Forecasting Realized Volatility with Changes of Regimes

Realized volatility of financial time series generally shows a slow–moving average level from the early 2000s to recent times, with alternating periods of turmoil and quiet. Modeling such a pattern has been variously tackled in the literature with solutions spanning from long–memory, Markov switching and spline interpolation. In this paper, we explore the extension of Multiplicative Error Models to include a Markovian dynamics (MS-MEM). Such a model is able to capture some sudden changes in volatility following an abrupt crisis and to accommodate different dynamic responses within each regime. The model is applied to the realized volatility of the S&P500 index: next to an interesting interpretation of the regimes in terms of market events, the MS-MEM has better in–sample fitting capability and achieves good out–of–sample forecasting performances relative to alternative specifications.