Learning to simulate realistic limit order book markets from data as a World Agent.#

1. Authors:#

Andrea Coletta, Aymeric Moulin, Svitlana Vyetrenko, Tucker Balch.

2. Affiliation:#

J.P. Morgan AI Research (for the first, third, and fourth authors), and Balyasny Asset Management, L.P. (for the second author).

3. Keywords:#

GANs, synthetic data, time-series, financial markets.

4. Urls:#

Paper: https://doi.org/10.1145/3533271.3561753, Github: None

5. Summary:#

  1. This paper aims to address the problem of multi-agent market simulators requiring careful calibration to emulate real markets and presents a world model simulator that accurately emulates a limit order book market.

  2. Previous work uses multi-agent modeling, which is a natural bottom-up approach to emulate financial markets, but modeling a realistic market through a multi-agent simulation is still a challenge as specifications about how the agents should behave and interact are not obvious, and there are unknown proprietary trading strategies. Publicly available historical data does not include attribution to the various market participants, making calibration of the agents difficult. The proposed approach in this paper is well-motivated as it introduces a world agent simulator that requires no agent calibration, but rather learns the simulated market behavior directly from historical data.

  3. The proposed approach introduces a unique “world” agent that is intended to emulate the overall trader population, without the need of making assumptions about individual market agent strategies. The world agent is learned from historical data, where the models for it are implemented as a Conditional Generative Adversarial Network (CGAN) and a mixture of parametric distributions.

  4. The proposed approaches consistently outperform previous work, providing more realism and responsiveness, and qualitative and quantitative evaluations show that the proposed models accurately simulate historical financial markets’ data. The performance achieved by the proposed methods supports their goal of accurately emulating a limit order book market.

7. Conclusion:#

  1. This work addresses the problem of multi-agent market simulators requiring careful calibration to emulate real markets and proposes a novel solution using a world agent simulator that accurately emulates a limit order book market through the use of GANs and synthetic data. The proposed approach has the potential to significantly advance the field of financial market simulation, as it eliminates the need for agent calibration and can accurately simulate historical financial market data.

  2. The innovation point of this article lies in the use of a world agent simulator that requires no agent calibration and learns the simulated market behavior directly from historical data using GANs and synthetic data. The performance of the proposed models is consistently better than previous work, providing more realism and responsiveness. The workload required for this approach is not discussed in detail in the article, which may be a potential weakness. Overall, this article is a valuable contribution to the field of financial market simulation, but further research is needed to assess the potential impact of this approach on practical applications.