statistical Component

The equilibrium equation between demand and supply is:

𝑅𝑅𝑅𝑅 + 𝐸𝐸𝐸𝐸 = 𝑁𝑁𝑁𝑁𝑁𝑁 − 𝐶𝐶𝐶𝐶𝐶𝐶 − 𝐺𝐺𝐺𝐺𝐺𝐺 + 𝑁𝑁𝑁𝑁𝑁𝑁 𝑂𝑂𝑂𝑂𝑂𝑂 𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆𝑆 + 𝑁𝑁𝑁𝑁𝑁𝑁 𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶𝐶 𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹𝐹. (1)

The required reserve (RR) and excess reserve (ER) at the current account of the central bank’s monetary
policy 1 can be viewed as the demand for reserves, while the remaining impacting factors form the basis of
reserve supply. Figure 1 plots an idealized liquidity demand curve, connecting the reserves with the short-
term interbank rate, and illustrates a scenario in which liquidity supply increases, causing a shift from
position (1) to (2). The short-term interbank rate will be drawn to the floor, away from the stable point. The
central bank can intervene with OMO to relieve the downward pressure and maintain the level of the
short-term interbank rate. Therefore, estimating changes in liquidity supply is essential as central banks
need to absorb equivalent amounts of liquidity to return the short-term rate to the policy target, or,
likewise, inject liquidity when the pressure is in the opposite direction.

Figure 1. Equilibrium of Reserve Demand and Supply through OMO

Source: IMF staff.

Some of the accounts in Table 1 are easier to predict because central banks can determine or have
advance information about certain movements. For example, the required reserve ratio is predetermined

1The monetary policy counterparties of the central bank with whom the central bank implement monetary policy are
usually commercial banks. Other financial actors could be included if they are critical for the transmission of monetary
policy.

IMF | MCM Technical Assistance Handbook 7

Liquidity Forecasting—Part II: The Statistical Component

by central banks, making these “forecasts” straightforward. However, several accounts are under the
control of non-monetary policy institutions, 2 and not under the direct control of central banks. Estimating
these factors therefore becomes a priority in liquidity management. These accounts, known as
autonomous factors, typically include the public’s demand for currency in circulation (CIC), the
government’s position at the central bank (GAB), and the volatility of the net foreign assets (NFA)
account.

Many central banks have established bespoke statistical models for liquidity forecasting, typically focusing
on CIC. In practice, these models can be difficult to maintain, due to their specialized and manual model
specifications. Maintenance can be constrained by software limitations and the modeling preferences and
skillset of the responsible analysts, who often change over time. Moreover, forecasting CIC alone is
insufficient for OMO calibration, with quantitative forecasting models for the remaining autonomous
factors being less common in central banks. Although these limitations are not universal across central
banks, there is a need for a systematic modeling methodology for liquidity forecasting.

In this work, we propose a modeling framework for developing daily forecasts for all autonomous factors.
Key benefits of the framework for the central bank users are: (i) a unified approach for all autonomous
factors, thereby reducing modeling overheads; (ii) automatic model calibration and selection; (iii) detailed
reporting to facilitate individualized interventions from analysts; and (iv) leverage hierarchical forecasting
to further improve forecasting performance of the autonomous factors. The modeling framework has been
extensively tested in various central banks and is flexible for adjustment to different market and
operational conditions. The audience for this chapter is central bank staff with a strong quantitative
background.

The chapter is organized as follows. Section 2 summarizes previous studies on autonomous factors and
aggregate liquidity forecasting. Section 3 provides illustrative examples of the liquidity autonomous
factors that we aim to model. Section 4 introduces the forecasting models and methodologies used to
generate the forecasts, with Section 5 detailing how the forecast evaluation and selection is done.
Section 6 elaborates on the technical assistance (TA) approach, explaining the liquidity forecasting
mission process and briefly describing the implementation of the framework into a toolbox. Section 7
discusses limitations and future avenues for research, followed by concluding remarks.

II. Literature Review

Literature on forecasting aggregate liquidity or autonomous factors is scarce. Most of the quantitative
modeling on the autonomous factors centers on specifying an Autoregressive Integrated Moving Average
(ARIMA) to forecast CIC. Following Bell et al. (1983) and Harvey et al. (1997) non-linear approaches to
quantify calendar and seasonal effects, Cabrero et al. (2009) used ARIMA and Structural Time Series
models to forecast CIC in the euro area, finding that a combination forecast was best. El Hamiani-Khatat
(2018) investigated short- and long-term money demand forecasting models. For long-term money
demand, macroeconomic factors were important, while short-term demand is dominated by seasonal
patterns or recurring events that are relevant for calibrating OMO. ARIMA with additional regressors

2Which non-monetary policy institutions are allowed to bank with the central bank depend on the circumstances but
usually include the government and the public (demand for banknotes via intermediaries).

IMF | MCM Technical Assistance Handbook 8

Liquidity Forecasting—Part II: The Statistical Component

(RegARIMA) is widely used to forecast the demand for currency and has been applied in many countries,
including Poland (Koziński et al., 2015), Ghana (Nasiru et al., 2013), and Nigeria (Ikoku, 2014). The
regressors are often used to model recurring events, seasonal patterns, and holidays. Hlaváček et al.
(2005) compared a neural network model with a RegARIMA on a task to predict Fed CIC and concluded
that the neural network model was slightly better. However, they also noted that the neural network model
will overlearn sparse holiday events.

The existing literature does not touch much upon the detailed modeling for the GBA and NFA. Gray
(2008), in his liquidity forecasting handbook, elaborated on the factors that influence aggregate liquidity
and each autonomous factor, respectively. However, there are no quantitative models discussed in detail.
For GAB, Williams (2010) acknowledged that it was a universal challenge to obtain good cash flow
forecasting and mentioned the importance of intelligence input from relevant spending or revenue
departments to improve the forecast quality of government cash balance. Iskandar et al. (2018)
investigate ARIMA, neural networks, and hybrid models to forecast expenditures for the Indonesian
government, constituting the only time series model investigation for non-CIC autonomous factors that we
identified.

III. Characteristics of the Autonomous Factors

To better understand the challenges and requirements for forecasting autonomous factors we provide a
brief exploratory analysis of typical CIC, GAB, NFA, and their aggregate (AGG), which captures the total
contribution of the autonomous factors on liquidity. Following Equation (1) the aggregate is calculated as:

𝐴𝐴𝐴𝐴𝐴𝐴 = 𝑁𝑁𝑁𝑁𝑁𝑁 − 𝐶𝐶𝐶𝐶𝐶𝐶 − 𝐺𝐺𝐺𝐺𝐺𝐺. (2)

Given the typical operations of central banks, and specifically OMO, for the modeling of liquidity, daily
data are used and often there are long time series for each of the autonomous factors. Some central
banks opt for data organized in a five-day week, while others use complete seven-day weeks.

Figure 2 provides examples of CIC, NFA, GAB, and AGG (for details, see Gallardo et al., 2024). Note that
all time series are non-stationary. Daily time series can exhibit multiple seasonal patterns as well as day
of the week, day of the month, dominant events like payday, and day of the year. We show this in
Figure 3, which plots the various seasonal cycles for CIC. The trend of the time series is first removed
using classical decomposition, and for each period of the seasonal cycle, the minimum, 10 percent, 25
percent, median (50 percent), 75 percent, 90 percent empirical quantiles, and maximum are plotted
(Kourentzes, 2023). Observe that in all cases, a seasonal pattern emerges. As the number of days in a
month is not constant, it is more convenient to consider the seasonal cycle of days in a quarter.

The CIC, in agreement with the literature, exhibits fairly canonical time series patterns, which should be
possible to approximate well with time series models. NFA and GAB demonstrate a mix of standard time
series components and additional exogenous effects. For instance, government actions may affect them,
and this information cannot be retrieved from past observations before the government action. This
necessitates high-quality exogenous information (Williams, 2010), with the net government balance often
considered by central bankers to be the most challenging series to forecast. The most significant factor
impacting liquidity in the NFA account is unsterilized foreign exchange intervention. In countries with a

IMF | MCM Technical Assistance Handbook 9

Liquidity Forecasting—Part II: The Statistical Component

floating exchange rate regime, foreign exchange operations are less common, leading to less drastic
changes in the NFA series. NFA rarely exhibits seasonality. GAB sometimes exhibits strong seasonal
patterns in both expenditure and revenue decomposition, such as biweekly payrolls for civil servants and
quarterly corporate tax collections (Gray, 2008). CIC is not immune to these effects, but typically we
observe smaller effects. Moreover, there may be additional effects due to special calendar events, such
as holidays and celebrations, as well as structural breaks that may be due to policy effects or
extraordinary long-lasting effects, such as the COVID-19 pandemic.

Figure 2. Examples of Autonomous Factors Series

CIC NFA

160000
80000
GTQ (millions)

GTQ (millions)
60000

130000
40000

100000
2018 2020 2022 2018 2020 2022

GAB AGG
GTQ (millions)

GTQ (millions)
75000
25000

55000 65000
15000
5000

2018 2020 2022 2018 2020 2022

Source: Gallardo et al., (2024).

Note: Guatemalan quetzal (GTQ).

Figure 3. Plots of Seasonal Cycles in CIC

1.06

Median Median Median

25%-75% 25%-75% 25%-75%
1.005

10%-90% 10%-90% 10%-90%

1.05

MinMax MinMax MinMax

1.02
1.000

1.00
0.98
0.995

0.95
0.94

1 2 3 4 5 1 5 9 14 20 26 32 38 44 50 56 62 1 20 42 64 86 111 139 167 195 223 251

Days Days Days

Source: Gallardo et al., (2024).

IMF | MCM Technical Assistance Handbook 10

Liquidity Forecasting—Part II: The Statistical Component

IV. Forecasting Models

In this section we briefly describe the various modeling alternatives that are provided by the proposed
liquidity forecasting framework. As the objective is to provide forecasts of the short-term liquidity
movements to support OMO, exogenous information such as macroeconomic variables are not
considered, as changes in the short-term are dominated by stochastic trends, seasonal components, and
other factors such as holidays, payroll dates, subsidy dates, regular foreign exchange auctions, etc.
(El-Hamiani Khatat 2018). Macroeconomic variables are expected to change much slower than the usual
range of forecast horizons that are required by OMO and, likewise, such variables are not typically
sampled at high frequencies. Nonetheless, this may impact the quality of the models, and we discuss this
further below.

To support OMO, the forecasting horizon shall align with the actual OMO schedule. For instance, if OMO
is scheduled on Wednesday every week, the forecasting exercise should be performed on Tuesday,
generating 7-day (or 5-day if the weekend is skipped) forecasts ahead from Wednesday to next Tuesday.
If the OMO happens biweekly, the forecast horizon should be 14 days (or 10 days if weekends are
skipped). However, the forecasters may need to be flexible with the forecast horizon due to special
factors such as delayed delivery of data or ad-hoc fine-tuning at the end of the reserve maintenance
period. When setting an appropriate forecasting horizon, one should also bear in mind that forecasting
errors increase over time.

The various models can be grouped into two categories: (i) extrapolative time series models; and
(ii) volatility models. OMO requires point forecasts, nonetheless, in all cases we provide predictive
distributions to help analysts weigh the uncertainty that the different forecasts entail. Moreover, we argue
that there are benefits in communicating the predictive distribution to counterparties and that it can lead to
better decisions when costs may be asymmetric. These points are expanded upon in later sections.
Similarly, volatility models are useful to capture any structure in the predictive distribution beyond the
mean, particularly when the latter exhibits minimal or no structure, as is often the case for NFA. Below we
provide a brief overview of the models, with relevant references. The reader is recommended to look at
Ord et al. (2017) and Hyndman and Athanasopoulos (2021) for a more thorough treatment of the various
models.

Equation (2) describes the connection between autonomous factors. Observationally, it will always hold.
However, when we forecast the autonomous factors separately there is no expectation that this will be the
case, and we anticipate that reconciliation errors, i.e., deviations from that equation, will occur. As these
reconciliation errors are connected with the forecast errors, we can take advantage of these to improve
forecasts further. Conversely, if there were no forecast errors, the forecasts of the autonomous factors
would be coherent with no reconciliation errors. This is done with the use of forecast reconciliation
methods (Athanasopoulos et al., 2023).

4.1 The Naive Method (Random Walk)

The naive method assumes that the time series has no structure, while at the same time requiring no
parameter estimation or any other modeling choices. The forecast is generated as:

IMF | MCM Technical Assistance Handbook 11

Liquidity Forecasting—Part II: The Statistical Component

𝑦𝑦�𝑡𝑡+ℎ = 𝑦𝑦𝑡𝑡 ,

where 𝑦𝑦𝑡𝑡 is the observation of the period 𝑡𝑡, 𝑦𝑦�𝑡𝑡+ℎ is the forecast for the period 𝑡𝑡 + ℎ, and ℎ is the forecast
horizon. To understand why the model assumes that there is no structure in the data, we can refer to the
underlying random walk model:

y𝑡𝑡+1 = 𝑦𝑦𝑡𝑡 + 𝜀𝜀𝑡𝑡 ,

𝜀𝜀𝑡𝑡 = y𝑡𝑡+1 − 𝑦𝑦𝑡𝑡 ,

where 𝜀𝜀𝑡𝑡 are i.i.d. normally distributed innovations, and therefore the change in the value of 𝑦𝑦𝑡𝑡 is only due
to the random 𝜀𝜀𝑡𝑡 . As the future random innovations are unknown, these are assumed to be zero when we
generate forecasts, resulting in all forecasted values to be equal to the last observation of the time series.
Arguably, this is an inappropriate model to forecast liquidity, but it is a useful benchmark. Any more
complex models must outperform the naive in forecasting performance to be considered predictively
valuable. This can guard us against overfit models and demonstrates an active preference toward simpler
models that are more transparent. Therefore, within the proposed forecasting framework the naive is
used as an upper bound of acceptable forecasting errors.

A helpful modification of the naive is its seasonal counterpart, where instead of repeating the last
observation, the last seasonal period is repeated:

𝑦𝑦�𝑡𝑡+ℎ = 𝑦𝑦𝑡𝑡−𝑠𝑠+ℎ ,

where 𝑠𝑠 is the seasonal period, corresponding to the number of days in the week, for instance, 5 when
weekends are excluded. The seasonal naive is a useful benchmark for highly seasonal time series.

4.1.2. Exponential Smoothing Family of Models

Exponential smoothing models operate by modeling the time series as a collection of patterns, namely
level, trend, and seasonality. Usually, exponential smoothing (ETS) is framed within a state-space model,
where each component of the time is a state, and together they produce the forecast 𝑦𝑦�𝑖𝑖 , as:

𝑦𝑦𝑖𝑖 = 𝑓𝑓(μ𝑖𝑖 , 𝜀𝜀𝑖𝑖 )

μ𝑖𝑖 = 𝑔𝑔(𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑒𝑒𝑖𝑖 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑛𝑛𝑖𝑖 )

The functions 𝑓𝑓(⋅) and 𝑔𝑔(⋅) can be either additive, multiplicative, or have some mixed form. Figure 4
provides an example of the decomposition of a time series into separate components by exponential
smoothing. Observe that the level, slope, and season components together can explain most of the time
series, with any unexplained part attributed to the noise component. The level tracks the local mean of
the time series, while the slope models how the level increases or decreases over time (e.g., a slope of
+2 suggests an upward movement by two units per period). Finally, the season component models any
periodic patterns in the data. Not all time series require all components to be modeled, as some may be
absent.

IMF | MCM Technical Assistance Handbook 12

Liquidity Forecasting—Part II: The Statistical Component

Figure 4. Decomposition of an Observed Time Series by Exponential Smoothing

Source: IMF staff calculations.

In the fully additive case, the model becomes:

𝑦𝑦𝑖𝑖 = μ𝑖𝑖 + 𝜀𝜀𝑖𝑖

μ𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 + 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑒𝑒𝑖𝑖 + 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑛𝑛𝑖𝑖

Each of the states (𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 , 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑒𝑒𝑖𝑖 , and 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑛𝑛𝑖𝑖 ) is structured similarly. For example, the additive 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 is:

𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖−1 + α𝑒𝑒𝑖𝑖−1 ,

where α is a smoothing parameter between 0 and 1, and 𝑒𝑒𝑖𝑖−1 is the previous period error. Intuitively, this
equation suggests that the current level estimate is updated by α times the last error. Given that the error
is the difference between the actuals (𝑦𝑦𝑖𝑖 ) and the forecast (𝑦𝑦�𝑖𝑖 ) for the case of exponential smoothing that
has only an additive level, the model can be written in two alternative forms to help explain its function:

𝑦𝑦𝑖𝑖 = μ𝑖𝑖 + 𝜀𝜀𝑖𝑖

μ𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖−1 + α𝑒𝑒𝑖𝑖−1

or equivalently:

y𝑖𝑖 = μ𝑖𝑖 + 𝜀𝜀𝑖𝑖

μ𝑖𝑖 = 𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖
𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖 = α ⋅ 𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑎𝑠𝑠𝑖𝑖−1 + (1 − α)𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑙𝑖𝑖−1 3

3
the level is a simple distributed lag of realized values.

IMF | MCM Technical Assistance Handbook 13

Liquidity Forecasting—Part II: The Statistical Component

The second set of equations suggest that the smoothing parameter α decides by how much to update the
previous level with the last observed actuals. Noting that 0 < α < 1, a percentage contribution
interpretation becomes possible. For example, if α = 0.2, the last estimated level is updated by
20 percent of the last observation. All other states operate similarly, requiring an additional parameter for
each additional state, and a model may have any of these states on their own or together. This provides
30 possible ETS models. However, typically we restrict the selection to a subset of those (see Table 2),
as some models can be unstable. For instance, combining an additive trend and multiplicative seasonality
can result in calculation issues when the additive trend pushes the observations to zero or negative
values.

Table 2. Subset of Permitted ETS Models

Source: Kourentzes and Athanasopoulos (2019).

The estimation of the model parameters is done using maximum likelihood. This also facilitates the
calculation of information criteria that enable automatic model selection, such as the Akaike Information
Criterion (AIC). A detailed exposition of ETS is provided by Ord et al., (2017) and the theoretical
underpinnings of the state space model are discussed by Hyndman et al., (2008). ETS models can be
further modified to include exogenous regressors (Kourentzes and Petropoulos, 2016), which can help
model additional factors that are relevant for liquidity forecasting, such as calendar events.

4.1.3. Autoregressive Integrated Moving Average Models

The Autoregressive Integrated Moving Average (ARIMA) family of models is a flexible class of models
used for time series forecasting in a wide range of settings. In general, the ARIMA model is defined as:

(1 − 𝜙𝜙(𝐵𝐵))(1 − 𝐵𝐵)𝑑𝑑 𝑦𝑦𝑡𝑡 = (1 + 𝜃𝜃(𝐵𝐵))𝜀𝜀𝑡𝑡

Here 𝐵𝐵 is the backshift operator that lags a variable, i.e., 𝐵𝐵𝑦𝑦𝑡𝑡 = 𝑦𝑦𝑡𝑡−1 , 𝐵𝐵2 𝑦𝑦𝑡𝑡 = 𝑦𝑦𝑡𝑡−2 , etc. The order of
differencing 𝑑𝑑 is typically equal to 1 (or in rare cases 2) for non-stationary series and 0 for stationary
series. The term �1 − 𝜙𝜙(𝐵𝐵)� = 1 − 𝜙𝜙1 𝐵𝐵 − 𝜙𝜙2 𝐵𝐵2 − ⋯ − 𝜙𝜙𝑝𝑝 𝐵𝐵𝑝𝑝 is known as the autoregressive polynomial (or
order 𝑝𝑝) and the term �1 + 𝜃𝜃(𝐵𝐵)� = 1 + 𝜃𝜃1 𝐵𝐵 + 𝜃𝜃2 𝐵𝐵2 + ⋯ + 𝜃𝜃𝑝𝑝 𝐵𝐵𝑝𝑝 is known as the moving polynomial (or
order 𝑞𝑞). The term 𝜀𝜀𝑡𝑡 is a random innovation term. The nomenclature ARIMA(p,d,q) is used to describe
an ARIMA model. For example, an ARIMA model with 𝑝𝑝 = 2, 𝑑𝑑 = 1 and 𝑞𝑞 = 2 would be referred to as an
ARIMA(2,1,2) model.

IMF | MCM Technical Assistance Handbook 14

Liquidity Forecasting—Part II: The Statistical Component

An important extension to ARIMA models is the seasonal ARIMA (SARIMA), which allows the modeling of
patterns that repeat themselves every 𝑚𝑚 observations. In general, SARIMA take the form:

(1 − 𝜙𝜙(𝐵𝐵))(1 − 𝛷𝛷(𝐵𝐵𝑚𝑚 ))(1 − 𝐵𝐵)𝑑𝑑 (1 − 𝐵𝐵𝑚𝑚 )𝐷𝐷 𝑦𝑦𝑡𝑡 = (1 + 𝜃𝜃(𝐵𝐵))(1 + 𝜃𝜃(𝐵𝐵𝑚𝑚 ))𝜀𝜀𝑡𝑡

where 𝑃𝑃, 𝐷𝐷, and 𝑄𝑄 are the orders of the seasonal autoregressive component, seasonal differencing, and
seasonal moving component. The nomenclature ARIMA(p,d,q)(P,D,Q)[m] is used to describe such
models. Seasonal ARIMA models of this form are only capable of explicitly capturing one seasonality of
periodicity 𝑚𝑚. It is possible to extend them by introducing additional seasonal differences and
polynomials, although this can quickly become very expensive in terms of data and complicate the
identification of model orders substantially. An alternative approach to include additional seasonal cycles
is by including regressors (ARIMAX). ARIMAX models are also useful to incorporate other information in
the models, such as holidays, paydays, etc., that are relevant for liquidity forecasting. Details about the
ARIMA model family can be found in Ord et al. (2017). Note that RegARIMA and ARIMAX are closely
connected, often being identical or simply implying estimation methodology differences.

For the specification of ARIMA we follow the methodology proposed by Hyndman and Khandakar (2008).
In brief, they propose to first make the time series stationary by testing for the appropriate order of
differencing. Then, start with a simple model and calculate a relevant information criterion, such as AIC.
Iterate the autoregressive and moving average orders by ±1 and calculate the resulting AIC. At each step,
choose the model with the lowest AIC and repeat the local search until the information criterion cannot be
improved further.

Finally, it is easy to show algebraically that there are equivalences between the purely additive ETS
models and ARIMA (Hyndman et al., 2008), while it is simple to obtain multiplicative forms by applying
ARIMA on log-transformed data. Mixed-form ETS models are not encompassed by ARIMA. Furthermore,
given the constrained modeling alternatives of ETS, it is often easier to identify a well-performing ETS
model when compared to ARIMA. Therefore, we use both modeling families in the liquidity forecasting
framework.

4.1.4. Trigonometric Seasonality Box-Cox, ARMA errors, Trend, and Seasonal Components

The Trigonometric Seasonality BOX-Cox, ARMA errors, Trend, and Seasonal Components (TBATS)
model incorporates many of the features of the models already introduced. With TBATS, seasonality and
trend are handled via exponential smoothing (using trigonometric terms for the former, see Section 4.1.5),
a Box-Cox transformation is used, and ARIMA errors are incorporated. A particularly attractive feature of
the TBATS model is its ability to handle multiple seasonalities, without any special treatment. For
additional details, the reader is referred to De Livera et al. (2011).

4.1.5. Modeling Multiple Seasonalities with Trigonometric Encoding

In a regression context, seasonality can also be introduced in models using binary indicator variables. For
example, day-of-week effects can be modeled using only four indicator variables (five-day week) of the
form:

IMF | MCM Technical Assistance Handbook 15

Liquidity Forecasting—Part II: The Statistical Component

(𝑆𝑆𝑆𝑆𝑆𝑆) 1 if day t is a Sunday,

𝐷𝐷𝑡𝑡 =�
0 otherwise.

Similar indicators (or dummies) can be defined for Mon, Tue, Wed, and Thur. These indicators are then
included in a vector of covariates 𝑥𝑥′𝑡𝑡 and the ARIMA model has the same specification as before, but with
𝑦𝑦𝑡𝑡 replaced by 𝑦𝑦𝑡𝑡 − 𝑥𝑥′𝑡𝑡 𝛽𝛽. A similar modification can be made for ETS. Of course, both ARIMA and ETS
can also model seasonality directly, without the inclusion of regressors.

Daily time series can exhibit multiple seasonal cycles that must be accounted for in the modeling. These
include day in the week, day in the month, and day in the year, corresponding to different cyclicities in the
data. This substantially complicates the creation of forecasts, as many models typically incorporate a
single seasonal periodicity. Three elements are of interest in modeling multiple seasonalities: the length
of the seasonal cycles, their encoding, and the efficiency of the latter, as we aim for parsimonious
models. To resolve questions raised by the first element, one counts how many days are in each
periodicity. For example, there are five days in the week (without weekends). However, day-of-the-month
seasonality is more challenging as months have a different number of days. To overcome this, quarterly
seasonality is used, as a quarter contains a fixed number of weeks, and, by extension, days.

The multiple seasonal cycles are encoded using trigonometric indicator variables. Given the length of a
season of 𝑠𝑠 periods, 𝑠𝑠/2 pairs of trigonometric variables are constructed, with 𝑖𝑖 = 1, … , 𝑠𝑠/2:

2𝑖𝑖π𝑡𝑡
𝑑𝑑𝑖𝑖 = 𝑐𝑐𝑐𝑐𝑐𝑐 � �,
𝑠𝑠

2𝑖𝑖𝑖𝑖𝑖𝑖
𝑑𝑑𝑖𝑖+𝑠𝑠/2 = 𝑠𝑠𝑠𝑠𝑠𝑠 � �,
𝑠𝑠

where 𝑡𝑡 = 1, … , 𝑛𝑛 (with 𝑛𝑛 being the sample size). When 𝑠𝑠 is an odd number, 𝑠𝑠/2 is rounded up to the
closest integer. This encoding is mathematically equivalent to using 𝑠𝑠 binary indicator variables, in which
case, each binary indicator would encode the level of a particular day in the season (Ghysels and
Osborn, 2001). Note that one of the indicators will correspond to a constant, resulting in 𝑠𝑠 − 1 informative
indicator variables.

A major advantage of trigonometric encoding over its binary counterpart is how it behaves when
indicators are removed to achieve sparsity. Consider a time series with a single seasonality and 𝑚𝑚 = 4,
i.e., a quarterly seasonality. This would require 3 binary indicators, or 3 trigonometric indicators (with the
4th being a constant and therefore excluded, as similarly, the 3 binary indicators assume the presence of
a constant in the model). A sparse encoding would use less than 3 indicators, and in our example, we
retain only the first indicator. Figure 5 visualizes the outcome. With full formulation, the resulting model fit
is identical from both binary and trigonometric encodings. With sparse formulation, the binary encoding
correctly captures the quarter for which the indicator is provided, while for the rest, the mean value of the
time series is estimated (from the constant in the model). In the case of trigonometric encoding, the model
is still able to approximate a simplified version of the seasonality. Eliminating terms with trigonometric
seasonality lowers the quality of the approximation but is still able to model the whole length of the
seasonal cycle. Therefore, with trigonometric encoding we can produce more parsimonious models that
are easier to estimate.

IMF | MCM Technical Assistance Handbook 16

Liquidity Forecasting—Part II: The Statistical Component

Figure 5. An Example of Sparse Binary and Trigonometric Seasonality Encoding

Binary encoding Trigonometric encoding

800

800
600

600
Data
Full
Sparse
400

400
200

200
1 2 3 4 5 1 2 3 4 5
Season Season

Source: IMF staff calculations.

A disadvantage of the trigonometric seasonality is that the corresponding coefficients are not interpretable
in the same way as with the binary encoding. For the latter, the coefficient signifies the vertical shift from
the coefficient of the constant term in the model to obtain the seasonal value. With the trigonometric
encoding, the coefficients should be interpreted as a Fourier decomposition of the time series.

For the liquidity forecasting framework, we propose to use binary encoding for the day-of-the-week
seasonality, as it requires minimal additional model terms and allows direct interpretation of the resulting
coefficients. For the remaining seasonalities we recommend using a trigonometric representation, which
is then made sparse. This is done using a regression, which can either rely on stepwise selection with
AIC, or, preferably, lasso regression. The lasso regression is tasked to find a good compromise between
how well the model fits the data and its complexity as measured by the number of parameters it has.
Models with more parameters (and therefore input variables) are better at modeling the observations, but
can potentially overfit, capturing the randomness in the time series instead of the underlying structure.
Lasso offers a better search of the model space than stepwise regression. More details about the lasso
regression can be found in Ord et al., (2017) and Kourentzes and Sagaert (2018).

To keep the complexity of the regression modeling low, first the trend of the time series is removed by
subtracting a centered moving average from the time series (Ord et al., 2017). The centered moving
average simply calculates the average of all values within a season that effectively models the trend in
the time series. This is subtracted from the data, and the residuals are then modeled with different
trigonometric indicator variables as explanatory variables.

Finally, it should be stressed that the seasonality modeled using binary or trigonometric indicators will be
deterministic. This means that the shape of the seasonality will not evolve over time. In contrast, the
seasonality modeled with autoregressive and moving average terms, in ETS, ARIMA, and TBATS, is
stochastic, which means that its shape can evolve over time. A detailed discussion of the differences
between the two and ways to identify them is provided by Ghysels and Osborn (2001). For the objectives
of the liquidity forecasting framework, we assume that longer period seasonalities evolve slower, such as
the annual seasonality, and therefore there is little loss of accuracy by modeling it as a deterministic

IMF | MCM Technical Assistance Handbook 17

Liquidity Forecasting—Part II: The Statistical Component

seasonality. On the other hand, for the day-of-the-week seasonality this can be a strong assumption.
Therefore, this seasonality is preferably treated by the models directly to retain its stochastic nature.

4.1.6. Level Shifts

Various effects can introduce level shifts to the autonomous factors. In recent years, in many countries,
the effect of Covid-19 resulted in very strong level shifts in liquidity. A level shift can be permanent or
transient, with the time series of interest returning to its original level. For models that can use regressors
(ETS and ARIMA) a level shift can be added with a binary indicator variable:

1 if t occurs after the level shift,

D𝑡𝑡 = �
0 otherwise.

If the analysts identify a level shift as transient, one can model this with two binary indicators, one
modeling the first level shift, and another modeling the termination of the shift. More economically, if the
resulting level is the same as the one before the shift, one can use a single binary indicator which takes
values of 1 only for the duration of the transient shift.

Note that models with autoregressive terms will tend to smooth the level shift over multiple periods,
depending on the autoregressive structure. This can help us model shifts that are happening over a
number of periods, without the need to include more complex encoding.

4.1.7. Calendar and Special Events

Both ETS and ARIMA models can be augmented by adding regressors in the same fashion as with
conventional regression modeling. Taking the series of currency in circulation (CIC) as an example,
Cabrero et al. (2009) identified the deterministic structure of the RegARIMA model as follows:

i. Fixed and Moving Festivals: These include holidays that have either fixed dates or change
annually, impacting consumer behavior and cash demand.

ii. Intramonthly Effects Related to Payrolls: These are regular fluctuations associated with payroll
cycles that influence cash withdrawals and deposits.

iii. Trading Day (Intraweek) Effects: This accounts for variations in economic activity based on
different days of the week, reflecting trading patterns and financial transactions.

These effects can be incorporated into the ARIMA and ETS models using binary indicators. We
distinguish a number of options in the modeling of these. First, these effects may be restricted solely to
the period of the special event. Second, if the model includes autoregressive terms, their effect can be
spread to the following periods. Third, we may want to introduce the leading and training effects of a
special event. This can be done with the introduction of multiple binary dummies, giving us full control of
the modeled response. Alternatively, a more economical approach is to use a parabolic indicator. Some
days before the holiday or event, the indicators start to increase quadratically and reach the maximum of
the parabolic curve with a value of 1. Then the indicators begin to decay quadratically and become 0.
Such encoding can best capture the real movement of liquidity caused by CIC. The public will start

IMF | MCM Technical Assistance Handbook 18

Liquidity Forecasting—Part II: The Statistical Component

withdrawing cash to prepare, for example, for a celebration right ahead of the holiday, while after the
holiday the demand for money will dwindle. These indicators are calculated as

𝑡𝑡− 𝑡𝑡ℎ𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜𝑜 2
𝐷𝐷𝑡𝑡 = max (0, 1 − ( ) ),
7

where the effect is spread across seven periods. In some cases, it may be advisable to use two parabolic
indicators, centered across two consecutive periods, to economically capture complex responses.

4.1.8. Augmented ETS and ARIMA Models

The various modeling options described in Sections 4.1.5–4.1.7 can be incorporated into ETS and ARIMA
models, allowing them to model the multiple complexities of the autonomous factors. Figure 6 provides
examples of using these options from cases in Guatemala and the United Arab Emirates (UAE). A
potential risk is that analysts oversaturate the models with additional regressors, which may or may not be
informative. This can make the model cumbersome to use and difficult to estimate. To avoid this, we
recommend evaluating the inclusion of the various indicators using lasso or stepwise based on AIC
regression. In the liquidity forecasting framework all regressors, apart from those modeling level shifts,
are evaluated in this way before being introduced to the models. Seasonality indicators and special event
regressors are evaluated separately, in two consecutive steps.

Finally, for the case of ARIMA, we consider alternatives with and without seasonality when considering
the additional regressors. This corresponds to modeling the day-of-the-week seasonality as stochastic,
deterministic, or mixed. The choice is resolved by comparing the AIC of the resulting ARIMA.

Figure 6. Modeling Deterministic Structure as Regressors: Example of CIC

Permanent or transitory structural breaks may exist due Increasing demand for currency may start right before
to economic, banking, or exchange rate crises. a holiday and gradually decrease right after, forming a
Continuous indicators can be used to codify structural parabolic arch. The figure below presents the case of
breaks. Eid-al-Fitr in the UAE, and the holiday may be moving.

Source: Haver (2024) Source: El Gemayel (2022)

IMF | MCM Technical Assistance Handbook 19

Liquidity Forecasting—Part II: The Statistical Component

An example of Guatemalan CIC demand, which tends Guatemalan CIC also demonstrate Intra-yearly
to increase close to weekends. This may be because seasonality, which reflects a huge increase of CIC due
wages are paid on Fridays, or the public are to the year-end holiday season. Trigonometric terms
withdrawing cash for weekends. Indicator variables can will be the more parsimonious way to model the intra-
be used to model such effects. yearly effect.

Source: Gallardo (2024) Source: Gallardo (2024)

4.2 Volatility Models

Volatility models are appropriate for forecasting series with high volatility. Normally, these models will be
applied to forecast NFA. Unlike CIC and GAB, the NFA series is usually volatile with little structural
information to be modeled. While the mean of NFA seems to follow a random walk, the first difference
does exhibit conditional heteroskedasticity, indicating that generalized autoregressive conditional
heteroskedasticity (GARCH)-type models probably will be more appropriate for NFA forecasts. Several
conditional volatility models are fitted to obtain the probabilistic forecasts of NFA, which will be later
transformed into point forecasting using bootstrapping simulation.

The most popular family of conditional volatility models is the GARCH model. The variance is modeled as:

𝑝𝑝 𝑝𝑝
2 2
𝜎𝜎𝑡𝑡2 = 𝜔𝜔 + � 𝛽𝛽𝑖𝑖 𝜎𝜎𝑡𝑡−𝑖𝑖 + � 𝛼𝛼𝑖𝑖 𝑒𝑒𝑡𝑡−𝑖𝑖
𝑖𝑖=1 𝑖𝑖=1

The proposed framework also makes available two variations of GARCH, which allows for asymmetric
effects of shocks on volatility: eGARCH and gjrGARCH. The specification of the exponential GARCH
model (eGARCH) is given by:

𝑝𝑝 𝑞𝑞
2
log𝜎𝜎𝑡𝑡2 = 𝜔𝜔 + � 𝛽𝛽𝑖𝑖 log𝜎𝜎𝑡𝑡−𝑖𝑖 + � 𝛼𝛼𝑖𝑖 𝑔𝑔(𝜖𝜖𝑡𝑡−𝑖𝑖 ).
𝑖𝑖=1 𝑖𝑖=1

IMF | MCM Technical Assistance Handbook 20

Liquidity Forecasting—Part II: The Statistical Component

where 𝑔𝑔(𝜖𝜖𝑡𝑡 ) = 𝜃𝜃𝜖𝜖𝑡𝑡 + 𝜆𝜆�|𝜖𝜖𝑡𝑡 | − 𝐸𝐸(|𝜖𝜖𝑡𝑡 |)�. An advantage of this specification is its asymmetry, since the sign
and magnitude of innovations have different effects on the variance.

The gjr (Glosten-Jagannathan-Runkle)-GARCH specification is given by:

𝜎𝜎𝑡𝑡2 = 𝜔𝜔 + 𝛿𝛿𝜎𝜎𝑡𝑡−1
2 2
+ 𝛼𝛼𝜖𝜖𝑡𝑡−1 2
+ 𝜙𝜙𝜖𝜖𝑡𝑡−1 𝐼𝐼𝑡𝑡−1

where 𝐼𝐼𝑡𝑡−1 = 0 if 𝜖𝜖𝑡𝑡−1 ≥ 0 and 𝐼𝐼𝑡𝑡−1 = 1 if 𝜖𝜖𝑡𝑡−1 < 0. Like eGARCH, this specification allows for asymmetric
effects. For more details on the volatility models the reader is referred to Tsay (2005).

Similar to the extrapolative time series models, we recommend using the random walk as a benchmark
for the volatility models.

4.3 Hierarchical Reconciliation Methods

To support OMO, the main quantity of interest is the net liquidity injection, AGG = NFA – CIC - GAB. One
approach would be to forecast each autonomous factor separately and aggregate them, which is common
practice in many central banks. An alternative would be to directly forecast AGG. As forecasts are
� 𝑡𝑡 , 𝐺𝐺𝐺𝐺𝐺𝐺
�𝑡𝑡 , 𝐶𝐶𝐶𝐶𝐶𝐶
imperfect, we anticipate that for a set of forecasts 𝐴𝐴𝐴𝐴𝐴𝐴 �𝑡𝑡 , and 𝑁𝑁𝑁𝑁𝑁𝑁
� 𝑡𝑡 the following will be true:

�𝑡𝑡 = 𝐴𝐴𝐴𝐴𝐺𝐺𝑡𝑡 + 𝑒𝑒𝐴𝐴𝐴𝐴𝐴𝐴,𝑡𝑡 ,

𝐴𝐴𝐴𝐴𝐺𝐺 (3)

�𝑡𝑡 = 𝐶𝐶𝐶𝐶𝐶𝐶𝑡𝑡 + 𝑒𝑒𝐶𝐶𝐶𝐶𝐶𝐶,𝑡𝑡 ,

𝐶𝐶𝐶𝐶𝐶𝐶 (4)

�𝑡𝑡 = 𝐺𝐺𝐺𝐺𝐵𝐵𝑡𝑡 + 𝑒𝑒𝐺𝐺𝐺𝐺𝐺𝐺,𝑡𝑡 ,

𝐺𝐺𝐺𝐺𝐵𝐵 (5)

�𝑡𝑡 = 𝑁𝑁𝑁𝑁𝐴𝐴𝑡𝑡 + 𝑒𝑒𝑁𝑁𝑁𝑁𝑁𝑁,𝑡𝑡 ,

𝑁𝑁𝑁𝑁𝐴𝐴 (6)

and

�𝑡𝑡 = 𝑁𝑁𝑁𝑁𝐴𝐴
𝐴𝐴𝐴𝐴𝐺𝐺 �𝑡𝑡 − 𝐶𝐶𝐶𝐶𝐶𝐶
�𝑡𝑡 − 𝐺𝐺𝐺𝐺𝐵𝐵
�𝑡𝑡 + 𝑒𝑒̃𝑡𝑡 , (7)

where 𝑒𝑒̃𝑡𝑡 is the reconciliation error, i.e., how much forecasting the disaggregate autonomous factors
disagrees with forecasting directly the AGG. By replacing in Equation (7) the expressions above that
include the forecast errors (3) – (6), we can easily see that the reconciliation error is connected with the
forecast errors. Therefore, if we were to eliminate the reconciliation error, we can anticipate some
reduction of the forecast errors. Intuitively, forecast reconciliation suggests that each forecast contains an
incomplete set of information, greater than what is available to the whole problem. By reconciling the
forecasts, this information is blended and therefore has the potential to improve the quality of the
forecasts. In fact, Panagiotelis et al. (2021) provide a proof that the total forecast errors across the
hierarchy, after reconciliation, will be reduced, irrespective of how the reconciliation is done. Note, that
this does not necessarily prove that the forecasts errors of AGG, the quantity of interest, will always be
better, and therefore empirical evaluation remains important.

Forecast reconciliation is mathematically achieved using a restricted forecast combination. First, we

define a vector that contains all the individual bottom-level (disaggregate) observations, 𝒃𝒃𝑡𝑡 =

IMF | MCM Technical Assistance Handbook 21

Liquidity Forecasting—Part II: The Statistical Component

(𝐶𝐶𝐶𝐶𝐶𝐶, 𝐺𝐺𝐺𝐺𝐺𝐺, 𝑁𝑁𝑁𝑁𝑁𝑁)′ and a vector that contains all individual observations across the hierarchy 𝒚𝒚𝑡𝑡 =
(𝐴𝐴𝐴𝐴𝐺𝐺𝑡𝑡 , 𝒃𝒃𝑡𝑡 )′. We can now write a summing matrix 𝑺𝑺, so that:

−1 −1 1
1 0 0
𝑺𝑺 = � �,
0 1 0
0 0 1

𝒚𝒚𝑡𝑡 = 𝑺𝑺𝒃𝒃𝑡𝑡 .

Observe that the first row of S matches the aggregation equation (2) while the rest of the rows simply
point to one of the bottom-level series. We can rewrite the expression for forecasts:

� 𝑡𝑡 + 𝒆𝒆�𝒕𝒕
�𝑡𝑡 = 𝑺𝑺𝒃𝒃
𝒚𝒚

that is a compact expression of Equation (7). Additionally, we can see that it resembles a regression
formulation. Wickramasuriya et al. (2019) showed that the reconciliation errors can be minimized by

�t = 𝑺𝑺(𝑺𝑺′𝑾𝑾−1 𝑺𝑺)−1 𝑺𝑺′𝑾𝑾−1 𝒚𝒚

𝒚𝒚 �𝑡𝑡 (8)

where 𝑾𝑾 is the variance-covariance matrix of the forecast errors of 𝒚𝒚 �. In practice, the estimation of 𝑾𝑾 can
be challenging and, instead, we opt for various approximations: (i) the OLS approximation, where 𝑾𝑾 = 𝑰𝑰,
the identity matrix; (ii) the structural (STR), where 𝑾𝑾 is a diagonal with each element being the row-wise
sum of 𝑺𝑺; (iii) the weighted least squares (WLS), where 𝑾𝑾 is a diagonal with each element being the MSE
of the corresponding forecasting model for that row; and (iv) the MinT, which estimates the complete
variance-covariance matrix shrinking the off-diagonal elements toward zero. A detailed discussion of the
approximations for the variance-covariance matrix and their implications is provided by Pritularga et al.
(2021), while a detailed overview of hierarchical reconciliation is given by Athanasopoulos et al. (2023).

Observe that Equation (8) can be written as 𝒚𝒚 �𝑡𝑡 where 𝑮𝑮 = (𝑺𝑺′𝑾𝑾−1 𝑺𝑺)−1 𝑺𝑺′𝑾𝑾−1 . This makes the
�t = 𝑺𝑺𝐆𝐆𝒚𝒚
forecast combination underlying forecast reconciliation apparent, where the forecasts across the whole
hierarchy are combined with the weights in 𝐆𝐆 to construct reconciled bottom-level forecasts, i.e., the
autonomous factors, that are multiplied by 𝑺𝑺 to provide reconciled forecasts for the complete hierarchy, as
per 𝒚𝒚𝑡𝑡 = 𝑺𝑺𝒃𝒃𝑡𝑡 . Moreover, note that the forecasts in 𝒚𝒚
�𝑡𝑡 are model independent, meaning that they can
originate from any of the models outlined in Sections 4.1. and 4.2. Or otherwise, and they may be further
adjusted, based on expert judgment by the analysts, before the reconciliation takes place.

As the forecasting performance of the reconciled AGG forecasts remains an empirical question, we
recommend the use of two benchmarks. First, a direct forecast of the AGG without any reconciliation, and
second, a direct aggregation from the three autonomous factors to the AGG (bottom-up).

V. Forecast Evaluation and Selection

The selection of the best forecast is an empirical question, for which we need to collect evidence of the
predictive performance of the various models on test data. To this end, we need to define appropriate
metrics and evaluation schemes. In this section, we introduce the various metrics that we use, the
evaluation scheme, and then provide heuristics to aid the users of the liquidity forecasting framework.

IMF | MCM Technical Assistance Handbook 22

Liquidity Forecasting—Part II: The Statistical Component

5.1 Evaluation Metrics

The most common practice to calibrate OMO is to rely on point forecasts of the mean. The proper score
for evaluating these forecasts is the Mean Squared Error (MSE) due to its quadratic loss (Gneiting and
Raftery, 2007). For scaling reasons, in practice, the Root Mean Squared Error (RMSE) is preferred over
the MSE for reporting. Liquidity can be subject to shocks, which are both difficult to forecast and not
representative of normal conditions. Therefore, it can be desirable to consider the Mean Absolute Error
(MAE), which is the proper score for the median of the predictive distribution, and is thus more robust to
extreme errors, which may occur due to shocks. Both metrics measure the magnitude of forecast errors
(accuracy). It is useful to measure the bias of forecasts, which captures their tendency to over or under
forecast. This can be done with the Mean Error (ME). Formally these metrics are defined as

2
1 𝑛𝑛
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅 = � � (𝑦𝑦𝑖𝑖 − 𝑦𝑦�)
𝚤𝚤 ,
𝑛𝑛 𝑖𝑖=1

𝑛𝑛
1
𝑀𝑀𝑀𝑀𝑀𝑀 = �|𝑦𝑦𝑖𝑖 − 𝑦𝑦�|
𝚤𝚤 ,
𝑛𝑛
𝑖𝑖=1

𝑛𝑛
1
𝑀𝑀𝑀𝑀 = �(𝑦𝑦𝑖𝑖 − 𝑦𝑦�)
𝚤𝚤 ,
𝑛𝑛
𝑖𝑖=1

where 𝑦𝑦𝑖𝑖 is the actual value, 𝑦𝑦�𝚤𝚤 is the forecast for the same period, and n is the number of periods
considered. The errors should be measured across forecasts of the same horizon, e.g., one period
ahead, and then, can be further averaged to provide an overall performance measurement. All RMSE,
MAE, and ME provide errors in the scale and units of the time series and cannot be used to summarize
across different time series. If a scale-independent version of the metrics is required, then we recommend
dividing the metrics by the in-sample RMSE or MAE (for ME as well) of the random walk. These metrics
are known in the literature as scaled metrics and avoid many of the pitfalls of other scale-independent
metrics. The reader is referred to Athanasopoulos and Kourentzes (2021) for more details.

Forecasts with very wide predictive distributions can imply large forecast uncertainty, and vice versa. Note
that this statement assumes that the models are well calibrated, as models can provide unreasonably
narrow prediction intervals, particularly when they are overfit to the training data. Forecasts are not
anticipated to be equal with future observations, but, instead, to follow their trajectory closely. This is due
to the randomness in all stochastic time series. Therefore, it is helpful to evaluate the forecasts
probabilistically. Intuitively, this evaluation investigates whether the expected percentage of observations
falls within the respective prediction intervals.

We argue that the central bank should forecast the complete predictive distribution to calibrate OMO and
publish the forecast. The predictive distribution can be interpreted as a representation of the risk that a
forecast carries. This risk can be due to the modeling risk, i.e., the ability of the forecasting models to
approximate well the underlying data-generating process, and the inherent stochasticity of the modeled
series. When using the point forecast, which corresponds to the mean of the predictive distribution, the
implicit assumption is that the consequences of under-allotting the operation are the same as those of

IMF | MCM Technical Assistance Handbook 23

Liquidity Forecasting—Part II: The Statistical Component

over-alloting. Note that this corresponds to the 50th percentile, the median, for symmetric distributions.
More generally, if this assumption is relaxed, which is more realistic, and the under- and over-allotting
have different costs, then a different percentile should be targeted, reflecting the risk preference of the
bank. Communicating the predictive distribution, rather than the point prediction, provides more
information to users, particularly for the associated forecast uncertainty. Users can thus factor in the
forecasting risk in their decisions. Additionally, providing the predictive distribution can mitigate potential
anchoring to the point forecast. We therefore recommend publishing the predictive distribution.

To assess the predictive distribution, we focus on the width of the prediction intervals. For this purpose,
one can use the Mean Interval Score (MIS). The MIS penalizes forecasts for the width of the prediction
intervals, as arbitrarily large intervals would include all observations, and for the number of observations
that fall outside the intervals. Therefore, it prefers forecasts with as narrow intervals as possible, while
containing the desired number of observations within these. For the evaluation, the 95 percent intervals
are considered, i.e., covering 95 percent of the observations. The metric is calculated as:

1 𝑛𝑛 2 2
𝑀𝑀𝑀𝑀𝑀𝑀 = � �(𝑈𝑈 − 𝐿𝐿) + (𝐿𝐿 − 𝑌𝑌𝑖𝑖 )1�𝑌𝑌𝑗𝑗 < 𝐿𝐿� + (𝑌𝑌𝑖𝑖 − 𝑈𝑈)1{𝑌𝑌𝑖𝑖 > 𝑈𝑈}�
𝑛𝑛 𝑖𝑖=1 α α

where α corresponds to the desired confidence level, e.g., 95 percent, U and L are the upper and lower
prediction intervals matching α, and 1(⋅) is an indicator function that takes the value of 1 when its
condition is true and 0 otherwise.

If a specific quantile is of interest, for example, due to an asymmetry of costs, the Pinball loss can be
used instead. The function of the Pinball is similar to MIS but only for one side of the interval. Other
metrics cater to different evaluations of the predictive distribution. A detailed treatment is provided by
Gneiting and Raftery (2007).

5.2 Evaluation Scheme

We collect errors on a test set to identify the appropriate forecast for a time series. Naturally, there is a
requirement to collect several error measurements to make a confident assessment of the performance of
a forecasting model. Likewise, the optimal performance remains unknown, as the data-generating
process of a time series is unknown. Nonetheless, benchmark model forecasts can help bound the
acceptable performance of forecasts. To this end, we organize forecasting models in different tiers of
complexity, as illustrated in Figure 7. For similar forecasting errors, simpler models are preferable, as they
require fewer modeling choices and they have less potential to overfit the data. Tier 1 includes the level
and seasonal random walks that provide an upper bound on the acceptable forecasting performance. Tier
2 relies on ETS, and ARIMA in level, seasonal, or fully automatic specifications, and TBATS. All of these
models, together with their specification methodologies, have been extensively researched and tested in
the literature. In practice, they require minimal intervention from the analyst, and they therefore provide a
set of benchmark forecasts that can capture key aspects of the liquidity time series. The last tier includes
ETS and ARIMA models augmented with the various regressors discussed in Section 4. Although these
models are the only ones that can capture various effects present in the liquidity time series, which are
well documented in the literature, they also have the largest potential to overfit, or to be overly complex
with potential estimation issues. Therefore, to provide reliable forecasts, we argue that these models

IMF | MCM Technical Assistance Handbook 24

Liquidity Forecasting—Part II: The Statistical Component

should be preferred only when there is strong evidence in their favor. Similarly, for volatility forecasts we
consider the Naive as a benchmark over the more complex GARCH models.

Figure 7. Forecasting Models Tiered in Groups of Increasing Complexity

Tier 1 Tier 2 Tier 3

ETS
Naive ETS with Regression
ARIMA
Seasonal Naive ARIMA with Regression
TBATS

Source: IMF staff.

To collect a sufficient sample of test errors we rely on a rolling origin evaluation scheme. For this, some
observations are retained aside as a test set, which is not used for model specification, and simulates
unseen data. Unlike in-sample evaluation, which measures the goodness of a model, out-of-sample
evaluation can indicate the predictive power, which aligns with our goal of forecasting. Out-of-sample
evaluation can help mitigate the potential overfitting of forecasting models. Relying on a single out-of-
sample measurement of the forecast accuracy of competing models is not adequate. Consider that each
observation of the time series contains structure and noise. Relying on a single measurement makes it
impossible to know whether the potentially high or low accuracy is due to modeling the structure, or due
to the randomness present in the time series. However, as the noise is randomly distributed around the
structure of the time series, averaging across multiple error measurements will tend to cancel out some of
this randomness and evaluate the performance of the forecasts against the underlying structure.

The rolling origin out-of-sample evaluation helps to achieve this. Figure 8 provides a representation of
how it operates. Using 𝑞𝑞 observations as in-sample to specify the forecasting models (blue dots),
forecasts for ℎ-steps ahead are generated for the out-of-sample period (red dots). The errors from this
first forecast origin are collected. In general, in the next step, the in-sample is increased by one period.
Using the new 𝑞𝑞 + 1 in-sample, the forecasting models are tuned again, and new ℎ-step forecasts are
generated from the second forecast origin. This process is repeated until all the available out-of-sample
have been exhausted. If 𝑢𝑢 is the size of the out-of-sample set, then 𝑢𝑢 − ℎ + 1 rolling forecasts and
respective errors are generated. The reader can find more details about the advantages of rolling origin
evaluation in Tashman (2000). Note that the rolling origin scheme is also referred to as cross-validation
over time.

IMF | MCM Technical Assistance Handbook 25

Liquidity Forecasting—Part II: The Statistical Component

Figure 8. The Out-of-sample Cross-validation Scheme

Source: IMF staff.

5.3 Testing the Forecasting Performance

Having collected forecast errors, the next question that needs to be resolved is what constitutes a
sufficient difference between the performance of forecasts, to prefer a model in a higher tier. We rely on
non-parametric statistical testing, namely, the Friedman and the post-hoc Nemenyi tests, which can
facilitate multiple comparisons without the need for repeated pairwise testing. The non-parametric nature
of the tests is convenient as it does not place any distributional assumptions on the forecast errors.
However, as such, non-parametric tests are typically weaker than their parametric counterparts, requiring
more samples. However, this is often not an issue in liquidity forecasting, as there is an abundance of
data.

The Friedman test first evaluates whether at least one of the methods is statistically different from the
rest. The Nemenyi test then groups the methods into subgroups. Briefly, the Nemenyi test calculates the
mean rank for each forecasting model and a critical distance. Any forecasting model that falls within the
mean rank ± critical distance is grouped together. The simpler model within the group with the lowest
mean rank is the preferred forecast. Additional details for these non-parametric tests can be found in
Hollander et al. (2003), and an example of how to apply them in forecasting comparisons in Kourentzes
and Athanasopoulos (2019). Derivations of the critical values for multiple models and various sample
sizes are provided by Kourentzes (2023).

Figure 9 provides an example of the use of the non-parametric tests. The vertical axis represents the
mean rank of the various forecasts. At each origin of the rolling origin evaluation the errors of the different
models are collected and ranked, with the lowest rank being best. This is done over all origins in the test
set and the mean is calculated. Forecasts that are connected with vertical lines belong to the same group.
From a group, the simplest method is preferred. In the example, we see evidence that ARIMA with
Regression is performing better than the rest of the forecasting models across multiple forecast horizons
of interest. If we focus on the t+1 week horizon, ARIMA with Regression, ETS with Regression, and
TBATS are all grouped together as best performing.

IMF | MCM Technical Assistance Handbook 26

Liquidity Forecasting—Part II: The Statistical Component

Figure 9. An Example Nemenyi Test Comparison for Three Different Target Forecast Horizons

Source: IMF staff calculations.

Balancing across multiple error metrics and comparisons can become cumbersome. Here we provide
some recommendations to help analysts. Note that these recommendations are heuristics and are based
on our experience with OMO and liquidity forecasting, and alternative rules can be formulated.

• If the focus is on the point forecast of the mean, we recommend that the analyst first rely on
RMSE comparisons. If a clear winning forecast does not emerge, then we recommend consulting
MAE results, and eventually ME. MIS can provide additional evidence, as it helps gauge the
associated uncertainty of the forecasts.

• If the focus is on the predictive distribution, MIS becomes necessary and should be consulted
first. Similarly, MIS is important for assessing volatility forecasts.

Note that the ranking of the forecasts may change across metrics, hence our recommendation is to
weight more the RMSE, and then the MAE, with the rest following. Moreover, although the forecasting
framework provides summary forecast errors, we recommend relying on the comparisons offered by the
statistical tests to support the choice of the appropriate forecast.

5.4 The Forecasting Process: Statistical Forecasts and Expert Judgment

A complete short-term liquidity forecasting framework should be a process of combining statistical

modeling and expert adjustments originating from communications with related stakeholders. The
expectation is that statistical forecasts can automatically make use of well-structured information to
provide baseline forecasts. The structured information includes historical observations and indicator
variables for holidays and special events. Additionally, through better communications with the
counterparties of the central bank, additional information such as budget revenues, gold purchase plans,
etc., can be obtained, which can be used to reduce the uncertainty of forecasts. This information can be

IMF | MCM Technical Assistance Handbook 27

Liquidity Forecasting—Part II: The Statistical Component

simply added to the forecasts (see Figure 10). More importantly, at each layer of forecasting, it is critical
to evaluate performance to guide any improvements.

Figure 10. A Schematic of the Process for Liquidity Forecasting

Policies

Historical Decisions by
liquidity actors (e.g.,
banks)

Trend &
season Statistical Expert Final
patterns forecasts knowledge forecast

Evaluation Evaluation
Known special
events (e.g.,
holidays) News

Source: IMF staff.

The literature has investigated in detail the efficacy of expert-judged adjustments in forecasts (Lawrence
et al., 2006; Arvan et al., 2019). The main findings are that experts are influenced by various behavioral
biases that can result in inconsistent performance of adjustments. This motivates our proposed separate
evaluation of the performance of expert adjustments, to provide feedback to analysts and guide them to
improve future adjustments. Sroginis et al., (2023) find that experts can be overwhelmed by contextual
information, which can lead to erroneous adjustments. To mitigate this, we recommend analysts use the
augmented ETS and ARIMA to incorporate all structured information in the models and limit their
adjustment to unstructured information that cannot be easily incorporated into a regression model.
Another helpful practice has been to decompose the expert-judged adjustment into its constituents, where
the experts have to account for all of the different factors that make up their total adjustment. This has
been found to help mitigate overly optimistic adjustments. The reader is referred to Sroginis et al. (2023)
and Arvan et al. (2019) for a detailed overview of the findings on behavioral adjustment in the literature.

The proposed framework for liquidity forecasting is implemented in R, an open-source mathematical

programming language. Figure 11 provides an overview of the forecasting workflow. Once the analysts
have prepared the data of the autonomous factors, and codified the various holidays, level shifts, and
other events, the implementation provides various automations to obtain rolling origin errors for the
various described forecasting models. Aided by the reports produced by the framework, the analyst is
asked to choose the best forecasts for each of the time series. Forecasts are then generated and
reconciled. The analyst is provided with the relevant reports to aid in the selection of which hierarchical
reconciliation is the most appropriate, leading to the final forecast output.

IMF | MCM Technical Assistance Handbook 28

Liquidity Forecasting—Part II: The Statistical Component

Figure 11. The Workflow for Statistical Modeling of Short-term Liquidity

Source: IMF staff.

The implementation offers various automations to enable the analysts to conduct the analysis easily. It is
not required to reconfigure the models at every forecasting cycle; however, it is strongly advised that this
be done periodically. The frequency at which reconfiguration is done should be a balance between the
workload of the analysts and computational resources.

5.5 Limitations and Extensions

The framework incorporates static and dynamic forecast combinations of pools of models. Forecast
pooling is considered an advantageous step in the forecasting process, aiding in both mitigating modeling
uncertainty and improving forecast accuracy (Elliot and Timmerman, 2013; Kourentzes et al., 2019).
Nonetheless, these are not currently part of the main forecasting workflow. Our experience with applying
the liquidity forecasting framework in various countries is that the forecast combinations offer marginal
improvements for the additional modeling complexity. Moreover, as the hierarchical reconciliation is a part
of the main workflow, the final forecasts incorporate a form of forecast combination. Our experiments
have indicated that there are diminishing returns to using two “layers” of forecast combination, which has
guided our choice to retain the forecast combination and pooling options as optional. During the
COVID-19 pandemic we found evidence that the additional dynamic forecast combination was helpful, as
the various forecasting models were failing in complimentary ways.

The liquidity forecasting framework does not automate any machine learning or artificial intelligence
methods. We have trialed the use of shallow neural network and random forest forecasts. We did not find
evidence of forecasting performance benefits, particularly given the additional computational cost. In
recent years, global forecasting methods have demonstrated good performance. A potential
implementation for the liquidity forecasting framework would be to train models across liquidity data from
multiple countries, to leverage the benefits of global learning. Although this entails data collection
complications, it is a fruitful avenue for future research.

Likewise, another modeling methodology that has the potential to further improve the forecasting
performance of the models is the use of Temporal Hierarchies (Athanasopoulos et al., 2017). Much like
the forecast reconciliation across the autonomous factors, Temporal Hierarchies enable reconciliation
over time scales. Beyond the forecasting accuracy gains, this modeling approach can help bridge the gap

IMF | MCM Technical Assistance Handbook 29

Liquidity Forecasting—Part II: The Statistical Component

between short- and long-term liquidity forecasting. At more aggregate time scales, it is simpler to
generate long-term forecasts, but also to incorporate additional macroeconomic variables. The contained
information can supplement the short-term forecasts in a mechanism similar to the one described in for
the reconciliation of forecasts across autonomous factors. Kourentzes and Athanasopoulos (2019)
propose a cross-temporal formulation that incorporates the reconciliation used here with temporal
hierarchies in a single modeling step and could be a useful extension of the liquidity forecasting
framework.

VI. Technical Assistance Approach

Below we provide a recommended process for using the liquidity forecasting framework in Technical
Assistance (TA) missions:

i. At least six weeks before the mission (T-6), request the time series. Ideally, the TA recipient
should provide at least three years of central bank daily balance sheet data and convert them into
a liquidity table, showing autonomous factors, banks’ reserves at the central bank, and monetary
operations. In the absence of this extensive set of data, use the standardized data request in
Appendix I. Share with the authorities the mission requirements in terms of software (R and
RStudio) and staffing (staff with coding capacity). 4

• At T-6, prepare the data for modeling. Follow up with the authorities on any problems
encountered in preparing the data. Run the models. Prepare a (PowerPoint) presentation
including: (i) time series descriptive analysis and model configuration; (ii) model selection
statistics; (iii) point forecasts output with predictive intervals; (iv) possible forecast
enhancement with reconciliation; and (v) alignment with existing liquidity table. The
standardized output template is presented in Appendix II.

ii. At T-2, the standardized results are presented to the TA recipient virtually. A Q&A session is
organized. The team will explain that the in-person mission will mainly consist of a workshop with
the staff of the TA recipient running the model on their own devices. The TA recipient is reminded
of the mission requirement, including staffing for the workshop and software.

iii. From T-2 to T-0, a draft report is prepared and circulated internally in MCMCO for preliminary
comments.

iv. At T-0, the draft report is shared with the TA recipient for its comments (due before the end of the
mission). The workshop session starts. The workshop consists of: (i) set-up of infrastructure and
working environment with relevant packages; (ii) R coding essentials; (iii) customization of model
configurations; (iv) execution of forecasting codes; (v) interpretation of the results; (vi) alignment
with existing liquidity table and OMO calibration; and (vii) discussion of qualitative input and other

4Liquidity forecasting should be in the policy implementation department. However, adequate quantitative and coding
skills may have to be drawn from the research or policy areas, creating the need for inter-departmental cooperation.

IMF | MCM Technical Assistance Handbook 30

Liquidity Forecasting—Part II: The Statistical Component

liquidity impacting scenarios. At T+1 (end of the mission), the authorities provide comments on
the draft reports and the draft is circulated for formal review at IMF Headquarters.

v. At T+3, the comments are included with the aide-mémoire and the report is finalized.

The TA should advocate in favor of publishing the forecast if forecast uncertainty is also disclosed. The
forecast would inform the central bank counterparties’ bidding at the central bank monetary operations,
which should help the central bank to achieve its operational target regardless of its operational
framework. However, the central bank should also publish the distribution forecast to allow the user of the
forecast to factor in forecasting risk.

Due to the complexity of the proposed models, the TA should prioritize skill development within central
banks. The TA experience is that coding skills are more broadly available inside central banks than
organically expected. That said, the TA should dedicate significant time to hand-on training in the context
of workshops using the central bank data and the code developed by the mission. The TA should ensure
that central bank staff could prepare the forecast independently before the end of mission. One approach
is to request the central bank colleagues to prepare and present the forecast at their senior management
at the end of mission meeting.

Except for the training session mentioned above, some other challenges encountered during the previous
TAs and their solutions include:

(1) Difficulties in inter- and/or intra-departmental data sharing. The mission team should assess the
data availability and make sure the data are sufficient for the modeling before the mission. In
addition, the mission team should emphasize the importance of data communication channels
with relevant counterparties and elaborate on how expert judgment is integral to the whole
forecasting process. If there is delayed delivery of data, the mission team should advise on
rescheduling the forecasting date or adopting flexible forecasting horizons.

(2) IT or hardware restrictions preventing the installation of RStudio and relevant packages. The
mission team should instruct the central bank counterparts on the installation of the infrastructure
and make sure there are no IT restrictions on running such open-source software and packages.
Central bank colleagues are usually advised to consult with their IT department in advance to
clear existing restrictions.

(3) Open-source packages version update causing bugs in the codes. Because the toolkit is
designed in a way to reduce the users’ learning curve, we have hidden underlying codes from the
users. Some package updates, such as the removal of a certain function, will undermine the
functionality of the toolbox. It requires our effort to modify the underlying code to make the
updated package become compatible again. The mission team will advise the forecaster not to
update any package or RStudio related to the forecasting exercise.

IMF | MCM Technical Assistance Handbook 31

Liquidity Forecasting—Part II: The Statistical Component

VII. Conclusions
In this document, we have proposed a comprehensive forecasting framework designed to tackle the
intricacies of predicting both autonomous factors and aggregate liquidity. Recognizing the limitations of
existing forecasting approaches, which often necessitate complex specifications and maintenance, our
framework is designed to be generic and adaptable. Moreover, it offers a high degree of automation,
thereby aiding teams of analysts with different degrees of expertise.

The framework is a holistic process consisting of not only statistical modeling, but also a modeling
process, helping to select and generate resilient forecasts and incorporate expert judgment. The
statistical component of the framework is detailed with two distinct families of models tailored to
accommodate the unique characteristics of autonomous factors. These are rigorously tested and
compared using cross-validation techniques so that the best-performing model can be identified.
Furthermore, we leverage forecast reconciliation methods to further reduce modeling uncertainty and
improve forecast accuracy.

In conclusion, our proposed forecasting framework represents a significant advancement in central bank
liquidity management, offering a generic yet efficient pipeline to predict short-term liquidity supply. The
application of this framework is expected to facilitate decision-making processes, optimize liquidity
interventions, and ultimately contribute to a more stable financial environment.

IMF | MCM Technical Assistance Handbook 32

Liquidity Forecasting—Part II: The Statistical Component

Appendix I. Data Request Template for Liquidity Forecasting

IMF: Monetary and Capital Markets Department

Central Bank Operations Division

Technical Assistance Mission: Liquidity Forecasting

INTRODUCTION

MCM will soon field a mission to assist the name of Central Bank (CB) with the conduct of its monetary
policy focusing on liquidity forecasting. In preparation for this mission, we ask for the following information
and data to be provided by date. Please provide all data in Excel format.

Institutional Arrangements

1. Other than commercial banks, what entities (financial and non-financial) hold accounts at CB?

a. Is there a Memorandum of Understanding with these entities that dictate how these accounts can
be operated, for example, imposing limits on maximum and minimum balances, or notification
periods for transfer of funds?

2. Is there a Memorandum of Understanding governing the cash and debt management relationship
between CB and the Ministry of Finance? If so, please provide a copy.

3. What information does the Ministry of Finance provide to CB regarding its cash management
activities?

Liquidity Forecasting

4. Please describe the current liquidity monitoring and forecasting process.

a. Which divisions within CB are involved?

b. What is the interval (e.g., daily) and horizon (e.g., one week) of the forecasts?

c. Please provide the liquidity template.

5. How are the autonomous factors forecasted (i.e., net foreign assets, government account, and
currency in circulation)? Please provide the specifications of all models used for forecasting the
autonomous factors.

6. Is there ex-post evaluation of forecasting accuracy undertaken? If so, please describe this process.

7. Are liquidity forecasts published?

IMF | MCM Technical Assistance Handbook 33

Liquidity Forecasting—Part II: The Statistical Component

8. Is there a model for estimating the (precautionary) demand for reserves and, if so, over what intervals
(e.g., daily, weekly)? Please provide the specification of any model used.

Data Requirements

Time Series or Textual Information Frequency Period

A sample of the Liquidity Table. Any Day
List of seasonal factors you think are relevant for
liquidity in your country, including not only the list of
dates but also the interval (e.g., +/- 6 days around a
religious holiday, etc.). Please differentiate between:
− Calendar seasonality: end of the week, end of
month, end of quarter, etc.
− Religious holidays.
− National non-religious holidays.
− Others (summer months, etc.).
Time series of the above list (e.g., the past 3 years),
but also for the current and next years.
List of any ad-hoc item that could be relevant as a
structural break, or to better understand the
specificities of your country (for instance, “after 2011,
the dynamic has changed due to a new policy,” etc.).
Liquidity Forecasting Details Done by CB or Counterparties
Past forecasts of currency in circulation, overall net
balance of government account at the central bank
(by the Treasury or CB), net foreign assets, and/or
other items that have an impact on liquidity. Please
include:
• Forecasted values.
• Dates when the forecasts were made
(e.g., forecast for April 5th based on March
30th information).
• Forecast horizon.
• If expert adjustments of the forecasts are
made, and these are stored separately, or
somehow indicated, please provide this
information as well.
• Any analysis done on the forecasting errors,
charts, etc.
• The main parameters of the model, as well as
the training sample.
• If possible, please share with us the codes
you use.

IMF | MCM Technical Assistance Handbook 34

Liquidity Forecasting—Part II: The Statistical Component

Currency in Circulation
Withdrawals of banknotes and coins (“increase of
currency in circulation”) by commercial banks from the Daily 2017–July 2023
central bank.
Deposits of banknotes and coins (“decrease of
currency in circulation”) by commercial banks to the Daily 2017–July 2023
central bank.
If the breakdown withdrawals/deposits are not
available, please provide time series of currency in Daily 2017–July 2023
circulation.
Time series of exogeneous factor(s) relevant for
forecasting currency in circulation in your country: for
Daily 2017–July 2023
instance, interest rate, exchange rate, high frequency
trade data, tourism data, etc.

Position of the Government at Central Bank

Overall net balance of the government account at the

Daily 2017–July 2023
central bank.
Daily credit and debit of government deposits at the
central bank. 1

If possible, please provide the decomposition of the

credit and debit of government deposits between:
− Revenues (taxes and tariffs).
Daily 2017–July 2023
− Current expenditures (public servant salaries,
pensions, etc.).
− Debt service.
− Maturing government debt.
− New issues of government debt.
− Other items.
Daily central bank claims on the government split
between:
− Securities: please provide the amount,
average duration, maturity, and interest rate.
Daily 2017–July 2023
− Loans and advances: please provide the
amount, average maturity, and interest rate.
− Others: please provide the amount, average
maturity, and interest rate.
Time series of exogeneous factors relevant for the
Daily 2017–July 2023
estimation: subsidies, grants, etc. (if relevant).

1 The Treasury can simultaneously withdraw and deposit at the Central Bank on the same day, which is why we ask
for the data separately.

IMF | MCM Technical Assistance Handbook 35

Liquidity Forecasting—Part II: The Statistical Component

Net Foreign Assets

Time series of net foreign assets at CB. Daily 2017–July 2023

Central bank purchase of foreign assets from

Daily 2017–July 2023
commercial banks.
Central bank sales of foreign assets to commercial
Daily 2017–July 2023
banks.
Time series of exogeneous factors relevant for the
estimation: exchange rate, interest rate spread, data
Daily 2017–July 2023
on sovereign fund/public pension fund, SWIFT
payments data, capital flows data (if relevant).

Other Items on Balance Sheet

Time series of other relevant items explaining the

Daily 2017–July 2023
systemic liquidity in your country.
Time series of any exogeneous factors relevant for
Daily 2017–July 2023
the estimation of these other items.
Details on the auctions of the last 3 years for liquidity-
injecting operations per maturity (if any):
− Instrument type (CB bill or repo).
− Number of participants.
− Amount announced.
− Amount submitted.
− Allocated amount.
Daily 2017–July 2023
− Minimum rate required.
− Minimum rate submitted.
− Maximum rate submitted.
− Minimum rate allocated (marginal rate).
− Weighted average rate.
− Percentage of allocation at minimum rate.
− Eligible collateral.

IMF | MCM Technical Assistance Handbook 36

Liquidity Forecasting—Part II: The Statistical Component

Appendix II. Standardized Output

The contents of this appendix are based on Gallardo (2024),

A. Time Series Descriptive Statistics Report: An Example of CIC

Models fit for Currency in Circulation over the period 2022-12-29 to 2023-07-20.

The following reports on time series pattern exploration for Currency in Circulation. Data are measured in
millions of Guatemalan quetzals (GTQ). The period used for testing the forecasting models is highlighted.
The long-term trend is captured using a centered moving average and is plotted in red.

Daily time series can exhibit multiple seasonal patterns: (i) day of the week; (ii) day of the month; and
(iii) day of the year. As the number of days in a month are not fixed, the seasonal profile within a quarter
is investigated.

The plots below are generated by de-trending the time series and then plotting all seasons as a
distribution. For instance, the “Week” seasonal plot shows the distribution of values for days of the week,
more specifically, the minimum, 10 percent, 25 percent, the median, 75 percent, 90 percent, and the
maximum values. Seasonality is identified as a clear pattern across the periods of the season.

IMF | MCM Technical Assistance Handbook 37

Liquidity Forecasting—Part II: The Statistical Component

B. Forecasting Models Cross-Validation and Comparison Report: An Example of CIC

Models fit for Currency in Circulation over the period 2022-12-29 to 2023-07-20.

The following reports on models fit to data from 2022-12-29 to 2023-07-20. Data are measured in GTQ
millions. Results are summarized for Currency in Circulation.

ACCURACY RESULTS

We provide Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) for forecasting up to one,
two, and four weeks ahead. RMSE is more sensitive to extreme values than the MAE. The lower the
value is, the more accurate the results are. Sorted by one-week accuracy value.

RMSE

MAE

IMF | MCM Technical Assistance Handbook 38

Liquidity Forecasting—Part II: The Statistical Component

BIAS RESULTS

Unbiased forecasts should have a Mean Error (ME) close to zero. Sorted by the one-week bias.

ME

PREDICTIVE DISTRIBUTION EVALUATION

The Mean Interval Score (MIS) assesses how well the 95 percent prediction intervals capture the real
distribution of the observed data. The lower the value is, the better the performance is, sorted by the 1-
week horizon.

MIS

IMF | MCM Technical Assistance Handbook 39

Liquidity Forecasting—Part II: The Statistical Component

VISUAL SUMMARY

Performance of the models across the various metrics is plotted below. The forecasts are ordered from
worst to best, according to each criterion, for producing forecasts for one-week ahead. The best forecast
for each horizon is highlighted with a green circle. When a Naive forecast is available, the errors are
provided relative to it. Any forecast less accurate than the Naive (performance equal to 1) should not be
considered.

IMF | MCM Technical Assistance Handbook 40

Liquidity Forecasting—Part II: The Statistical Component

STATISTICAL TESTING

We test whether the reported differences between forecasts are statistically significant. As the errors are
tracked over different periods in the test set, the results will vary as more evidence is collected. Statistical
tests attempt to account for this in comparing the models.

We use the Friedman and the Nemenyi non-parametric tests and report their results at a 95 percent
significance level. In the visualizations below, the models are ranked according to their mean rank (the
lower, the better), which is shown next to the model's name. Models that are connected by a vertical line
are grouped, and statistically speaking, there is no evidence of statistical differences. From a group, we
are statistically indifferent about which model to prefer. Nonetheless, simpler models are more resilient
and simpler to maintain and understand.

Note that as more test periods are accumulated, the results will become clearer, and for a limited number
of periods these tests have limited power to distinguish between the competing forecasts.

IMF | MCM Technical Assistance Handbook 41

Liquidity Forecasting—Part II: The Statistical Component

IMF | MCM Technical Assistance Handbook 42

Liquidity Forecasting—Part II: The Statistical Component

IMF | MCM Technical Assistance Handbook 43

Liquidity Forecasting—Part II: The Statistical Component

C. Forecasting Results: An Example of Aggregated Liquidity

The following contains details for forecasting of AGG. All models were forecasted on 2023-08-31.

LATEST FORECASTS (SELECTED MODEL)

The model selected for AGG is ARIMA with Regression. This selection can be changed in the
configuration file. The plot below shows forecasts with prediction intervals. The darker bands indicate an
80 percent prediction interval, while the lighter bands indicate a 95 percent prediction interval. The
observed data is shown in black.

IMF | MCM Technical Assistance Handbook 44

Liquidity Forecasting—Part II: The Statistical Component

LATEST FORECASTS (MODEL AVERAGES)

The forecasts below correspond to the equally weighted average of all models.

IMF | MCM Technical Assistance Handbook 45

Liquidity Forecasting—Part II: The Statistical Component

FORECASTS FOR ALL MODELS

The chart below shows the forecasts for all models.

D. Enhancement of Aggregated Liquidity Forecast through Reconciliation Report

Models fit over the period 2023-01-05 to 2023-07-27

The following contains details of reconciling liquidity forecasts. The models chosen for each autonomous
factor are as follows:

• State Account Balance: ETS with Regression.

• Net Foreign Accounts: ETS with Regression.

• Net Other Assets: ARIMA with Regression.

• Currency in Circulation: ARIMA with Regression.

• Aggregate: ARIMA with Regression.

These selections can be changed in the configuration file.

IMF | MCM Technical Assistance Handbook 46

Liquidity Forecasting—Part II: The Statistical Component

Most Recent Forecast for 2023-08-31

One-step-ahead forecasts for the most recent origin date (2023-08-31) are tabulated below for both
reconciled and unreconciled forecasts. The method selected for reconciliation is the STR method
(shaded), while the best method based on past performance (RMSE) is the OLS method (yellow).

Forecast Evaluation

Validation period used for the evaluation: 2023-01-05 to 2023-07-27.

RMSE (ACCURACY)

The table below summarizes RMSE for forecasts using different reconciliation methods. The chosen
method (STR) is shown shaded while the best method based on average RMSE across horizons (OLS) is
shown in yellow.

MAE (ACCURACY)

The table below summarizes MAE for forecasts using different reconciliation methods. The chosen
method (STR) is shown shaded while the best method based on average MAE across horizons (OLS) is
shown in yellow.

ME (BIAS)

The table below summarizes ME for forecasts using different reconciliation methods. The chosen method
(STR) is shown shaded while the best method based on average ME across horizons (MinT) is shown in
yellow.

IMF | MCM Technical Assistance Handbook 47

Liquidity Forecasting—Part II: The Statistical Component

MIS (PREDICTIVE DISTRIBUTION)

The table below summarizes MIS for forecasts using different reconciliation methods. The chosen method
(STR) is shown shaded while the best method based on average MIS across horizons (Base
(Unreconciled)) is shown in yellow.

VISUAL SUMMARY

Performance of the models across the various metrics is plotted below. The forecasts are ordered from
worst to best, according to each criterion, for producing forecasts for one week ahead. The best forecast
for each horizon is highlighted with a green circle. All errors are presented as relative to the unreconciled
(base) forecast.

IMF | MCM Technical Assistance Handbook 48

Liquidity Forecasting—Part II: The Statistical Component

IMF | MCM Technical Assistance Handbook 49

Liquidity Forecasting—Part II: The Statistical Component

E. Enhanced Forecasts for Aggregated Liquidity Report

The changes in Net Liquidity for one, two, and four weeks ahead are provided below.

Forecasts or the net liquidity position by the chosen method (STR) are plotted below.

IMF | MCM Technical Assistance Handbook 50

Liquidity Forecasting—Part II: The Statistical Component

References
Arvan, M., Fahimnia, B., Reisi, M., and Siemsen, E. “Integrating human judgement into quantitative
forecasting methods: A review,” Omega, vol. 86, 2019, pp. 237–252.

Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., and Petropoulos, F. “Forecasting with temporal
hierarchies,” European Journal of Operational Research, vol. 262, no. 1, 2017, pp. 60–74.

Athanasopoulos, G., Hyndman, R. J., Kourentzes, N., and Panagiotelis, A. “Forecast reconciliation: A
review,” International Journal of Forecasting, vol. 40, no. 2, 2023, pp. 430-456.

Athanasopoulos, G., and Kourentzes, N. “On the evaluation of hierarchical forecasts.” International
Journal of Forecasting, vol. 39, no. 4, 2021, pp. 1502-1511.

Bell, W. R., & Hillmer, S. C. "Modeling time series with calendar variation," Journal of the American
Statistical Association, vol. 78, no. 383, 1983, pp. 526–534.

Cabrero, A., Camba‐Mendez, G., Hirsch, A., and Nieto, F. “Modelling the daily banknotes in circulation in
the context of the liquidity management of the European Central Bank”. Journal of
Forecasting, vol. 28, no.3, 2009, pp. 194-217.

De Livera, A. M., Hyndman, R. J., and Snyder, R. D. “Forecasting time series with complex seasonal
patterns using exponential smoothing,” Journal of the American Statistical Association, vol. 106,
no. 496, 2011, pp. 1513–1527.

El Gemayel, J., Lafarguette, R., MCM, J. F., ITD, K. M., del Castillo Penna, R. A., Merrouche, O., and
Panagiotelis, A. N. United Arab Emirates: Technical Assistance Report – Liquidity Management
and Forecasting, International Monetary Fund, 2022.

El Hamiani Khatat, M. Monetary policy and models of currency demand, International Monetary Fund,
2018.

Elliott, G., and Timmermann, A., eds. Handbook of economic forecasting, Newnes, 2013.

Gallardo, D., Chen, Z., and Veyrune, R. Guatemala: Technical Assistance Report—The Statistical
Component of Liquidity Forecasting, International Monetary Fund, 2024.

Ghysels, E., and Osborn, D. R. The econometric analysis of seasonal time series, Cambridge University
Press, 2001.

Gneiting, T., and Raftery, A. E. “Strictly proper scoring rules, prediction, and estimation,” Journal of the
American Statistical Association, vol. 102, no. 477, 2007, pp. 359–378.

Gray, S. Liquidity forecasting, Bank of England, 2008.

Harvey, A., Koopman, S. J., and Riani, M. "The modeling and seasonal adjustment of weekly
observations," Journal of Business & Economic Statistics, vol. 15, no. 3, 1997, pp. 354–368.

IMF | MCM Technical Assistance Handbook 51

Liquidity Forecasting—Part II: The Statistical Component

Hlaváček, M., Čada J., and Hakl F. "The application of structured feedforward neural networks to the
modelling of the daily series of currency in circulation," International Conference on Natural
Computation. Berlin, Heidelberg: Springer Berlin Heidelberg, 2005.

Hollander, M., Wolfe, D. A., & Chicken, E. Nonparametric statistical methods. John Wiley & Sons, 2003.

Hyndman, R.J., and Athanasopoulos, G. “Forecasting: principles and practice,” 3rd edition, OTexts:
Melbourne, Australia, 2021. OTexts.com/fpp3.

Hyndman, R. J., and Khandakar, Y. “Automatic time series forecasting: the forecast package for
R,” Journal of Statistical Software, vol. 27, 2008, pp. 1–22.

Hyndman, R., Koehler, A. B., Ord, J. K., and Snyder, R. D. Forecasting with exponential smoothing: the
state space approach. Springer Science & Business Media, 2008.

Ikoku, A. "Modeling and forecasting currency in circulation for liquidity management in Nigeria," CBN
Journal of Applied Statistics, vol. 5, no. 1, 2014, pp. 79–104.

Iskandar, I., Willett, Roger., and Xu, S. "The development of a government cash forecasting model,"
Journal of Public Budgeting, Accounting & Financial Management, vol. 30, no. 4, 2018,
pp. 368–383.

Kourentzes, N. “tsutils: Time Series Exploration, Modelling and Forecasting,” R package version 0.9.4,
https://CRAN.R-project.org/package=tsutils, 2013.

Kourentzes, N., and Athanasopoulos, G. “Cross-temporal coherent forecasts for Australian

tourism,” Annals of Tourism Research, vol. 75, 2019, pp. 393–409.

Kourentzes, N., Barrow, D., and Petropoulos, F. “Another look at forecast selection and combination:
Evidence from forecast pooling,” International Journal of Production Economics, vol. 209, 2019,
pp. 226–235.

Kourentzes, N., and Petropoulos, F. “Forecasting with multivariate temporal aggregation: The case of
promotional modelling” International Journal of Production Economics, vol. 181, 2016,
pp. 145–153.

Kourentzes, N., and Sagaert ,Y. “Incorporating Leading Indicators into Sales Forecasts,” Foresight: The
International Journal of Applied Forecasting, vol. 48, 2018, pp. 24–40.

Koziński, W., and Świst, T. "Short-term currency in circulation forecasting for monetary policy purposes–
the case of Poland," Financial Internet Quarterly, vol. 11, no. 1, 2015, pp. 65–75.

Lawrence, M., Goodwin, P., O'Connor, M., and Önkal, D. "Judgmental forecasting: A review of progress
over the last 25 years," International Journal of Forecasting, vol. 22, no. 3, 2006, pp. 493–518.

IMF | MCM Technical Assistance Handbook 52

Liquidity Forecasting—Part II: The Statistical Component

Nasiru, S., Luguterah, A., and Anzagra, L. "The efficacy of ARIMAX and SARIMA models in predicting
monthly currency in circulation in Ghana," Mathematical Theory and Modeling, vol. 3, no. 5, 2013,
pp. 73–81.

Ord, K., Fildes, R., and Kourentzes, N. Principles of Business Forecasting. 2nd ed. Wessex Press
Publishing Co, 2017.

Panagiotelis, A., Athanasopoulos, G., Gamakumara, P., and Hyndman, R. J. “Forecast reconciliation: A
geometric view with new insights on bias correction,” International Journal of Forecasting, vol. 37,
no. 1, 2021, pp. 343–359.

Pritularga, K. F., Svetunkov, I., and Kourentzes, N. “Stochastic coherency in forecast

reconciliation,” International Journal of Production Economics, vol. 240, 2021, article 108221.

Sroginis, A., Fildes, R., and Kourentzes, N. “Use of contextual and model-based information in adjusting
promotional forecasts,” European Journal of Operational Research, vol. 307, no. 3, 2023, pp.
1177–1191.

Tashman, L. J. “Out-of-sample tests of forecasting accuracy: An analysis and review,” International

Journal of Forecasting, vol. 16, no. 4, 2000, pp. 437–450.

Tsay, R. S. Analysis of Financial Time Series. John Wiley & Sons, 2005.

Wickramasuriya, S. L., G. Athanasopoulos, and R. J. Hyndman. “Optimal Forecast Reconciliation for

Hierarchical and Grouped Time Series through Trace Minimization,” Journal of the American
Statistical Association, vol. 114, no. 526, 2019, pp. 804–819.

Williams, M. Government cash management: Its interaction with other financial policies, International
Monetary Fund, 2010.

IMF | MCM Technical Assistance Handbook 53