Forecasting asset volatility using autoregressive support vector regression model incorporating the intraday range measure and price information

Volatility is a measure of the instantaneous variability of a financial asset. High-volatility assets is often associated with high risk, highlighting the importance of precisely estimating the volatility. This paper proposes an autoregressive support vector regression (SVR) model integrating the lagged range-based Parkinson volatility measure and four lagged logarithmic prices ( SVR LagPK _ LagPrices ) jointly as predictor variables to capture the dynamics of volatility of asset returns. An empirical analysis based on the Standard and Poor’s 500 was adopted. We performed extensive comparisons among SVR models to determine the significance of integrating the predictor variables encompassing the lagged range-based Parkinson volatility measure and four lagged logarithmic prices, both jointly and singly in the autoregressive SVR models with different kernel settings. Additionally, the conditional autoregressive range (CARR) models were also evaluated. The in-sample results based on the two realised volatility measures that act as proxies for the unobserved true volatility, revealed two important findings: (i) Although the volatility estimates based on CARR models outperformed other SVR models in terms of root mean squared error (RMSE) and mean absolute error (MAE), the goodness-of-fit analysis results show that these models did not fulfil the underlying model assumptions, (ii) The SVR LagPK _ LagPrices   model generally predominates other SVR models for the in-sample model fit based on the RMSE and MAE. An examination of the SVR LagPK _ LagPrices model with linear kernel yielded the best out-of-sample forecasts, characterised by the smallest RMSE and MAE which were tested based on the mean squared error loss function using Hansen’s model confidence set.