Regression-based Forecast Combination Methods

Least squares combinations (Granger & Ramanathan, 1984) are an important development in the forecast combination literature. However, ordinary least squares methods often perform poorly in real application due to the variability of coefficient/weight estimations. In this work, on one hand, we propose sequential subset selections to reduce the variability during combinations. On the other hand, we propose a novel method to simultaneously stabilize and shrink the coefficient/weights estimates. The proposed methods can be applied to various combination methods to improve prediction as long as their weights are determined based on ordinary least squares.