Lu Zhang, Drake University United College
Lu Zhang
Drake University United College
Unlocking the Secrets of High-Frequency Financial Data: Innovative Approaches to Modeling and Analysis
Join this talk via Zoom: Click to join
This dissertation investigates the predictability of intraday stock price movements in the final 30 minutes of U.S. equity trading, a period characterized by heightened volatility and strategic trading behavior. To move beyond the limitations of traditional Gaussian-based approaches, I develop a Bayesian regression framework with Student-t error terms, enabling more robust inference under heavy-tailed return distributions. Empirical results reveal a weakening of several established predictors alongside the emergence of new market dynamics, including a post–Federal Reserve “tug-of-war” effect among market participants.
The research further advances high-frequency financial modeling using 30-second interval data by constructing a comprehensive data engineering pipeline and applying state-of-the-art deep learning architectures, including DeepAR, Mamba, and Transformer-based models. Among these, a student-t–enhanced DeepAR specification delivers the strongest predictive performance. To improve interpretability, I introduce a game-theoretic framework that identifies order flow imbalances as a dominant driver of short-horizon price movements, providing actionable guidance for feature selection.
Overall, the findings highlight how robust statistical modeling, modern deep learning, and disciplined data engineering can jointly reshape empirical research and practical forecasting in high-frequency financial markets.