SIM-GAN: Adversarial Calibration of Multi-Agent Market Simulators.

We look at the problem of how the simulation of a financial market should be configured so that it most accurately emulates the behavior of a real market. In particular, we address agent-based simulations of markets that are composed of many hundreds or thousands of trading agents. A solution to this problem is important because it provides a credible test bed for evaluating potential trading algorithms (e.g., execution strategies). Simple backtesting of such algorithms suffers from a critical weaknesses, chiefly that the overall market is not responsive to the candidate trading algorithm. Multi-agent simulations address this weakness by simulating {\it market impact} via interaction between market participants. Calibration of such multi-agent simulators to ensure realism, however, is a challenge. In this paper, we present SIM-GAN -- a multi-agent simulator calibration method that allows to tune simulator parameters and to support more accurate evaluations of candidate trading algorithm. Our calibration focus is on high level parameters such as the relative proportions of the various types of agents that populate the simulation. SIM-GAN is a two-step approach: first, we train a discriminator that is able to distinguish between ``real'' and ``fake'' market data as a part of GAN with self-attention, and then utilize it within an optimization framework to refine simulation parameters. The paper concludes with quantitative examples of applying SIM-GAN to improve simulator realism.