Combining and selecting forecasting models using rule based induction

As inaccurate forecasts can lead to lost business and inefficient operations, it is imperative that forecasts be as accurate as possible. A major problem, however, is that no single forecasting method is the most accurate for every data time series. Thus, generating a forecast is often an uncertain affair, involving the use of heuristics by human experts and/or the consistent use of forecasting models whose accuracy may or may not be the most accurate for that time series. To compound matters, the best forecasts are often produced by combining forecasting models. This research describes the use of an Artificial Intelligence-based technique, rule-based induction, to improve forecasting accuracy. By using training sets of time series (and their features), induced rules were created to predict the most appropriate forecasting method or combination of methods for new time series. The results of this experiment, which appear promising, are presented, together with guidelines for its practical application. Potential benefits include dramatic reductions in the effort and cost of forecasting; the provision of an expert ‘assistant’ for specialist forecasters; and increases in forecasting accuracy.