Learning Adaptive Trading Strategies for Market Models
Conference Year
January 2021
Abstract
Agent-based models (ABMs) are a natural choice for understanding many sociotechnical systems. In particular, stock markets are an excellent target. The constrained space of agent interactions should simplify model development, while the wealth of data produced by real-world markets should aid in model validation and verification. Additionally, a successful agent-based market model (ABMM) could serve as a valuable tool for understanding the impacts of system redesigns and policy changes. Despite these advantages, ABMMs have remained an academic novelty for the most part.We hypothesize that two primary factors limit the usefulness of ABMMs. First, failure to capture major mechanisms present in the real world system reduces the accuracy of such models. Second, simple agents with static strategies do not display the breadth of behaviors seen in real markets.We investigate these issues through the development of a detailed ABMM. Our model features a fragmented market structure, communication infrastructure with propagation delays, realistic auction mechanisms, and more. We populate this model with a variety of hand coded strategies, which serve as a baseline, along with learned strategies created by meta-reinforcement learning.The combination of more detailed market mechanisms and more intelligent agents leads to higher quality models that may be better able to inform design and policy decisions.
Primary Faculty Mentor Name
Safwan Wshah
Secondary Mentor Name
Brian Tivnan
Status
Graduate
Student College
Graduate College
Program/Major
Complex Systems
Primary Research Category
Engineering & Physical Sciences
Learning Adaptive Trading Strategies for Market Models
Agent-based models (ABMs) are a natural choice for understanding many sociotechnical systems. In particular, stock markets are an excellent target. The constrained space of agent interactions should simplify model development, while the wealth of data produced by real-world markets should aid in model validation and verification. Additionally, a successful agent-based market model (ABMM) could serve as a valuable tool for understanding the impacts of system redesigns and policy changes. Despite these advantages, ABMMs have remained an academic novelty for the most part.We hypothesize that two primary factors limit the usefulness of ABMMs. First, failure to capture major mechanisms present in the real world system reduces the accuracy of such models. Second, simple agents with static strategies do not display the breadth of behaviors seen in real markets.We investigate these issues through the development of a detailed ABMM. Our model features a fragmented market structure, communication infrastructure with propagation delays, realistic auction mechanisms, and more. We populate this model with a variety of hand coded strategies, which serve as a baseline, along with learned strategies created by meta-reinforcement learning.The combination of more detailed market mechanisms and more intelligent agents leads to higher quality models that may be better able to inform design and policy decisions.