An Improved Feature-Based Forecast Combination Method Using Rolling Origin Evaluation

Combining forecasts from diverse models via weighted sums enhances accuracy. Feature-based methods, which link data features to forecasting model performance, offer a promising approach to assigning weights. Currently, limited attention has been paid to the importance of reliable performance measurement. In this work, we propose an improved feature-based forecast combination method, particularly using the rolling origin evaluation to obtain reliable performance measurements of forecasting models. Experimental results based on the M4 competition data show that our method outperforms the state-of-the-art. The proposed method exhibits robustness to parameter variations.