Forecast accuracy of small and large scale dynamic factor models in developing economies

• Author: German Lopez‐Buenache
• Published date: 01 August 2018
• Date: 01 August 2018
• DOI: http://doi.org/10.1111/rode.12392

REGULAR ARTICLE

Forecast accuracy of small and large scale dynamic

factor models in developing economies

German Lopez-Buenache

Universidad Carlos III de Madrid, Spain

Correspondence

German Lopez Buenache

Department of Statistics, Instituto Flores

de Lemus, 28903 Getafe-Madrid, Spain.

Email: glbmerlin@gmail.com

Funding Information

This work was supported by the

MINECO/FEDER (Grant ECO2015-

70331-C2-2-R).

Abstract

Developing economies usually present limitations in the

availability of economic data. This constraint may affect

the capacity of dynamic factor models to summarize large

amounts of information into latent factors that reflect

macroeconomic performance. This paper addresses this

issue by comparing the accuracy of two kinds of dynamic

factor models at GDP forecasting for six Latin American

countries. Each model is based on a dataset of different

dimensions: a large dataset composed of series belonging

to several macroeconomic categories (large scale dynamic

factor model) and a small dataset with a few prescreened

variables considered as the most representative ones

(small scale dynamic factor model). Short-term pseudo

real time out-of-sample forecast of GDP growth is carried

out with both models reproducing the real time situation

of data accessibility derived from the publication lags of

the series in each country. Results (i) confirm the impor-

tant role of the inclusion of latest released data in the

forecast accuracy of both models, (ii) show better preci-

sion of predictions based on factors with respect to

autoregressive models and (iii) identify the most adequate

model for each country according to availability of the

observed data.

1

|

INTRODUCTION

The information contained in some key macroeconomic aggregates is of crucial relevance for econ-

omists. They provide a general assessment about the performance of a given economy, allowing

DOI: 10.1111/rode.12392

Rev Dev Econ. 2018;22:e63–e78. wileyonlinelibrary.com/journal/rode ©2018 John Wiley & Sons Ltd

|

e63

expectations about other specific indicators to be constructed and results of strategies deployed by

policymakers and central bankers to be evaluated.

The increasing differences between developing and developed economies in economic perfor-

mance, the unusual monetary and fiscal policies implemented in advanced economies after the

financial crisis, and their spillover effects on emerging countries all shape a challenging scenario

of global uncertainty. In such a context, the policymakers of less developed countries require early

evaluation of these key macroeconomic aggregates in real time in order to measure the conse-

quences of these global events and adapt their responses accordingly.

Unfortunately, the burdensome accounting task needed to compute these key indicators causes

considerable delay in the release of the data. For instance, gross domestic product (GDP), widely

considered to be the main indicator of the current economic situation, is usually published at a

quarterly frequency and released with more than 2 months delay; while, in contrast, there are hun-

dreds of more specific indicators that involve easier computation, which are earlier relea sed at a

higher frequency.

Dynamic factor models (DFMs) take advantage of this increasing availability of data. Given

that macroeconomic series are very collinear, it is assumed that they can be decomposed into two

orthogonal parts: a reduced set of latent common factors, which capture most of the comovements

in the data, and an idiosyncratic component that only affects a specific series or a reduced set of

them. In addition to other applications, this factor decomposition has been implemented with fore-

casting purposes. Because of the lower number of factors with respect to the amount of available

data, factors can be included in a forecast equation for a targeted variable, such as GDP, with a

reduced set of regressors containing the relevant information, while keeping a parsimonious speci-

fication.

Previous literature has shown a clear improvement in short-term forecasting using DFMs. In

fact, these models have become a key tool for several economic institutions such as the European

Central Bank and the Federal Reserve, among others. However, their implementation and the eval-

uation of their performance have been carried out mainly in developed economies where a large

amount of macroeconomic information is available. It is important to note that the quantity as well

as the quality of existing data affects the choice of different DFMs. Depending on the number of

series included in the model, DFMs can be classified into two clearly distinguishable strands of

the literature: large-scale DFMs (LS-DFMs), where factors are estimated from a huge dataset,

under the premise that there is no reason to discard any information and small-scale DFMs (SS-

DFMs), where the common factor is estimated from a reduced set of indicators prescreened by the

forecaster as those with the highest informational content and which are considered as sufficient

for a complete characterization of macroeconomic behavior. Depending on the number of series

used for estimation of the factors, these two DFMs present different theoretical assumptions, com-

putational limitations and estimation procedures. In this context, the constraints in data availability

in developing economies, such as the lower amount of time series, which are usually shorter, later

released or with missing values in many cases, play an important role in the performance of

DFMs. These limitations can make one of the two models more appealing for the forecaster,

depending on the amount, quality and informational content of accessible information. Thus, the

main contribution of this paper is to assess whichever of these two methodologies performs bett er

in the particular context of developing economies. In order to highlight what the effects of the

properties of the dataset are on the estimation of the latent factors, we review the main characteris-

tics of both methodologies next.

Stock and Watson (1991) pioneered the literature on SS-DFMs. In their seminal paper, they use

this method to compute a single factor that closely mirrors the Index of Coincident Economic

e64

|

BUENACHE