ArticlePDF Available

Forecasting Inflation in a Data-Rich Environment: The Benefits of Machine Learning Methods

July 2019
Journal of Business and Economic Statistics 39(1):1-45

DOI:10.1080/07350015.2019.1637745

Authors:

Marcelo C Medeiros

Pontifícia Universidade Católica do Rio de Janeiro

Gabriel Vasconcelos

University of California, Irvine

Alvaro Veiga

Pontifícia Universidade Católica do Rio de Janeiro

Inflation forecasting is an important but difficult task. Here, we explore advances in machine learning (ML) methods and the availability of new datasets to forecast US inflation. Despite the skepticism in the previous literature, we show that ML models with a large number of covariates are systematically more accurate than the benchmarks. The ML method that deserves more attention is the random forest model, which dominates all other models. Its good performance is due not only to its specific method of variable selection but also the potential nonlinearities between past key macroeconomic variables and inflation.

Example of a regression tree. Reproduction of part of Figure 9.2 in Hastie, Tibshirami, and Friedman (2001).

…

Figures - uploaded by Gabriel Vasconcelos

Content may be subject to copyright.

Content uploaded by Gabriel Vasconcelos

Content may be subject to copyright.

Full Terms & Conditions of access and use can be found at

https://amstat.tandfonline.com/action/journalInformation?journalCode=ubes20

Journal of Business & Economic Statistics

ISSN: 0735-0015 (Print) 1537-2707 (Online) Journal homepage: https://amstat.tandfonline.com/loi/ubes20

Forecasting Inflation in a Data-Rich Environment:

The Benefits of Machine Learning Methods

Marcelo C. Medeiros, Gabriel F. R. Vasconcelos, Álvaro Veiga & Eduardo

Zilberman

To cite this article: Marcelo C. Medeiros, Gabriel F. R. Vasconcelos, Álvaro Veiga

& Eduardo Zilberman (2019): Forecasting Inflation in a Data-Rich Environment: The

Benefits of Machine Learning Methods, Journal of Business & Economic Statistics, DOI:

10.1080/07350015.2019.1637745

To link to this article: https://doi.org/10.1080/07350015.2019.1637745

View supplementary material

Accepted author version posted online: 01

Jul 2019.

Published online: 19 Aug 2019.

Submit your article to this journal

Article views: 247

View related articles

View Crossmark data

Supplementary materials for this article are available online. Please go to http://tandfonline.com/r/JBES

Forecasting Inﬂation in a Data-Rich

Environment: The Beneﬁts of Machine Learning

Methods

Marcelo C. MEDEIROS

Department of Economics, Pontiﬁcal Catholic University of Rio de Janeiro, Rua Marqu ˆ

es de S˜

ao Vicente 225,

G´

avea, Rio de Janeiro 22451-900, Brazil (

mcm@econ.puc-rio.br

)

Gabriel F. R. VASCONCELOS

Department of Economics, University of California, Irvine, 3201 Social Sciences Plaza B, Irvine, CA 92617

(

gabriel.vasconcelos@uci.edu

)

Alvaro VEIGA

Department of Electrical Engineering, Pontiﬁcal Catholic University of Rio de Janeiro, Rua Marquˆ

es de S˜

Vicente 225, G´

avea, Rio de Janeiro 22451-900, Brazil (

alvf@ele.puc-rio.br

)

Eduardo ZILBERMAN

Central Bank of Chile and Department of Economics, Pontiﬁcal Catholic University of Rio de Janeiro (PUC-Rio),

Agustinas 1180, Santiago 867, Chile (

ezilberman@bcentral.cl

)

Inﬂation forecasting is an important but difﬁcult task. Here, we explore advances in machine learning

(ML) methods and the availability of new datasets to forecast U.S. inﬂation. Despite the skepticism in the

previous literature, we show that ML models with a large number of covariates are systematically more

accurate than the benchmarks. The ML method that deserves more attention is the random forest model,

which dominates all other models. Its good performance is due not only to its speciﬁc method of variable

selection but also the potential nonlinearities between past key macroeconomic variables and inﬂation.

Supplementary materials for this article are available online.

KEY WORDS: Big data; Inﬂation forecasting; LASSO; Machine learning; Random forests.

1. INTRODUCTION

It is difﬁcult to overemphasize the importance of forecasting

inﬂation in rational economic decision-making. Many contracts

concerning employment, sales, tenancy, and debt are set in

nominal terms. Therefore, inﬂation forecasting is of great

value to households, businesses, and policymakers. In addition,

central banks rely on inﬂation forecasts not only to inform

monetary policy but also to anchor inﬂation expectations and

thus enhance policy efﬁcacy. Indeed, as part of an effort

to improve economic decision-making, many central banks

release inﬂation forecasts on a regular basis.

Despite the beneﬁts of forecasting inﬂation accurately,

improving simple models has proved challenging. As Stock

and Watson (2010) emphasized, “it is exceedingly difﬁcult

to improve systematically upon simple univariate forecasting

models, such as the Atkeson and Ohanian (2001) random

walk model [...] or the time-varying unobserved components

model in Stock and Watson (2007).” This conclusion is

supported by a large literature (Faust and Wright 2013),

1See Mullainathan and Spiess (2017) for discussions of ML methods and big

data in economics. In this article, an ML model is any statistical model that

is able to either handle a large set of covariates and/or describe nonlinear

mappings nonparametrically. Some of these methods even predate “machines.”

but this literature has largely ignored the recent machine

learning (ML) and “big data” boom in economics.1With a

few exceptions, previous works either considered a restrictive

set of variables or were based on a small set of factors

computed from a larger number of predictors known as

“diffusion indexes” (Stock and Watson 2002). In addition,

most of these works focused on a time period when inﬂation

was very persistent, which favors models that treat inﬂation as

nonstationary.

“Big data” and ML methods are not passing fads, and

investigating whether the combination of the two is able to pro-

vide more accurate forecasts is of paramount importance. Gu,

Kelly, and Xiu (2018), for example, showed that ML methods

coupled with hundreds of predictors improve substantially out-

of-sample stock return predictions. In a similar spirit, despite

the previous skepticism, we argue that these methods lead to

more accurate inﬂation forecasts. We ﬁnd that the gains of using

ML methods can be as large as 30% in terms of mean squared

2019 American Statistical Association

Journal of Business & Economic Statistics

Month, 2019, Vol. 00, No. 0

DOI: 10.1080/07350015.2019.1637745

Color versions of one or more of the ﬁgures in the article can be

found online at

www.tandfonline.com/r/jbes

2Journal of Business & Economic Statistics, Month 2019

errors.2Moreover, this new set of models can help uncover the

main predictors for future inﬂation, possibly shedding light on

the drivers of price dynamics.

These ﬁndings are practically important in light of the large

forecast errors that many central banks, international institu-

tions and other forecasters have made recently. A striking exam-

ple is the European Central Bank (ECB), whose projections

have been systematically and substantially above realized inﬂa-

tion recently. Effective monetary policy depends on accurate

inﬂation forecasts; otherwise, the policy stance will be tighter or

looser than necessary. In addition, systematic forecasting errors

may undermine central banks’ credibility and their ability

to anchor inﬂation expectations. Furthermore, Svensson and

Woodford (2004) argued that optimal monetary policy can

be implemented through an inﬂation forecast-targeting regime

in which central banks target inﬂation forecasts over a given

horizon. Taken together, these arguments suggest potentially

large welfare costs associated with failures to forecast inﬂation.

Not surprisingly, such recent failures have fostered a debate on

inﬂation forecasting practices.3This article contributes to this

debate by providing a comprehensive guide and assessment of

ML methods for one important case study, US inﬂation.

1.1. Main Takeaways

We contribute to the literature in a number of ways. First,

contrary to the previous evidence in Stock and Watson (1999,

2007), Atkeson and Ohanian (2001), and many others, our

results show that it is possible to consistently beat univariate

benchmarks for inﬂation forecasting, namely, random walk

(RW), autoregressive (AR), and unobserved components

stochastic volatility (UCSV) models. We consider several ML

models in a data-rich environment from FRED-MD, a monthly

database compiled by McCracken and Ng (2016), to forecast

U.S. consumer price index (CPI) inﬂation during more than 20

years of out-of-sample observations. We show that the gains

can be as large as 30% in terms of mean squared errors. Our

results are valid for different subsamples. Forecasting inﬂation

is important to inform rational economic decision-making.

However, as these decisions are made in real time, we check the

robustness of our results by considering a real-time experiment

from 2001 to 2015. The superiority of ML methods persists

even in real time.

2To gauge the relevance of such gains, consider the rough welfare calculations

in Dellas et al. (2018). The authors use a textbook macro model with nominal

rigidities, in which inﬂation forecasting errors lead to inefﬁcient output gap

volatility. If relative risk aversion is equal to two, agents are willing to

forgo 0.16% (0.34%) of their steady state consumption to avoid a forecasting

deterioration of 20% (50%) in terms of mean squared errors. These ﬁgures

are relatively large. As a ground for comparison, Lucas (1987), who originally

proposed this method to measure the welfare cost of business cycles, found it

to be 0.10%.

3The roles of ML methods and big data were discussed at the Norges Bank’s

workshop on “big data, machine learning and the macroeconomy” in 2017, as

well as at the ECB workshop on “economic forecasting with large datasets” in

2018. Banca D’Italia and Deutsche Bundesbank promoted similar workshops

in 2018 and 2019, respectively. Finally, staff of central banks have just started

to produce working papers with applications that combine big data, ML and

forecasting, for example, Chakraborty and Joseph (2017) from Bank of England

and Hall (2018) from Kansas City Fed.

Second, we highlight the main set of variables responsible

for these forecast improvements. Our results indicate that this

set of variables is not sparse, which corroborates the ﬁndings

of Giannone, Lenza, and Primiceri (2018). Indeed, we ﬁnd

that ML models that do not impose sparsity are the best-

performing ones. By contrast, the high level of aggregation

of factor models, which have been among the most popular

models for macroeconomic forecasting, is not adequate. Fur-

thermore, either replacing standard principal component factors

with target factors, as advocated by Bai and Ng (2008), or using

boosting to select factors as in Bai and Ng (2009) improves the

results only marginally.

Finally, we pay special attention to a particular ML model,

the random forest (RF) of Breiman (2001), which systemati-

cally outperforms the benchmarks and several additional ML

and factor methods: the least absolute shrinkage and selection

operator (LASSO) family, which includes LASSO, adaptive

LASSO (adaLASSO), elastic net (ElNet), and the adaptive

elastic net (adaElNet); ridge regression (RR); Bayesian vec-

tor autoregressions (BVAR); and principal component fac-

tors, target factors (T. Factor), and linear ensemble methods

such as bagging, boosted factors (B. Factor), complete subset

regressions (CSR), and jackknife model averaging (JMA).

RF models are highly nonlinear nonparametric models that

have a tradition in statistics but have only recently attracted

attention in economics. This late success is partly due to

the new theoretical results developed by Scornet, Biau, and

Vert (2015) and Wagner and Athey (2018). As robustness

checks, we also compare the RF models with four nonlinear

alternatives: deep neural networks (Deep NN), boosted trees

(BTrees), and a polynomial model estimated either by LASSO

or adaLASSO. We also compare our results to a linear model

estimated with nonconcave penalties (SCAD) as in Fan and Li

(2001).

There are several reasons why our results differ from those

of Stock and Watson (1999,2007), and Atkeson and Oha-

nian (2001). The combination of ML methods with a large

dataset not considered by these authors is an obvious one.

The choice of a different sample can also explain the dif-

ferences. The above papers work in an environment where

inﬂation is clearly integrated, whereas inﬂation is stationary

in the period considered in this article. This fact alone makes

both the RW and UCSV models less attractive, as clearly

they are not suitable for stationary data. Nevertheless, if the

gains of the ML methods are due solely to the fact that

inﬂation is much less persistent, we would observe compet-

itive performance of the AR or factor models, which is not

the case. Although the performance of the RW and UCSV

models improves when accumulated inﬂation is considered,

the ML methods still achieve superior results. By construction,

accumulated inﬂation is much more persistent compared to

the monthly ﬁgures. Given all of these reasons, ML methods

coupled with large datasets merit serious consideration for

forecasting inﬂation.

To open the black box of ML methods, we compare the

variables selected by the adaLASSO, RR, and RF models.

Following McCracken and Ng (2016), we classify the variables

into eight groups: (i) output and income; (ii) labor market;

(iii) housing; (iv) consumption, orders, and inventories; (v)

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 3

money and credit; (vi) interest and exchange rates; (vii) prices;

and (viii) stock market. In addition, we consider AR terms

and the factors computed from the full set of predictors.

The most important variables for RR and RF models are

stable across horizons but are quite different between the

two speciﬁcations. For RR, AR terms, prices and employ-

ment are the most important predictors, resembling a sort

of backward-looking Phillips curve, whereas RF models give

more importance to disaggregated prices, interest-exchange

rates, employment and housing. The RF model resembles a

nonlinear Phillips curve in which the backward-looking terms

are disaggregated. The adaLASSO selection is quite different

across forecasting horizons and is, by construction and in

opposition to RF and RR models, sparse. Only AR terms retain

their relative importance independent of the horizon, and prices

gradually lose their relevance until up to 6 months ahead but

partially recover for longer horizons. Output-income variables

are more important for medium-term forecasts. Finally, none

of the three classes of models selects either factors or stocks,

not even RR or RF, which produces a nonsparse solution.

This result may indicate that the high level of cross-section

aggregation of the factors is one possible cause for its poor

performance.

To disentangle the effects of variable selection from nonlin-

earity, we consider two alternative models. The ﬁrst uses the

variables selected by RF and estimates a linear speciﬁcation

by OLS. The second method estimates RF with only the

regressors selected by adaLASSO. Both models outperform

RF only for one-month-ahead forecasting. For longer horizons,

the RF model is still the winner, which provides evidence that

both nonlinearity and variable selection play a key role in the

superiority of the RF model.

There are many sources of nonlinearities that could justify

the superiority of the RF model. For instance, the relationship

between inﬂation and employment is nonlinear to the extent

that it depends on the degree of slackness in the economy.

Another source of nonlinearity is economic uncertainty, as this

uncertainty increases the option value of economic decision

delays if they have an irreversible component (Bloom 2009).

For example, if it is expensive to dismiss workers, hiring

should be nonlinear on uncertainty. In addition, this real option

argument also makes households and businesses less sensitive

to changes in economic conditions when uncertainty is high.

Hence, the responses of employment and inﬂation to interest

rate decisions are arguably nonlinear on uncertainty. The pres-

ence of a zero lower bound on nominal interest rates and the

implications of this bound for unconventional monetary policy

is another source of nonlinearity among inﬂation, employment

and interest rate variables (Eggertsson and Woodford 2003).

Finally, to the extent that houses serve as collateral for loans, it

interacts with monetary policy (Iacoviello 2005) and ﬁnancial

intermediation (Mian and Suﬁ 2009). As in the Great Reces-

sion, a housing bubble can form, resulting in a deep credit crash

(Shiller 2014). Needless to say, these interactions are highly

nonlinear and arguably have nonlinear effects on inﬂation,

employment and interest rates. In line with these arguments, we

show that the gains of the RF model are larger during recessions

and periods of high uncertainty, especially during and after the

Great Recession.

1.2. A Brief Comparison to the Literature

The literature on inﬂation forecasting is vast, and there is

substantial evidence that models based on the Phillips curve do

not provide good forecasts. Although Stock and Watson (1999)

showed that many production-related variables are potential

predictors of US inﬂation, Atkeson and Ohanian (2001)showed

that in many cases, the Phillips curve fails to beat even sim-

ple naive models. These results inspired researchers to seek

different models and variables to improve inﬂation forecasts.

Among the variables used are ﬁnancial variables (Forni et al.

2003), commodity prices (Chen, Turnovsky, and Zivot 2014),

and expectation variables (Groen, Paap, and Ravazzolo 2013).

However, there is no systematic evidence that these models

outperform the benchmarks.

Recently, due to advancements in computational power,

theoretical developments in ML, and the availability of large

datasets, researchers have started to consider the usage of

high-dimensional models in addition to the well-established

(dynamic) factor models. However, most of these studies have

either focused only on a very small subset of ML models or

presented a restrictive analysis. For example, Inoue and Kilian

(2008) considered bagging, factor models, and other linear

shrinkage estimators to forecast US inﬂation with a small set of

real economic activity indicators. Their study is more limited

than ours both in terms of the pool of models and richness

of the set of predictors. Nevertheless, the authors are among

the few voices suggesting that ML techniques can deliver

nontrivial gains over univariate benchmarks. Medeiros and

Mendes (2016) provided evidence that LASSO-based models

outperform both factor and AR benchmarks in forecasting US

CPI. However, their analysis was restricted to a single ML

method for only one-month-ahead forecasting.

The literature has mainly explored linear ML models. One

explanation for this limitation is that several of the papers in

the early days considered only univariate nonlinear models that

were, in most cases, outperformed by simple benchmarks; see

Ter ¨

asvirta, van Dijk, and Medeiros (2005). An exception is

Nakamura (2005), who showed that neural networks outper-

form univariate autoregressive models for short horizons.

Recently, Garcia, Medeiros, and Vasconcelos (2017) applied

a large number of ML methods, including RFs, to real-time

inﬂation forecasting in Brazil. The results indicated a superior-

ity of the CSR method of Elliott, Gargano, and Timmermann

(2013). However, an important question is whether this is

a particular result for Brazil or if similar ﬁndings can be

replicated for the U.S. economy. The ﬁrst difference between

the results presented here and those in Garcia, Medeiros, and

Vasconcelos (2017) is that the RF model robustly outperforms

its competitors and CSR does not perform particularly well.

With respect to the set of models considered, on the one hand, in

this article we employ more ML models than Garcia, Medeiros,

and Vasconcelos (2017), but on the other hand, we do not have a

real-time daily database of inﬂation expectations as in the case

of Brazil. We also provide a much richer discussion of variable

selection and the nature of the best-performing models.

Finally, it is important to contextualize our work in light

of the criticisms of Makridakis, Spiliotis, and Assimakopoulos

(2018) with respect to the ability of ML methods to produce

4Journal of Business & Economic Statistics, Month 2019

reliable forecasts. The methods considered here are much

different than those in the study of Makridakis, Spiliotis, and

Assimakopoulos (2018). While we consider modern ML tools

such as RF, shrinkage, and bagging, the authors focus on simple

regression trees, shallow neural networks, and support vector

machines. Furthermore, the models considered in Makridakis,

Spiliotis, and Assimakopoulos (2018) are univariate in the

sense that no other variables apart from lags are used as

predictors.

1.3. Organization of the Article

Section 2 gives an overview of the data. Section 3 describes

the forecasting methodology. Section 4 describes the models

used in the article. Section 5 provides the main results. Sec-

tion 6 concludes. The online supplementary materials provide

additional results. Tables and ﬁgures labeled with an “S” refer

to this supplement.

2. DATA

Our data consist of variables from the FRED-MD database,

which is a large monthly macroeconomic dataset designed for

empirical analysis in data-rich environments. The dataset is

updated in real time through the FRED database and is available

from Michael McCraken’s webpage.4For further details, we

refer to McCracken and Ng (2016).

We use the vintage as of January 2016. Our sample extends

from January 1960 to December 2015 (672 observations), and

only variables with all observations in the sample period are

used (122 variables). In addition, we include as potential pre-

dictors the four principal component factors computed from this

set of variables. We consider four lags of all variables, as well

as four autoregressive terms. Hence, the analysis contemplates

508 potential predictors. The out-of-sample window is from

January 1990 to December 2015. All variables are transformed

to achieve stationarity as described in the supplementary mate-

rials. πtis the inﬂation in month tcomputed as πt=log(Pt)−

log(Pt−1), and Ptis a given price index in period t. The baseline

price index is the CPI, but in the supplementary materials we

report results for the PCE and the core CPI inﬂation. Figure S.1

in this supplement displays the evolution of inﬂation measures

during the full sample period. The noncore inﬂation measures

have a large outlier in November 2008 associated with a

large decline in oil prices. This outlier can certainly affect the

estimation of the models considered in this article. To attenuate

its effects, we include a dummy variable for November 2009

in all models estimated after that date. We do not include any

look-ahead bias, as the dummy is added only after the outlier

has been observed by the econometrician.

We compare performance across models in the out-of-

sample window and in two different subsample periods,

namely, January 1990 to December 2000 (132 observations)

and January 2001 to December 2015 (180 observations). The

ﬁrst sample corresponds to a period of low inﬂation volatility

(σ=0.17%), while in the second sample, inﬂation is more

4https://research.stlouisfed.org/econ/mccracken/fred-databases/.

volatile (σ=0.32%). However, on average, inﬂation is higher

during 1990–2000 than 2001–2015 and much more persistent

as well. Relative to the 1990–2000 period, inﬂation was more

volatile near the recession in the early 1990s. Table S.9 in the

supplementary material provides descriptive statistics and gives

an overview of the economic scenario in each subsample.

As a robustness check, we also report the results of a real-

time experiment using our second subsample, from 2001 to

2015. We choose this particular period because the real-time

vintages are easily available from the FRED-MD database. For

the real-time experiment, the number of potential regressors

varies according to the vintage.

3. METHODOLOGY

Consider the following model

πt+h=Gh(xt)+ut+h,h=1, ...,H,t=1, ...,T,(1)

where πt+his the inﬂation in month t+h;xt=(x1t,...,xnt)

is a n-vector of covariates possibly containing lags of πtand/or

common factors as well as a large set of potential predictors;

Gh(·)is the mapping between covariates and future inﬂation;

and utis a zero-mean random error. The target function Gh(xt)

can be a single model or an ensemble of different speciﬁcations.

There is a different mapping for each forecasting horizon.

The direct forecasting equation is given by

πt+h|t=

Gh,t−Rh+1:t(xt),(2)

where 

Gh,t−Rh+1:tis the estimated target function based on data

from time t−Rh+1uptotand Rhis the window size, which

varies according to the forecasting horizon and the number of

lagged variables in the model. We consider direct forecasts as

we do not make any attempt to predict the covariates. The only

exception is the case of the BVAR model, where joint forecasts

for all predictors are computed in a straightforward manner

following the procedure described in Ba´

nbura, Giannone, and

Reichlin (2010).

The forecasts are based on a rolling-window framework

of ﬁxed length. However, the actual in-sample number of

observations depends on the forecasting horizon. For example,

for the 1990–2000 period, the number of observations is Rh=

360 −h−p−1, where pis the number of lags in the model.

For 2001–2015, Rh=492 −h−p−1. We choose to work in a

rolling-window scheme for two reasons: to attenuate the effects

of potential structural breaks and outliers and to avoid problems

of running superior predictive performance tests among nested

models; see the discussion in Giacomini and White (2006).

For instance, with a rolling-window scheme, the unconditional

Giacomini–White (GW) test is equivalent to the traditional

Diebold–Mariano (DM) test.

In addition to three benchmark speciﬁcations (RW, AR, and

UCSV models), we consider factor-augmented AR models,

sparsity-inducing shrinkage estimators (LASSO, adaLASSO,

ElNet, and adaElNet), other shrinkage methods that do not

induce sparsity (RR and BVAR with Minnesota priors), aver-

aging (ensemble) methods (bagging, CSR, and JMA) and

RF. Bagging and CSR can be viewed as nonsparsity-inducing

shrinkage estimators. With respect to the factor-augmented AR

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 5

models, we consider in addition to the standard factors com-

puted with principal component analysis a set of target factors

as in Bai and Ng (2008) and boosted factors as in Bai and Ng

(2009). We also include in the comparison three different model

combination schemes, namely, the simple average, the trimmed

average, and the median of the forecasts. For both the shrinkage

and factor-based methods, the set of predictors are standardized

before estimation.

We ﬁnd that RF, a highly nonlinear method, robustly outper-

forms the other methods. To disentangle the role of variable

selection from nonlinearity, we also consider a linear model

where the regressors are selected by RFs (RF/ordinary least

squares, OLS) and an RF model with regressors preselected by

adaLASSO (adaLASSO/RF).

Forecasts for the accumulated inﬂation over the following

3, 6, and 12 months are computed, with the exception of the

RW and UCSV models, by aggregating the individual forecasts

for each horizon. In the case of the RW and UCSV models,

a different speciﬁcation is used to construct the forecasts (see

below).

4. MODELS

4.1. Benchmark Models

The ﬁrst benchmark is the RW model, where for h=

1, ..., 12, the forecasts are as πt+h|t=πt. For the accumu-

lated h-month forecast, we set πt+1:t+h|t=πt−(h−1):t, where

πt−(h−1):tis the accumulated inﬂation over the previous h

months.

The second benchmark is the autoregressive (AR) model of

order p, where pis determined by the Bayesian information

criterion (BIC) and the parameters are estimated by OLS. The

forecast equation is πt+h|t=

φ0,h+

φ1,hπt+···+

φp,hπt−p+1.

There is a different model for each horizon. The accumulated

forecasts are computed by aggregating the individual forecasts.

Finally, the third benchmark is the UCSV model, which is

described as follows

πt=τt+eht/2εt,τt=τt−1+ut,ht=ht−1+vt,

where {εt}is a sequence of independent and normally dis-

tributed random variables with zero mean and unit variance

and utand vtare also normal with zero mean and variance

given by inverse-gamma priors. τ1∼N(0, Vτ), and h1∼

N(0, Vh), where Vτ=Vh=0.12. The model is estimated

by Markov chain Monte Carlo (MCMC) methods. The h-steps-

ahead forecast is computed as πt+h=τt|t. For accumulated

forecasts, the UCSV is estimated with the accumulated h-month

inﬂation as the dependent variable.

4.2. Shrinkage

In this article, we estimate several shrinkage estimators for

linear models where Gh(xt)=β

hxtand



βh=arg min

βhT−h



t=1yt+h−β

hxt2+



i=1

p(βh,i;λ,ωi).(3)

p(βh,i;λ,ωi)is a penalty function that depends on the penalty

parameter λand on a weight ωi>0. We consider different

choices for the penalty functions.

4.2.1. Ridge Regression (RR). RR was proposed by

Hoerl and Kennard (1970). The penalty is given by



i=1

p(βh,i;λ,ωi):=λ



i=1

β2

h,i.(4)

RR has the advantage of having an analytic solution that is

easy to compute, and it also shrinks the coefﬁcients associated

with less-relevant variables to nearly zero. However, the coefﬁ-

cients rarely reach exactly zero for any size of λ.

4.2.2. Least Absolute Shrinkage and Selection Operator

(LASSO). LASSO was proposed by Tibshirani (1996), where

the penalty is



i=1

p(βh,i;λ,ωi):=λ



i=1|βh,i|.(5)

LASSO shrinks the irrelevant variables to zero. However,

model selection consistency is achieved only under very strin-

gent conditions.

4.2.3. Adaptive Least Absolute Shrinkage and Selection

Operator (adaLASSO). adaLASSO was proposed by Zou

(2006) to achieve model selection consistency. adaLASSO uses

the same penalty as LASSO with the inclusion of a weighting

parameter that comes from a ﬁrst-step estimation. The penalty

is given by



i=1

p(βh,i;λ,ωi):=λ



i=1

ωi|βh,i|,(6)

where ωi=|β∗

h,i|−1and β∗

h,iis the coefﬁcient from the ﬁrst-step

estimation. adaLASSO can deal with many more variables than

observation and works well in non-Gaussian environments and

under heteroscedasticity (Medeiros and Mendes 2016).

4.2.4. Elastic Net (ElNet). ElNet is a generalization that

includes LASSO and RR as special cases. It is a convex

combination of the 1and 2norms (Zou and Hastie 2005).

The ElNet penalty is deﬁned as



i=1

p(βh,i;λ,ωi):=αλ



i=1

β2

h,i+(1−α)λ



i=1|βh,i|;(7)

where α∈[0, 1]. We also consider an adaptive version of ElNet

(adaElNet), which works in the same way as adaLASSO.

4.3. Factor Models

The idea of the factor model is to extract common compo-

nents from all predictors, thus reducing the model dimension.

Factors are computed as principal components of a large set of

variables ztsuch that Ft=Azt, where Ais a rotation matrix

and Ftis the vector of the principal components. Consider

Equation (1). In this case, xtis given by πt−j,j=0, 1, 2, 3

plus ft−j,j=0, 1, 2, 3, where ftis the vector with the ﬁrst four

principal components of zt. The theory behind factor models

can be found in Bai and Ng (2003).

6Journal of Business & Economic Statistics, Month 2019

4.3.1. Target Factors. To improve the forecasting per-

formance of factor models, Bai and Ng (2008) proposed tar-

geting the predictors. The idea is that if many variables in

ztare irrelevant predictors of πt+h, factor analysis using all

variables may result in noisy factors with poor forecasting

ability. The idea is to compute the principal components only

of the variables with high prediction power for future inﬂation.

Let zi,t,i=1, ...,qbe the candidate variables and wta set of

controls. We follow Bai and Ng (2008) and use lagged values

of πtas controls. The procedure is described as follows.

1. For i=1, ...,q,regressπt+hon wtand zi,tand compute the

t-statistics for the coefﬁcient corresponding to zi,t.

2. Choose a signiﬁcance level αand select all variables that are

signiﬁcant using the computed t-statistics.

3. Let zt(α) be the selected variables from Steps 1–2. Estimate

the factors Ftfrom zt(α) by the principal components.

4. Regress πt+hon wtand ft−j,j=0, 1, 2, 3, where ft⊂Ft.

The number of factors in ftis selected using the BIC.

The same procedure was used by Medeiros and Vasconcelos

(2016) to forecast US CPI inﬂation. The authors showed that

the target factors slightly reduce the forecasting errors.

4.3.2. Factor Boosting. The optimal selection of factors

for predictive regressions is an open problem in the literature.

Even if the factor structure is clear in the data, it is not obvious

that only the most relevant factors should be included in the

predictive regression. We adopt the boosting algorithm as in Bai

and Ng (2008) to select the factors and the number of lags in the

model. Deﬁne zt∈Rq,thesetofallnfactors computed from

the original nvariables plus four lags of each factor. Therefore,

q=5n.

The algorithm is deﬁned as follows:

1. Let t,0 =¯πfor each t, where ¯π=1

tt

i=1πi.

2. For m=1, ...,M: (a) Compute 

ut=πt−t−h,m−1.(b)

For each candidate variable i=1, ...,q, regress the current

residual on zi,tto obtain 

bi, and compute 

et,i=

ut−zi,t

bi.

Calculate SSRi=

e



ei. (c) Select i∗

mas the index of the

variable that delivers the smallest SSR, and deﬁne 

φt,m=

zi∗

m,t

bi∗

m. (d) Update 

t,m=

t,m−1+vφt,m, where vis the

step length. We set v=0.2.

3. Stop the algorithm after the Mth iteration or when the BIC

starts to increase.

4.4. Ensemble Methods

Ensemble forecasts are constructed from a (weighted) aver-

age of the predictions of an ensemble of methods.

4.4.1. Bagging. The term “bagging” comes from boot-

strap aggregation, which was proposed by Breiman (1996).

The idea is to combine forecasts from several unstable models

estimated for different bootstrap subsamples. Normally, there

is much more to gain from combinations of models if they are

very different. The bagging steps are as follows:

1. For each bootstrap sample, run an OLS regression with

all candidate variables and select those with an absolute t-

statistic above a certain threshold c.

2. Estimate a new regression only with the variables selected in

the previous step.

3. The coefﬁcients from the second regression are ﬁnally used

to compute the forecasts on the actual sample.

4. Repeat the ﬁrst three steps for Bbootstrap samples and

compute the ﬁnal forecast as the average of the Bforecasts.

Note that in our case, the number of observations may

be smaller than the number of variables, which makes the

regression in the ﬁrst step unfeasible. We solve this issue by,

for each bootstrap subsample, randomly dividing all variables

in groups and running the pretesting step for each one of the

groups.

4.4.2. Complete Subset Regressions (CSR). CSR was

developed by Elliott, Gargano, and Timmermann (2013,2015).

The motivation was that selecting the optimal subset of xtto

predict πt+hby testing all possible combinations of regressors

is computationally very demanding and, in most cases, unfea-

sible. Suppose that we have ncandidate variables. The idea is

to select a number q≤nvariables and run regressions using all

possible combinations qof nvariables. The ﬁnal forecast is the

average over all forecasts.

CSR handles a small number of variables. For large sets, the

number of regressions to be estimated increases rapidly. For

example, with n=25 and q=4, there are 12,650 regressions.

Therefore, we adopt a pretesting procedure similar to that used

with the target factors. We start ﬁtting a regression of πt+h

on each of the candidate variables (including lags) and save

the t-statistics of each variable.5The t-statistics are ranked by

absolute value, and we select the ˜nvariables that are more

relevant in the ranking. The CSR forecast is calculated on these

variables.

4.4.3. Jackknife Model Averaging (JMA). JMA is a

method to combine forecasts from different models. Instead of

using simple average, JMA uses leave-one-out cross-validation

to estimate optimal weights; see Hansen and Racine (2012) and

Zhang, Wan, and Zou (2013).

Suppose we have Mcandidate models that we want to

average from and write the forecast of each model as π(m)

t+h,

m=1, ...,M. Set the ﬁnal forecast as

πt+h=



m=1

ωmπ(m)

t+h,

where 0 ≤ωm≤1 for all m∈{1, ...,M}and M

m=1ωm=1.

The JMA procedure is as follows:

1. For each observation of (xt,πt+h): (a) Estimate all candidate

models leaving the selected observation out of the estima-

tion. Since we are in a time series framework with lags in

the model, we also remove four observations before and four

observations after (xt,πt+h). (b) Compute the forecasts from

each model for the observations that were removed in the

previous step.

2. Choose the weights that minimize the cross-validation errors

subject to the constraints previously described.

Each candidate model in the JMA has four lags of the

inﬂation and four lags of one candidate variable.

5We do not use a ﬁxed set of controls, wt, in the pretesting procedure like we

did for the target factors.

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 7

4.5. Random Forests

The RF model was proposed by Breiman (2001) to reduce

the variance of regression trees and is based on bootstrap aggre-

gation (bagging) of randomly constructed regression trees. In

turn, a regression tree is a nonparametric model that approx-

imates an unknown nonlinear function with local predictions

using recursive partitioning of the space of the covariates

(Breiman 1996).

To understand how a regression tree works, an example from

Hastie, Tibshirami, and Friedman (2001) is useful. Consider

a regression problem in which X1and X2are explanatory

variables, each taking values in some given interval, and Yis

the dependent variable. We ﬁrst split the space into two regions

at X1=s1, and then the region to the left (right) of X1=s1

is split at X2=s2(X1=s3). Finally, the region to the right

of X1=s3is split at X2=s4. As illustrated in the left plot

of Figure 1, the ﬁnal result is a partitioning into ﬁve regions:

Rk,k=1, ..., 5. In each region Rk, we assume that the model

predicts Ywith a constant ck, which is estimated as the sample

average of realizations of Ythat “fall” within region Rk.Akey

advantage of this recursive binary partition is that it can be

represented as a single tree, as illustrated in the right plot of

Figure 1. Each region corresponds to a terminal node of the

tree. Given a dependent variable πt+h, a set of predictors xt

and a number of terminal nodes K, the splits are determined to

minimize the sum of squared errors of the following regression

model

πt+h=



k=1

ckIk(xt;θk),

where Ik(xt;θk)is an indicator function such that

Ik(xt;θk)=1ifxt∈Rk(θk),

0 otherwise.

θkis the set of parameters that deﬁne the kth region. Ik(xt;θk)

is in fact a product of indicator functions, each of which deﬁnes

one of the splits that yields the kth region.

RF is a collection of regression trees, each speciﬁed in a

bootstrap sample of the original data. Since we are dealing

with time series, we use a block bootstrap. Suppose there are

Bbootstrap samples. For each sample b,b=1, ...,B,atree

with Kbregions is estimated for a randomly selected subset of

the original regressors. Kbis determined to leave a minimum

number of observations in each region. The ﬁnal forecast is the

average of the forecasts of each tree applied to the original data

πt+h=1



b=1Kb



k=1

ck,bIk,b(xt;

θk,b).

4.6. Hybrid Linear-Random Forest Models

RF/OLS and adaLASSO/RF are adaptations speciﬁcally

designed to disentangle the relative importance of variable

selection and nonlinearity to forecast US inﬂation. RF/OLS is

estimated using the following steps:

1. For each bootstrap sample b=1, ...,B: (a) Grow a single

tree with knodes (we used k=20), and save the N≤ksplit

variables. (b) Run an OLS on the selected splitting variables.

t+h.

2. The ﬁnal forecast will be ˆπt+h=B−1B

b=1πb

t+h.

The main objective of RF/OLS is to check the performance

of a linear model using variables selected by the RF model.

If the results are very close to those of the RF model, we

understand that nonlinearity is not an issue, and the RF model

is superior solely because of variable selection. However, if

we see some improvement compared to other linear models,

especially bagging,6but RF/OLS is still less accurate than RF,

we have evidence that both nonlinearity and variable selection

play important roles.

The second adapted model is adaLASSO/RF, where we

use adaLASSO for variable selection and then estimate a

fully grown RF with the variables selected by adaLASSO. If

adaLASSO/RF performs similarly to RF, we understand that

the variable selection in RF is less relevant and nonlinearity is

more important.

6Bagging and RF are bootstrap-based models; the former is linear, and the latter

is nonlinear.

Figure 1. Example of a regression tree. Reproduction of part of Figure 9.2 in Hastie, Tibshirami, and Friedman (2001).

8Journal of Business & Economic Statistics, Month 2019

adaLASSO/RF and RF/OLS together create an “if and only

if” situation where we test the importance of variable selection

and nonlinearity from both sides. The results indicate that

nonlinearity and variable selection are both important to explain

the performance of RF.

4.7. Computer Codes and Tuning Parameters

All ML methods are estimated in Rusing, in most

cases, standard and well-established packages. For the

linear ML methods, the following packages are used:

HDEconometrics and glmnet. The RF models are

estimated using the randomForest package. The deep

networks and boosted trees used for robustness checks

are estimated, respectively, with the h2o and xgboost

packages.7The Rcodes are available online at https://github.

com/gabrielrvsc/ForecastInﬂation.8

For all methods within the LASSO family, the penalty

parameter λis chosen by the BIC as advocated by Kock

and Callot (2015) and Medeiros and Mendes (2016). The α

parameter for the ElNet penalty in (7) is set to 0.5. We also tried

selecting it by using the BIC, but the results are quite similar,

and the computational burden is much higher. The weights

of the adaptive versions of LASSO and ElNet are chosen as

ωi=1



βi+1

√T

, where 

βiis the estimate from the nonadaptive

version of the method. The additional term 1

√Tgives variables

excluded by the initial estimator a second chance for inclusion

in the model.

The numbers of factors and lags in the factor models are

set to four. Data-driven methods for selecting the number of

factors and the lags usually yield smaller quantities, delivering

in most cases worse forecasts. For this reason, we decided

to ﬁx the number of factors and lags at four. For the target

factors, we adopt the 5% signiﬁcance level (α=0.05). For

the factor-boosting algorithm, we make the following choices.

The maximum number of iterations is set to M=10 ×

number of variables ≈5000. However, the boosting algorithm

is stopped whenever the BIC of the model starts to increase.

The shrinkage parameter is set to ν=0.2 as advocated by Bai

and Ng (2009).

The number of bootstrap replications for the bagging method

is B=100. We experimented with other values, and the results

are very stable. The pretesting procedure is conducted at the 5%

level as in Inoue and Kilian (2008). Given the large number of

variables, the pretesting was performed in two steps. First, for

each bootstrap sample, we divide the variables into 10 groups

and perform pretesting on each group. Then, selected variables

from each group are used in the next pretesting, which selects

the ﬁnal variables.

For the CSR, we set ˜n=20 and q=4. These choices are

made to avoid a huge computational burden. We tried varying

both ˜nand q, and the results do not differ much. As in the

7HDEconometrics is available at https://github.com/gabrielrvsc/

HDeconometrics/. The remaining packages are available at https://cran.

r-project.org/web/packages/.

8See also http://www.econ.puc-rio.br/mcm/codes/.

bagging algorithm and the target factors, the initial pretesting

is carried out at the 5% level.

Each individual tree in the RF model is grown until there are

only ﬁve observations in each leaf. The proportion of variables

randomly selected in each split is 1/3. This is the default setting

in the randomForest package in R. The number of bootstrap

samples, B, is set to 500. We also experimented with other

combinations of parameters, and the ﬁnal results are reasonably

stable.

5. RESULTS

The models are compared according to three different statis-

tics, namely, the root mean squared error (RMSE), the mean

absolute error (MAE), and the median absolute deviation from

the median (MAD), which are deﬁned as follows

RMSEm,h=



1

T−T0+1



t=T0

t,m,h,

MAEm,h=1

T−T0+1



t=T0

et,m,h, and

MADm,h=median 

et,m,h−median 

et,m,h,

where

et,m,h=πt−πt,m,hand πt,m,his the inﬂation forecast for

month tmade by model mwith information up to t−h.The

ﬁrst two measures above are the usual ones in the forecasting

literature. MAD, which is less commonly used, is robust to

outliers and asymmetries. Reporting both MAE and MAD in

addition to RMSE is important for conﬁrming that the results

are not due to a few large forecasting errors.

To test whether the forecasts from distinct models are

different, we consider a number of tests: the model conﬁdence

sets (MCSs) of Hansen, Lunde, and Nason (2011), the superior

predictive ability (SPA) tests of Hansen (2005), and the multi-

horizon SPA test of Quaedvlieg (2017).

5.1. Overview

Table 1 reports a number of summary statistics across all

forecasting horizons. The results concern the sample from

1990 to 2015. Additional results for the different subsamples

can be found in Tables S.10 and S.11 in the supplementary

materials. Columns (1), (2), and (3) report the average RMSE,

the average MAE, and the average MAD. Columns (4), (5), and

(6) report, respectively, the maximum RMSE, MAE, and MAD

over the forecasting horizons. Columns (7), (8), and (9) report,

respectively, the minimum RMSE, MAE, and MAD over the

15 different horizons considered. We normalize these statistics

for the benchmark RW model to one. Columns (10), (11), and

(12) report the number of times (across horizons) each model

achieved the lowest RMSE, MAE, and MAD, respectively.

Columns (13) and (14) show the average p-values of the SPA

test proposed by Hansen (2005). The SPA test of Hansen (2005)

compares a collection of models against a benchmark where the

null hypothesis is that no other model in the pool of alternatives

has superior predictive ability. In the present context, for each

forecasting horizon, we run Hansen’s SPA test by setting each

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 9

Table 1. Forecasting results: summary statistics for the out-of-sample period from 1990 to 2015

Forecasting precision Sup. pred. ability Model conﬁdence set

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) (12) (13) (14) (15) (16) (17)

ave. ave. ave. max. max. max. min. min. min. # min. # min. # min. ave. p.v. ave. p.v. ave. p.v. ave. p.v. p.v.

Model RMSE MAE MAD RMSE MAE MAD RMSE MAE MAD RMSE MAE MAD sq abs Tmax sq Tmax abs m.h. sq.

RW 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 0 0 0 0.02 0.00 0.13 0.07 0.00

AR 0.84 0.86 0.78 1.22 1.22 1.16 0.75 0.76 0.60 0 0 0 0.05 0.01 0.48 0.21 0.00

UCSV 0.82 0.82 0.85 0.95 0.91 1.04 0.77 0.78 0.76 0 0 0 0.07 0.04 0.52 0.37 0.00

LASSO 0.78 0.79 0.74 0.98 1.04 0.91 0.73 0.71 0.61 0 0 0 0.14 0.17 0.72 0.56 0.73

adaLASSO 0.78 0.78 0.76 0.96 0.96 0.99 0.72 0.71 0.63 0 0 0 0.15 0.31 0.68 0.66 0.18

ElNet 0.78 0.80 0.73 0.98 1.05 0.89 0.73 0.71 0.61 0 0 2 0.15 0.15 0.72 0.55 0.89

adaElnet 0.78 0.78 0.76 0.96 0.97 0.96 0.73 0.71 0.61 0 0 2 0.18 0.33 0.70 0.67 0.05

RR 0.76 0.77 0.79 0.89 0.93 1.03 0.70 0.71 0.67 0 0 0 0.43 0.46 0.77 0.69 0.40

BVAR 0.80 0.82 0.80 1.07 1.09 1.14 0.74 0.73 0.64 0 0 0 0.12 0.13 0.66 0.51 0.04

Bagging 0.79 0.83 0.90 0.83 0.89 1.29 0.74 0.78 0.76 0 0 0 0.20 0.06 0.63 0.39 0.04

CSR 0.82 0.82 0.80 1.13 1.11 1.09 0.76 0.74 0.67 0 0 0 0.13 0.06 0.58 0.37 0.00

JMA 0.86 0.91 1.00 0.99 0.99 1.39 0.76 0.83 0.78 0 0 0 0.06 0.01 0.29 0.08 0.00

Factor 0.84 0.87 0.87 1.17 1.21 1.25 0.78 0.78 0.71 0 0 0 0.04 0.01 0.34 0.13 0.00

T. Factor 0.83 0.88 0.89 1.17 1.23 1.26 0.77 0.80 0.70 0 0 0 0.01 0.00 0.30 0.09 0.00

B. Factor 0.83 0.90 0.99 1.17 1.32 1.60 0.74 0.75 0.73 0 0 0 0.02 0.00 0.58 0.25 0.00

RF 0.73 0.73 0.70 0.84 0.81 0.84 0.68 0.67 0.58 11 13 8 0.94 0.95 0.95 0.97 1

Mean 0.77 0.77 0.77 0.95 0.97 1.01 0.71 0.70 0.66 0 0 0 0.37 0.40 0.75 0.64 0.89

T.Mean 0.77 0.77 0.75 0.95 0.96 0.95 0.71 0.70 0.62 0 0 0 0.35 0.48 0.74 0.71 0.83

Median 0.77 0.77 0.75 0.94 0.97 0.99 0.71 0.70 0.63 0 0 0 0.32 0.44 0.73 0.70 0.89

RF/OLS 0.77 0.78 0.81 0.94 0.97 1.05 0.71 0.72 0.63 1 1 0 0.46 0.47 0.76 0.70 0.96

adaLASSO/RF 0.75 0.75 0.73 0.85 0.82 0.87 0.70 0.68 0.58 3 1 3 0.53 0.59 0.82 0.79 0.89

NOTE: The table reports for each model a number of different summary statistics across all the forecasting horizons, including the accumulated three-, six-, and twelve-month horizons. Columns (1), (2), and (3) report the average root mean square error

(RMSE), the average mean absolute error (MAE), and the average median absolute deviation (MAD). Columns (4), (5), and (6) report, respectively, the maximum RMSE, MAE, and MAD over the forecasting horizons. Columns (7), (8), and (9) report,

respectively, the minimum RMSE, MAE, and MAD over the 15 different horizons considered. Columns (10), (11), and (12) report the number of times (across horizons) each model achieved the lowest RMSE, MAE, and MAD, respectively. Columns

(13) and (14) show the average p-values of the superior predictive ability (SPA) test proposed by Hansen (2005). Columns (15) and (16) present for square and absolute losses, the average p-values for the model conﬁdence sets (MCSs) based on the tmax

statistic as described in Hansen, Lunde, and Nason (2011). Column (17) displays the p-value of the multi-horizon MCS proposed by Quaedvlieg (2017). The test is based on the squared errors only.

10 Journal of Business & Economic Statistics, Month 2019

one of the models as the benchmark. A rejection of the null

indicates that the reference model is outperformed by one or

more competitors. Columns (15) and (16) present for square

and absolute losses, respectively, the average p-values for the

MCSs based on the tmax statistic as described in Hansen,

Lunde, and Nason (2011). An MCS is a set of models that is

built such that it will contain the best model with a given level of

conﬁdence. An MCS is analogous to a conﬁdence interval for a

parameter. The MCS procedure also yields p-values for each of

the models considered. The best model has the highest p-value.

We construct MCSs for each forecasting horizon. To build a

more general MCS that contemplates all horizons, column (17)

displays the p-value of the multi-horizon MCS proposed by

Quaedvlieg (2017). The test is based on the squared errors only.

The following facts emerge from the tables: (1) ML models

and the use of a large set of predictors systematically improve

the quality of inﬂation forecasts over traditional benchmarks.

This is a robust and statistically signiﬁcant result. (2) The RF

model outperforms all the other alternatives in terms of point

statistics. The superiority of RF is due both to the variable selec-

tion mechanism induced by the method as well as the presence

of nonlinearities in the relation between inﬂation and its predic-

tors. RF has the lowest RMSEs, MAEs, and MADs across the

horizons and the highest MCS p-values. The RF model also has

the highest p-values in the SPA test and the multi-horizon MCS.

The improvements over the RW in terms of RMSE, MAE, and

MAD are almost 30% and are more pronounced during the

second subsample, when inﬂation volatility is much higher.

(3) Shrinkage methods also produce more precise forecasts

than the benchmarks. Sparsity-inducing methods are slightly

worse than nonsparsity-inducing shrinkage methods. Overall,

the forecasting performance among shrinkage methods is very

similar, and ranking them is difﬁcult. (4) Factor models are

strongly outperformed by other methods. The adoption of

boosting and target factors improves the quality of the forecasts

only marginally. The poor performance of factor models is

more pronounced during the ﬁrst subsample (low-volatility

period). (5) CSR and JMA do not perform well either and

are comparable to the factor models. (6) Forecast combination

schemes (simple average, trimmed average and median) do not

bring any signiﬁcant improvements in any of the performance

criteria considered. (7) Among the benchmark models, both AR

and UCSV outperform the RW alternative. Furthermore, the

UCSV model is slightly superior to the AR speciﬁcation.

5.2. Random Forests Versus Benchmarks

Tables 2–4show the results of the comparison between RF

and the benchmarks. Table 2 presents the RMSE, MAE, and

MAD ratios of the AR, UCSV, and RF models with respect to

the RW alternative for all forecasting horizons as well as for the

accumulated forecasts over 3, 6, and 12 months. The models

with the smallest ratios are highlighted in bold. It is clear that

the RF model has the smallest ratios for all horizons.

To check whether this is a robust ﬁnding across the out-

of-sample period, we compute rolling RMSEs, MAEs, and

MADs over windows of 48 observations. Table 3 shows the

frequency with which each model achieved the lowest RMSEs,

MAEs, and MADs as well as the frequency with which each

model was the worst-performing alternative among the four

competitors. The RF model is the winning speciﬁcation and

is superior to the competitors for the majority of periods,

including the Great Recession. By contrast, the RW model

delivers the worst forecasts most of the time. Figures S.4–S.6 in

the supplementary materials show the rolling RMSEs, MAEs,

and MADs over the out-of-sample period. The performance

of the RW deteriorates as the forecasting horizon increases.

However, the accomplishments of the RF seem rather robust.

Table 4 reports the p-values of the unconditional Giacomini

and White (2006) test for superior predictive ability for squared

and absolute errors and the multi-horizon superior predictive

ability test of Quaedvlieg (2017). The latter test compares all

horizons jointly. Rejections of the null mean that the forecasts

are signiﬁcantly different. It is clear that the RF has forecasts

that are signiﬁcantly different from and superior to the three

benchmarks.

5.3. The Full Picture

In this section, we compare all models. Table 5 presents the

results for the full out-of-sample period, whereas Tables S.12

and S.13 present the results for the 1990–2000 and 2001–

2015 periods, respectively. The tables report the RMSEs and,

in parentheses, the MAEs for all models relative to the RW

speciﬁcation. The error measures were calculated from 132

(180) rolling windows covering the 1990–2000 (2001–2015)

period. Values in bold denote the most accurate model in each

horizon. Cells in gray (blue) show the models included in the

50% MCS using the squared error (absolute error) as the loss

function. The MCSs were constructed based on the maximum

t-statistic. The last column in the table reports the number of

forecast horizons in which the model was included in the MCS

for the square (absolute) loss. The last two rows in the table

report the number of models included in the MCS for the square

and absolute losses.

We start by analyzing the full out-of-sample period.

Apart from a few short horizons, where either RF/OLS or

adaLASSO/RF is the winning model, RF delivers the smallest

ratios in most of the cases. RF is followed closely by shrinkage

models, where RR seems be superior to the other alternatives.

RR, RF, and the hybrid linear-RF models are the only ones

included in the MCS for all horizons. Neither RF nor RR

impose sparsity, which may corroborate the conclusions of

Giannone, Lenza, and Primiceri (2018), who provide evidence

against sparsity in several applications. Factor models have

very poor results and are almost never included in the MCS.

When factors are combined with boosting, there is a small

gain, but the results are still greatly inferior to those from

the RF and shrinkage models. This is particularly curious as

there is a correspondence between the factor models and RR:

RR predictions are weighted combinations of all principal

component factors of the set of predictors. Several reasons

might explain the difference. The ﬁrst potential explanation

is a lack of a clear factor structure in the regressors. This

is not the case as shown in Figure S.2 in the supplementary

materials, where we display the eigenvalues of the correlation

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 11

Table 2. Forecasting results: RMSE, MAE and MAD Ratios (1990–2015)

Panel (a): RMSE ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.90 0.81 0.79 0.81 0.79 0.79 0.78 0.76 0.78 0.82 0.84 0.75 0.86 0.97 1.22

UCSV 0.95 0.82 0.80 0.81 0.78 0.78 0.78 0.78 0.77 0.80 0.83 0.78 0.87 0.86 0.91

RF 0.84 0.73 0.71 0.74 0.71 0.72 0.72 0.71 0.72 0.76 0.77 0.68 0.71 0.71 0.77

Panel (b): MAE ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.87 0.79 0.78 0.81 0.80 0.81 0.78 0.76 0.81 0.85 0.86 0.76 0.89 1.04 1.22

UCSV 0.91 0.82 0.79 0.80 0.80 0.79 0.80 0.79 0.78 0.80 0.85 0.78 0.86 0.90 0.89

RF 0.81 0.72 0.71 0.75 0.73 0.73 0.70 0.68 0.72 0.75 0.77 0.67 0.74 0.77 0.77

Panel (c): MAD ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.74 0.70 0.82 0.82 0.83 0.75 0.66 0.68 0.77 0.70 0.77 0.60 0.87 1.16 0.89

UCSV 0.88 0.77 0.83 0.91 0.88 0.79 0.76 0.83 0.86 0.83 0.88 0.78 0.87 1.04 0.88

RF 0.70 0.63 0.77 0.84 0.75 0.73 0.65 0.64 0.73 0.69 0.71 0.58 0.71 0.80 0.59

NOTE: The table reports, for each forecasting horizon, the root mean squared error (RMSE), mean absolute error (MAE), and median absolute deviation from the median (MAD) ratios with respect to the random walk model for the full out-of-sample

period (1990–2015). The last three columns represent, respectively, the ratios for the accumulated three, six, and twelve-month forecasts. The statistics for the best-performing model are highlighted in bold.

12 Journal of Business & Economic Statistics, Month 2019

Table 3. Forecasting results: ranking of models (1990–2015)

Panel (a): Lowest rolling RMSE

Forecasting horizon

Model1234567891011123m6m12m

RW 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.04 0.00 0.00 0.06 0.08

AR 0.08 0.05 0.00 0.16 0.01 0.01 0.10 0.18 0.12 0.13 0.11 0.00 0.00 0.00 0.00

UCSV 0.02 0.05 0.21 0.10 0.18 0.11 0.00 0.03 0.19 0.11 0.09 0.00 0.21 0.31 0.24

RF 0.89 0.90 0.79 0.74 0.81 0.88 0.90 0.79 0.69 0.76 0.75 1.00 0.78 0.63 0.68

Panel (b): Lowest rolling MAE

Forecasting horizon

Model1234567891011123m6m12m

RW 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.09 0.00 0.03 0.11 0.04

AR 0.17 0.03 0.00 0.05 0.00 0.02 0.10 0.14 0.13 0.06 0.05 0.02 0.00 0.00 0.00

UCSV 0.15 0.18 0.26 0.23 0.15 0.15 0.00 0.07 0.24 0.23 0.09 0.02 0.21 0.24 0.20

RF 0.68 0.79 0.74 0.72 0.85 0.82 0.90 0.79 0.63 0.71 0.76 0.95 0.76 0.65 0.76

Panel (c): Lowest rolling MAD

Forecasting horizon

Model1234567891011123m6m12m

RW 0.14 0.00 0.03 0.00 0.09 0.05 0.00 0.00 0.03 0.05 0.03 0.05 0.02 0.15 0.02

AR 0.23 0.16 0.23 0.23 0.12 0.23 0.26 0.32 0.11 0.15 0.42 0.23 0.04 0.03 0.04

UCSV 0.04 0.19 0.27 0.33 0.09 0.12 0.03 0.01 0.22 0.09 0.03 0.02 0.10 0.05 0.05

RF 0.59 0.65 0.46 0.44 0.69 0.60 0.70 0.67 0.64 0.71 0.52 0.70 0.84 0.77 0.89

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 13

Table 3. Continued

Panel (d): Highest rolling RMSE

Forecasting horizon

Model1234567891011123m6m12m

RW 0.82 0.98 0.94 0.99 0.98 1.00 1.00 1.00 0.86 0.80 0.71 0.85 0.61 0.26 0.00

AR 0.00 0.00 0.06 0.01 0.02 0.00 0.00 0.00 0.14 0.19 0.29 0.15 0.38 0.74 0.97

UCSV 0.18 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.02 0.00 0.03

RF 0.00 0.01 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 0.00 0.00 0.00

Panel (e): Highest rolling MAE

Forecasting horizon

Model1234567891011123m6m12m

RW 0.90 0.94 0.86 1.00 0.99 0.97 1.00 0.94 0.82 0.73 0.69 0.77 0.63 0.26 0.03

AR 0.08 0.03 0.13 0.00 0.01 0.03 0.00 0.06 0.18 0.27 0.28 0.23 0.36 0.74 0.86

UCSV 0.02 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.11

RF 0.00 0.02 0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.03 0.00 0.00 0.00 0.00

Panel (f): Highest rolling MAD

Forecasting horizon

Model1234567891011123m6m12m

RW 0.68 0.94 0.85 0.80 0.80 0.66 0.81 0.94 0.77 0.80 0.79 0.92 0.68 0.30 0.31

AR 0.05 0.03 0.12 0.10 0.10 0.07 0.04 0.02 0.10 0.03 0.02 0.01 0.18 0.57 0.51

UCSV 0.22 0.03 0.03 0.10 0.04 0.26 0.14 0.04 0.11 0.14 0.19 0.07 0.12 0.13 0.17

RF 0.05 0.00 0.00 0.00 0.06 0.02 0.01 0.00 0.02 0.02 0.00 0.00 0.02 0.01 0.00

NOTE: The table reports the frequency with which each model achieved the best (worst) performance statistics over a rolling window period of four years (48 observations). The last three columns represent, respectively, the ratios for the accumulated

three, six, and twelve-month forecasts. The statistics for the model with the highest ﬁgures are highlighted in bold.

14 Journal of Business & Economic Statistics, Month 2019

Table 4. Forecasting results: superior predictive ability test (1990–2015)

Panel (a): Squared errors

Giacomini–White test—Forecasting horizon Quaedvlieg test

Model 1 2 3 4 5 6 7 8 9 10 11 12 3m 6m 12m Unif. Avg.

RW 0.00 0.00 0.00 0.00 0.01 0.01 0.01 0.00 0.00 0.03 0.02 0.00 0.00 0.05 0.05 0.00 0.00

AR 0.00 0.01 0.02 0.04 0.02 0.02 0.06 0.08 0.05 0.06 0.01 0.00 0.01 0.02 0.02 0.00 0.00

UCSV 0.00 0.00 0.01 0.06 0.06 0.02 0.00 0.00 0.00 0.04 0.00 0.00 0.04 0.06 0.07 0.00 0.00

Panel (b): Absolute errors

Giacomini–White test—Forecasting horizon Quaedvlieg test

Model 1 2 3 4 5 6 7 8 9 10 11 12 3m 6m 12m Unif. Avg.

RW 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.05 0.05 0.00 0.00

AR 0.00 0.00 0.00 0.01 0.00 0.00 0.01 0.02 0.01 0.00 0.00 0.00 0.01 0.02 0.02 0.00 0.00

UCSV 0.00 0.00 0.01 0.03 0.00 0.01 0.00 0.00 0.01 0.05 0.01 0.00 0.04 0.06 0.07 0.00 0.00

NOTE: The table reports the p-values of the unconditional Giacomini-White test for superior predictive ability between the random forest models and each of the benchmark models for each forecasting horizon as well as for the three accumulated

horizons. The test is based on the full out-of-sample period. Panel (a) presents the results for squared errors, while panel (b) shows the results for absolute errors. The GW statistics are computed with Heteroscedastic-Autocorrelation (HAC) robust

variances with the quadratic spectral kernel and bandwidth selected by Andrew’s automatic method. The table also reports the p-values of the uniform and average multi-horizon superior predictive ability test proposed by Quaedvlieg (2017).

matrix of regressors over the forecasting period. As shown

in the ﬁgure, there is a small number of dominating factors.

Second, there might be factors that explain only a small portion

of the total variance of the regressors but have high predictive

power on inﬂation. Again, we do not think this is the case,

as target factors as well as boosting are speciﬁcally designed

to enhance the quality of the predictions but do not bring any

visible improvement in this case. Furthermore, we allow the

ML methods to select factors as well, and as shown below,

they are never selected. Finally, we believe the most probable

explanation is that although sparsity can be questioned, factor

models are an overly aggregated representation of the potential

predictors. The results of JMA are not encouraging either.

All of the competing models outperform RW for almost all

horizons. Finally, the forecast combination does not provide

any signiﬁcant gain due to the empirical fact that most of the

forecasts are positively correlated; see Figure S.3.

To check whether this is a robust ﬁnding across the out-of-

sample period, we compute rolling RMSEs, MAEs, and MADs

over windows of 48 observations as shown in Figures S.7–

S.18 in the supplementary materials. The results corroborate

the superiority of the RF model, particularly for long as well as

aggregated horizons.

Focusing now on the two subsamples, the following conclu-

sions stand out from the tables in the supplementary materials.

The superiority of RF is more pronounced during the 2001–

2015 period, when inﬂation is much more volatile. During this

period, RF achieves the smallest RMSE and MAE ratios for

almost all horizons. From 1990 to 2000, the linear shrinkage

methods slightly outperform RF for short horizons. However,

RF dominates for long horizons and for the twelve-month

forecasts. Among the shrinkage models and during the ﬁrst

period, there is no clear evidence of a single winner. Depending

on the horizon, different models perform the best. Another

salient fact is that there are fewer models included in the MSC

during the ﬁrst subsample.

Finally, we test whether the superiority of the RF model

depends on the state of the economy. We consider two cases,

namely, recessions versus expansions according to the NBER

classiﬁcation and high versus low macroeconomic uncertainty.9

We consider the following regressions



t+h,RF −

t+h,other model =α0It+h+α1(1−It+h)+error, (8)

where 

t+h,RF is the squared forecasting error of the RF for

horizon h,

t+h,other model is the squared forecasting error of the

competing model, and Itis an indicator function that equals one

for periods of recession (or high macroeconomic uncertainty).

The results are presented in Tables S.14 and S.15 in the supple-

mentary materials for periods of expansion versus recession and

high versus low macroeconomic uncertainty, respectively. The

9Periods of high (low) macroeconomic uncertainty are those where uncertainty

is higher (lower) than the historical average. Since the results barely change if

we consider either ﬁnancial or real, rather than macroeconomic, uncertainty,

we do not report them for brevity. They are available upon request. These

measures of macroeconomic, ﬁnancial and real uncertainty are computed

as in Jurado, Ludvigson, and Ng (2015) and are the conditional volatility

of the unforecastable part of macroeconomic, ﬁnancial and ﬁrm-level vari-

ables, respectively. They are available at Sydney C. Ludvigson’s webpage

(https://www.sydneyludvigson.com/).

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 15

Table 5. Forecasting errors for the CPI from 1990 to 2015

Consumer price index 1990–2015

Forecasting horizon

RMSE/(MAE)1234567891011123m6m12m

RMSE count

(MAE count)

RW 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1.00 1

(1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (1.00) (0)

AR 0.90 0.81 0.79 0.81 0.79 0.79 0.78 0.76 0.78 0.82 0.84 0.75 0.86 0.97 1.22 8

(0.87) (0.79) (0.78) (0.81) (0.80) (0.81) (0.78) (0.76) (0.81) (0.85) (0.86) (0.76) (0.89) (1.04) (1.22) (0)

UCSV 0.95 0.82 0.80 0.81 0.78 0.78 0.78 0.78 0.77 0.80 0.83 0.78 0.87 0.86 0.91 9

(0.91) (0.82) (0.79) (0.80) (0.80) (0.79) (0.80) (0.79) (0.78) (0.80) (0.85) (0.78) (0.86) (0.90) (0.89) (4)

LASSO 0.83 0.75 0.73 0.76 0.74 0.75 0.75 0.73 0.75 0.80 0.82 0.73 0.75 0.79 0.98 13

(0.82) (0.74) (0.73) (0.78) (0.77) (0.75) (0.74) (0.71) (0.76) (0.81) (0.84) (0.74) (0.79) (0.89) (1.04) (11)

adaLASSO 0.84 0.76 0.74 0.77 0.75 0.75 0.76 0.75 0.76 0.80 0.83 0.72 0.76 0.80 0.96 13

(0.81) (0.75) (0.72) (0.77) (0.75) (0.74) (0.73) (0.71) (0.75) (0.79) (0.84) (0.73) (0.79) (0.86) (0.96) (13)

ElNet 0.83 0.75 0.73 0.76 0.75 0.74 0.75 0.74 0.76 0.81 0.82 0.73 0.75 0.79 0.98 13

(0.82) (0.74) (0.73) (0.78) (0.78) (0.76) (0.75) (0.71) (0.77) (0.81) (0.85) (0.75) (0.78) (0.89) (1.05) (11)

adaElnet 0.84 0.75 0.73 0.77 0.75 0.75 0.75 0.74 0.76 0.80 0.81 0.73 0.76 0.81 0.96 13

(0.82) (0.74) (0.72) (0.76) (0.75) (0.74) (0.73) (0.71) (0.75) (0.79) (0.83) (0.75) (0.79) (0.86) (0.97) (13)

RR 0.85 0.73 0.72 0.75 0.74 0.75 0.75 0.73 0.74 0.77 0.78 0.70 0.73 0.77 0.89 14

(0.83) (0.72) (0.72) (0.77) (0.76) (0.76) (0.73) (0.71) (0.74) (0.77) (0.79) (0.71) (0.77) (0.86) (0.93) (14)

BVAR 0.86 0.76 0.75 0.77 0.74 0.76 0.77 0.76 0.77 0.82 0.83 0.74 0.79 0.85 1.07 12

(0.87) (0.73) (0.75) (0.79) (0.78) (0.78) (0.76) (0.76) (0.81) (0.83) (0.85) (0.76) (0.82) (0.93) (1.09) (9)

Bagging 0.83 0.76 0.76 0.80 0.78 0.79 0.83 0.81 0.78 0.82 0.83 0.74 0.74 0.76 0.82 12

(0.84) (0.78) (0.79) (0.87) (0.86) (0.85) (0.83) (0.80) (0.80) (0.84) (0.86) (0.78) (0.79) (0.89) (0.88) (7)

CSR 0.85 0.77 0.76 0.79 0.77 0.79 0.79 0.77 0.79 0.83 0.84 0.76 0.79 0.87 1.13 10

(0.84) (0.76) (0.75) (0.79) (0.79) (0.79) (0.76) (0.74) (0.79) (0.83) (0.84) (0.77) (0.82) (0.94) (1.11) (4)

JMA 0.99 0.82 0.84 0.85 0.84 0.81 0.91 0.86 0.84 0.95 0.92 0.80 0.76 0.80 0.88 4

(0.99) (0.85) (0.89) (0.94) (0.96) (0.90) (0.91) (0.87) (0.93) (0.96) (0.96) (0.83) (0.86) (0.96) (0.91) (0)

Factor 0.87 0.78 0.78 0.79 0.78 0.78 0.80 0.81 0.82 0.84 0.84 0.78 0.82 0.90 1.17 4

(0.88) (0.80) (0.80) (0.82) (0.82) (0.80) (0.78) (0.80) (0.87) (0.87) (0.87) (0.82) (0.89) (1.02) (1.21) (1)

T. Factor 0.88 0.79 0.78 0.80 0.77 0.79 0.79 0.80 0.80 0.82 0.83 0.78 0.82 0.91 1.17 3

(0.87) (0.82) (0.81) (0.84) (0.83) (0.84) (0.80) (0.80) (0.84) (0.87) (0.86) (0.80) (0.90) (1.06) (1.23) (0)

16 Journal of Business & Economic Statistics, Month 2019

Table 5. Continued

Consumer price index 1990–2015

Forecasting horizon

RMSE/(MAE)1234567891011123m6m12m

RMSE count

(MAE count)

B. Factor 0.95 0.77 0.76 0.78 0.77 0.79 0.79 0.78 0.79 0.83 0.84 0.74 0.82 0.91 1.17 10

(0.96) (0.80) (0.81) (0.85) (0.84) (0.86) (0.84) (0.82) (0.85) (0.86) (0.86) (0.75) (0.92) (1.13) (1.32) (1)

RF 0.84 0.73 0.71 0.74 0.71 0.72 0.72 0.71 0.72 0.76 0.77 0.68 0.71 0.71 0.77 15

(0.81) (0.72) (0.71) (0.75) (0.73) (0.73) (0.70) (0.68) (0.72) (0.75) (0.77) (0.67) (0.74) (0.77) (0.77) (15)

Mean 0.83 0.75 0.73 0.76 0.74 0.74 0.75 0.74 0.75 0.77 0.78 0.71 0.75 0.80 0.95 14

(0.81) (0.74) (0.73) (0.76) (0.76) (0.75) (0.73) (0.71) (0.75) (0.76) (0.78) (0.70) (0.78) (0.87) (0.97) (14)

T.M ea n 0.84 0.74 0.73 0.75 0.74 0.74 0.75 0.73 0.74 0.78 0.79 0.71 0.75 0.79 0.95 14

(0.81) (0.74) (0.72) (0.76) (0.75) (0.74) (0.72) (0.70) (0.74) (0.77) (0.79) (0.70) (0.78) (0.86) (0.96) (14)

Median 0.84 0.75 0.72 0.76 0.74 0.74 0.75 0.73 0.74 0.78 0.79 0.71 0.75 0.79 0.94 14

(0.81) (0.74) (0.72) (0.76) (0.76) (0.74) (0.73) (0.70) (0.74) (0.77) (0.79) (0.71) (0.78) (0.86) (0.97) (14)

RF/OLS 0.81 0.73 0.72 0.75 0.74 0.75 0.75 0.74 0.74 0.78 0.79 0.71 0.73 0.78 0.94 14

(0.79) (0.73) (0.72) (0.76) (0.76) (0.76) (0.73) (0.72) (0.75) (0.78) (0.81) (0.72) (0.78) (0.86) (0.97) (14)

adaLASSO/RF 0.85 0.76 0.72 0.73 0.73 0.72 0.72 0.71 0.72 0.79 0.82 0.70 0.73 0.73 0.80 15

(0.82) (0.73) (0.72) (0.74) (0.74) (0.73) (0.71) (0.68) (0.72) (0.79) (0.82) (0.68) (0.76) (0.80) (0.82) (15)

RMSE count 14 15 13 17 19 19 17 15 17 18 19 7 13 17 5

MAE count (12) (13) (12) (11) (10) (12) (14) (12) (10) (15) (16) (7) (13) (13) (4)

NOTE: The table shows the root mean squared error (RMSE) and, between parenthesis, the mean absolute errors (MAE) for all models relative to the random walk (RW). The error measures were calculated from 132 rolling windows covering the

1990–2000 period and 180 rolling windows covering the 2001–2015 period. Values in bold show the most accurate model in each horizon. Cells in gray (blue) show the models included in the 50% model conﬁdence set (MCS) using the squared error

(absolute error) as loss function. The MCSs were constructed based on the maximum t-statistic. The last column in the table reports in how many horizons the row model was included in the MCS for square (absolute) loss. The last two rows in the table

report how many models were included in the MCS for square and absolute losses.

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 17

tables report the estimated values of α0and α1in Equation (8)

as the respective standard errors. For conciseness, we display

only the results for the most relevant models.

Inspecting the tables, it is clear that the majority of the

statistics are negative, which indicates that the RF model is

superior to its competitors. Of the 72 entries in each table,

the values of the statistics are positive only in four and seven

cases in Tables S.14 and S.15, respectively. However, the

differences are not statistically signiﬁcant during recessions.

This result is not surprising, as only 34 of the 312 out-of-sample

observations are labeled as recessions. Nevertheless, the mag-

nitudes of the differences are much higher during recessions.

Turning attention to periods of low and high macroeconomic

uncertainty, it is evident that the RF model is statistically

superior to the benchmarks and the differences are higher

in periods of greater uncertainty. The gains from using RF

are particularly large during and after the Great Recession

(see Figures S.4–S.18). As argued above, both the degrees of

slackness and uncertainty might be sources of nonlinearities

in the economy. The fact that the RF model outperforms

competitors in these states of the economy suggests that allow-

ing for nonlinearities is key to improving macroeconomic

forecasts.

5.4. Real-Time Experiment

To test the robustness of our results, we also conduct an

experiment with the real-time database provided by FRED-MD.

Due to data limitations, we compute the real-time forecasts only

for the second subsample considered here. We compare the fol-

lowing models against the three benchmarks: RR, adaLASSO,

RF, RF/OLS, and adaLASSO/RF. Table 6 reports, for each

forecasting horizon, the RMSE, MAE, and MAD ratios with

respect to the RW model. The following conclusions emerge

from the table. The ML methods clearly outperform the three

benchmarks. Furthermore, for the three accumulated horizons

considered, RF is the best model. For the monthly forecasts,

both RF and adaLASSO/RF achieve the best performance for

most of the horizons. For one-month-ahead, RR seems to be

the best model. The results are robust to the choice between the

RMSE and MAE criteria. For MAD, RR is also competitive.

These results corroborate our conclusions that ML methods

should be considered seriously to forecast inﬂation.

5.5. Opening the Black Box: Variable Selection

We compare the predictors selected by some of the ML

methods: adaLASSO, RR, and RF. We select these three mod-

els for two reasons. First, they are generally the three best-

performing models. Second, they have quite different charac-

teristics. While adaLASSO is a true sparsity-inducing method,

RR and RF models are only approximately sparse. In addition,

RR is a linear model, whereas RF is a highly nonlinear speciﬁ-

cation.

This analysis is straightforward for sparsity-inducing shrink-

age methods such as adaLASSO, as the coefﬁcients of poten-

tially irrelevant variables are automatically set to zero.10 For

the other ML methods, the analysis is more complex. To

ensure that the results among models are comparable, we adopt

the following strategy. For RR and adaLASSO, the relative

importance measure is computed as the average coefﬁcient

size (multiplied by the respective standard deviations of the

regressors). To measure the importance of each variable for the

RF models, we use out-of-bag (OOB) samples.11 When the bth

tree is grown, the OOB samples are passed down the tree, and

the prediction accuracy is recorded. Then, the values of the jth

variable are randomly permuted in the OOB sample, and the

accuracy is again computed. The decrease in accuracy due to

the permutation is averaged over all trees and is the measure of

the importance of the jth variable in RF.

Due to space constraints, we cannot show the relative impor-

tance for each variable, each lag, each horizon, or each esti-

mation window. Therefore, as described in the supplementary

materials and following McCracken and Ng (2016), we cat-

egorize the variables, including lags, into the following eight

groups: (i) output and income; (ii) labor market; (iii) housing;

(iv) consumption, orders, and inventories; (v) money and credit;

(vi) interest and exchange rates; (vii) prices; and (viii) stock

market. We also consider two additional groups, namely, the

principal component factors computed from the full set of

potential predictors and autoregressive terms. Furthermore, the

results are averaged across all estimation windows.

Figure 2 shows the importance of each variable group for the

adaLASSO, RR, and RF methods for each horizon. The values

in the plots are rescaled to sum to one.

The set of the most relevant variables for the RR and

RF models is quite stable across forecasting horizons but is

remarkably different between them. While for RR, AR terms,

prices and employment are the most important predictors, RF

models give more importance to prices, interest-exchange rates,

employment and housing. For adaLASSO, selection is quite

different across forecasting horizons, and only AR terms retain

their relative importance independent of the horizon. Other

prices gradually lose their relevance until up to six months

ahead and partially recover relevance when longer horizons are

considered. Output-income variables are more important for

medium-term forecasts. Finally, none of the three classes of

models selects either factors or stocks. This result may indicate

that the high level of cross-section aggregation of the factors is

responsible for the poor performance of factor models.

To compare the degree of sparsity between adaLASSO and

RF, we report word clouds of the selected variables in the

supplementary materials. A word cloud is an image composed

of the names of variables selected by a speciﬁc model across the

estimation windows; in the word cloud, the size of each word

indicates its frequency or importance. The names displayed in

the clouds are as deﬁned in the third columns of Tables S.1–S.8.

10Medeiros and Mendes (2016) showed, for example, that under sparsity

conditions, the adaLASSO model selection is consistent for high-dimensional

time series models in very general settings, that is, the method correctly selects

the “true” set of regressors.

11For a given data point (yt,x

t), the OOB sample is the collection of all

bootstrap samples that do not include (yt,x

t).

18 Journal of Business & Economic Statistics, Month 2019

Table 6. Real-time forecasting results: RMSE, MAE, and MAD ratios

Panel (a): RMSE ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.91 0.81 0.77 0.78 0.79 0.80 0.80 0.77 0.79 0.82 0.83 0.77 0.84 0.95 1.21

UCSV 0.97 0.82 0.79 0.80 0.78 0.79 0.81 0.80 0.79 0.81 0.81 0.77 0.86 0.88 0.87

RR 0.84 0.74 0.72 0.75 0.76 0.78 0.79 0.76 0.77 0.79 0.80 0.74 0.72 0.80 0.96

adaLASSO 0.93 0.92 1.08 1.01 1.56 0.96 0.96 0.92 0.88 0.93 0.94 0.87 0.86 1.03 1.11

RF 0.88 0.74 0.70 0.72 0.72 0.75 0.76 0.74 0.74 0.79 0.80 0.72 0.69 0.64 0.65

RF/OLS 0.84 0.75 0.72 0.75 0.76 0.78 0.78 0.76 0.76 0.78 0.79 0.73 0.73 0.79 0.95

adaLASSO/RF 0.85 0.73 0.70 0.71 0.74 0.73 0.75 0.72 0.74 0.77 0.78 0.72 0.70 0.72 0.74

Panel (b): MAE ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.87 0.80 0.73 0.75 0.80 0.81 0.79 0.75 0.77 0.80 0.84 0.78 0.83 0.98 1.19

UCSV 0.91 0.82 0.77 0.77 0.79 0.80 0.81 0.80 0.79 0.80 0.83 0.78 0.85 0.91 0.86

RR 0.81 0.72 0.68 0.72 0.77 0.80 0.77 0.73 0.76 0.76 0.82 0.74 0.72 0.81 0.94

adaLASSO 0.88 0.91 0.89 0.98 1.14 1.00 1.00 0.94 0.88 0.92 1.01 0.92 0.87 1.11 1.09

RF 0.83 0.72 0.66 0.72 0.74 0.76 0.75 0.72 0.72 0.75 0.81 0.73 0.68 0.65 0.63

RF/OLS 0.82 0.73 0.69 0.74 0.77 0.80 0.79 0.74 0.75 0.76 0.80 0.74 0.74 0.80 0.90

adaLASSO/RF 0.81 0.71 0.65 0.68 0.73 0.74 0.75 0.70 0.72 0.75 0.80 0.74 0.69 0.71 0.71

Panel (c): MAD ratio

Forecasting horizon

Model1234567891011123m6m12m

AR 0.80 0.78 0.66 0.80 0.76 0.76 0.73 0.74 0.81 0.77 0.83 0.73 0.83 1.13 0.95

UCSV 0.86 0.82 0.82 0.81 0.80 0.76 0.90 0.86 0.84 0.75 0.92 0.70 0.93 1.03 0.88

RR 0.77 0.70 0.60 0.76 0.74 0.75 0.75 0.72 0.77 0.69 0.75 0.67 0.72 0.90 0.75

adaLASSO 0.75 0.94 0.86 1.01 1.07 1.04 1.09 1.05 1.05 0.90 1.08 0.94 0.85 1.22 0.97

RF 0.74 0.72 0.62 0.73 0.75 0.77 0.75 0.73 0.76 0.65 0.81 0.68 0.72 0.73 0.53

RF/OLS 0.77 0.71 0.66 0.76 0.77 0.78 0.81 0.77 0.83 0.72 0.85 0.69 0.76 0.84 0.77

adaLASSO/RF 0.70 0.68 0.60 0.66 0.74 0.73 0.78 0.64 0.71 0.76 0.81 0.67 0.69 0.80 0.56

The last three columns represent, respectively, the ratios for the accumulated three, six, and twelve-month forecasts. The statistics for the best-performing model are highlighted in bold. The forecasts are computed in real-time.

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 19

Figure 2. Variable importance. The picture shows the importance of each variable group for the adaLASSO, RR, and RF methods for all

the twelve forecasting horizons. For all different methods, the values in the plots are rescaled to sum one. For RR and adaLASSO, the relative

importance measure is computed as the average coefﬁcient size (multiplied by the respective standard deviations of the regressors). To measure

the importance of each variable for the RF models, we use out-of-bag (OOB) samples. When the bth tree is grown, the OOB samples are passed

down the tree and the prediction accuracy is recorded. Then, the values of the jth variable are randomly permuted in the OOB sample, and the

accuracy is again computed. The decrease in accuracy due to the permutation is averaged over all trees and is the measure of the importance of

the jth variable in the RF.

It is evident that the RF models are much less sparse than

adaLASSO.

RR, adaLASSO, and RF select different variables, which

suggests nontrivial interactions among variable selection, spar-

sity, and nonlinearity. If the econometrician is only interested

in forecasting, variable selection is less of a concern. One

should select (a combination of) the best-performing models. If

she/he is also interested in the precise mechanisms underlying

price dynamics, careful identiﬁcation schemes should be con-

sidered, which is beyond the scope of this article. Nonetheless,

compared to adaLASSO and RR, the best-performing method,

RF, uses disaggregated inﬂation as a substitute for lags of

CPI inﬂation, thus resembling an unusual Phillips curve with

heterogeneous backward-looking terms. This ﬁnding may shed

light on price dynamics and can be useful for future research

aimed at uncovering such mechanisms.

5.6. Opening the Black Box: Nonlinearities

The role of nonlinearity in the relative performance

of the RF model is highlighted in the plots of recursive

RMSEs, MAEs, and MADs in Figures S.7–S.18. First, RF

is compared with the adaLASSO/RF and RF/OLS models

in Figures S.16–S.18. For h=1, the performances of

the three models are almost identical, indicating that there

are no beneﬁts in introducing nonlinearity for very short-

term forecasts. On the other hand, the superiority of the

RF model becomes more evident for longer horizons and

for the accumulated inﬂation over six and twelve months.

Furthermore, the results do not seem to be due to a few

outliers, as both the rolling MAEs and MADs, which are less

sensitive to extreme observations, conﬁrm the overperformance

of the RF model. These ﬁndings are also corroborated when

20 Journal of Business & Economic Statistics, Month 2019

Table 7. Forecasting results (alternative models)

Panel (a): RMSE ratio

Forecasting horizon

Model 1 2 3 4 5 6 7 8 9 10 11 12 3m 6m 12m

RF 0.84 0.73 0.71 0.74 0.71 0.72 0.72 0.71 0.72 0.76 0.77 0.68 0.71 0.71 0.77

SCAD 0.85 0.76 0.77 0.79 0.77 0.77 0.81 0.77 0.78 0.83 0.84 0.73 0.75 0.79 0.96

BTrees 0.88 0.78 0.74 0.76 0.73 0.74 0.74 0.73 0.73 0.77 0.79 0.71 0.74 0.70 0.71

Deep NN 1.00 0.84 0.78 0.80 0.78 0.85 0.79 0.77 0.78 0.84 0.81 0.75 0.81 0.84 0.90

LASSO 0.89 0.76 0.74 0.76 0.74 0.72 0.74 0.75 0.75 0.79 0.83 0.74 0.79 0.83 1.04

adaLASSO 0.91 0.73 0.71 0.73 0.71 0.70 0.72 0.73 0.74 0.76 0.80 0.69 0.74 0.73 0.80

Panel (b): MAE ratio

Forecasting horizon

Model 1 2 3 4 5 6 7 8 9 10 11 12 3m 6m 12m

RF 0.81 0.72 0.71 0.75 0.73 0.73 0.70 0.68 0.72 0.75 0.77 0.67 0.74 0.77 0.77

SCAD 0.84 0.77 0.78 0.81 0.81 0.78 0.80 0.74 0.79 0.84 0.87 0.76 0.80 0.87 0.99

BTrees 0.84 0.75 0.74 0.78 0.76 0.77 0.74 0.70 0.73 0.75 0.78 0.70 0.77 0.75 0.71

Deep NN 0.97 0.84 0.83 0.85 0.84 0.87 0.79 0.79 0.83 0.85 0.84 0.79 0.83 0.93 0.91

LASSO 0.94 0.78 0.75 0.79 0.79 0.74 0.74 0.76 0.80 0.82 0.89 0.78 0.88 1.01 1.21

adaLASSO 0.86 0.71 0.71 0.75 0.73 0.71 0.69 0.71 0.75 0.76 0.82 0.69 0.75 0.80 0.85

Panel (c): MAD ratio

Forecasting horizon

Model 1 2 3 4 5 6 7 8 9 10 11 12 3m 6m 12m

RF 0.70 0.63 0.77 0.84 0.75 0.73 0.65 0.64 0.73 0.69 0.71 0.58 0.71 0.80 0.59

SCAD 0.80 0.75 0.79 0.83 0.89 0.77 0.71 0.69 0.80 0.78 0.84 0.70 0.82 1.02 0.81

BTrees 0.73 0.70 0.79 0.86 0.80 0.78 0.69 0.71 0.76 0.68 0.77 0.65 0.79 0.88 0.61

Deep NN 0.85 0.73 0.90 0.97 0.94 0.83 0.78 0.86 0.95 0.83 0.87 0.83 0.82 1.15 0.87

LASSO 0.83 0.68 0.70 0.80 0.73 0.68 0.66 0.64 0.76 0.73 0.73 0.60 0.77 0.88 0.65

adaLASSO 0.81 0.65 0.78 0.81 0.81 0.76 0.63 0.63 0.81 0.78 0.82 0.66 0.75 0.90 0.76

The table reports, for each forecasting horizon, the root mean squared error (RMSE), mean absolute error (MAE), and median absolute deviation from the median (MAD) ratios with respect to the random walk model for the full out-of-sample period

(1990–2015). The last three columns represent, respectively, the ratios for the accumulated three, six, and twelve-month forecasts. The statistics for the best-performing model are highlighted in bold.

Medeiros et al.: Forecasting Inﬂation in a Data-Rich Environment: The Beneﬁts of Machine Learning Methods 21

RF is compared with the other linear models; see Figures

S.7–S.15.

Although is clear that nonlinearity is present in the dynamics

of inﬂation, it is very difﬁcult to completely uncover the

nonlinear mechanism driving the forecasts, as in our experiment

we estimated 12 different RF models (one for each horizon) for

each rolling window. As we have 312 windows, there are total

of 3744 different models to analyze.

5.7. Robustness: Alternative Nonlinear Models and

Other Penalties

As a robustness check, we compute forecasts from other

nonlinear alternatives. The ﬁrst is a boosted regression tree

(BTree) as in Friedman (2001). The second one is based

on the results of Gu, Kelly, and Xiu (2018) and is a Deep

Neural Network (DeepNN) with three hidden layers and 32,

16, and 8 rectiﬁed linear units (ReLU) in each layer, respec-

tively. The DeepNN model is estimated by the stochastic

gradient descent algorithm. We also consider two second-order

polynomial models with interaction terms estimated either by

LASSO or adaLASSO. Finally, in addition to these nonlinear

alternatives, we estimate a linear model with a SCAD penalty.

Table 7 reports, for each forecasting horizon, the RMSE, the

MAE, and the MAD ratios with respect to the RW model for

the full out-of-sample period (1990–2015). In terms of RMSE,

RF is the best model in 11 of 15 cases. Among these, in

ﬁve cases there is a tie with the polynomial model estimated

with adaLASSO. In general, the DeepNN model has the worst

performance. SCAD is also outperformed by the RF model.

The results for MAE are quite similar. On the other hand, there

are some differences when MAD is considered. Although the

RF model is still the best alternative, the polynomial model

estimated with LASSO overperforms the one estimated with

adaLASSO.

These results corroborate the superiority of RF. Furthermore,

the fact that the polynomial models with interactive terms are

competitive sheds some light on the nature of the nonlinearity

captured by the tree-based models. We should bear in mind

that regression trees are models that explore interactions among

covariates.

6. CONCLUSIONS

We show that with the recent advances in ML methods

and the availability of new and rich datasets, it is possible

to improve inﬂation forecasts. Models such as LASSO, RF,

and others are able to produce more accurate forecasts than

the standard benchmarks. These results highlight the beneﬁts

of ML methods and big data for macroeconomic forecasting.

Although our article focuses on inﬂation forecasting in the US,

one can easily apply ML methods to forecast other macroeco-

nomic series in a variety of countries.

The RF model deserves special attention, as it robustly

delivers the smallest errors. Its good performance is due to both

potential nonlinearities and its variable selection mechanism.

The selection of variables for RF models is stable across

horizons. These variables are mostly selected from the follow-

ing groups: prices, exchange and interest rates, and the housing

and labor markets. Although it is difﬁcult to disentangle the

precise sources of the nonlinearities that the RF model uncov-

ers, the variable selection may shed light on them. In fact, there

are many theoretical arguments justifying nonlinear relation-

ships among inﬂation, interest rate, labor market outcomes,

and housing. For example, the relationship between inﬂation

and employment depends on the degree of slackness in the

economy. Uncertainty might also induce nonlinearities. Finally,

part of the out-of-sample window encompasses quarters when

the zero lower bound on nominal interest rates is binding, which

is another source of nonlinearity. This out-of-sample window

also encompasses a period in which a housing bubble led to a

credit crunch, events with highly nonlinear consequences.

RF is also the winning method in the periods of expansion

and recession as well as in the periods of low uncertainty and

high uncertainty. The gains from using RF are larger in periods

of recession and high uncertainty. RF also outperforms other

methods during and after the Great Recession, when uncer-

tainty skyrocketed and when the zero lower bound was binding.

Taken together, these results suggest that the relationships

among key macroeconomic variables can be highly nonlinear. If

this is the case, the linear methods applied in the profession not

only to forecast variables but also to achieve other objectives

such as approximate DSGE models might lead to inaccurate

results, especially during bad states of the economy.

SUPPLEMENTARY MATERIALS

The supporting material consists of a complete description on the data and

required transformations, descriptive statistics, more detailed results for the

CPI inﬂation as well as results concerning the PCE and the CPI core inﬂation

measures.

ACKNOWLEDGMENTS

We are very grateful to the coeditor, Christian B. Hansen; the associate

editor; and two anonymous referees for very helpful comments. We thank

Federico Bandi, Anders B. Kock, and Michael Wolf for comments during

the workshop “Trends in Econometrics: Machine Learning, Big Data, and

Financial Econometrics,” Rio de Janeiro, October 2017. The participants

of the “Third International Workshop in Financial Econometrics,” Arraial

d’Ajuda, October 2017, the workshop “Big Data, Machine Learning and

the Macroeconomy,” Oslo, October 2017, and the National Symposium on

Probability and Statistics (SINAPE), S˜

ao Pedro, September 2018, are also

gratefully acknowledged. We also thank Alberto Cavallo for an insightful

discussion during the workshop “Measuring and Analyzing the Economy using

Big Data” at the Central Bank of Brazil, November 2017. We also acknowledge

the suggestions received during seminars at the Universit`

a Ca’Foscari, Venice;

at the School of Applied Mathematics at FGV/RJ; and at the Department

of Economics at Duke University. Special thanks also to Tim Bollerslev, Jia

Li, Andrew Patton, Rogier Quaedvlieg, Peter R. Hansen, Lucciano Villacorta,

Mark Watson, and Brian Weller for inspiring discussions. Finally, we are

grateful to Michael McCracken for his help with the real-time dataset. The

views expressed herein do not necessarily reﬂect the position of the Central

Bank of Chile.

FUNDING

The research of the ﬁrst author is partially supported by CNPq and FAPERJ.

[Received November 2018. Revised April 2019.]

22 Journal of Business & Economic Statistics, Month 2019

REFERENCES

Atkeson, A., and Ohanian, L. (2001), “Are Phillips Curves Useful for Forecast-

ing Inﬂation?,” Federal Reserve Bank of Minneapolis Quarterly Review,

25, 2–11. [1,2,3]

Bai, J., and Ng, S. (2003), “Inferential Theory for Factor Models of Large

Dimensions,” Econometrica, 71, 135–171. [5]

——— (2008), “Forecasting Economic Time Series Using Targeted Predic-

tors,” Journal of Econometrics, 146, 304–317. [2,5,6]

——— (2009), “Boosting Diffusion Indexes,” Journal of Applied Economet-

rics, 24, 607–629. [2,5,8]

Ba´

nbura, M., Giannone, D., and Reichlin, L. (2010), “Large Bayesian Vector

Autoregressions,” Journal of Applied Econometrics, 25, 71–92. [4]

Bloom, N. (2009), “The Impact of Uncertainty Shocks,” Econometrica,

77, 623–685. [3]

Breiman, L. (1996), “Bagging Predictors,” Machine Learning, 24, 123–140.

[6,7]

——— (2001), “Random Forests,” Machine Learning, 45, 5–32. [2,7]

Chakraborty, C., and Joseph, A. (2017), “Machine Learning at Central Banks,”

Working Paper 674, Bank of England Staff Working Paper. [2]

Chen, Y.-C., Turnovsky, S., and Zivot, E. (2014), “Forecasting Inﬂation Using

Commodity Price Aggregates,” Journal of Econometrics, 183, 117–134. [3]

Dellas, H., Gibson, H., Hall, S., and Tavlas, G. (2018), “The Macroeconomic

and Fiscal Implications of Inﬂation Forecast Errors,” Journal of Economic

Dynamics and Control, 93, 203 –217. [2]

Eggertsson, G. B., and Woodford, M. (2003), “Zero Bound on Interest Rates

and Optimal Monetary Policy,” Brookings Papers on Economic Activity,

2003, 139–233. [3]

Elliott, G., Gargano, A., and Timmermann, A. (2013), “Complete Subset

Regressions,” Journal of Econometrics, 177, 357–373. [3,6]

——— (2015), “Complete Subset Regressions With Large-Dimensional Sets

of Predictors,” Journal of Economic Dynamics and Control, 54, 86–110.

[6]

Fan, J., and Li, R. (2001), “Variable Selection via Nonconcave Penalized

Likelihood and Its Oracle Properties,” Journal of the American Statistical

Association, 96, 1348–1360. [2]

Faust, J., and Wright, J. (2013), “Forecasting Inﬂation,” in Handbook of

Economic Forecasting (Vol. 2A), eds. G. Elliott and A. Timmermann,

Amsterdam: Elsevier. [1]

Forni, M., Hallin, M., Lippi, M., and Reichlin, L. (2003), “Do Financial

Variables Help Forecasting Inﬂation and Real Activity in the Euro Area?,”

Journal of Monetary Economics, 50, 1243–1255. [3]

Friedman, J. (2001), “Greedy Function Approximation: A Gradient Boosting

Machine,” The Annals of Statistics, 29, 1189–1232. [21]

Garcia, M., Medeiros, M., and Vasconcelos, G. (2017), “Real-Time Inﬂation

Forecasting With High-Dimensional Models: The Case of Brazil,” Interna-

tional Journal of Forecasting, 33, 679–693. [3]

Giacomini, R., and White, H. (2006), “Tests of Conditional Predictive Ability,”

Econometrica, 74, 1545–1578. [4,10]

Giannone, D., Lenza, M., and Primiceri, G. (2018), “Economic Predictions

With Big Data: The Illusion of Sparsity,” Working Paper, Northwestern

University. [2,10]

Groen, J., Paap, R., and Ravazzolo, F. (2013), “Real-Time Inﬂation Forecasting

in a Changing World,” Journal of Business and Economic Statistics, 31, 29–

44. [3]

Gu, S., Kelly, B., and Xiu, D. (2018), “Empirical Asset Pricing With Machine

Learning,” Working Paper, University of Chicago. [1,21]

Hall, A. S. (2018), “Machine Learning Approaches to Macroeconomic Fore-

casting,” The Federal Reserve Bank of Kansas City Economic Review, 103,

63. [2]

Hansen, B. E., and Racine, J. S. (2012), “Jackknife Model Averaging,” Journal

of Econometrics, 167, 38–46. [6]

Hansen, P. (2005), “A Test for Superior Predictive Ability,” Journal of Business

and Economic Statistics, 23, 365–380. [8,9]

Hansen, P. R., Lunde, A., and Nason, J. M. (2011), “The Model Conﬁdence

Set,” Econometrica, 79, 453–497. [8,9,10]

Hastie, T., Tibshirami, R., and Friedman, J. (2001), The Elements of Statistical

Learning: Data Mining, Inference and Prediction, New York: Springer. [7]

Hoerl, A. E., and Kennard, R. W. (1970), “Ridge Regression: Biased Estimation

for Nonorthogonal Problems,” Technometrics, 12, 55–67. [5]

Iacoviello, M. (2005), “House Prices, Borrowing Constraints, and Monetary

Policy in the Business Cycle,” American Economic Review, 95, 739–764.

[3]

Inoue, A., and Kilian, L. (2008), “How Useful Is Bagging in Forecasting

Economic Time Series? A Case Study of U.S. CPI Inﬂation,” Journal of

the American Statistical Association, 103, 511–522. [3,8]

Jurado, K., Ludvigson, S., and Ng, S. (2015), “Measuring Uncertainty,”

American Economic Review, 105, 1177–1215. [14]

Kock, A., and Callot, L. (2015), “Oracle Inequalities for High Dimensional

Vector Autoregressions,” Journal of Econometrics, 186, 325–344. [8]

Lucas, R. (1987), Models of Business Cycles, Oxford: Blackwell. [2]

Makridakis, S., Spiliotis, E., and Assimakopoulos, V. (2018), “Statistical and

Machine Learning Forecasting Methods: Concerns and Ways Forward,”

PLoS One, 13, e0194889. [3,4]

McCracken, M., and Ng, S. (2016), “FRED-MD: A Monthly Database for

Macroeconomic Research,” Journal of Business and Economic Statistics,

34, 574–589. [2,4,17]

Medeiros, M., and Mendes, E. (2016), “1-Regularization of High-Dimensional

Time-Series Models With Non-Gaussian and Heteroskedastic Errors,” Jour-

nal of Econometrics, 191, 255–271. [3,5,8,17]

Medeiros, M., and Vasconcelos, G. (2016), “Forecasting Macroeconomic

Variables in Data-Rich Environments,” Economics Letters, 138, 50–52.

[6]

Mian, A., and Suﬁ, A. (2009), “The Consequences of Mortgage Credit Expan-

sion: Evidence From the U.S. Mortgage Default Crisis,” Quarterly Journal

of Economics, 124, 1449–1496. [3]

Mullainathan, S., and Spiess, J. (2017), “Machine Learning: An Applied

Econometric Approach,” Journal of Economic Perspectives, 31, 87–106.

[1]

Nakamura, E. (2005), “Inﬂation Forecasting Using a Neural Network,” Eco-

nomics Letters, 86, 373–378. [3]

Quaedvlieg, R. (2017), “Multi-Horizon Forecast Comparison,” Working Paper,

Erasmus School of Economics. [8,9,10,14]

Scornet, E., Biau, G., and Vert, J.-P. (2015), “Consistency of Random Forests,”

The Annals of Statistics, 43, 1716–1741. [2]

Shiller, R. J. (2014), “Speculative Asset Prices,” American Economic Review,

104, 1486–1517. [3]

Stock, J., and Watson, M. (1999), “Forecasting Inﬂation,” Journal of Monetary

Economics, 44, 293–335. [2,3]

——— (2002), “Macroeconomic Forecasting With Diffusion Indexes,” Journal

of Business and Economic Statistics, 20, 147–162. [1]

——— (2007), “Why has US Inﬂation Become Harder to Forecast?,” Journal

of Money, Credit and Banking, 39, 3–33. [1,2]

——— (2010), “Modeling Inﬂation After the Crisis,” Technical Report,

National Bureau of Economic Research. [1]

Svensson, L. E. O., and Woodford, M. (2004), “Implementing Optimal Policy

Through Inﬂation-Forecast Targeting,” in The Inﬂation-Targeting Debate,

eds. B. S. Bernanke and M. Woodford, Chicago: University of Chicago

Press, pp. 19–92. [2]

Ter ¨

asvirta, T., van Dijk, D., and Medeiros, M. (2005), “Linear Models, Smooth

Transition Autoregressions and Neural Networks for Forecasting Macroe-

conomic Time Series: A Reexamination,” (with discussion), International

Journal of Forecasting, 21, 755–774. [3]

Tibshirani, R. (1996), “Regression Shrinkage and Selection via the LASSO,”

Journal of the Royal Statistical Society, Series B, 58, 267–288. [5]

Wagner, I., and Athey, S. (2018), “Estimation and Inference of Heterogeneous

Treatment Effects Using Random Forests,” Journal of the American Statis-

tical Association, 113, 1228–1242. [2]

Zhang, X., Wan, A. T., and Zou, G. (2013), “Model Averaging by Jackknife

Criterion in Models With Dependent Data,” Journal of Econometrics,

174, 82–94. [6]

Zou, H. (2006), “The Adaptive Lasso and Its Oracle Properties,” Journal of the

American Statistical Association, 101, 1418–1429. [5]

Zou, H., and Hastie, T. (2005), “Regularization and Variable Selection via the

Elastic Net,” Journal of the Royal Statistical Society, Series B, 67, 301–320.

[5]

Forecasting inflation using disaggregates and machine learning

Preprint

Full-text available

Aug 2023

This paper examines the effectiveness of several forecasting methods for predicting inflation, focusing on aggregating disaggregated forecasts - also known in the literature as the bottom-up approach. Taking the Brazilian case as an application, we consider different disaggregation levels for inflation and employ a range of traditional time series techniques as well as linear and nonlinear machine learning (ML) models to deal with a larger number of predictors. For many forecast horizons, the aggregation of disaggregated forecasts performs just as well survey-based expectations and models that generate forecasts using the aggregate directly. Overall, ML methods outperform traditional time series models in predictive accuracy, with outstanding performance in forecasting disaggregates. Our results reinforce the benefits of using models in a data-rich environment for inflation forecasting, including aggregating disaggregated forecasts from ML techniques, mainly during volatile periods. Starting from the COVID-19 pandemic, the random forest model based on both aggregate and disaggregated inflation achieves remarkable predictive performance at intermediate and longer horizons.

Econometrics of Machine Learning Methods in Economic Forecasting

Preprint

Full-text available

Aug 2023

This paper surveys the recent advances in machine learning method for economic forecasting. The survey covers the following topics: nowcasting, textual data, panel and tensor data, high-dimensional Granger causality tests, time series cross-validation, classification with economic losses.

Transform Network Modeling: A New Paradigm in Macroeconomic Research

Article

Full-text available

Jul 2023

Emad Ghadiri

In this article, we explore the foundations and applications of transform network modeling, a groundbreaking approach that leverages the capabilities of neural networks to analyze complex systems, such as macroeconomic dynamics, in unprecedented ways. We begin by understanding the fundamental principles of neural networks and the intricate mathematics behind their functioning. With a grasp of these foundational concepts, we then explore how this transformative modeling approach enhances the forecasting performance and validity of macroeconomic models through the ingenious use of pooled country data. Moving forward, we embark on an enlightening journey into the realm of network structures and heterogeneous agents, two vital aspects of the macroeconomic puzzle. By employing neural networks, we can unravel the significance of network topologies and how the diverse characteristics of economic agents collectively shape macroeconomic outcomes. Moreover, we illuminate the practical implications of transform network modeling across various domains within macroeconomics, such as financial crisis analysis, monetary policy evaluation, trade dynamics, labor market insights, environmental economics, public health and economic growth, financial inclusion and poverty alleviation, and innovation and economic development. We conclude by emphasizing the boundless prospects that lie ahead for transform network modeling in macroeconomics. As technology continues to advance and data availability grows, the potential of neural networks in deciphering the complexities of macroeconomic systems remains virtually untapped.

A Review of Machine Learning Methods Applied to Structural Dynamics and Vibroacoustic

Article

Full-text available

Oct 2023
MECH SYST SIGNAL PR

The use of Machine Learning (ML) has rapidly spread across several fields of applied sciences, having encountered many applications in Structural Dynamics and Vibroacoustic (SD&V). An advantage of ML algorithms compared to traditional techniques is that physical phenomena can be modeled using only sampled data from either measurements or simulations. This is particularly important in SD&V when the model of the studied phenomenon is either unknown or computationally expensive to simulate. This paper presents a survey on the application of ML algorithms in three classical problems of SD&V: structural health monitoring, active control of noise and vibration, and vibroacoustic product design. In structural health monitoring, ML is employed to extract damage-sensitive features from sampled data and to detect, localize, assess, and forecast failures in the structure. In active control of noise and vibration, ML techniques are used in the identification of state-space models of the controlled system, dimensionality reduction of existing models, and design of controllers. In vibroacoustic product design, ML algorithms can create surrogates that are faster to evaluate than physics-based models. The methodologies considered in this work are analyzed in terms of their strength and limitations for each of the three considered SD&V problems. Moreover, the paper considers the role of digital twins and physics-guided ML to overcome current challenges and lay the foundations for future research in the field.

Big Data for Financial Analysis: Inflation Rate Forecasting using Random Forest

Conference Paper

Aug 2023

Analysis Factors Affecting Egyptian Inflation Based on Machine Learning Algorithms

Article

Full-text available

Sep 2023

Mohamed Abd El-Aal

Given that inflation is one of the most important problems facing the Egyptian economy, determining the factors affecting it is very important. Thus, we aim to use machine learning algorithms like Support vector machine (SVM), K-nearest neighbor (KNN), Random Forest (R.F.), Neural Network (ANN), Gradient boosting (G.B.) and decision tree (D.T.) to determine the accurate algorithm and analyze the factors affecting on Egypt inflation. The study found that the G.B. algorithm is the most accurate among the used algorithms and showed that The major significant variables determining inflation in Egypt are the exchange rate (30.5%), gross fixed formation (24.5%) and government expenditure (12.3%). We also found a positive relationship between the inflation rate and government expenditure, money supply, gross domestic product (GDP) growth, gross fixed formation, foreign direct investment, GDP per capita and exchange rate. Furthermore, there is a negative relationship between the inflation rate, household expenditure and the external trade balance.

Boosting and Predictability of Macroeconomic Variables: Evidence from Brazil

Article

Full-text available

Aug 2023
Comput Econ

This paper aims to elaborate a treated data set and apply the boosting methodology to monthly Brazilian macroeconomic variables to check its predictability. The forecasting performed here consists in using linear and nonlinear base-learners, as well as a third type of model that has both linear and nonlinear components in the estimation of the variables using the history itself with lag up to 12 periods. We want to investigate which models and for which forecast horizons we have the strongest performance. The results obtained here through different evaluation approaches point out that, on average, the performance of boosting models using P-Splines as base-learner are the ones that have the best results, especially the methodology with two components: two-stage boosting. In addition, we conducted an analysis on a subgroup of variables with data available until 2022 to verify the validity of our conclusions. We also compared the performance of boosted trees with other models and evaluated model parameters using both cross-validation and Akaike Information Criteria in order to check the robustness of the results.

Forecasting oil prices with random forests

Article

Full-text available

Aug 2023
EMPIR ECON

Emanuel Kohlscheen

This study analyzes oil price movements through the lens of an agnostic random forest model, which is based on 1000 regression trees. It shows that this highly disciplined, yet flexible computational model reduces in-sample root-mean-square errors (RMSEs) by 65% relative to a standard linear least square model that uses the same comprehensive set of 11 high-frequency explanatory factors. In 1–3 months ahead price forecasting exercises the RMSE reduction relative to OLS ranges exceeds 50%, highlighting the relevance of non-linearities in oil markets. The results underscore the importance of incorporating financial factors into oil models: US interest rates, the dollar and the VIX together account for 39% of the models’ RMSE reduction in the post-2010 sample, rising to 48% in the post-2020 sample.

Nowcasting World Trade with Machine Learning: A Three-Step Approach

Article

Full-text available

Jan 2023

Nowcasting World Trade with Machine Learning: a Three-Step Approach

Article

Full-text available

Jan 2023

We nowcast world trade using machine learning, distinguishing between tree-based methods (random forest, gradient boosting) and their regression-based counterparts (macroeconomic random forest, gradient linear boosting). While much less used in the literature, the latter are found to outperform not only the tree-based techniques, but also more “traditional” linear and non-linear techniques (OLS, Markov-switching, quantile regression). They do so significantly and consistently across different horizons and real-time datasets. To further improve performances when forecasting with machine learning, we propose a flexible three-step approach composed of (step 1) pre-selection, (step 2) factor extraction and (step 3) machine learning regression. We find that both pre-selection and factor extraction significantly improve the accuracy of machine-learning-based predictions. This three-step approach also outperforms workhorse benchmarks, such as a PCA-OLS model, an elastic net, or a dynamic factor model. Finally, on top of high accuracy, the approach is flexible and can be extended seamlessly beyond world trade.

Economic Predictions with Big Data: The Illusion of Sparsity

Article

Full-text available

Jan 2018

Statistical and Machine Learning forecasting methods: Concerns and ways forward

Article

Full-text available

Mar 2018
PLOS ONE

Machine Learning (ML) methods have been proposed in the academic literature as alternatives to statistical ones for time series forecasting. Yet, scant evidence is available about their relative performance in terms of accuracy and computational requirements. The purpose of this paper is to evaluate such performance across multiple forecasting horizons using a large subset of 1045 monthly time series used in the M3 Competition. After comparing the post-sample accuracy of popular ML methods with that of eight traditional statistical ones, we found that the former are dominated across both accuracy measures used and for all forecasting horizons examined. Moreover, we observed that their computational requirements are considerably greater than those of statistical methods. The paper discusses the results, explains why the accuracy of ML models is below that of statistical ones and proposes some possible ways forward. The empirical results found in our research stress the need for objective and unbiased ways to test the performance of forecasting methods that can be achieved through sizable and open competitions allowing meaningful comparisons and definite conclusions.

The Macroeconomic and Fiscal Implications of Inflation Forecast Errors

Article

Full-text available

Feb 2018
J ECON DYN CONTROL

The accuracy of inflation forecasts has important implications for macroeconomic stability and real interest rates in economies with nominal rigidities. Erroneous forecasts destabilize output, undermine the conduct of monetary policy under inflation targeting and affect the cost of both short and long-term government borrowing. We propose a new method for forecasting inflation that combines individual forecasts using time-varying-coefficient estimation along with an alternative method based on neural nets. Its application to forecast data from the US and the euro area produces superior performance relative to the standard practice of using individual or linear combinations of individual forecasts, especially during periods marked by structural changes.

Regression Shrinkage and Selection Via the Lasso

Article

Jan 1996

Robert Tibshirani

We propose a new method for estimation in linear models. The ‘lasso’ minimizes the residual sum of squares subject to the sum of the absolute value of the coefficients being less than a constant. Because of the nature of this constraint it tends to produce some coefficients that are exactly 0 and hence gives interpretable models. Our simulation studies suggest that the lasso enjoys some of the favourable properties of both subset selection and ridge regression. It produces interpretable models like subset selection and exhibits the stability of ridge regression. There is also an interesting relationship with recent work in adaptive function estimation by Donoho and Johnstone. The lasso idea is quite general and can be applied in a variety of statistical models: extensions to generalized regression models and tree‐based models are briefly described.

Estimation and Inference of Heterogeneous Treatment Effects using Random Forests

Article

Oct 2015

Many scientific and engineering challenges---ranging from personalized medicine to customized marketing recommendations---require an understanding of treatment effect heterogeneity. In this paper, we develop a non-parametric causal forest for estimating heterogeneous treatment effects that extends Breiman's widely used random forest algorithm. Given a potential outcomes framework with unconfoundedness, we show that causal forests are pointwise consistent for the true treatment effect, and have an asymptotically Gaussian and centered sampling distribution. We also discuss a practical method for constructing asymptotic confidence intervals for the true treatment effect that are centered at the causal forest estimates. Our theoretical results rely on a generic Gaussian theory for a large family of random forest algorithms that, to our knowledge, is the first set of results that allows any type of random forest, including classification and regression forests, to be used for provably valid statistical inference. In experiments, we find causal forests to be substantially more powerful than classical methods based on nearest-neighbor matching, especially as the number of covariates increases.

Machine Learning Approaches to Macroeconomic Forecasting

Article

Nov 2018

Aaron Smalter Hall

Empirical Asset Pricing Via Machine Learning

Article

Jan 2018

Multi-Horizon Forecast Comparison

Article

Jan 2017

Rogier Quaedvlieg

Machine Learning at Central Banks

Article

Jan 2017

Machine Learning: An Applied Econometric Approach

Article

May 2017
J ECON PERSPECT

Machines are increasingly doing "intelligent" things. Face recognition algorithms use a large dataset of photos labeled as having a face or not to estimate a function that predicts the presence y of a face from pixels x. This similarity to econometrics raises questions: How do these new empirical tools fit with what we know? As empirical economists, how can we use them? We present a way of thinking about machine learning that gives it its own place in the econometric toolbox. Machine learning not only provides new tools, it solves a different problem. Specifically, machine learning revolves around the problem of prediction, while many economic applications revolve around parameter estimation. So applying machine learning to economics requires finding relevant tasks. Machine learning algorithms are now technically easy to use: you can download convenient packages in R or Python. This also raises the risk that the algorithms are applied naively or their output is misinterpreted. We hope to make them conceptually easier to use by providing a crisper understanding of how these algorithms work, where they excel, and where they can stumble—and thus where they can be most usefully applied.

Forecasting Inflation in a Data-Rich Environment: The Benefits of Machine Learning Methods

Abstract and Figures

Recommendations

FORECASTING INFLATION IN A DATA-RICH ENVIRONMENT: THE BENEFITS OF MACHINE LEARNING METHODS

FORECASTING INFLATION IN A DATA-RICH ENVIRONMENT: THE BENEFITS OF MACHINE LEARNING METHODS

Forecasting macroeconomic variables in data-rich environments

Real-time inflation forecasting with high-dimensional models: The case of Brazil