Forecasting Bond Risk Premia with Forecast Combinations using Many Predictors

In this report the application of forecast combinations is examined for construction of excess bond return forecasts using a large set of predictor variables. The analysis is done by applying a large number of forecast combination methods to forecast one-year U.S. government excess bond returns for maturities ranging from two to five years. For every maturity 753 forecast sets are constructed using different forecast combinations methods, predictor variables and cluster methods. These are then compared using the model confidence set procedure. The forecast combination methods are subsequently subjected to a significance test and to a test over different time periods. The conclusion that this report draws is that two sets of forecast sets outperform the other forecast sets as well as the historical average benchmark and achieve a higher out-of-sample R2 then was previously found in the literature. These are the forecast sets that are constructed when either the recursive OLS weighting scheme or the complete subset regression method is applied to the set of forecast sets which are constructed with the macroeconomic predictors and which are first clustered according to their economic background. These two forecast combination methods continue to outperform the others for every maturity and also if the time period is changed.