A test of a minimalist rule-based forecasting system

Abstract

Following on the work of Adya et al. (2000) where 35 percent of the 99 rules of the RBF system of Collopy and Armstrong (1992) were removed without noticeably affecting forecast accuracy, we undertook to further refine the RBF system by asking what would be the result if a minimum number of rules and a very simple model weighting scheme were used to develop the forecasts? Further, we tested this minimalist model with non-expert users. The results were that the minimalist RBF model did show some improvement in the reduction of forecast error compared to simple eyeball judgments for the APE and the RAE measures; however, these improvements were far off the results reported by Adya et al. (2000) suggesting that we removed too many rules.