Forecasting Realized Volatility Using Neural Networks

This diploma thesis investigates the effectiveness of neural network models, encompassing both feedforward and LSTM architectures, in forecasting volatility across three major forex currency pairs. Traditional econometric models often struggle to capture the intricate dynamics of financial data, whereas NNs show the ability to cope with these complexities. The analysis focuses on applying both NN frameworks to volatility prediction and juxtaposing their performance against the commonly employed Heterogeneous Autoregressive models. The modeling phase entails deploying both feedforward and LSTM NNs while taking into account also supplementary parameters such as jumps, quarticities and semivariance. The findings indicate that simpler models like HAR demonstrate superior performance in out-of-sample forecasting owing to their enhanced generalization capabilities, whereas NNs exhibit strong in-sample performance. Upon examining the distribution of out-of-sample data, NNs demonstrate superiority up to the 95th percentile, particularly during extreme events, while HAR models excel in the most extreme scenarios. Notably, jumps and quarticities emerge as valuable additional parameters during extreme events.