A0178
Title: Forecast combination and interpretability using random subspace
Authors: Boris Kozyrev - Halle Institute for Economic Research (IWH) (Germany) [presenting]
Abstract: Forecast aggregation is investigated via random subspace regressions (RS), and the potential link between RS and the Shapley value decomposition (SVD) is explored using the US GDP growth rates. This combination of techniques enables handling high-dimensional data and reveals the relative importance of each individual forecast. First, we demonstrate that in certain practical instances, it is possible to enhance forecasting performance by randomly selecting smaller subsets of individual forecasts and obtaining a new set of predictions based on a regression-based weighting scheme. The optimal value of selected individual forecasts is also empirically studied. Then, we propose a connection between RS and the SVD, enabling the examination of each individual forecast's contribution to the final prediction, even when the number of forecasts is relatively large. This approach is model-agnostic (can be applied to any set of forecasts) and facilitates understanding of how the aggregated prediction is obtained based on individual forecasts, which is crucial for decision-makers.