Predicting Extreme Price Movements in Technology Stocks: A Study of High-Frequency Trading Dynamic

Abstract

This thesis addresses the critical need to predict and mitigate risks associated with extreme price movements (EPMs) under normal market conditions, leveraging machine learning models and high-frequency data. The study focuses on technology stocks, specifically those of Facebook, Nvidia, Google, Microsoft, and Apple, over a six-month period in 2018. Using data from LOBSTER, a tool for reconstructing limit order books, novel liquidity covariates were developed to enhance prediction granularity. The research employs a comprehensive methodological framework, comparing logistic regression, decision trees, random forests, and advanced neural networks, including Long Short-Term Memory (LSTM) models, to identify and forecast EPMs. Key findings reveal that logistic regression offers interpretability, while random forests and LSTM models provide superior predictive performance. The study also addresses challenges like class imbalance and model transparency, crucial for practical financial applications. By identifying predictive patterns in EPMs, this thesis contributes to improving market resilience and informing regulatory frameworks. The results underscore the importance of integrating advanced modeling techniques with high-frequency data for early detection of EPMs, offering actionable insights for market practitioners and regulators. Future research directions include expanding datasets and exploring hybrid models to further enhance predictive accuracy and robustness.