Forecasting Realized Volatility With Tree-Based HAR-Type Models Incorporating Macroeconomic Uncertainty

Abstract

This paper extends heterogenous autoregressive (HAR) models of realized volatility to allow for regime dependence for the purpose of volatility forecasting. We construct two series representing “good” and “bad” macroeconomic uncertainties and allow them to shape the regime structure. The coefficients are estimated with the CART method and can be updated during the forecast horizon. An empirical application to 20 liquid US securities shows that our approach, which imposes economically motivated structure on realized volatility models, can deliver a significant improvement in forecasting realized volatility, especially at long forecast horizons. This improvement is quite robust to alternative realized volatility measures and some variations in the tuning parameters. The documented improvement in volatility forecasting translates into economic gains when we forecast Value-at-Risk.