APARCH Models Estimated by Support Vector Regression

Abstract

This thesis presents a comprehensive study of asymmetric power autoregressive conditional heteroschedasticity (APARCH) models for modelling volatility in financial return data. The goal is to estimate and forecast volatility in financial data with excess kurtosis, volatility clustering and asymmetric distribution. Models based on maximum likelihood estimation (MLE) will be compared to the kernel based support vector regression (SVR). The popular Gaussian kernel and a wavelet based kernel will be used for the SVR. The methods will be tested on empirical data, including stock index prices, credit spreads and electric power prices. The results indicate that asymmetric power models are needed to capture the asseymtry in the data. Furthermore, SVR models are able to improve estimation and forecasting accuracy, compared with the APARCH models based on MLE.