A0820
Title: Factor models and forecast combinations
Authors: Rachida Ouysse - University of New South Wales (Australia) [presenting]
Andrey Vasnev - University of Sydney (Australia)
Abstract: Principal components (PC) forecasts are more accurate than many single econometric models. However, the principal components are computed from the predictors without accounting for their relationship with the forecast target variables. An alternative approach is the factor combination of forecasts. An alternative scenario is considered where multiple forecasters have access to subsets (possibly overlapping) of the full information set. Forecasters use a single-factor model to construct factor forecasts. The performance of combining these partial information forecasts with a single factor forecast from the full set of predictors is analysed. Combination methods include equal weight, optimal weight, shrinkage weights, and principal component combination are considered. An application to forecasting the monthly growth rate of U.S. industrial production, where the full set of predictors consists of 130 economic indicators, shows that combining forecasts outperforms the full information factor forecasts. The Shrinkage weights combination performed better than equal weights. More data is not better for forecast performance than a weighted combination of selective consensus from its snippets.
