An extreme bias-penalized forecast combination approach to commodity price forecasting

Abstract

Forecast combination, a well-established technique for improving forecasting accuracy, investigates the integration of competing forecasts to produce a composite superior to individual forecasts. In this study, we propose a novel forecast combination method that would reduce overfitting risk and improve forecast’s generalization ability. To capture the extreme bias of a forecast in combination process, we define a measurement PaR for forecast combination. A novel PaR-based loss function with an elastic net is proposed that can effectively trade off the sparsity of weights to mitigate the risk of underfitting or overfitting. An improved artificial bee colony algorithm-based optimization method is introduced to achieve the optimal weights. The experimental results on gold, silver and crude oil price data demonstrate that the proposed forecast combination approach can outperform not only individual models but also combination approaches like simple averaging and other competitive benchmarks. The MAPE achieved by the presented method could decrease by 10.98%, 5.03% and 10.28% in gold, silver and crude oil price forecasting respectively, compared to the best individual model.