Rule-based forecasting: Development and validation of an expert systems approach to time series extrapolation

The accuracy of extrapolation methods varies greatly from one time series to another and across forecast horizons. Selecting the most accurate method for a particular series and horizon is one of the forecaster\u27s principal tasks. This dissertation discusses the development and validation of a rule-based system to produce extrapolation forecasts. Rule-based forecasting uses domain knowledge, forecasting expertise, the features of each series, and research to combine quantitative forecasts. The rule-base was developed using a review of the literature on forecasting accuracy, a survey of forecasting experts, direct assessments from five experts in forecasting, and protocol analyses from the same five experts. The resulting rule-base consisted of 87 rules, which make use of 18 features to weight four component methods to produce forecasts. An error measure (the relative absolute error) is developed in this dissertation to aid in the calibration of rule-based forecasting systems. The measure is shown to be sensitive, reliable, and valid. Rule-based forecasting produced substantially more accurate forecasts than could be obtained by the best prior formal approach, which was to combine forecasts using equal weights. For six-year-ahead ex ante forecasts of 90 annual series, the median absolute percentage error for rule-based forecasting was 42% less than that from combining forecasts. The improvement in accuracy of the rule-based forecasts over combining was significant at p $\u3c$.01. Rule-based forecasting was most accurate in situations involving good domain expertise, stability, significant trends, and low uncertainty