Integration of Statistical Methods and Judgment for Time Series Forecasting: Principles from Empirical Research

Abstract

We consider how judgment and statistical methods should be integrated for time-series forecasting. Our review of published empirical research identified 47 studies, all but four published since 1985. Five procedures were identified: revising judgment; combining forecasts; revising extrapolations; rule-based forecasting; and econometric forecasting. This literature suggests that integration generally improves accuracy when the experts have domain knowledge and when significant trends are involved. Integration is valuable to the extent that judgments are used as inputs to the statistical methods, that they contain additional relevant information, and that the integration scheme is well structured. The choice of an integration approach can have a substantial impact on the accuracy of the resulting forecasts. Integration harms accuracy when judgment is biased or its use is unstructured. Equal-weights combining should be regarded as the benchmark and it is especially appropriate where series have high uncertainty or high instability. When the historical data involve high uncertainty or high instability, we recommend revising judgment, revising extrapolations, or combining. When good domain knowledge is available for the future as well as for the past, we recommend rule-based forecasting or econometric methods.

Keywords: Forecasting, time series forecasting, integration