High‐frequency data and stock–bond investing

Abstract

Understanding the comovements between stock and bond returns is crucial in asset allocation. This paper employs a new class of multivariate covariance models with realized covariance measures for modeling the joint distribution of returns. Estimation results indicate that high-frequency data not only enhances explanatory power, but also sorts out the features of heteroskedasticity with a short response time and short-run momentum effects in describing the covariance dynamics. To confirm the efficiency of the models in covariance predictions, multiperiod volatility-timing strategies are implemented to evaluate the benefits of incorporating the distinguishing features into the modeling. Out-of-sample forecasting results indicate that the models outperform conventional models in finding optimal portfolio shares and thus considerably reduce the conditional volatility in a portfolio. Consequently, investors with high risk aversions are willing to pay pronounced performance fees to obtain the economic value of volatility timing.