Statistics and Probability Seminar
Speaker: Dr. Koushik Bera, IIT Guwahati
Host: Radhendushka Srivastava
Title: Statistical and Deep Learning Approaches to Volatility Forecasting in Financial Time Series
Time, day and date: 3:00:00 PM – 4:00:00 PM, Wednesday, July 29
Venue: Online (https://meet.google.com/zac-sksd-szo)
Abstract: Volatility measures the time-varying uncertainty in financial markets and plays a central role in risk assessment, portfolio allocation, and investment decisions. Its forecasting is challenging because financial time series are noisy, non-stationary, heteroscedastic, and driven by dynamics operating at different time scales. In this talk, I will discuss statistical and deep learning approaches developed to address these features. I will first describe how decomposition methods can be combined with Realized GARCH models to capture multi-scale volatility dynamics using high-frequency information. I will then present a leakage-aware forecasting framework based on optimized Variational Mode Decomposition and state-space learning, designed to avoid the use of future information. Finally, I will discuss heteroscedastic deep learning models that integrate GARCH-type volatility signals with extended LSTM architectures. The usefulness of these forecasts will also be examined through Value-at-Risk, stress testing, and volatility-based portfolio optimization. In particular, I will discuss how improved volatility forecasts can be translated into dynamic covariance estimates and used in minimum-variance and maximum-Sharpe-ratio portfolio strategies. This provides a way to evaluate not only statistical forecasting accuracy but also the economic relevance of the proposed models.