Model selection for forecasting

摘要

This paper presents empirical comparisons of forecast accuracy resulting from variety of model selection procedures. Models of monthly sales of residential electricity are estimated, and used to forecast three years into the future for twenty states in the U.S. Models are selected by a variety of complexity criteria and by upward and downward F-tests at various significance levels. Forecast accuracy was measured by one-step and multistep conditional root-mean-square forecast errors. Overall the selection criteria which most heavily penalized overparametrized models performed best: the Schwarz criterion and 1% size sequential F-testing.