Forecasting realized betas using predictors indicating structural breaks and asymmetric risk effects

This paper studies the importance of structural breaks and asymmetric risk effects for accurate forecasts of the realized beta. Specifically, structural breaks in the realized beta are detected by Iterated Cumulative Sum of Square (ICSS) algorithm and asymmetric risk effects are captured by decomposing the realized beta further into various components following Ang et al. (2006) and Bollerslev et al. (2021). We propose a set of Heterogeneous Autoregressive (HAR) model variants by incorporating these new predictors. To achieve model parsimony and to keep only the predictors with significant power, we employ Least Absolute Shrinkage and Selection Operator (LASSO) method for variable selection. Our proposed LASSO-HAR model with estimators of structural breaks and asymmetric risk effects is found to yield more accurate out-of-sample beta forecasts than a variety of alternative models in terms of both statistical and economic criteria. In particular, our model successfully achieves the long-memory feature of realized betas in a tractable and parsimonious way. These empirical findings are robust across different data sampling frequencies, different estimation windows, different sub-samples, different quantiles of the beta distribution and different industrial sectors.