Forecasting Realised Volatility Using Regime-Switching Models

The long-memory and nonlinearity coexist in realised volatility. This paper incorporates the linear AR and HAR models with regime-switching models, including the smooth transition and Markov-switching approaches, to assess the forecasting performance of realized volatility. In-sample results show that the regime-switching models provide higher estimation accuracy than the linear models. For the out-of-sample results, although the regime-switching models have limited forecasting ability over the daily horizon, the regime-switching models do outperform the linear HAR model over weekly and monthly horizons, in which the Markov-switching model with time-varying transition function is the best and consistently exhibits the most accurate forecasting error over time. Empirically, the abrupt transition technique is better suited to real financial market changes and therefore outperforms smoothed changes in terms of forecasting performance. Further, the results have implications for risk management in terms of the Value at Risk and Expected Shortfall