A Machine Learning Approach to the Forecast Combination Puzzle

Résumé 0

Forecast combination algorithms provide a robust solution to noisy data andshifting process dynamics. However in practice, sophisticated combinationmethods often fail to consistently outperform the simple mean combination.This “forecast combination puzzle” limits the adoption of alternative com-bination approaches and forecasting algorithms by policy-makers. Throughan adaptive machine learning algorithm designed for streaming data, this pa-per proposes a novel time-varying forecast combination approach that retainsdistribution-free guarantees in performance while automatically adapting com-binations according to the performance of any selected combination approachor forecaster. In particular, the proposed algorithm offers policy-makers theability to compute the worst-case loss with respect to the mean combinationex-ante, while also guaranteeing that the combination performance is neverworse than this explicit guarantee. Theoretical bounds are reported with re-spect to the relative mean squared forecast error. Out-of-sample empiricalperformance is evaluated on the Stock and Watson seven-country dataset and the ECB Sur-vey of Professional Forecasters.