Machine learning approaches for market manipulation detection insurveillance systems

This thesis explores the use of machine learning for detecting market manipulation, in collaboration with the French financial regulator (Autorité des Marchés Financiers). Detecting manipulation poses unique challenges, including noisy high-frequency data, scarce labeled cases, evolving behaviors, and industrial constraints. Our approach combines a review of definitions and detection methods and the design of a modular ensemble-based system for market surveillance. From an application standpoint, we exploited historical order data and manipulation cases documented by the regulator, focusing on the Layering pattern. Experiments highlight both the potential and the limitations of ML in this setting: case-specific models outperform existing rule-based algorithms, but scaling to a global surveillance system requires advances in rule aggregation, evaluation protocols, and system integration.