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

Abstract

Forecast combinations support production and inventory decisions, but inaccurate or redundant candidates can dilute crowd wisdom. We develop a modular forecast-trimming framework based on robustness, accuracy, and diversity, introducing autoAD and autoRAD. Both adaptively balance error magnitude and complementarity, with autoRAD additionally applying robustness screening. Relative Diversity (RelDiv) provides a scale-normalised diagnostic of the initial pool. Across 103,826 M-competition series, autoAD performs comparably to accuracy-only trimming overall, with significant MASE-rank gains for monthly series and moderate-RelDiv pools. On 5,000 Royal Air Force spare-parts series, it significantly outperforms accuracy-only trimming in rank at all three horizons. Allowing a zero diversity weight lets autoAD revert to accuracy-based selection when validation evidence does not support complementarity. Together with the weaker performance of alignment-only trimming, these findings support diversity as a selective addition to accuracy. In the spare-parts application, autoAD achieves interval scores comparable to the full-pool and accuracy-only benchmarks, while autoRAD attains the lowest scores considered. Both substantially reduce pool size while maintaining similar service levels, with modest inventory cost savings under moderate backorder penalties. These savings disappear when penalties are high. The framework provides a practical approach to constructing parsimonious forecasting crowds for operational decision support.