Machine learning in finance: from earnings data analysis to algorithmic trading
File(s)
Author(s)
Ye, Zhengxin
Type
Thesis
Abstract
Traditional statistical methods often fail to capture the complexity of modern stock price movements. This doctoral thesis develops advanced machine learning approaches to improve intra-day algorithmic trading, with a focus on earnings data. A hybrid model combining Extreme Gradient Boosting (XGBoost) and Genetic Algorithms is introduced to analyze Post-Earnings Announcement Drift (PEAD). This framework effectively models non-linear links between earnings metrics—such as EPS, revenue, and return on equity—and stock behavior. Findings reveal that trading directly on earnings releases is limited by timing constraints, but a “delayed entry” strategy yields stronger profitability. The models also enable portfolio construction with higher positive returns and reduced downside risk, supporting market-neutral long-short strategies.
To enhance practical trading applications, the Human Aligned Trading (HAT) model is proposed. HAT integrates machine efficiency with human-like decision-making by combining Deep Q-Network reinforcement learning, supervised learning, multi-step Q learning, and imitation learning. A rigorous backtesting and profit evaluation framework ensures realistic performance assessment. Across diverse stocks, HAT consistently outperforms benchmark models.
The thesis further introduces the Contrastive Earnings Transformer (CET), which merges contrastive self-supervised learning with Transformer architecture. CET manages mid-frequency trading data and earnings information, delivering robust predictive accuracy across multiple industries.
Overall, this research provides a comprehensive guide to applying machine learning in financial markets, particularly around earnings-driven trading. By presenting innovative models such as HAT and CET, it advances both academic understanding and practical strategy design. The work lays a foundation for future studies to incorporate richer real-world data and address realistic trading constraints, highlighting the growing role of artificial intelligence in shaping financial market dynamics.
To enhance practical trading applications, the Human Aligned Trading (HAT) model is proposed. HAT integrates machine efficiency with human-like decision-making by combining Deep Q-Network reinforcement learning, supervised learning, multi-step Q learning, and imitation learning. A rigorous backtesting and profit evaluation framework ensures realistic performance assessment. Across diverse stocks, HAT consistently outperforms benchmark models.
The thesis further introduces the Contrastive Earnings Transformer (CET), which merges contrastive self-supervised learning with Transformer architecture. CET manages mid-frequency trading data and earnings information, delivering robust predictive accuracy across multiple industries.
Overall, this research provides a comprehensive guide to applying machine learning in financial markets, particularly around earnings-driven trading. By presenting innovative models such as HAT and CET, it advances both academic understanding and practical strategy design. The work lays a foundation for future studies to incorporate richer real-world data and address realistic trading constraints, highlighting the growing role of artificial intelligence in shaping financial market dynamics.
Version
Open Access
Date Issued
2024-08-24
Date Awarded
01/12/2025
License URL
Advisor
Schuller, Bjoern
Publisher Department
Department of Computing
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)