Combining Forecasts: An Application to U.S. Presidential Elections
with Alfred Cuzán, Randy Jones, and Scott Armstrong, .
Abstract. Prior research suggested that accuracy gains from combining forecasts are particularly high if one uses forecasts from different methods that draw upon different data. We tested this assumption by combining forecasts from three component methods (polls, econometric models, and experts) used for predicting the five U.S. presidential elections between 1992 and 2008. The gains from combining were substantially larger than the 12% reduction obtained from a prior meta-analysis. Mean error reduction from combining within component methods ranged from 18% to 21%. Combining across component methods yielded further error reductions, ranging from 37% to 40%. Compared to the typical component forecast, combining reduced forecast error by 40% and performed about as well as the best component forecast. Average error reduction compared to the worst component forecast was 59%. Compared to the typical uncombined individual forecast, our combining procedure yielded error reductions ranging from 42% to 50%. Combining is probably the most cost efficient way to improve forecast accuracy and to prevent large errors.
Full text available here.
