Presenting Models for Realised Volatility at 12:15-13:30 in QA406.
On 30th January 2019 we have Dario Palumbo of the University of Cambridge visiting us. He is going to give a seminar titled Models for Realised Volatility at 12:15-13:30 in room A406 in Building Q, Budapest University of Technology and Economics, Faculty of Economic and Social Sciences, Magyar tudósok körútja 2, 1117 Budapest. Sandwiches will be provided. Please register.
Abstract: This paper sets up a statistical framework for modeling daily realised volatility (RV) data using Dynamic Conditional Scores models. It shows how a simple two component DCS scale model can be used to model efficiently long memory in strictly positive and fat tailed time series like RV. This approach differs from the popular Heterogeneous Autoregression (HAR) model for volatility, which is based on simply modeling the RV through regressing its current value on its past daily value, its past 7 days average, and its past 30 days average (daily, weekly and monthly average respectively). Despite its good forecasting performance, the HAR model fails to capture features of the RV data series, like the fatness f the tails of its empirical distribution, which can have a huge impact on Risk Management applications. Fitting the two component specification with leverage is then carried out assuming that the data are conditionally generated by a Generalised Beta of the second kind (GB2) distribution. This is also equivalent to estimating a model for logRV with an Exponential Generalised Beta of the second kind (EGB2) conditional distribution, of which the normal distribution is a limiting case. Moreover, given the sore’s robustness to outliers, fittinch such a model allows to identify in the residuals correlation other characteristics of the data series, like the presence of a daily effects. In addition the analysis of log RV also indicates presence of heteroscedasticity in the residuals. Moreover the relationship between the GB2 and EGB2 distributions suggests that this heteroscedasticity may be due to a dynamic shape parameters of the GB2 distribution, which in the EGB2 governs the scale. The DCS model is extended to allow for this possibility. Ultimately the forecasting power of the DCS model is compared with the HAR revealing similar forecasting performance besides its higher descriptive power.