[Review written with Fred Collopy]
This study examined annual earnings forecasts for 205 firms for the 8 years from 1978 to 1985. The forecast horizons ranged from one to four quarters. That is, a 1-year ahead forecast was made, and then forecasts for the same year were made after the results for quarters 1, 2 and 3 were available. Forecasts were obtained from an average of at least five analysts forecasts and from three quantitative extrapolations (univariate ARIMA models). Three combined forecasts were prepared by using equal weights for the analysts average forecast and each of the quantitative extrapolations. This provided 6560 forecasts for each of seven forecasting procedures. A trimmed MAPE was used as the criterion for accuracy. The pattern of results was similar for each combined forecast. To examine the error reduction, we took a simple average of gains provided by each of the three combined procedures. The average error reduction, compared with the average of the components was 11.3%. In fact, the combined forecasts were more accurate than either component alone for each of the three combined forecasts and for each of the four forecast horizons. Not surprisingly, the combined forecast improved accuracy most when there was little agreement among the analysts forecasts. (Differences among analysts were highly correlated with the level of error.) Lobo concluded that combining was more useful for the longer forecast horizons, a finding that disagreed with Lawrence, Edmundson, and OConnor (1986). However, Lobos interpretation was based on the absolute error reduction. Because errors tend to be smaller for shorter horizons, the potential for gain is smaller. Consistent with prior research, we calculated the percentage error reduction in the MAPE as the forecast horizon was increased from one to four quarters; the reductions were 13.6%, 11.1%, 10.7%, and 9.9%, respectively. Thus, we conclude that improvements from combining were greatest for the shorter forecast horizons.