Realized semicovariances: Empirical applications to volatility forecasting and portfolio optimization

Abstract

We propose a two-fold empirical study applying the concept of realized semicovariances as introduced by Bollerslev et al. (2020): in the first part of the paper we aim to estimate and forecast the realized volatility of an equally weighted portfolio formed by Brazilian B3 asset returns, whereas in the second part we search and find an optimum portfolio for these returns. In both parts we use high frequency data of ten assets from different segments and among the most negotiated in B3 financial market from July 2018 to January 2021. In addition, we investigate whether a Markov Switching strategy fits well to our volatility modeling approach considering that our observed data starts some time before the Covid-19 pandemic and spans well into the pandemic period. Machine Learning Regularization (LASSO) methods are employed to select covariates and potentially improve volatility estimation and forecasting. In the portfolio optimization analysis we see that under higher frequency rebalancing periods, minimum variance portfolios using the negative semicovariance matrices present better performances in terms of risk-adjusted returns compared to those that use the standard realized covariance matrices. In general we see that the realized semicovariances bring improvements to the solutions of our two problems.