On the generalization of machine learning models in finance: five essays on bridging the empirical gap
This thesis details advanced machine learning methodologies in high-frequency trading, featuring anomaly detection, box-target strategies, and end-to-end pipelines.
What it examines
The thesis presents innovative ML approaches to financial trading and anomaly detection in high-frequency data. It applies deep learning, boosting, and data augmentation techniques to improve predictions in market making, fraud detection, and automated trading pipelines, addressing challenges such as non-stationarity and transaction costs.
What it concludes
The research demonstrates that integrating ML techniques such as deep semi-supervised anomaly detection, era splitting, and enhanced data augmentation improves predictive performance in financial markets. Potential applications include fraud detection, automated trading, and risk management, with future work focusing on refining targets and addressing market-specific limitations.
Evidence objects
The thesis unveils a surprising 'negative drift' effect in limit order fills, revealing hidden costs when markets reverse, significantly impacting market makers and challenging traditional high-frequency trading assumptions using tests.
key_findings bullet 1 · key_findings · validation V0
Innovative era splitting and directional era splitting criteria enhance decision tree performance by incorporating environment-specific information, reducing overfitting and improving out-of-sample predictability for gradient boosted models amid market distributional shifts.
key_findings bullet 2 · key_findings · validation V0
A novel data augmentation method reorganizes limit order book data into image formats for convolutional neural networks, significantly accelerating training and prediction, while walk-forward cross-validation exposes profitability and computational challenges.
key_findings bullet 3 · key_findings · validation V0
Integrating novel theories and empirical analyses, this paper reinterprets limit order fill dynamics in high-frequency trading and introduces innovative splitting criteria for gradient boosted decision trees. It offers a fresh perspective through invariant learning, data augmentation, and anomaly detection. The work, while evolving existing paradigms, captivates readers with original insights.
key_findings bullet 4 · key_findings · validation V0
Raw abstract and provenance
- … Many ML/AI papers for finance come short of implementing an end-to-end trading … We propose that any high frequency finance order fill simulator needs to account for …
Source row: 1467 · abstract type: snippet