Volatility Forecasting with Artificial Neural Networks: Can we trust them?

Abstract: This thesis investigates how two types of artificial neural network models (ANN), feedforwardneural networks (FNN) and long short-term memory (LSTM), used for realized volatility (RV) forecasting, perform during high and low volatility regimes in comparison to the heterogeneousautoregressive (HAR) model. This is done for 23 stocks, constituents of the Swedish index OMXS30, between the 8th of February 2010 and the 31st of January 2022 using ten exogenous and three endogenous input variables. We find the ANNs generally superior to the HAR model, but also a lack of robustness when investigating ANNs performance in different volatilityregimes. The study shows that HAR and ANN models have differing forecasting performances across the volatility range and that the variation is dependent on the regularization regime inplace. Where lower regularization supports enhanced accuracy during high-volatility days while higher regularization promotes performance during low-volatility days. In addition, the existence of a trade-off between model complexity and performance during high versus low volatility for LSTM models are confirmed, and it is concluded that this relation is conditioned upon the regularization.