To test the predictability of U.S. bond risk premia at different quantile levels, we propose a robust hypothesis testing procedure for the predictability of multiple predictors that could be highly persistent. Our method extends the popular extended instrumental variable (IVX) in mean regression to a two-step quantile regression framework. We address the size distortion induced by two higher-order terms of the aforementioned inference procedure. To tackle the problem of power loss, we construct a power-enhanced test statistic. When constructing the final test statistic, we introduce an accurate density estimator of the error term in quantile regression. The final test, therefore, obtains good size and power performance simultaneously, especially at tails and in multivariate models with many predictors.
Numerical simulations and an empirical study on the predictability of bond returns are provided to demonstrate the effectiveness of the newly proposed approach. Using U.S. monthly macroeconomic data from 1960-2022, we find that the heterogeneous predicting power of the macroeconomic variables.
Joint work with Xiaosai Liao and Xinjue Li. |
