Forecast robustness in macroeconometric models

• Published date: 01 September 2017
• Author: Gunnar Bårdsen,Ragnar Nymoen,Dag Kolsrud
• Date: 01 September 2017
• DOI: http://doi.org/10.1002/for.2459

Received: 12 May 2014 Revised: 13 October 2016 Accepted: 7 January 2017

DOI: 10.1002/for.2459

RESEARCH ARTICLE

Forecast robustness in macroeconometric models

Gunnar Bårdsen1Dag Kolsrud2Ragnar Nymoen3

1Department of Economics, Norwegian University

of Science and Technology,Trondheim, Norway;

and Norges Bank

2Research Department, Statistics Norway,Oslo,

Norway

3Department of Economics, University of Oslo,

Norway

Correspondence

Dag Kolsrud, Statistics Norway,P. O. Box 8131

Dep, NO-0033 Oslo, Norway.

Email: dok@ssb.no

Abstract

This paper investigates potential invariance of mean forecast errors to structural

breaks in the data generating process. From the general forecasting literature, such

robustness is expected to be a rare occurrence. With the aid of a stylized macro

model we are able to identify some economically relevant cases of robustness and

to interpret them economically. We give an interpretation in terms of co-breaking.

The analytical results resound well with the forecasting record of a medium-scale

econometric model of the Norwegian economy.

KEYWORDS

co-breaking, dynamic homogeneity, equilibrium correction, forecasting, structural

breaks

1INTRODUCTION

Structural breaks represent a main challenge to economic

forecasting. Such breaks are an important source of forecast

failures in macroeconometric models (see, e.g., Clements &

Hendry, 2002; Eitrheim, Husebø, & Nymoen, 1999; Elliott &

Timmermann, 2008). The vulnerability of model-based fore-

casting is generic. Important parts of the literature therefore

focus on ways of making economic forecasts more robust

to breaks. Forecasts can be robustified either individually

or combined, either automatically (like differencing prior to

modeling) or by means of manual intervention (as with inter-

cept correction). The damage done by structural breaks can

also be reduced if there are effective ways of detecting a

break as early as possible after it has occurred (see, e.g.,

Bårdsen, Eitrheim, Jansen, & Nymoen, 2005, chapter 11;

Castle, Fawcett, & Hendry,2011; Clements & Hendr y, 2011;

Giacomini & Rossi, 2009).

In this paper we consider the properties of forecasts pro-

duced by “traditional” equilibrium-correction models. We

demonstrate how mechanisms in the economy can partially

neutralize effects of future breaks on forecasts generated by a

relevant model of that economy.In terms of econometric con-

cepts, we interpret the mechanisms as examples co-breaking

(see Clements & Hendry, 1999, chapter 9; Hendry &

Massmann, 2007).

A concrete motivation for this paper is conveyed by

Figure 1. The panels show real-time forecasts from March

2007 for some key variables in a macroeconometric model

for Norway (cf. Bårdsen & Nymoen, 2009).1The forecasts

(dashed graphs, with prediction bands) can be compared with

the outcomes (solid graphs) until the end of 2011. Some

forecasts fail after the outbreak of the financial crisis, but

most variables are well forecast after the shocks of 2008

and 2009. Most interesting, though, is the behavior of the

actual data series. While the interest rate drops dramatically

(lower right panel)—corresponding to a postforecast struc-

tural break—the outcomes for inflation and the unemploy-

ment rate (upper left and right panel, respectively) are mostly

unaffected, while other variables such as wagegrowth, import

price inflation and gross domestic product (GDP) growth

converge back toward their prebreak predicted paths.

Several variables in the displayed set of forecasts seem to

be more immune to postforecast breaks than what is taken to

be typical (see Clements & Hendry, 2008). There are many

possible explanations for such seemingly partial robustness

to breaks, for example related to the size of the public sector

and the country’s role as an oil exporter. Also, discretionary

policy might have mitigated the effects of the financial crisis,

since in the standard case one would expect an interest rate

forecast failure to go together with large forecast errors for

unemployment for example, but this is not confirmed by the

graphs.

This paper tries to answer the question whether the

empirical partial robustness in Figure 1 can be established

1The model used is NAM. See http://www.svt.ntnu.no/iso/gunnar.bardsen/

nam/evaluation/index.html for an evaluationof the forecasts.

Journal of Forecasting.2017;36:629–639. wileyonlinelibrary.com/journal/for Copyright © 2017 John Wiley & Sons, Ltd. 629