Academic journal article Atlantic Economic Journal

An Analysis of RSQE Forecasts: 1971-1992

Academic journal article Atlantic Economic Journal

An Analysis of RSQE Forecasts: 1971-1992

Article excerpt

I. Introduction

The Research Seminar in Quantitative Economics (RSQE) at the University of Michigan was established in 1951. One of the first research projects of the seminar was the construction of an econometric model of the U.S. economy. A version of what has come to be known as the Klein-Goldberger model was used in November 1952 to produce the first RSQE forecast. Each year since then, a macroeconomic forecast has been presented at the Annual Conference on the Economic Outlook which is held in November.

The purpose of this paper is to analyze the recent RSQE forecasting record. Such an analysis is of interest for several reasons. Macroeconometric forecasting models have been under attack for a number of years. As Sims [1980, p. 1] notes,

"though large-scale statistical macroeconomic models exist and are by some criteria successful, a deep vein of skepticism about the value of these models runs through that part of the economics profession not actively engaged in constructing or using them."

Lucas and Sargent [1981, p. 303] are less refrained:

"the track record of the major econometric models is, on any dimension other than very short-term unconditional forecasting, very poor.... macroeconometric models were subjected to a decisive test in the 1970s.... the inflationary bias on average of monetary and fiscal policy in this period, according to all of these models, should have produced the lowest average unemployment rates for any decade since the 1940s. In fact, as we know, they produced the highest unemployment rates since the 1930s. This was econometric failure on a grand scale."

A systematic examination of the forecasting record may shed some light on whether this skepticism and pessimism are warranted.

More recently, interest has centered on the combination of forecasts from alternative models as a way to obtain improved forecasts. Nelson [1984] found that a composite forecast obtained by combining a single equation, autoregressive, integrated, moving average model for real GNP growth with five different macroeconometric models produced more accurate forecasts than the econometric models alone for two-, three-, or four-quarter ahead forecasts for the period 1976-82. Lupoletti and Webb [1986], using forecasts from Data Resources Inc., Chase Econometrics, and Wharton Econometric Forecasting Associates for the period 1970-83, found statistically significant evidence that the combination of econometric model forecasts with VAR forecasts resulted in an improved composite forecast. Both of these studies point to the conclusion that simple univariate and multivariate time series models contain useful information for forecasting that is often overlooked by well-known commercial forecasters, at least over the 1970-83 time period.

Various types of forecast encompassing tests have been proposed in the literature, including both univariate and multivariate procedures. However, most of the empirical results that have been reported are based on univariate analysis. Fair and Shiller [1989, 1990], for example, confine their analysis to GNP forecasts obtained from a set of models. Both univariate and multivariate tests are examined in this paper to see if such tests provide useful information for model evaluation.

The analysis in this paper is consistent with two themes that were emphasized in Professor Harold Hotelling's work on prediction. In his paper with Holbrook Working, Hotelling [1929, p. 73] argued that "the probable error of a trend is always appropriate, and furthermore, it is frequently necessary if sound conclusions are to be formed." An examination of the RSQE forecasting record provides a basis for assigning a measure of probable error to the forecasts. If econometric forecasts were accompanied by such probable error estimates, unrealistic expectations of forecast precision might be avoided.

Hotelling [1940] also took up the issue of selection of one variate from a set of variates that are available to predict a variable of interest. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed

Oops!

An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.