Enhancing Volatility Forecasts with Neural Networks: a Comparison of GARCH, LSTM and their Hybrid Models for IMOEX Stock Market Index

This thesis aims to improve the accuracy of 1-day ahead conditional volatility forecasts for the IMOEX index, a major stock market index in Russia that represents the country risk and serves as a benchmark for investors. Accurate predictions of conditional volatility allow market participants to make informed decisions, adjusting trading portfolios and implementing effective hedging strategies. Traditional approaches involve the estimation of GARCH-family models, while recent advancements in technology have proposed an alternative method which relies on deep learning algorithms. To leverage both approaches, recent studies have focused on the development of hybrid models. However, the Russian market have been largely underrepresented in this analysis. In this paper, we will close the country gap by introducing the first hybrid volatility model for Russia. This model will be based on the LSTM algorithm with a new specification. Our analysis considers a wide historical period of daily data from 2015 to 2023, accounting for the Covid-19 pandemic and changes in the geopolitical environment in Russia. We compare the performance of a hybrid model with a standalone LSTM and GARCH-family structures, considering three rolling window scenarios. Our results suggest the absolute superiority of LSTM-based models over traditional econometric approaches in both MSE and QLIKE loss functions. However, we find that the choice between LSTM and a hybrid depends on the size of the rolling window used to make predictions. The best results were achieved for estimation in 7-day periods, and we suggest its use for risk management purposes.