Algorithmic Trading with Machine Learning

Abstract

More than 70% of today's stocks trade volume is attributable to automatic order execution by trading algorithms. Yet, despite convincing experimental results, machine learning has so far played only a minor role in algorithmic trading compared to traditional rule-based methods. One reason for practitioners' reluctance is that algorithmic trading with machine learning requires a deep understanding of what the underlying model learns and how it behaves when market conditions change. This dissertation therefore addresses two research questions: (1) How can machine learning algorithms be optimized for algorithmic trading? (2) What do these optimized algorithms learn and what is their advantage over traditional rule-based models? The dissertation contributes to the transition from rule-based to machine learning techniques in algorithmic trading by introducing new machine learning-based trading algorithms that are superior to rule-based models in directional trading as well as in order execution and market making. Specifically, it provides methods for modeling the market tailored to firms' individual trading objectives using novel target variables and loss functions for the learning process, and empirically demonstrates that market participants can benefit equally from improved market quality.