Generalized Gibbs Ensemble Weighting for Forecast Combination

This paper proposes Generalized Gibbs Ensemble Weighting, an online forecast-combination framework for aggregating multiple forecasting experts. The method uses stabilized Gibbs-style exponential weighting, diversity-aware score corrections, and Local-UCB adaptation to select Gibbs parameter states using only past observed performance. We evaluate the approach on official M4 forecast-submission matrices and rolling-origin experiments using Monash Traffic Hourly, Electricity Hourly, and Solar Weekly datasets. The results show that Gibbs-family methods are competitive across heterogeneous forecasting settings and often achieve leading aggregate performance, especially in sequential deployment experiments. The findings suggest that adaptive Gibbs-style weighting is a useful practical framework for online forecast combination, while also showing that classical aggregation methods such as median, trimmed mean, and inverse-loss weighting remain strong in some regimes.