Forecasting with Factor Models

This paper evaluates a modication done on the DI framework of Stock and Watson (2002). The regressors used for PCA are rst evaluated for their predictive power. As there are several techniques to create such subsets, ve are used of which two are hard thresholding and three are soft thresholding methods. The 1, 6, 12 and 24 month ahead point-forecasts are then compared to a 3 factor benchmark with factors constructed from the complete dataset. Results show that selecting relevant predictors before applying principal component analysis yields improved forecasts.