Incorporating Model Uncertainty and Model Instability in Forecasting Bond Risk Premia and Term Structure of Government Bond Yield – A Bayesian Model Averaging Approach

Abstract

This research investigates the macro factors for forecasting (1) bond risk premia and (2) term structure of government bond yields by using Bayesian Model Averaging (BMA) based on empirical prior. Different from the traditional variable selection approach which advocates finding an “optimized” variable subset, BMA combines all model specifications with their posterior probability to handle model uncertainty. Our result shows strong empirical evidence that BMA outperforms the traditional model selection criteria. As market environment keep changing, the relationship between response variable and explanatory variable(s) changes too. The BMA based on empirical prior does not need subjective elicitation of priors and can be more adaptive to dynamic economic condition. Through explicitly incorporating model uncertainty and model instability, the BMA based on empirical prior can easily match artificial neural network, the state-of-the-art universal approximating method, in forecasting accuracy. Although we find there is no significant difference in forecasting performance between BMA and neural networks trained with Bayesian regularization, we find each method does offer unique information which could further improve the other method’s forecasting performance. The performance of using BMA to forecast bond excess return is tested with both statistical measures and economic measures. Many existing yield curve models do not incorporate the role of macro factors (Diebold et al 2006). Diebold and Li (2006) demonstrated no existing model can outperform the random walk model in forecasting one-month-ahead yield curve. We apply BMA to forecast the government bond yield change and indicate BMA model can significantly outperform the random walk model at one-month-ahead horizon.