EdotEnv, founded by Rui and Michael, introduces self-improving Reinforcement Learning (RL) environments derived from quantitative trading workflows. They address the saturation of static benchmarks by leveraging the continuously evolving nature of financial markets, which naturally increase in difficulty as models improve.

These environments immerse LLMs in professional quant trading tasks, evaluating their ability to build predictive features, design portfolios, backtest strategies, and adapt to changing market regimes. Unlike synthetic benchmarks, EdotEnv uses real-world data, incorporates noise and trade-offs, and provides verifiable rewards. Initial tests with SOTA models revealed struggles with deep research iteration and a lack of understanding of trading dynamics.

EdotEnv aims to teach transferable research skills like applied ML, long-horizon planning, and continual learning. They have open-sourced a sample task repository and plan to sell these continuously improving environments to AI labs and enterprises. They seek feedback on agents tested within their environments and welcome discussions on the future of LLMs in trading.