Consistent time travel for realistic interactions with historical data: reinforcement learning for market making

Reinforcement learning works best when the impact of the agent's actions on its environment can be perfectly simulated or fully appraised from available data. Some systems are however both hard to simulate and very sensitive to small perturbations. An additional difficulty arises when a RL agent is trained offline to be part of a multi-agent system using only anonymous data, which makes it impossible to infer the state of each agent, thus to use data directly. Typical examples are competitive systems without agent-resolved data such as financial markets. We introduce consistent data time travel for offline RL as a remedy for these problems: instead of using historical data in a sequential way, we argue that one needs to perform time travel in historical data, i.e., to adjust the time index so that both the past state and the influence of the RL agent's action on the system coincide with real data. This both alleviates the need to resort to imperfect models and consistently accounts for both the immediate and long-term reactions of the system when using anonymous historical data. We apply this idea to market making in limit order books, a notoriously difficult task for RL; it turns out that the gain of the agent is significantly higher with data time travel than with naive sequential data, which suggests that the difficulty of this task for RL may have been overestimated.