ETD

• Deep Learning
• Econometrics
• Interpretability
• Machine Learning
• Realized Volatility
• XAI

Forecasting realized volatility has a central role in risk management, portfolio optimization,
and derivative pricing. While traditional econometric models often struggle to capture
the nonlinearities and complex interactions inherent in financial markets, on the other side
recent advances in deep learning may produce stronger predictive performance, but their
lack of interpretability limits practical adoption in finance.
This thesis implements an interpretable deep learning architecture for realized volatility
forecasting that integrates not only historical volatility patterns but also a rich set of
external predictors, including macroeconomic variables, technical indicators, and sentiment
indicators obtained through NLP techniques. The proposed model leverages modern
attention based interpretability techniques in order to improve transparency into the drivers
of volatility dynamics while maintaining high predictive accuracy. Moreover, adoption
of a stacking ensemble architecture with HAR model, a classical econometric model for
volatility modeling, is able, in some cases, to improve forecasting accuracy still retaining
overall interpretability of the system.
Empirical evaluation is conducted on European stock data, with comparisons against
benchmark econometrics and machine learning models. The results indicate that the
proposed model and architecture outperforms traditional baselines in forecast accuracy and
yields valuable insights on relative importance of economic, technical, and sentiment-based
predictors.
This research project contributes practically, by offering a transparent and robust tool
for financial decision-making leveraging both deep learning architecture and classical
econometrics models and by evaluating models on EU stock market.