A comparison of Statistical and Neural Network models in volatility modeling

Abstract

This thesis compares statistical and neural network–based models for financial volatility forecasting. A rolling GJR–GARCH(1,1) model, a pure Long Short-Term Memory (LSTM) network and a hybrid GARCH–LSTM model are evaluated across gold futures, the EUR/ USD exchange rate and the OMXS30 equity index using five-day realized volatility. The results show that the pure LSTM performs best for two assets, while the hybrid model performs best for the equity index, highlighting asset-dependent benefits of hybridization.