Recent years have seen a dramatic increase in algorithmic trading, to the point that the majority of orders in major equity exchanges today are generated by machines without direct human control. Experience has shown that this automation--particularly at the extremes of high-frequency trading (HFT)--makes a qualitative difference, due to the unprecedented speed and lack of direct human control. The emergence of HFT raises fundamental issues for the efficiency, fairness, and stability of financial markets. The practice is highly controversial, yet the lack of scientific understanding of HFT's implications impedes informed public debate bearing on the question.
Algorithmic trading can also be viewed as herald of a wave of automated behavior with far-reaching effects in a plethora of domains. Methodological improvements from this project advance our ability to anticipate effects of autonomous agents in other areas of major economic and societal impact.
This project conducts a systematic computational study of algorithmic trading. The investigation combines online learning and optimization techniques from the point of view of theoretical machine learning and agent-based modeling (ABM) approaches to develop models of financial trading substantially more comprehensive and robust than heretofore possible. Modeling financial markets as multiagent systems affords heterogeneity: traders differing in objectives, information (access to data and observability of the environment), and response capability (processing and execution speed). Learning and decision-theoretic methods provide a principled basis for defining adaptive strategies that are effective across a broad range of operating conditions and possess guarantees in adversarial environments. Evidence on algorithmic trading implications is derived through systematic computational experimentation.
The project contributes both to scientific knowledge about algorithmic trading, and to agent-based methodology for analyzing complex strategic domains. One particularly novel feature of this study is its emphasis on the effect of temporal structure (e.g., communication latency patterns, adaptive strategies) on the dynamics of algorithm interaction. The agent-based methodology developed here provides a unifying framework for selecting among candidate behaviors based on specified solution concepts, such as game-theoretic or evolutionary equilibria. It exploits ideas from several fields, including simulation modeling, stochastic search, statistical analysis, and machine learning.