SPECS-Transformer: A Transformer Architecture for Volatility Forecasting with Signal-Spectrum Dynamics

Modeling volatility in financial markets has been a major challenge for many years, playing an essential role in portfolio risk management and in the pricing of derivative products such as options. This project proposes a methodological, theoretical, and practical approach, beginning with feature engineering applied to financial time series. The study is initially based on the use of a Variational Autoencoder (VAE) which, through probabilistic latent modeling, makes it pos- sible to generate the dynamics and potential trajectories of future volatility. As the study progressed, more complex architectures were introduced to improve the models’ sensitivity.