Finding Wisdom in the Crowd: Adaptive Accuracy-Diversity Trimming for Forecast Combinations

Abstract

Forecast combinations support production and inventory decisions, but including every candidate can dilute crowd wisdom through inaccurate or redundant information. Forecast trimming addresses this risk by retaining an informative subset. We develop a modular framework that evaluates candidates along three dimensions—robustness, accuracy, and diversity—and introduce autoAD, an adaptive accuracy-diversity rule balancing individual error magnitude against error complementarity. We also propose Relative Diversity (RelDiv), a scale-normalized diagnostic of the initial pool. Tests on exponential-smoothing (ETS) pools for 103,826 M1, M3, and M4 series and a heterogeneous pool for 5,000 Royal Air Force (RAF) spare-parts demand series show that autoAD attains the lowest series-weighted mean MASE rank in the M data, although its mean rank is not statistically distinguishable from that of accuracy-only trimming overall. Its rank gains are clearest for monthly series and moderate-RelDiv pools; it also significantly improves within-series MASE ranks at all three RAF forecast horizons. By selecting a diversity weight of zero when validation performance does not improve, autoAD avoids imposing diversity indiscriminately. Alignment-only trimming is less reliable, and the evaluated robustness screen provides little systematic benefit. The framework offers interpretable evidence and practical guidance for constructing parsimonious forecast pools in production settings.