An Empirical Evaluation of Alternative Forecasting Combinations

Abstract

Combining naive forecasting models as an alternative to using any one model has shown promise for improving forecasting accuracy at reasonable additional cost. This study extends previous combination modeling by selectively including models in combination rather than using all individual models in a set, evaluating six new model combinations and comparing them to three previously investigated model combinations, investigating the impact of forecasting time horizon, and investigating alternative error measures to mean square error (MSE). Results support using a model-combination that (1) selects the best 3 to 5 models from the 10 models studied and (2) weights the selected models based upon the inverse proportion of their individual accuracy as measured by MSE.