Abstract:

The determination of optimal individual model sets and the setting of ensemble weights are two critical problems for ensemble forecasting， which are related to the prediction performance of the ensemble model. On the one hand， the prediction performance of individual model is unstable and the static ensemble prediction model cannot fully exploit the prediction advantages of individual models； on the other hand， ergodic method to determine the optimal model subset is faced with high computational complexity. To this end， a dynamic ensemble forecasting model is proposed with a regime-switching regression method. First， the optimal individual model set is determined by calculating the mutual information between the individual forecasts and the original data； second， the regime-switching regression is used to ensemble the individual forecasts and get the final prediction values. Through the prediction experiments on the sovereign credit default swaps in nine sample countries， it is found that the proposed regime-switching regression ensemble model performs well， not only better than the individual and combination prediction models but also better than the sliding window technology dynamic combination forecasting model.

Key words: combination forecast, dynamic ensemble, regime-switching, mutual information