Gaussian process-based algorithmic trading strategy identification

Many market participants now employ algorithmic trading, commonly defined as the use of computer algorithms, to automatically make certain trading decisions, submit orders and manage those orders after submission. Identifying and understanding the impact of algorithmic trading on financial markets has become a critical issue for market operators and regulators. Advanced data feeds and audit trail information from market operators now allow for the full observation of market participants' actions. A key question is the extent to which it is possible to understand and characterize the behaviour of individual participants from observations of trading actions. In this paper, we consider the basic problems of categorizing and recognizing traders (or, equivalently, trading algorithms) on the basis of observed limit orders. These problems are of interest to regulators engaged in strategy identification for the purposes of fraud detection and policy development. Methods have been suggested in the literature for describing trader behaviour using classification rules defined over a feature space consisting of summary trading statistics of volume and inventory, along with derived variables that reflect the consistency of buying or selling behaviour. Our principal contribution is to suggest an entirely different feature space that is constructed by inferring key parameters of a sequential optimization model that we take as a surrogate for the decision-making process of the traders. In particular, we model trader behaviour in terms of a Markov decision process. We infer the reward (or objective) function for this process from observations of trading actions using a process from machine learning known as inverse reinforcement learning (IRL). The reward functions learned through IRL then constitute a feature space that can be the basis for supervised learning (for classification or recognition of traders) or unsupervised learning (for categorization of traders). Making use of a real-world data-set from the E-Mini futures contract, we compare two principal IRL variants, linear IRL and Gaussian Process IRL, against a method based on summary trading statistics. Results suggest that IRL-based feature spaces support accurate classification and meaningful clustering. Further, we argue that, because they attempt to learn traders' underlying value propositions under different market conditions, the IRL methods are more informative and robust than the summary statistic-based approach and are well suited for discovering new behaviour patterns of market participants.