Forecasting key US macroeconomic variables with a factor‐augmented Qual VAR

• Published date: 01 September 2017
• Author: Mark E. Wohar,Eric Olson,Rangan Gupta
• Date: 01 September 2017
• DOI: http://doi.org/10.1002/for.2460

RESEARCH ARTICLE

Forecasting key US macroeconomic variables with a

factor‐augmented Qual VAR

Rangan Gupta

1

| Eric Olson

2

| Mark E. Wohar

3,4

1

Department of Economics, University of

Pretoria, Pretoria, South Africa

2

College of Business and Economics, West

Virginia University, Morgantown,West

Virginia, USA

3

College of Business Administration,

University of Nebraska at Omaha, Omaha,

Nebraska, USA

4

School of Business and Economics,

Loughborough University, Loughborough,

UK

Correspondence

Rangan Gupta, Department of Economics,

University of Pretoria, Pretoria 0002, South

Africa.

Email: rangan.gupta@up.ac.za

Abstract

In this paper, we first extract factors from a monthly dataset of 130 macroeco-

nomic and financial variables. These extracted factors are then used to construct

a factor‐augmented qualitative vector autoregressive (FA‐Qual VAR) model to

forecast industrial production growth, inflation, the Federal funds rate, and the

term spread based on a pseudo out‐of‐sample recursive forecasting exercise over

an out‐of‐sample period of 1980:1 to 2014:12, using an in‐sample period of

1960:1 to 1979:12. Short‐, medium‐, and long‐run horizons of 1, 6, 12, and

24 months ahead are considered. The forecast from the FA‐Qual VAR is compared

with that of a standard VAR model, a Qual VAR model, and a factor‐augmented

VAR (FAVAR). In general, we observe that the FA‐Qual VAR tends to perform

significantly better than the VAR, Qual VAR and FAVAR (barring some excep-

tions relative to the latter). In addition, we find that the Qual VARs are also well

equipped in forecasting probability of recessions when compared to probit models.

KEYWORDS

business cycleturning points, factors, forecasting, vectorautoregressions

1|INTRODUCTION

In a seminal contribution, Sims (1980) proposed the vector

autoregressive (VAR) model, which is essentially a system

of equations whereby, in a specific equation, a specific

variable is regressed on the past values of itself and past

values of other variable(s) in this system, with possible

allowance for deterministic terms (e.g., constant and trend).

By design, this linear framework treats all variables in the

system as endogenous. Although atheoretical, the VAR

model has been shown to perform exceptionally well when

compared to various other econometric models, including

its popular predecessor, namely the simultaneous equations

approach of the Cowles Foundation (Dueker &

Assenmacher‐Wesche, 2010).

However, one substantial issue regarding forecasting

macroeconomic variables in VARs is that the variables

behave nonlinearly, rather than linearly, with regard to their

recent past values around peaks and troughs in the business

cycle. Given this issue, Hamilton (1989) developed a non-

linear Markov‐switching model whereby changes in

coefficients correspond to changes in the business cycle.

In these models, the regime switches are usually based on

a latent state variable or time‐varying regime probabilities

that are logistic functions of lagged endogenous variables.

An additional form of nonlinear modeling of macroeco-

nomic variables was developed by Leamer and Potter

(2004), who allow a latent business cycle index to alter

the dynamics of the threshold variable as in a threshold

autoregressive model. More recently, Auerbach and

Gorodnichenko (2013) estimate a nonlinear smooth‐transition

VAR for government spending, tax revenue, and output in

which they impose that fiscal stimulus has differing effects

throughout the business cycle. Mittnik and Semmler (2012)

estimate a bivariate threshold VARin which the threshold var-

iable is deterministic and set equal to the average of output.

Fazzari, Morley, and Panovska (2015) and Donayre and

Panovska (2016) allow for nonlinear state‐dependent effects

Received: 13 January 2016 Revised: 23 December 2016 Accepted: 18 January 2017

DOI: 10.1002/for.2460

640 Copyright © 2017 John Wiley & Sons, Ltd. Journal of Forecasting. 2017;36:640–650.wileyonlinelibrary.com/journal/for