Show HN: RL agent trained only in synthetic markets got better on real 2020crash (github.com)

A reinforcement learning (RL) agent, trained exclusively in synthetic market environments, demonstrated its ability to hedge options effectively during real market crashes in 2020 and 2022. This model, which never encountered actual price data, was evaluated by comparing its hedging errors against those of untrained models and standard hedging strategies. Remarkably, the trained agent achieved a median hedging error that was significantly lower than that of the untrained counterpart, showcasing its capacity to transfer learned skills from synthetic training to real-world crises. This development is significant for the AI/ML community, particularly in the fields of finance and algorithmic trading, as it suggests that synthetic training environments can successfully impart valuable skills that carry over to live settings. The study's rigorous methodology—pre-registering the test parameters and ensuring reproducibility—adds to its credibility. The implications are profound: synthetic markets could serve as effective training grounds for developing RL agents capable of navigating complex and unpredictable real-world financial scenarios, thereby enhancing the robustness of automated trading systems.