Structural Regional Factors That Determine Absolute and Relative Accuracy of U.S. Regional Labor Market Forecasts
West, Carol Taylor, Journal of Agricultural and Applied Economics
Panel data on regional employment forecasts from structural equation econometric models and time-series models are used to examine whether accuracy of the forecasts can be predicted, both absolutely and relatively. Specification of accuracy includes the time forecast was made, forecast horizon, and regional economic/demographic characteristics. The estimated model is able to predict accuracy of each forecast set at high step lengths but is less successful at low step lengths and is not successful at all in predicting relative accuracy. Regional characteristics are significant determinants of accuracy for both sets of forecasts, but the significant characteristics differ across methodologies and step lengths.
Key Words: forecasting accuracy, panel data, regional economies models
The historical development of U.S. regional forecasting models has long assumed that regional economic structure affects attainable forecast accuracy (e.g., Ballard and Glickman; Glennon, Lane and Johnson; Latham, Lewis, and Landon; Rubin and Erickson; Shoesmith 1990; Taylor 1982a,b). Typically, the assumption is a general cautionary note on comparing forecast accuracy results across regions. The note focuses vaguely on "size" or "diversity" or "growth stability," with higher values of these characteristics implying greater attainable accuracy. Despite the long-standing assumption, the role of regional structure in determining regional economic forecast accuracy has not been systematically examined. Most convincing are the comparative results of Taylor (1982a,b) and Shoesmith (1990), which simultaneously control for model methodology and forecast test period in comparing across regions. However, Taylor's six regions, Shoesmith's three, and the limited test periods of both are small samples that permit only suggestive observation, not statistical analysis.
Lack of more precise understanding of the role of regional characteristics in determining forecast accuracy has inhibited comparative analysis. New methodologies are typically tested on one or only a few regional areas. If a new method tested on Area A looks promising on the basis of accuracy tests, is it because the method is indeed promising or is it because Area A is "easy" to forecast? If a new method tested on Areas A, B, and C appears to "work well" for A and B, but not C, are the results indicative of generally questionable methodological accuracy or do the comparative results reasonably reflect what might be expected given the differences in regional characteristics of areas A, B, and C?
Predicting regional economic forecast accuracy contributes to research beyond regional economics per se. A separate literature, evolving over approximately the same period as regional economic forecasting, has addressed the issue of combination of forecasts.1 It is a literature fraught with frustrated attempts to determine ex ante reliable weights for combining two or more sets of forecasts. Recent studies have called for greater emphasis on attempting to directly forecast accuracy and comparative accuracy so that the evolution of optimal weights can be projected (e.g., Diebold and Mariano; West and Fullerton 1996). Are there economic characteristics that can "explain" relative forecast accuracy? The issue is similar to the one raised in the regional forecasting studies, but, like the latter topic, it has not been systematically examined.
In response to this question, which has emerged from both the regional forecasting and the combination of forecasts literatures, the present study uses a unique data set to statistically analyze determinants of ex ante regional economic forecast accuracy. Regional characteristics are explicitly included as explanatory variables, and the model additionally allows for traditional impacts of period effects. Projection horizon effects are treated by estimation of separate models for each of six different step lengths, one-quarter to six-quarters out. …