Estimation and Inference of Impulse Responses

by Local Projections

By ÒSCAR JORDÀ*

This paper introduces methods to compute impulse responses without specification

and estimation of the underlying multivariate dynamic system. The central idea
consists in estimating local projections at each period of interest rather than
extrapolating into increasingly distant horizons from a given model, as it is done
with vector autoregressions (VAR). The advantages of local projections are numer-
ous: (1) they can be estimated by simple regression techniques with standard
regression packages; (2) they are more robust to misspecification; (3) joint or
point-wise analytic inference is simple; and (4) they easily accommodate experi-
mentation with highly nonlinear and flexible specifications that may be impractical
in a multivariate context. Therefore, these methods are a natural alternative to
estimating impulse responses from VARs. Monte Carlo evidence and an application
to a simple, closed-economy, new-Keynesian model clarify these numerous advan-
tages. (JEL C32, E47, C53)

In response to the rigid identifying assump- well characterized by a VAR, Arnold Zellner and
tions used in theoretical macroeconomics dur- Franz Palm (1974) and Kenneth F. Wallis (1977)
ing the 1970s, Christopher A. Sims (1980) show that the macroeconomy’s subset of variables
provided what has become the standard in em- that practitioners can analyze at one time will
pirical macroeconomic research: VARs. Since follow a vector autoregressive-moving average
then, researchers in macroeconomics often (VARMA) model instead. From a different angle,
compute dynamic multipliers of interest, such Thomas F. Cooley and Mark Dwyer (1998) show
as impulse responses and forecast-error vari- that the dynamics of basic real business cycle
ance decompositions, by specifying a VAR models often follow VARMA representations that
even though the VAR per se is often of no are incompatible with VARs. These two observa-
particular interest. tions often explain the relatively long lag lengths
There is no specific reason, however, to expect required in practice to properly calculate impulse
that the data are generated by a VAR. In fact, even responses with a VAR. Additionally, new second-
assuming that a macroeconomy’s variables are and higher-order accurate solution techniques for
nonlinear dynamic stochastic general equilibrium
models (see, e.g., Jinill Kim et al., 2003) deliver
* Department of Economics, University of California,
Davis, One Shields Ave., Davis, CA 95616-8578 (e-mail: equilibrium conditions that are polynomial (rather
[email protected], URL: www.econ.ucdavis.edu/faculty/ than linear) difference equations. VARs may in-
jorda). This version of the paper has benefited from the deed be a significantly misspecified representation
suggestions of three anonymous referees and the editor. I of the data generating process (DGP).
thank Paolo Angelini, Colin Cameron, John Cochrane, Tim-
othy Cogley, James Hamilton, Peter Hansen, Kevin Hoover,
Impulse responses (and variance decomposi-
Monika Piazzesi, Simon Potter, Peter Robinson, Shinichi tions) are important statistics in their own right:
Sakata, Aaron Smith, Daniel Thornton, and seminar partic- they provide the empirical regularities that sub-
ipants at the Federal Reserve Bank of Kansas City, Indiana stantiate theoretical models of the economy and
University, University of California, Davis, and University are therefore a natural empirical objective. This
of California, Santa Cruz, for many useful suggestions.
Massimiliano de Santis provided outstanding research as- paper introduces methods for computing im-
sistance. I am thankful to the Institute of Governmental pulse responses for a vector time series based on
Affairs at U.C. Davis for financial support. local projections (a term defined precisely in the

161
162 THE AMERICAN ECONOMIC REVIEW MARCH 2005

next section) that do not require specification iterated forecasts for autoregressive models
and estimation of the unknown true multivariate whose lag length is too short—a typical sce-
dynamic system itself. nario when a VAR is used to approximate a
The advantages of local projections are nu- VARMA model, for example. Bhansali (2002)
merous: they can be estimated by simple least provides a nice review on this recent literature.
squares; they provide appropriate inference (in- Direct forecasting seeks an optimal multi-
dividual or joint) that does not require asymp- step forecast whereas the local projections pro-
totic delta-method approximations or numerical posed here seek a consistent estimate of the
techniques for its calculation; they are robust to corresponding impulse response coefficients.
misspecification of the DGP; and they easily Obviously, these objectives are not disjointed in
accommodate experimentation with highly non- much the same way that they are not when
linear specifications that are often impractical or estimating a VAR.
infeasible in a multivariate context. Since local The paper contains ample Monte Carlo evi-
projections can be estimated by univariate equa- dence illustrating the basic consistency and ef-
tion methods, they can be easily calculated with ficiency properties of local projections under
available standard regression packages and thus ideal conditions and under several forms of
become a natural alternative to estimating im- linear and nonlinear misspecifications, all rela-
pulse responses from VARs. tive to fixed parameter VARs and the more
The key insight is that estimation of a model recent time-varying Bayesian VARs used in
based on the sample, such as a VAR, represents Timothy W. Cogley and Thomas J. Sargent
a linear global approximation to the DGP ideal (2001). As an illustration, I estimate impulse
and is optimally designed for one-period ahead responses for a simple new-Keynesian model
forecasting. Even when the model is misspeci- (see Jordi Galı́, 1992; Jeffrey C. Fuhrer and
fied, it may still produce reasonable one-period George R. Moore, 1995a, 1995b; and references
ahead forecasts (see James H. Stock and Mark in John B. Taylor, 1999) based on cubic poly-
W. Watson, 1999). An impulse response, how- nomial projections with threshold effects. The
ever, is a function of forecasts at increasingly results are supportive of the view that the Fed-
distant horizons, and therefore misspecification eral Reserve faced a changing economic envi-
errors are compounded with the forecast hori- ronment from the 1970s to mid-1980s (a view
zon. This paper suggests that it is preferable to supported by, among others, Cogley and Sar-
use a collection of projections local to each gent, 2001) rather than attributing the inflation-
forecast horizon instead, thus matching design unemployment outcomes of the time to bad
and evaluation. policy, as J. Bradford DeLong (1997) and
Local projections are based on sequential re- Christina D. Romer and David H. Romer (2002)
gressions of the endogenous variable shifted have suggested.
several steps ahead and therefore have many
points of commonality with direct multi-step I. Estimation and Inference
forecasting. The ideas behind direct forecasting
(sometimes also called adaptive forecasting or A. Estimation
dynamic estimation) go back to at least David
R. Cox (1961). Andrew A. Weiss (1991) estab- Impulse responses are almost universally es-
lishes consistency and asymptotic normality of timated from the Wold decomposition of a lin-
the direct forecasts under general conditions. ear multivariate Markov model such as a VAR.
The accuracy of direct forecasting has been However, this two-step procedure consisting of
evaluated in several papers. Ruey S. Tsay first estimating the model and then inverting its
(1993) and Jin-Lung Lin and Tsay (1996) show estimates to find the impulse responses is justi-
that direct forecasting performs very well even fied only if the model coincides with the DGP.
relative to models where cointegrating restric- Furthermore, deriving correct impulse re-
tions are properly incorporated. Rajendra J. sponses from cointegrated VARs can be ex-
Bhansali (1996, 1997) and Ching-Kang Ing tremely complicated (see Bruce E. Hansen,
(2003) show that direct forecasts outperform 2003). Instead, impulse responses can be de-
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 163

fined without reference to the unknown DGP, dictions. These can be calculated by recursively
even when its Wold decomposition does not iterating on an estimated model optimized to
exist (see Gary Koop et al., 1996; Simon M. characterize the dependence structure of succes-
Potter, 2000). Specifically, an impulse response sive observations, of which a VAR is an exam-
can be defined as the difference between two ple. While this approach is optimal if indeed the
forecasts (see James D. Hamilton, 1994; Koop postulated model correctly represents the DGP,
et al., 1996): better multi-step predictions can often be found
with direct forecasting models that are reesti-
(1) IR共t, s, d i 兲 ⫽ E共yt ⫹ s兩vt ⫽ di ; Xt 兲 mated for each forecast horizon. Therefore, con-
sider projecting yt ⫹ s onto the linear space
⫺ E共yt ⫹ s兩vt ⫽ 0ᠪ ; Xt 兲 s ⫽ 0, 1, 2, ... generated by (yt ⫺ 1, yt ⫺ 2, ... , yt ⫺ p)⬘, specifi-
cally

where the operator E(.兩.) denotes the best, mean (2) y t ⫹ s ⫽ ␣s ⫹ B1s ⫹ 1yt ⫺ 1 ⫹ B2s ⫹ 1yt ⫺ 2
squared error predictor; yt is an n ⫻ 1 random
vector; Xt ⬅ (yt ⫺ 1,yt ⫺ 2, ...)⬘; 0ᠪ is of dimension
⫹ ... ⫹ Bps⫹ 1yt ⫺ p ⫹ uts⫹ s s ⫽ 0, 1, 2, ... , h
n ⫻ 1; vt is the n ⫻ 1 vector of reduced-form
disturbances; and D is an n ⫻ n matrix, whose
columns di contain the relevant experimental where ␣s is an n ⫻ 1 vector of constants, and
shocks. the Bs⫹1
i are matrices of coefficients for each
Time provides a natural arrangement of the lag i and horizon s ⫹ 1 (this timing convention
dynamic causal linkages among the variables in will become clear momentarily). I denote the
yt, but does not identify its contemporaneous collection of h regressions in (2) as local pro-
causal relations. The VAR literature has often jections, a term aptly evocative of nonparamet-
relied on assuming a Wold-causal order for ric considerations.
the elements of yt to organize the triangular According to definition (1), the impulse re-
factorization of the reduced-form, residual vari- sponses from the local-linear projections in (2)
ance-covariance matrix, ⍀ ⫽ PP⬘. Such an are
identification mechanism, for example, is equiva-
lent to defining the experimental matrix as D ⫽ ˆ
P⫺1 so that its ith column, di, then represents the (3) IR 共t, s, d i 兲 ⫽ B̂1s di s ⫽ 0, 1, 2, ... , h
“structural shock” to the ith element in yt (in the
usual parlance of the VAR literature). with the obvious normalization B01 ⫽ I. An
Statistical-based structural identification of extensive literature (see Bhansali, 2002, and
contemporaneous causal links is so far elusive.1 references therein) on the direct, multi-step
Further, a one-time shock to a given variable in forecasts implied by (2) establishes their con-
the system may not be the only type of experi- sistency and asymptotic normality properties
ment of interest. For these reasons, and to en- (see Weiss, 1991). However, here we are inter-
compass broad designs, the remainder of the ested in establishing the consistency and distri-
analysis accommodates a generic choice of ex- butional properties of the estimates B̂s1—the
periment D without loss of generality. Identifi- impulse response coefficients. This is rather
cation is an important issue but not one that is straightforward: as the next section shows, the
explored in this paper. residuals uts⫹ s in (2) are a moving average of the
Expression (1) shows that the statistical ob- forecast errors from time t to t ⫹ s and therefore
jective in calculating impulse responses is to uncorrelated with the regressors, which are
obtain the best, mean-squared, multi-step pre- dated t ⫺ 1 to t ⫺ p.
A few final practical comments conclude this
section. First, the maximum lag p (which can be
determined by information criteria, for exam-
1
Exceptions are the work by Granger and Swanson ple) need not be common to each horizon s (to
(1997), and Demiralp and Hoover (2003), for example. see this consider a VMA(q) DGP, for example).
164 THE AMERICAN ECONOMIC REVIEW MARCH 2005

Second, the lag length and the dimension of the B. Relation to VARs and Inference
vector yt will impose degree of-freedom con-
straints on the maximum practical horizon h for A VAR specifies that the n ⫻ 1 vector yt
very small samples. Third, consistency does not depends linearly on Xt ⬅ (yt ⫺ 1, yt ⫺ 2, ... ,
require that the sequence of h system regres- yt ⫺ p)⬘, through the expression
sions in (2) be estimated jointly—the impulse
response for the jth variable in yt can be esti- (5) y t ⫽ ␮ ⫹ ⌸⬘Xt ⫹ vt
mated by a univariate regression of yjt onto
Xt ⬅ (yt ⫺ 1, yt ⫺ 2, ... , yt ⫺ p)⬘. where vt is an i.i.d. vector of disturbances and
Finally, local projections are also useful in ⌸⬘ ⬅ [⌸1 ⌸2 ... ⌸p]. The VAR(1) companion
computing the forecast-error variance decom- form to this VAR can be expressed by defining2
position. By definition, the error in forecasting

冤 冥
yt ⫹ s is given from expression (2) by yt ⫺ ␮
yt ⫺ 1 ⫺ ␮
(6) W t ⬅ ⯗
y t ⫹ s ⫺ E共yt ⫹ s兩Xt 兲 ⫽ uts⫹ s s ⫽ 0, 1, 2, ...
yt ⫺ p⫹1 ⫺ ␮

冤 冥
from which the non-normalized mean squared ⌸1 ⌸2 ... ⌸p ⫺ 1 ⌸p

冤冥
error (MSEu) is ... vt
I 0 0 0
... 0
F⬅ 0 I 0 0 ; ␯t ⬅
⯗
MSE u 共E共y t ⫹ s兩Xt 兲兲 ⫽ E共uts⫹ suts⬘⫹ s兲 ⯗ ⯗ ⯗ ⯗ ⯗ 0
0 0 ... I 0
s ⫽ 0, 1, 2, ... , h
and then realizing that according to (5) and (6),
The choice experiment D renormalizes MSEu
into (7) W t ⫽ FWt ⫺ 1 ⫹ ␯t

(4) MSE共E共y t ⫹ s兩Xt 兲兲 ⫽ D⫺1E共uts⫹ suts⬘⫹ s兲D⬘⫺1 from which s-step ahead forecasts can be easily
computed since
s ⫽ 0, 1, 2, ... , h
W t ⫹ s ⫽ ␯t ⫹ s ⫹ F␯t ⫹ s⫺1 ⫹ ... ⫹ Fs␯t

from which the traditional variance decomposi- ⫹ F s ⫹ 1Wt ⫺ 1

tions can be calculated by directly plugging in
the sample-based equivalents from the projec- and therefore
tions in (2). For comparison, in traditional
VARs the non-normalized MSE is (8) y t ⫹ s ⫺ ␮ ⫽ vt ⫹ s ⫹ F11vt ⫹ s⫺1 ⫹ ... ⫹ F1s vt

MSEu 共E共yt ⫹ s兩Xt 兲兲 ⫽ E共u0t u0t ⬘兲 ⫹ ⌿1 E共u0t u0t ⬘兲⌿⬘1 ⫹ F1s ⫹ 1共yt ⫺ 1 ⫺ ␮兲 ⫹ ... ⫹ Fps⫹ 1共yt ⫺ p ⫺ ␮兲
where Fsi is the ith upper, n ⫻ n block of the
⫹ ... ⫹ ⌿s E共u0t u0t ⬘兲⌿⬘s s ⫽ 0, 1, 2, ... , h
matrix Fs (i.e., F raised to the power s).
Assuming Wt is covariance-stationary (or, in
where the ⌿i and E(u0t u0⬘
t ) are derived from other words, that the eigenvalues of F lie inside
the moving-average representation and the the unit circle), the infinite vector moving-
residual variance-covariance matrix of the
estimated VAR. The quality of the variance
decompositions will therefore depend on 2
For a more detailed derivation of some of the expres-
how well the ⌿i are approximated by the sions that follow, the reader should consult Hamilton
VAR. (1994), Chapter 10.
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 165

average representation of the original VAR in ture of the residuals implied by the VAR as-
expression (5) is sumption and estimate the following system
jointly:
(9) y t ⫽ ␥ ⫹ vt ⫹ F11vt ⫺ 1 ⫹ F12vt ⫺ 2
(13) Y t ⫽ Xt⌿ ⫹ Vt⌽
⫹ ... ⫹ F1s vt ⫺ s ⫹ ...
where (ignoring the constant terms) the param-
and the impulse response function is given by eter matrices are constrained as follows

冤 冥
IR共t, s, d i 兲 ⫽ F1s di F11 F12 ... F1h
F21 F22 ... F2h
Expression (8) establishes the relationship be- ⌿⫽
⯗ ⯗ ... ⯗
tween the impulse responses calculated by local ... Fph
projection and with a VAR. Specifically, compar- Fp1 Fp2
ing expression (2), repeated here for convenience,

冤 冥
In F11 ... F1h
(10) yt ⫹ s ⫽ ␣ ⫹ B
s s⫹1
y t⫺1 ⫹B s⫹1
yt⫺2 0 In ... F1h ⫺ 1
1 2
⌽⫽
⯗ ⯗ ... ⯗
⫹ ... ⫹ Bps⫹ 1yt ⫺ p ⫹ uts⫹ s 0 0 ... In

s ⫽ 0, 1, 2, ... , h Defining E(vtv⬘t ) ⫽ ⍀v, then E(VtV⬘t) ⫽ ⌽(Ih R

⍀v) ⌽⬘ ⬅ ⌺.
with expression (8) rearranged, Maximum likelihood estimation of this sys-
tem can then be accomplished by standard GLS
(11) y t ⫹ s ⫽ 共I ⫺ F1s ⫺ ... ⫺ Fps兲␮ ⫹ F1s ⫹ 1yt ⫺ 1 formulas according to

⫹ ... ⫹ Fps⫹ 1yt ⫺ p (14) vec共⌿̂兲 ⫽ 关共I 丢 X兲⬘⌺⫺1共I 丢 X兲兴⫺1

⫹ 共vt ⫹ s ⫹ F11vt ⫹ s⫺1 ⫹ ... ⫹ F1s vt 兲 ⫻ 共I 丢 X兲⬘⌺⫺1vec共Y兲

it is obvious that, The usual impulse responses would then be

given by rows 1 through n and columns 1
␣ s ⫽ 共I ⫺ F1s ⫺ ... ⫺ Fps兲␮ through (nh) of ⌿̂ and standard errors are pro-
vided directly from the regression output. Fur-
B 1s ⫹ 1 ⫽ F1s ⫹ 1 ther simplification is available due to the special
structure of the variance-covariance matrix ⌺,
(12) u ts⫹ s ⫽ 共vt ⫹ s ⫹ F11vt ⫹ s⫺1 ⫹ ... ⫹ F1s vt 兲. which allows GLS estimation of the system
block by block.
In fact, ML estimation of (13) delivers as-
Hence, when the DGP is the VAR in (5), ex- ymptotically exact formulas for single and joint
pressions (10) and (11) establish the equiva- inference on the impulse response coefficients
lence between impulse responses calculated by of the implied VAR, rather than the usual point-
local projections and with this VAR. Expression wise, analytic, delta-method approximations
(12) shows that the error terms of the local (see Hamilton, 1994, Ch. 11), or simulation-
projection, uts⫹ s, are a moving average of the based estimates based on Monte Carlo or boot-
forecast errors from time t to t ⫹ s, which strap replications.3 In general the true DGP is
knowledge can be used to improve efficiency.
Specifically, define Yt ⬅ (yt ⫹ 1, ... , yt ⫹ h),
Vt ⬅ (vt⫹1, ... , vt⫹h), and Xt ⬅ (yt⫺1, yt⫺2, ... , 3
Sims and Tao Zha (1999) provide methods for joint
yt ⫺ p), so that we stack the h local projections in inference in impulse responses but they involve complicated
expression (10) and take advantage of the struc- and rather computationally intensive Bayesian methods.
166 THE AMERICAN ECONOMIC REVIEW MARCH 2005

unknown, as is the specific structure of ⌽; The major selling point is that the error series
hence the GLS restrictions described above are v̂t is “fundamental” in some sense. The argu-
unavailable. This poses no difficulty, however, ment is that because forecast errors are
as the error terms uts⫹ s in expression (13) will constructed from market-based (rather than
follow some form of moving-average structure econometric) expectations, all available infor-
whose order is a function of the horizon s. Thus, mation is appropriately incorporated and, in ad-
impulse responses can be calculated by univar- dition, one can dispense with the thorny issue of
iate regression methods with a heteroskedastic- identification. These two features make these
ity and autocorrelation (HAC) robust estimator, methods attractive.
with little loss of efficiency. In principle, the There are some trade-offs to be considered,
efficiency of these estimators can be improved however. In general, it is perilous to disassoci-
upon by recursively including the residuals of ate the series of “shocks” from the model that
the stage s ⫺ 1 local projection as regressors in generated them, especially in a multivariate
the stage s local projection–an improvement context. The Wold decomposition theorem (see
whose details are reserved for another paper. Peter J. Brockwell and Richard A. Davis, 1991)
In practice the DGP is unknown and it is pref- ensures that any covariance-stationary process
erable to make as few assumptions as possible on can be expressed as an infinite moving average
its specification. Thus valid inference for local of forecast errors that are optimal in the mean-
projection impulse responses can be obtained with square sense. It does not guarantee, however,
HAC robust standard errors. For example, let ⌺̂L that these “shocks” are structural in the sense of
be the estimated HAC, variance-covariance ma- representing the residual series that describes
trix of the coefficients B̂s1 in expression (2); then a the DGP.
95-percent confidence interval for each element of Additionally, market-based expectations are
the impulse response at time s can be constructed available for a limited number of variables.
approximately as 1.96⫾ (d⬘i ⌺̂Ldi). Monte Carlo Econometrically, except for Thapar (2002),
experiments in Section III suggest that even when the second stage regression includes moving-
the true underlying model is a VAR, unrestricted average terms involving information dated t ⫺
local projections experience small efficiency 1, t ⫺ 2, ... which is problematic for consis-
losses. tency. (To see this, substitute the Wold decom-
position of yt into the second stage regressions.)
C. Comparison with Recent Impulse Response Finally, it is difficult to produce correct infer-
Estimators ence as the second stage uses generated regres-
sors, thus requiring bootstrap methods.
Pao-Li Chang and Shinichi Sakata (2002), Impulse responses characterize the partial de-
John H. Cochrane and Monika Piazzesi (2002), rivatives between different elements in yt over
and Aditi Thapar (2002) recently introduced time in the multidimensional process that re-
impulse response estimators that proceed in two lates yt to its past. Thus, while small variations
stages: in the first stage a forecast-error series, in the specification of this multidimensional
v̂t, is created, which is then used in a second process may do little to alter the “slopes” that
stage regression involving the original data yt.4 measure such trade-offs, they may well generate
Chang and Sakata (2002) calculate v̂t with an residual series that are relatively uncorrelated
autoregression, Cochrane and Piazzesi (2002) with each other. A similar point was raised by
with forecast errors implied by financial prices, Sims (1998) in response to a critique of VARs
and Thapar (2002) with errors in surveys of by Glenn D. Rudebusch (1998).
forecasts. Hence, all of these methods can be This argument can be underscored by an ad-
seen as a truncated version of Robert J. Barro’s ditional observation, that while it is perfectly
(1977, 1978) well-known regressions. coherent to think of impulse responses in the
context of a nonlinear, non-Gaussian model for
yt (such as when the data are transition data),
4
The ensuing discussion is in the univariate context, there may not always be a natural series of
hence the lowercase notation. “shocks” that can be manufactured for such a
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 167

model. On the other hand, it is not conceptually Under mild assumptions, a nonlinear time
difficult to see that one could obtain the impulse series process yt can be expressed as a generic
responses by computing the sequence of first- function of past values of a white noise process
order marginal effects in models that seek to vt in the form
explain yt⫹s as a function of information dated
t ⫺ 1 and beyond. y t ⫽ ⌽共vt , vt ⫺ 1 , vt ⫺ 2 , ...兲.
II. Flexible Local Projections
Assuming ⌽(.) is sufficiently well behaved, so
Linear models, such as VARs, bestow four that it can be approximated by a Taylor series
restrictive properties on their implied impulse expansion around some fixed point, say 0 ⫽
responses:5 (1) symmetry, where responses to (0, 0, 0, ...), then the closest equivalent to the
positive and negative shocks are mirror images Wold representation in nonlinear time series is
of each other; (2) shape invariance, where re- the Volterra series expansion (see Maurice B.
sponses to shocks of different magnitudes are Priestley, 1988),
scaled versions of one another; (3) history in-
dependence, where the shape of the responses is
冘 ⌽v 冘 冘⌽v
⬁ ⬁ ⬁
independent of the local conditional history;
and (4) multidimensionality, where responses (15) y t ⫽ i t⫺i ⫹ v
ij t ⫺ i t ⫺ j
i⫽0 i⫽0 j⫽0
are high-dimensional nonlinear functions of pa-
rameter estimates which complicate the calcu-
冘 冘 冘⌽ v
⬁ ⬁ ⬁
lation of standard errors and quickly compound
misspecification errors. For example, there is no ⫹ v v
ijk t ⫺ i t ⫺ j t ⫺ k ⫹ ...
i⫽0 j⫽0 k⫽0
reason to expect that the output losses due to
higher interest rates will be equivalent to the
output gains when interest rates are lowered; which is a polynomial extension of the Wold
that the output losses will simply double when decomposition in expression (9) with the con-
interest rates double as well; or that the same in- stant omitted for simplicity. Similarly, consider
crease in interest rates will have the same effect extending the local projections in expression (2)
on output whether we are coming out of a with polynomial terms on yt ⫺ 1:6
recession or just plunging into one.
Although local-linear projection methods dis- (16)
pense with the fourth of these problems, they
are indeed linear and thus constrained by prop- y t ⫹ s ⫽ ␣s ⫹ B1s ⫹ 1yt ⫺ 1 ⫹ Q1s ⫹ 1yt2⫺ 1 ⫹ C1s ⫹ 1yt3⫺ 1
erties (1)–(3). In a traditional multivariate,
model-based setting, investigation of nonlin- ⫹ B 2s ⫹ 1yt ⫺ 2 ⫹ ... ⫹ Bps⫹ 1yt ⫺ p ⫹ uts⫹ s
earities is limited by at least three consider-
ations: (1) the ability to estimate jointly a s ⫽ 0, 1, 2, ... , h
nonlinear system of equations with its inherent
computational difficulties; (2) the complexity in
generating multiple-step ahead forecasts from a where I do not allow for cross-product terms so
multivariate nonlinear model (which, at a min- that yt2⫺ 1 ⫽ ( y1,t
2 2 2
⫺ 1, y2,t ⫺ 1, ... , yn,t ⫺ 1)⬘, as a
imum, requires simulation methods since there matter of choice and parsimony rather than as a
are no closed forms available); and (3) the com-
plication in computing appropriate standard er-
rors for multiple-step ahead forecasts, and thus
6
the impulse responses. Hence, it is natural to Since the impulse response coefficients involve the
explore alternatives based on local projections. terms yt ⫺ 1 only, it is more parsimonious to restrict nonlin-
earities to these terms alone. In practice, if degrees of
freedom are not a consideration, they can be extended to the
remaining regressors, although the gain of doing so is prob-
5
For a detailed discussion, see Koop et al. (1996). ably small.
168 THE AMERICAN ECONOMIC REVIEW MARCH 2005

requirement. It is readily apparent that the im- yt ⫹ s ⫽ ms共yt ⫺ 1 ; Xt ⫺ 1 兲 ⫹ uts⫹ s s ⫽ 0, 1, 2, ... , h

pulse response at time s now becomes

(17) where ms(.) may include any parametric, semi-

parametric, and non-parametric approximation,
IR共t, s, d i 兲 ⫽ 兵B̂1s 共yt ⫺ 1 ⫹ di 兲 ⫹ Q̂1s 共yt ⫺ 1 ⫹ di 兲2 and for which there is a rather extensive list of
possible specifications to choose from. Monte
⫹ Ĉ1s 共yt ⫺ 1 ⫹ di 兲3} ⫺ 兵B̂1s yt ⫺ 1 ⫹ Q̂1s 共yt ⫺ 1 兲2 Carlo experiments in Section III show some of
the benefits of the local-cubic projection exam-
⫹ Ĉ1s 共yt ⫺ 1 兲3} ple just discussed, while the application in Sec-
tion IV shows how to compute impulse
⫽ 兵B̂1s di ⫹ Q̂1s 共2yt ⫺ 1di ⫹ d2i 兲 responses based on polynomial projections with
threshold effects.
⫹ Ĉ1s 共3yt2⫺ 1di ⫹ 3yt ⫺ 1d2i ⫹ d3i 兲}

s ⫽ 0, 1, 2, ... , h III. Monte Carlo Evidence

and with the obvious normalizations, B01 ⫽ I, This section contains two main simulations
Q01 ⫽ 0n, and C01 ⫽ 0n. These nonlinear esti- that evaluate the performance of local projec-
mates can be easily calculated by least squares, tions for impulse response estimation and infer-
equation by equation, with any conventional ence. The first experiment is based on a
econometric software. When some of the terms standard monetary VAR that appears in Law-
Qsi and Csi are non-zero, the impulse response rence J. Christiano et al. (1996) and Charles L.
function will vary according to the sign and Evans and David A. Marshall (1998), among
with the size of the experimental shock defined many other papers. The experiment illustrates
by di, thus dispensing with the symmetry and that local projections deliver impulse responses
shape invariance restrictions. In addition, the that are robust to lag length misspecification,
impulse response depends on the local history consistent, and only mildly inefficient relative
yt ⫺ 1 at which it is evaluated. In particular, im- to the responses from the true DGP. The second
pulse responses comparable to local-linear or experiment simulates a SVAR-GARCH (see
VAR-based impulse responses can be achieved Jordà and Kevin D. Salyer, 2003) to show that
by evaluation at the sample mean, i.e., yt ⫺ 1 ⫽ flexible local projections do a reasonable job at
yt ⫺ 1. Different responses will be obtained if a approximating the inherent nonlinearities of this
different experimental value for yt ⫺ 1 is chosen model, and compares its performance relative to
and one can consider a 3-D plot of the impulse a Bayesian VAR with time-varying parameters
response for a range of values for yt ⫺ 1.7 Finally, and volatilities—a natural flexible alternative to
the 95-percent confidence interval for the cubic conventional VARs.
approximation in expression (16) can be easily
calculated. Define the scaling ␭i ⬅ (di, 2yt⫺1di ⫹
d2i , 3yt2⫺ 1di ⫹ 3yt ⫺ 1d2i ⫹ d3i )⬘ and denote ⌺̂C A. Consistency and Efficiency
the HAC, variance-covariance matrix of the coef-
ficients B̂s1,Q̂s1, and Ĉs1 in (16). Then, a 95-percent This Monte Carlo simulation is based on
confidence interval for the impulse response at monthly data from January 1960 to February
time s is approximately 1.96 ⫾ (␭⬘i ⌺̂C␭i) . 2001 (494 observations). First, I estimate a
Natural extensions of this principle would VAR of order 12 on the following variables:
consist in formulating a flexible specification EM, log of non-agricultural payroll employ-
for the terms yt ⫺ 1 in expression (2), that is, ment; P, log of personal consumption expendi-
tures deflator (1996 ⫽ 100); PCOM, annual
7
Potter (2000) contains a detailed and more formal dis-
growth rate of the index of sensitive materials
cussion of alternative ways of defining and reporting non- prices issued by the Conference Board; FF,
linear impulse responses in general. federal funds rate; NBRX, ratio of nonborrowed
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 169

reserves plus extended credit to total reserves; tion that the VAR(2) is only mildly misspeci-
and ⌬M2, annual growth rate of M2 stock. I fied. However, both the local-linear and -cubic
then save the coefficient estimates from this projections are much more accurate at capturing
VAR and simulate 500 series of 494 observa- detailed patterns of the true impulse response
tions using multivariate normal residuals and over time, even at medium and long horizons.
the variance-covariance matrix from the esti- In one case, the departure from the true im-
mation stage, and use the first 12 observations pulse response was economically meaningful:
from the data to initialize all 500 runs. Infor- the response of the variable P. The response
mation criteria based on the original data sug- based on the VAR(2) is statistically different
gest the lag-length to be 12 when using from the true response for the first 17 periods
Hirotugu Akaike’s AIC and Clifford Hurvich and suggests that prices increase in response to
and Chih-Ling Tsai’s8 AICc, or two when using an increase in the federal funds rate over 23 out
Gideon Schwarz’s SIC. These choices are very of the 24 periods displayed. Many researchers
consistent across the 500 simulated runs.9 have previously encountered this type of coun-
The first experiment compares the impulse terintuitive result and dubbed it the “price puz-
responses that would result from fitting a VAR zle.” Sims (1992) suggested this behavior is
of order two (as SIC would suggest) with local- probably related to unresolved endogeneity is-
linear and -cubic projections of order two as sues and proposes including a materials price
well. Although a reduction from 12 to 2 lags index, as it is done here with PCOM. In con-
may appear severe, this is a very mild form trast, the local-linear projection is virtually
misspecification in practice. The results are dis- within the true two standard error bands
played graphically in Figure 1 rather than re- throughout the 24 periods depicted, and is
porting tables of root mean-squared errors, strictly negative for the last 7 periods.
which are less illuminating. Each panel in Fig- The second experiment shows that local pro-
ure 1 displays the impulse response of a variable jection methods are consistent under true spec-
in the VAR due to a shock in the variable ification by calculating impulse responses with
FF,10 calculated as follows: the thick-solid up to 12 lags. The results are reported in Figure
line is the true VAR(12) impulse response 2, also for a shock to FF only. Thus, the thick
with two standard-error bands displayed in line is the true impulse response, along with two
thick-dashed lines (these are based on the standard error bands displayed in thick-dashed
Monte Carlo simulations of the true model). lines. The responses based on local linear pro-
The responses based on a VAR(2) are dis- jections are displayed with the dashed line and
played by the line with squares; the responses the responses based on local cubic projections
from the local linear approximation are dis- are displayed by the line with circles. Generally,
played by the dashed line; and the responses the responses by either approximation literally
from the cubic local approximation are dis- lie on top of the true response11 with occasional
played by the line with circles. minor differences that disappeared with slightly
Several results deserve comment. The VAR(2) bigger samples, not reported here.
responses often fall within the two standard- The final set of experiments evaluates the
error bands of the true response and have the standard error estimates of the impulse response
same general shape. This supports the observa- coefficients (which are commonly used to dis-
play error bands around impulse responses). In
8
order to stack the odds against local projection
Hurvich and Tsai (1993) provide a correction to AIC methods and because in practice we never
specifically designed for VARs and with superior properties
to either AIC or SIC. know the true multivariate DGP describing the
9
Although the true DGP contains 12 lags, the coeffi- data, I consider standard errors calculated from
cients used in the Monte Carlo are based on the estimated
VAR and it is plausible that many of these coefficients are
not significantly different from zero in practice.
10 11
Responses to shocks in all the variables are available This is also true for the responses to all the remaining
upon request and are not reported here in the interest of shocks that are not reported here but are available upon
space. request.
170 THE AMERICAN ECONOMIC REVIEW MARCH 2005

FIGURE 1. IMPULSE RESPONSES TO A SHOCK IN FF. LAG LENGTH: 2

Notes: Evans and Marshall (1998) VAR(12) Monte Carlo Experiment. The thick line is the
true impulse response based on a VAR(12). The thick dashed lines are Monte Carlo two
standard error bands. Three additional impulse responses are compared, based on estimates
involving two lags only: (1) the response calculated by fitting a VAR(2) instead, depicted by
the line with squares; (2) the response calculated with a local-linear projection, depicted by
the dashed line; and (3) the response calculated with a local-cubic projection, depicted by the
line with circles. 500 replications.

univariate projections, equation by equation. linear and -cubic projections with 12 lags as
Specifically, I generated 500 runs of the original well. Then I computed Monte Carlo standard
series and then fitted a VAR(12) and local- errors for the VAR(12) to give a measure of the
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 171

FIGURE 2. IMPULSE RESPONSES TO A SHOCK IN FF. LAG LENGTH: 12

Notes: Evans and Marshall (1998) VAR(12) Monte Carlo Experiment. The thick line is the
true impulse response based on a VAR(12). The thick dashed lines are Monte Carlo two
standard error bands. Two additional impulse responses are compared: (1) the response
calculated with a local-linear projection with 12 lags, depicted by the dashed line; and (2) the
response calculated with a local-cubic projection, depicted by line with circles. 500 replications.

true standard errors, and calculated Newey- projections. Table 1 reports these results for
West12 corrected standard errors for the local each variable in response to a shock in FF as well.
In Section I, I argued that local projection
12
The Newey-West lag correction is selected to in-
estimates of impulse responses are less efficient
crease with s, the horizon of the impulse response being than VAR-based estimates when the VAR is
considered. correctly specified and it is the true model.
172 THE AMERICAN ECONOMIC REVIEW MARCH 2005

TABLE 1—STANDARD ERRORS FOR IMPULSE RESPONSES

EM P PCOM FF NBRX ⌬M2

Newey- Newey- Newey- Newey- Newey- Newey- Newey- Newey- Newey- Newey- Newey- Newey-
True- West West True- West West True- West West True- West West True- West West True- West West
s MC (Linear) (Cubic) MC (Linear) (Cubic) MC (Linear) (Cubic) MC (Linear) (Cubic) MC (Linear) (Cubic) MC (Linear) (Cubic)

1 0.000 0.007 0.008 0.000 0.007 0.007 0.000 0.089 0.096 0.000 0.022 0.024 0.0005 0.0005 0.0005 0.014 0.012 0.014
2 0.008 0.011 0.012 0.007 0.010 0.011 0.094 0.146 0.161 0.027 0.036 0.041 0.0007 0.0006 0.0007 0.025 0.023 0.026
3 0.013 0.015 0.016 0.012 0.014 0.015 0.155 0.191 0.212 0.044 0.046 0.052 0.0008 0.0007 0.0008 0.035 0.032 0.035
4 0.018 0.019 0.021 0.015 0.017 0.018 0.202 0.224 0.250 0.054 0.053 0.060 0.0008 0.0008 0.0009 0.044 0.039 0.043
5 0.022 0.023 0.025 0.018 0.020 0.022 0.240 0.255 0.284 0.061 0.058 0.065 0.0009 0.0008 0.0009 0.050 0.045 0.050
6 0.027 0.026 0.030 0.021 0.023 0.025 0.267 0.279 0.311 0.064 0.062 0.069 0.0009 0.0008 0.0009 0.056 0.050 0.056
7 0.031 0.030 0.033 0.025 0.026 0.029 0.296 0.301 0.335 0.067 0.064 0.072 0.0009 0.0008 0.0009 0.061 0.056 0.062
8 0.035 0.033 0.037 0.028 0.029 0.032 0.325 0.322 0.357 0.072 0.066 0.074 0.0009 0.0008 0.0009 0.066 0.060 0.067
9 0.038 0.036 0.040 0.031 0.032 0.035 0.350 0.340 0.376 0.073 0.067 0.075 0.0009 0.0009 0.0010 0.070 0.064 0.072
10 0.041 0.039 0.043 0.035 0.035 0.039 0.361 0.356 0.392 0.074 0.069 0.077 0.0009 0.0009 0.0010 0.074 0.069 0.076
11 0.044 0.042 0.046 0.038 0.038 0.042 0.377 0.371 0.407 0.075 0.072 0.080 0.0009 0.0009 0.0010 0.078 0.073 0.081
12 0.046 0.044 0.048 0.042 0.042 0.045 0.390 0.380 0.416 0.077 0.075 0.083 0.0009 0.0009 0.0010 0.082 0.077 0.085
13 0.048 0.046 0.050 0.046 0.045 0.049 0.402 0.385 0.423 0.079 0.078 0.087 0.0009 0.0009 0.0010 0.084 0.080 0.088
14 0.050 0.048 0.053 0.049 0.048 0.052 0.402 0.389 0.427 0.079 0.080 0.089 0.0009 0.0009 0.0010 0.085 0.082 0.090
15 0.051 0.050 0.055 0.052 0.052 0.056 0.399 0.392 0.430 0.080 0.082 0.090 0.0008 0.0009 0.0010 0.084 0.084 0.092
16 0.053 0.052 0.057 0.055 0.055 0.059 0.393 0.394 0.434 0.080 0.083 0.091 0.0008 0.0009 0.0010 0.085 0.085 0.093
17 0.054 0.054 0.058 0.059 0.058 0.063 0.393 0.396 0.437 0.081 0.084 0.092 0.0008 0.0009 0.0010 0.085 0.086 0.094
18 0.055 0.055 0.060 0.062 0.062 0.066 0.386 0.399 0.441 0.081 0.084 0.093 0.0008 0.0009 0.0010 0.085 0.087 0.095
19 0.057 0.057 0.061 0.066 0.065 0.070 0.381 0.402 0.444 0.079 0.085 0.093 0.0007 0.0009 0.0010 0.084 0.088 0.096
20 0.059 0.058 0.062 0.070 0.068 0.073 0.380 0.405 0.448 0.079 0.086 0.093 0.0007 0.0009 0.0010 0.083 0.088 0.096
21 0.060 0.059 0.064 0.074 0.071 0.076 0.378 0.409 0.453 0.077 0.086 0.094 0.0007 0.0009 0.0010 0.082 0.088 0.096
22 0.061 0.061 0.065 0.078 0.075 0.080 0.377 0.415 0.462 0.077 0.087 0.094 0.0007 0.0009 0.0010 0.081 0.088 0.096
23 0.063 0.062 0.066 0.082 0.078 0.083 0.377 0.423 0.472 0.077 0.087 0.095 0.0006 0.0009 0.0010 0.080 0.088 0.096
24 0.064 0.063 0.068 0.086 0.081 0.086 0.371 0.431 0.484 0.077 0.087 0.095 0.0006 0.0009 0.0010 0.078 0.088 0.096

Notes: True-MC refers to the Monte Carlo (500 replications) standard errors for the impulse response coefficients due to a shock in FF in a VAR(12) with the variables
EM, P, PCOM, FF, NBRX, ⌬M2. Similarly, Newey-West (linear) refers to standard errors calculated from local-linear projections and their Newey-West corrected
standard errors, while Newey-West (cubic) refers to the local-cubic projections instead.

Table 1 confirms this statement but also shows jection methods with a traditional and a flexibly
that this loss of efficiency is not particularly big. parametrized VAR when the DGP is nonlinear.
The Newey-West corrected standard errors The specific DGP for this experiment is based
based on single equation estimates of the local on the SVAR-GARCH model in Jordà and
linear projections are virtually identical to the Salyer (2003), which is a multivariate version of
Monte Carlo standard errors from the VAR, a traditional GARCH-M model. Here, I exper-
particularly for the variables EM and P. The iment with the following specification:
biggest discrepancy is for the variable NBRX,

冋册 冋 册 冋 册
because the VAR Monte Carlo standard errors
actually decline as the horizon increases (spe- y 1t y1t ⫺ 1 冑h1t␧1t
cially after the fourteenth period). This anom- (18) y 2t ⫽ A y2t ⫺ 1 ⫹ Bh1t ⫹ ␧2t ,
aly, which is explained in Sims and Zha (1999), y 3t y3t ⫺ 1 ␧3t
is not a feature of the local projection standard
errors, which incorporate the additional uncer- ␧ t ⬃ N共0, I 3 兲
tainty existing in long-horizon forecasts. Alto-
gether, these results suggest that the efficiency
losses are rather minor, even for a system that h 1t ⫽ 0.5 ⫹ 0.3u 1,t ⫺ 1 ⫹ 0.5h 1,t ⫺ 1 ;
contains as many as 6 variables, 12 lags, and
horizons of 24 periods. u 1t ⫽ 冑h 1t ␧ 1t

冋 册 冋 册
B. Impulse Responses and Nonlinearities
0.5 ⫺0.25 0.25 ⫺1.75
The following Monte Carlo experiment com- A⫽ 0.75 0.25 0.25 ; B ⫽ ⫺1.5
pares impulse responses calculated by local pro- ⫺0.25 ⫺0.25 0.75 1.75
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 173

and a sample size of 300, replicated 500 times. Calculating impulse responses for each of
This DGP is advantageous for several reasons. these five selected dates requires an additional
First, the SVAR-GARCH nests a linear Monte Carlo simulation since the parameters of
VAR(1), and in fact impulse responses to the model are varying over time stochastically.
shocks to ␧2t or ␧3t, or to “small” shocks to ␧1t Hence, I generate 100 sequences of 1– 8-step
are equivalent in both models. Discrepancies ahead forecasts, conditional on the parameter
arise with larger shocks to ␧1t since there is a values at each of the five dates previously se-
revision of their conditional variance (due to the lected and the driving processes estimated from
GARCH term) that affects the conditional mean the data. The average over these 100 histories is
and makes the responses more nonlinear. Sec- used to produce the impulse responses.15
ond, since I also specify a time-varying param- Figure 3 displays the impulse responses from
eter/volatility VAR13 (TVPVAR) as Cogley and a shock to ␧1t of unit in size.16 The thick solid
Sargent (2001) as a flexible approximator to the lines in each graph represent the true impulse
DGP, it is useful that the nonlinearity be of a responses with and without GARCH effects
smooth nature (say, relative to a model with (i.e., B ⫽ 0ᠪ ), the less variable of the two rep-
structural breaks or switching-regimes). Notice resenting the latter case. Indistinguishable from
that the DGP will have time-varying volatility the impulse responses generated when the
with some effects on the conditional mean and GARCH effects are switched off, both the linear
one would expect that the TVPVAR is well VAR(1) and the linear projection responses are
suited to capture these features. displayed by thin dashed lines. Finally, the thick
As a foil to the cubic projection, the TVPVAR dashed line with crosses displays the cubic pro-
is estimated with Bayesian methods for each of jection responses, whereas the thick line with
the 500 Monte Carlo replications using the first squares displays the TVPVAR responses, aver-
100 observations to calibrate the prior, leaving aged over the five selected days. (It will become
the remaining 200 for inference. The estimator clear momentarily why the averaging.) Standard
is based on a Gibbs-sampler initialized with errors are omitted from the figure to improve
2,000 draws and allowed to run for an addi- clarity. Suffice it to say that conventional error
tional 5,000 iterations to ensure convergence. bands are very narrow and clearly separate im-
This produces a history of 200 observations for pulse responses from the DGP with and without
each estimated parameter in the model. To cal- GARCH effects, except at crossing points.
culate the impulse responses, I select the quin- Several results deserve comment. The
tiles of the distribution of the residual for the VAR(1), the linear projections, and the cubic
first equation (the one with GARCH effects) projections evaluated at the sample mean (not
which identify five dates from the last 200 ob- reported) precisely capture the shape of the im-
servations in the sample (the ones with time- pulse responses from the DGP without GARCH
varying parameters). This allows the TVPVAR effects, even though these are estimated from a
to tailor the impulse responses to different val- sample generated without this constraint. The
ues of the conditional variance and to better true impulse responses with GARCH effects are
capture any resulting nonlinearities.14 far more variable, and to capture this feature I
consider cubic projections evaluated at five
13
I thank Tim Cogley for all his advice on the numerous standard deviations away from the mean and
intricacies of this model and Massimiliano de Santis for his responses from the TVPVAR evaluated at the
invaluable assistance in estimating the model with his five dates selected previously. In the end, the
GAUSS code. For further details on the specification and
estimation of the model, the reader is referred to Cogley and
TVPVAR responses displayed no variability for
Sargent (2001, 2003). The GAUSS code for estimating the the first six to seven periods. After that, they fan
time-varying parameter VAR can be obtained by e-mailing
Massimiliano de Santis directly at: [email protected].
14 15
A provision in the code discards any Monte Carlo The complete Monte Carlo took nine days, two hours,
draws for which the stationarity condition for the distribu- and 17 minutes on a SUN Sunfire server with eight 900 Mhz
tion of the parameters is violated. If a draw is discarded, the processors and 16GB of RAM memory.
Monte Carlo runs for an additional draw. In the end, 5 to 10 16
Shocks to ␧2t and ␧3t would simply produce the usual
percent of the draws had to be replaced. linear VAR responses.
174 THE AMERICAN ECONOMIC REVIEW MARCH 2005

FIGURE 3. IMPULSE RESPONSES TO A SHOCK IN Y1 FROM A SVAR-GARCH

Notes: The thick solid lines describe the true impulse response in the SVAR-GARCH model
with and without GARCH effects (i.e., B ⫽ 03), the less variable referring to the latter. The
thin dashed lines are the responses from a VAR(1) and from local-linear projections. The thick
dashed line with crosses is the local-cubic projection whereas the thick dashed line with
squares is the impulse response from the Bayesian VAR.

out in different directions, much like the picture didate policy rules is often evaluated in the
of a forecast confidence interval. Hence, to context of a simple, new-Keynesian, closed-
simplify the figure, I report the average over economy model which, at a minimum, can be
the five dates. As Figure 3 clearly shows, the summarized by three fundamental expressions:
TVPVAR responses were unable to capture the an IS equation, a Phillips relation, and the can-
nonlinearities in the model, whereas the cubic didate policy rule itself. While models may dif-
projections provided a much closer fit to the true fer in their degree of micro-foundation and
impulse responses. Overall, local polynomial forward-looking behavior (see Taylor’s [1999]
projections seem to afford better control over edited volume for examples), they share the
smooth nonlinearities since they nest linearity need to reproduce the fundamental dynamic
and their complexity can be easily controlled. properties of actual economies with some de-
gree of accuracy.
IV. Application: Inflation-Output Trade-Offs Consequently, it is natural to investigate the
dynamic properties of inflation, the output gap,
Pioneering work by Bennett T. McCallum and interest rates to provide a benchmark for
(1983) and Taylor (1993) inspired a remarkable competing theoretical models. The specific def-
amount of research on the efficacy, optimality, initions of the variables I consider correspond to
credibility, and robustness of interest rate rules the definitions in Rudebusch and Lars E.O.
for monetary policy. The performance of can- Svensson (1999) and are relatively standard for
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 175

FIGURE 4. TIME SERIES PLOTS OF THE OUTPUT GAP, INFLATION, AND THE
FEDERAL FUNDS RATE
Notes: All variables in annual percentage rates. Shaded areas indicate NBER-dated reces-
sions. Output gap is defined as the percentage difference between real GDP and potential GDP
(Congressional Budget Office); inflation is defined as the percentage change in the GDP,
chain-weighted price index at annual rate; and the federal funds rate is the quarterly average
of daily rates, in annual percentage rate. The solid horizontal lines display the thresholds
detected by Hansen’s (2000) test for inflation and the federal funds rate.

this literature: yt is the percentage gap between and Moore (1995a and 1995b), use four lags for
real GDP and potential GDP (as measured by variables analyzed in the levels. This selection
the Congressional Budget Office); ␲t is quar- is confirmed by AICc, while AIC suggests six
terly inflation in the GDP, chain-weighted price lags and SIC suggests two lags. Therefore, Fig-
index in percent at annual rate;17 and it is the ure 5 displays the impulse responses based on a
quarterly average of the federal funds rate in VAR(4) and local-linear and -cubic projections,
percent at an annual rate. The sample of quar- all identified with a standard Cholesky decom-
terly data runs over the period 1955:I–2003:I position18 and the Wold-causal order yt,␲t,
and is displayed in Figure 4. and it.
A good starting point for the analysis is to The VAR(4) responses are depicted with a
calculate impulse responses with a VAR and thick line and the solid line with crosses, and the
local-linear and -cubic projections. The lag- two accompanying dashed lines depict the re-
length is determined by information criteria, sponses from local-linear projections and the
allowing for a maximum lag-length of eight. corresponding Newey-West corrected, two
Similar studies, such as Galı́ (1992) and Fuhrer standard-error bands. The solid line with circles

17 18
I thank an anonymous referee for spotting that I had This Cholesky ordering is consistent with the litera-
used the quantity index in the previous version of this paper. ture and facilitates replicability.
176 THE AMERICAN ECONOMIC REVIEW MARCH 2005

FIGURE 5. IMPULSE RESPONSES FOR THE NEW KEYNESIAN MODEL BASED ON A VAR, AND LINEAR AND CUBIC PROJECTIONS
Notes: The thick line is the response calculated from a VAR; the solid line with crosses is the response calculated by linear
projection; the two dashed lines are 95-percent confidence level error bands for the individual coefficients of the linear
projection response; and the solid line with circles is the response calculated by cubic projection evaluated at the sample mean.
All responses calculated with four lags.

is the response from local-cubic projections sponse is also almost twice as aggressive (at
evaluated at the sample mean. Each row repre- 0.75 percent versus 0.4 percent) 12 quarters
sents the responses of yt,␲t, and it to orthogo- after impact. The responses of the system to
nalized shocks, starting with yt,␲t, and then it, interest rate shocks are perhaps more interesting
all measured in percentages. Generally, there is from an economic point of view because they
broad correspondence among the responses cal- give us an idea of the relative output and infla-
culated by the different methods, with a few tion trade-offs in response to monetary policy.
exceptions. The VAR response of the output gap suggests a
The cubic projection responses show that in- loss of output of 0.25 percent in response to a
flation is considerably more persistent to its own 0.8-percent increase in the federal funds rate, 12
shocks than what is reflected by the responses quarters after impact. This loss is approximately
calculated by either linear method. Perhaps not half what linear and cubic projections predict (at
surprisingly, the associated interest rate re- around 0.5 percent). On the other hand, the
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 177

response of inflation to this increase in the Fed TABLE 2—HANSEN’S (2000) TEST FOR THE PRESENCE OF

funds rate is mostly positive but not signifi- THRESHOLD EFFECTS. THRESHOLD ESTIMATES AND
BOOTSTRAP p-VALUES
cantly different from zero. The linear projection
is similar for the first seven quarters but is Dependent variable
significantly negative thereafter, whereas the
Threshold Output gap Inflation Fed funds
cubic projections show a more positive initial variable (yt) (␲t) (it)
inflation response with a dramatic decline
around quarter seven as well. yt ⫺ 1 ⫺0.85 ⫺1.31 ⫺0.09
(0.74) (0.62) (0.24)
Based on this preliminary analysis, we inves- yt ⫺ 2 ⫺1.97 ⫺2.07 ⫺0.85
tigate for further nonlinearities in the impulse (0.73) (0.33) (0.33)
responses. It seems of considerable importance yt ⫺ 3 0.37 ⫺1.42 ⫺2.34
to determine whether the inflation-output gap (0.20) (0.28) (0.24)
yt ⫺ 4 ⫺1.20 ⫺1.25 ⫺2.09
trade-offs that the monetary authority faces vary (0.50) (0.18) (0.84)
with the business cycle, or during periods of
high inflation, or when interest rates are close to ␲t ⫺ 1 4.68 4.00 4.93
(0.03) (0.39) (0.10)
the zero bound, for example. Although the poly- ␲t ⫺ 2 4.66 4.54 4.24
nomial terms in local projections approximate (0.09) (0.02) (0.18)
smooth nonlinearities, they are less helpful in ␲t ⫺ 3 3.91 5.31 4.24
detecting the type of nonlinearity implicit in (0.30) (0.02) (0.00)
these examples. Therefore, I tested all linear ␲t ⫺ 4 2.82 3.25 3.98
(0.13) (0.59) (0.04)
projections19 for evidence of threshold effects
due to all four lags of yt, ␲t, and it using Hansen’s it ⫺ 1 5.94 6.52 7.88
(2000) test20. For example, a typical regression is, (0.53) (0.46) (0.01)
it ⫺ 2 6.02 5.94 5.56
(0.19) (0.92) (0.21)
(19) z t ⫽ ␳⬘LXt ⫺ 1 ⫹ ␧ L
t if wt ⫺ j ⱕ ␦ it ⫺ 3 6.27 8.16 5.82
(0.04) (0.95) (0.05)
z t ⫽ ␳⬘HXt ⫺ 1 ⫹ ␧Ht if wt ⫺ j ⬎ ␦ it ⫺ 4 5.64 5.28 5.09
(0.55) (0.37) (0.07)
where zt is respectively yt,␲t, and it and wt ⫺ j can Notes: The equation for each dependent variable contains
be any of yt ⫺ j, ␲t ⫺ j, and it ⫺ j, j 僆 {1, 2, 3, 4}. four lags of each of the dependent variables in the system.
Xt ⫺ 1 collects lags 1 through 4 of the variables The test is an LM-type test for the null hypothesis that there
yt,␲t, and it and ␳k, k ⫽ L, H collects the is no threshold effect. Each cell contains the estimate of the
optimal threshold value estimate and a bootstrap-based p-
coefficients and L stands for “low” and H stands value (in parenthesis) calculated with 1,000 draws and a
for “high.” The test is an F-type test that se- 20-percent trimming value of the sample to allow for suf-
quentially searches for the optimal threshold ␦ ficient degrees of freedom. The test corrects for left-over
and adjusts the corresponding distribution via heteroskedasticity. The results on this table were calculated
with the GAUSS code that Bruce Hansen makes available
1,000 bootstrap replications. on his Web site based on his 2000 Econometrica paper.
Table 2 summarizes the results of these tests Entries in bold and italic signify evidence of a threshold at
by reporting the estimated thresholds and p- the conventional 95-percent confidence level.
values (in parenthesis) for all possible combi-
nations of endogenous and threshold variables.
Several results deserve comment. First, there
are no “business-cycle” asymmetries associated the null of linearity is rejected across equations
with threshold effects in the output gap. Second, for several lags of both inflation and the federal
funds rate. Third, there is considerable corre-
spondence between the estimated threshold val-
19
I used the local linear projections for the test for ues for the lags of a given variable across
parsimony although the final analysis is based on cubic all equations. Fourth, the reported estimated
projections.
20
The GAUSS routines to perform the test are available
threshold values correspond to the value that
directly from Bruce Hansen’s Web site (www.ssc.wisc.edu/ maximizes the likelihood. Note that when the
⬃bhansen/progs/progs_threshold.html). null of linearity is rejected, however, it is often
178 THE AMERICAN ECONOMIC REVIEW MARCH 2005

rejected for an interval around this optimal several dimensions. Although the evidence is
value. not definitive, these results support the view
Despite the apparent complexity of these re- held by Cogley and Sargent (2001), among oth-
sults, the overall message that emerges is ers, that the adverse inflation-unemployment
straightforward: the estimated thresholds are di- outcomes of the 1970s were not the result of bad
viding the data into the turbulent period of the policy (as advocated by DeLong, 1997, and
1970s to mid-1980s (the “high-inflation” re- Romer and Romer, 2002) but rather the result of
gime) and the rest of the sample (the “low- a changing economic environment. Perhaps one
inflation” regime). Consequently, it is natural to argument that could undermine these results
consolidate these results by conducting two ex- would suggest that the Federal Reserve had
periments. In the first experiment I allow for a become less credible during the 1970s, although
threshold in the third lag of inflation at 4.75 it seems clear that it had not become any less
percent. In the second, the threshold is in the vigilant (this is especially evident in the re-
third lag of the federal funds rate at 6 percent sponses depicted in the right column and last
instead. Figure 4 displays these two thresholds row of Figure 6). The results in Figure 6 also
in reference to the raw data to illustrate that the suggest that the “prize puzzle” (the common
main difference is that the threshold in the fed- finding in the VAR literature that inflation ac-
eral funds rate extends the high-inflation regime tually increases in response to an increase in
up to the late 1980s. interest rates) does not characterize the current
Figure 6 compares the responses to a shock in economic environment. In fact, the current re-
the federal funds rate for these two experi- gime is characterized by rather effective re-
ments.21 In particular, the left column displays sponses of the output gap and inflation to an
the graphs corresponding to the inflation thresh- increase in interest rates, an observation with
old, while the right column displays the graphs important implications in the design of contem-
for the federal funds rate threshold. The solid porary optimal policy responses that are not
thick line and the dashed lines are the cubic unduly contractionary. Finally, notice that while
projections evaluated at the sample mean and we have estimated flexible impulse responses
the corresponding Newey-West– corrected, two that allow for threshold effects, the entire anal-
standard-error bands. The solid line with ysis was conducted by means of simple least
crosses corresponds to the low-inflation regime squares regressions—an ostensible simplifica-
responses, whereas the solid line with circles tion relative to any multivariate alternative
represents the high-inflation regime responses. based on a flexible nonlinear model.
All experiments are normalized to a 0.8-percent
shock in the federal funds rate to facilitate com-
parability. (This is a one standard-error shock V. Conclusion
which is the one most often reported in standard
econometric packages.) The first-order Taylor series expansion of a
Figure 6 clearly shows that the nature of the function at a given point gives a reasonable
inflation-output trade-offs varies quite substan- approximation to the function in a neighbor-
tially depending on regime but does not really hood of that point. The more nonlinear the
depend on which variable is used as a threshold. function, however, the more the quality of the
Generally, inflation and output are far more approximation deteriorates as we move farther
responsive to interest rates in the low-inflation away from the original evaluation point. Simi-
regime than in the high-inflation regime, even larly, a VAR linearly approximates the DGP to
though the federal funds rate responds some- produce optimal, one-period ahead forecasts,
what more aggressively in the latter. but impulse responses are functions of forecasts
This empirical application is illustrative in at ever-increasing horizons for which a VAR
may provide a poor approximation.
This paper shows that impulse responses can
21
The responses to the remaining shocks are omitted for be calculated by a sequence of projections of the
brevity but are available upon request. endogenous variables shifted forward in time
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 179

Inflation Threshold - 4.75% Fed Funds Threshold - 6%

Response of Y-Gap to Fed Funds Shock Response of Y-Gap to Fed Funds Shock
0.4 0.4

0.2 0.2

0.0 0.0

-0.2 -0.2

-0.4 -0.4

-0.6 -0.6

-0.8 -0.8

-1.0 -1.0
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12

Response of Inflation to Fed Funds Shock Response of Inflation to Fed Funds Shock
0.8 .8

.6
0.4
.4

0.0 .2

.0
-0.4
-.2

-.4
-0.8
-.6
-1.2 -.8
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12

Response of Fed Funds to Fed Funds Shock Response of Fed Funds to Fed Funds Shock
1.2 1.2

0.8 0.8

0.4 0.4

0.0 0.0

-0.4 -0.4

-0.8 -0.8
1 2 3 4 5 6 7 8 9 10 11 12 1 2 3 4 5 6 7 8 9 10 11 12

FIGURE 6. IMPULSE RESPONSES FROM NEW KEYNESIAN MODEL. CUBIC PROJECTIONS WITH
THRESHOLD EFFECTS IN INFLATION AT 4.75 PERCENT VERSUS THE FEDERAL FUNDS RATE
AT 6 PERCENT

Notes: The thick line is the response calculated by cubic projection at the sample mean; the
dashed lines are two standard error bands for the individual coefficients of the cubic response;
the solid line with crosses is the response by cubic projection below the threshold; and the
solid line with circles is the response by cubic projection above the threshold.

onto its lags. Hence, these projections are local Inference for impulse responses from VARs
to each forecast horizon and therefore more is difficult because impulse response coeffi-
robust to misspecification of the unknown DGP. cients are high-dimensional nonlinear functions
Local projections are therefore a natural and of estimated parameters. By contrast, local pro-
preferable alternative to VARs when the object jections directly estimate impulse response co-
of interest is to calculate impulse responses. efficients so that standard errors from traditional
180 THE AMERICAN ECONOMIC REVIEW MARCH 2005

HAC regression routines provide appropriate jections and optimal GMM-type weights to pro-
joint or point-wise inference. duce more efficient estimates and standard
The principles presented in this paper open a errors for a wide range of rational expectations
number of new avenues for research. The se- models.
quential nature of the local projections allow us
to take advantage of the stage s ⫺ 1 forecast REFERENCES
errors to improve inference in the stage s pro-
jections. Preliminary Monte Carlo evidence not Amato, Jeffery D. and Laubach, Thomas. “Esti-
reported here shows significant gains in using mation and Control of an Optimization-
this procedure, whose formal derivations are Based Model with Sticky Prices and Wages.”
left for a different paper. The same sequential Journal of Economic Dynamics and Control,
feature of local projections can be used to im- 2003, 27(7), pp. 1181–1215.
prove structural identification, since any con- Barro, Robert J. “Unanticipated Money Growth
temporaneous structure among the endogenous and Unemployment in the United States.”
variables must remain as we shift time forward American Economic Review, 1977, 67(2), pp.
through each local projection. 101–15.
Panel data applications in macroeconomics, Barro, Robert J. “Unanticipated Money, Output,
where dynamics dominate cross-sectional con- and the Price Level in the United States.”
siderations, are likely to become more preva- Journal of Political Economy, 1978, 86(4),
lent. However, while the high dimensionality of pp. 549 – 80.
VARs make impulse response estimation pro- Bhansali, Rajendra J. “Asymptotically Efficient
hibitive in this context, local projections offer a Autoregressive Model Selection for Multi-
natural and simple alternative for estimation step Prediction.” Annals of the Institute of
and inference of the dynamics of different treat- Statistical Mathematics, 1996, 48(3), pp.
ment effects. 577– 602.
Recent sophisticated solution methods have Bhansali, Rajendra J. “Direct Autoregressive
opened the doors to increasingly complex nonlin- Predictors for Multistep Prediction: Order
ear economic models. It is often impractical to Selection and Performance Relative to the
calculate impulse response functions from multi- Plug-In Predictors.” Statistica Sinica, 1997,
variate nonlinear models (for reasons explained in 7(2), pp. 425– 49.
Section II) or simply impossible for non-Gaussian Bhansali, Rajendra J. “Multi-Step Forecasting,”
data whose multivariate density is unknown, yet in Michael P. Clements and David F. Hendry,
impulse responses can still be calculated relatively eds., A companion to economic forecasting.
simply by local projection methods. Oxford: Blackwell Publishing, 2002, pp.
Finally, local projection methods can be used 206 –21.
to formalize the estimation of deep parameters Brockwell, Peter J. and Davis, Richard A. Time
in rational expectations models in the manner series: Theory and methods, 2nd ed. Berlin:
proposed in Julio J. Rotemberg and Michael Springer-Verlag, 1991.
Woodford (1997) and used in several papers Chang, Pao-Li and Sakata, Shinichi. “A Mis-
thereafter (Christiano et al., 2005, Jeffrey D. specification-Robust Impulse Response Esti-
Amato and Thomas Laubach, 2003, and so on). mator.” Unpublished Paper, 2002.
The technique consists of conjecturing a so- Christiano, Lawrence J.; Eichenbaum, Martin
lution path represented by an infinite moving and Evans, Charles L. “Identification and the
average (MA) and then matching the deep pa- Effects of Monetary Policy Shocks,” in
rameters of the model to the MA coefficients Mario I. Blejer, Zvi Eckstein, Zvi Hercowitz,
with the method of undetermined coefficients. and Leonardo Leiderman, eds., Financial
A minimum distance estimator between data- factors in economic stabilization and
based impulse responses and the theoretically Growth. Cambridge: Cambridge University
constrained MA coefficients thus produces es- Press, 1996, pp. 36 –74.
timates of these deep parameters. In work in Christiano, Lawrence J.; Eichenbaum, Martin
progress, Sharon Kozicki and I use local pro- and Evans, Charles L. “Nominal Rigidities
VOL. 95 NO. 1 JORDÀ: IMPULSE RESPONSES BY LOCAL PROJECTIONS 181

and the Dynamic Effects of a Shock to Mon- Galı́, Jordi. “How Well Does the IS-LM Model
etary Policy.” Journal of Political Economy, Fit Postwar U.S. Data?” Quarterly Journal of
2005, 113(1), pp. 1– 45. Economics, 1992, 107(2), pp. 709 –38.
Cochrane, John H. and Piazzesi, Monika. “The Hamilton, James D. Time series analysis. Prince-
Fed and Interest Rates—A High-Frequency ton: Princeton University Press, 1994.
Identification.” American Economic Review, Hansen, Bruce E. “Sample Splitting and Thresh-
2002 (Papers and Proceedings), 92(2), pp. old Estimation.” Econometrica, 2000, 68(3),
90 –95. pp. 575– 603.
Cogley, Timothy W. and Sargent, Thomas J. Hansen, Peter R. “Granger’s Representation
“Evolving Post World War II U.S. Inflation Theorem: A Closed-Form Expression for I(1)
Dynamics.” in Ben S. Bernanke and Kenneth Processes.” University of California, San Di-
Rogoff, eds., NBER macroeconomics annual, ego, Department of Economics Discussion
Vol. 16. Cambridge: MIT Press, 2002, pp. Paper: No. 2000 –17, 2000.
331–73. Hurvich, Clifford M. and Tsai, Chih-Ling. “A
Cogley, Timothy W. and Sargent, Thomas J. Corrected Akaike Information Criterion for
“Drifts and Volatilities: Monetary Policies Vector Autoregressive Model Selection.”
and Outcomes in the Post WWII U.S.” Uni- Journal of Time Series Analysis, 1993, 14(3),
versity of California, Davis, Department of pp. 271–79.
Economics Working Paper: No. 2003–25, Ing, Ching-Kang. “Multistep Prediction in Au-
2003. toregressive Processes.” Econometric The-
Cooley, Thomas F. and Dwyer, Mark. “Business ory, 2003, 19(2), pp. 254 –79.
Cycle Analysis without Much Theory: A Jordà, Òscar and Salyer, Kevin D. “The Re-
Look at Structural VARs.” Journal of Econo- sponse of Term Rates to Monetary Policy
metrics, 1998, 83(1–2), pp. 57– 88. Uncertainty.” Review of Economic Dynam-
Cox, David R. “Prediction by Exponentially ics, 2003, 6(4), pp. 941– 62.
Weighted Moving Averages and Related Kim, Jinill; Kim, Sunghyun; Schaumburg, Ernst
Methods.” Journal of the Royal Statistical and Sims, Christopher A. “Calculating and
Society, 1961, Series B, 23(2), pp. 414 –22. Using Second Order Accurate Solutions of
Demiralp, Selva and Hoover, Kevin D. “Search- Discrete Time Dynamic Equilibrium Mod-
ing for the Causal Structure of a Vector Au- els.” Board of Governors of the Federal Re-
toregression.” Oxford Bulletin of Economics serve System (U.S.), Finance and Economics
and Statistics, 2003, 65(S1), pp. 745– 67. Discussion Series: 2003– 61, 2003.
DeLong, J. Bradford. “America’s Only Peace- Koop, Gary; Pesaran, M. Hashem and Potter,
time Inflation: The 1970’s.” in Christina Ro- Simon M. “Impulse Response Analysis in
mer and David Romer, eds., Reducing Nonlinear Multivariate Models.” Journal of
inflation: Motivation and strategy. NBER Econometrics, 1996, 74(1), pp. 119 – 47.
Studies in Business Cycles, Vol. 30. Chicago: Lin, Jin-Lung and Tsay, Ruey S. “Co-Integration
University of Chicago Press, 1997, Chapter 6. Constraint and Forecasting: An Empirical
Evans, Charles L. and Marshall, David A. “Mon- Examination.” Journal of Applied Economet-
etary Policy and the Term Structure of Nom- rics, 1996, 11(5), pp. 519 –38.
inal Interest Rates: Evidence and Theory.” McCallum, Bennett T. “Robustness Properties of
Carnegie-Rochester Conference Series on a Rule for Monetary Policy: Reply.” Carne-
Public Policy, 1998, 49(0), pp. 53–111. gie-Rochester Conference Series on Public
Fuhrer, Jeffrey C. and Moore, George R. “Infla- Policy, 1988, 29(0), pp. 213–14.
tion Persistence.” Quarterly Journal of Eco- Potter, Simon M. “Nonlinear Impulse Response
nomics, 1995a, 110(1), pp. 127–59. Functions.” Journal of Economic Dynamics
Fuhrer, Jeffrey C. and Moore, George R. “Mon- and Control, 2000, 24(10), pp. 1425– 46.
etary Policy Trade-Offs and the Correlation Priestley, Maurice B. Non-linear and non-
between Nominal Interest Rates and Real stationary time series analysis. London: Ac-
Output.” American Economic Review, 1995b, ademic Press, 1988.
85(1), pp. 219 –39. Romer, Christina D. and Romer, David H. “The
182 THE AMERICAN ECONOMIC REVIEW MARCH 2005

Evolution of Economic Understanding and L. White, eds., Cointegration, causality and

Postwar Stabilization Policy” in Rethinking Forecasting: A festschrift in honor of Clive
Stabilization Policy, Federal Reserve Bank of W. J. Granger. Oxford: Oxford University
Kansas City: Symposium Proceedings, Au- Press, 1999, pp. 1– 44.
gust 29 –31, 2000 Jackson Hole Conference, Swanson, Norman R. and Granger, Clive W. J.
2003, pp. 11–78. “Impulse Response Functions Based on a
Rotemberg, Julio J. and Woodford, Michael. “An Causal Approach to Residual Orthogonaliza-
Optimization-Based Econometric Frame- tion in Vector Autoregressions.” Journal of
work for the Evaluation of Monetary Policy.” the American Statistical Association, 1997,
In Ben S. Bernanke and Julio J. Rotemberg, 92(437), pp. 357– 67.
eds., NBER Macroeconomics Annual. Cam- Taylor, John B. “Discretion versus Policy Rules
bridge: MIT Press, 1997, pp. 297–346. in Practice.” Carnegie-Rochester Conference
Rudebusch, Glenn D. “Do Measures of Mone- Series on Public Policy, 1993, 39(0), pp.
tary Policy in a VAR Make Sense?” Interna- 195–214.
tional Economic Review, 1998, 39(4), pp. Taylor, John B., ed. Monetary policy rules.
907–31. NBER Research Studies in Business Cycles
Rudebusch, Glenn D. and Svensson, Lars E. O. 1999. Chicago: University of Chicago Press,
“Policy Rule for Inflation Targeting,” in John 1999.
B. Taylor, ed., Monetary policy rules. NBER Thapar, Aditi. “Using Private Forecasts to Esti-
Research Studies in Business Cycles 1999. mate the Effects of Monetary Policy.” Un-
Chicago: University of Chicago Press, 1999, published Paper, 2002.
pp. 203– 46. Tsay, Ruey S. “Calculating Interval Forecasts:
Sims, Christopher A. “Macroeconomics and Re- Comment: Adaptive Forecasting.” Journal of
ality.” Econometrica, 1980, 48(1), pp. 1– 48. Business and Economic Statistics, 1993,
Sims, Christopher A. “Interpreting the Macro- 11(2), pp. 140 – 42.
economic Time Series Facts: The Effects of Tsay, Ruey S. “Testing and Modelling Multivar-
Monetary Policy.” European Economic Re- iate Threshold Models.” Journal of the Amer-
view, 1992, 36(5), pp. 975–1000. ican Statistical Association, 1998, 93(443),
Sims, Christopher A. “Do Measures of Monetary pp. 1188 –202.
Policy in a VAR Make Sense? A Reply.” Wallis, Kenneth F. “Multiple Time Series and
International Economic Review, 1998, 39(4), the Final Form of Econometric Models”
pp. 943– 48. Econometrica, 1977, 45(6), pp. 1481–97.
Sims, Christopher A. and Zha, Tao. “Error Bands Weiss, Andrew A. “Multi-Step Estimation and
for Impulse Responses.” Econometrica, Forecasting in Dynamic Models.” Journal
1999, 67(5), pp. 1113–55. of Econometrics, 1991, 48(1–2), pp. 135– 49.
Stock, James H. and Watson, Mark W. “A Com- Zellner, Arnold and Palm, Franz. “Time Series
parison of Linear and Non-linear Univariate Analysis and Simultaneous Equation Econo-
Models for Forecasting Macroeconomic metric Models.” Journal of Econometrics,
Time Series,” in Robert F. Engle and Halbert 1974, 2(1), pp. 17–54.