LASSO principal component averaging: A fully automated approach for point forecast pooling

This paper develops a novel, fully automated forecast averaging scheme which combines LASSO estimation with principal component (PCA). LASSO-PCA (LPCA) explores pool of predictions based on single model but calibrated to windows different sizes. It uses information criteria select tuning parameters and hence reduces the impact researchers’ ad hoc decisions. The method is applied average hourly day-ahead electricity prices over 650 point forecasts obtained various lengths calibration windows. evaluated four European American markets an out-of-sample period almost two half years compared other semi- methods, such as simple mean, AW/WAW, LASSO, PCA. results indicate that very efficient in terms error reduction, whereas PCA robust selection specification parameter. LPCA inherits advantages both methods outperforms approaches mean absolute error, remaining insensitive choice