Launch HN: EdotEnv (YC S26) – Quant Trading RL Envs to Teach LLMs Research
edotenv.com Mzzzzz
Two ex-quants built RL training environments that use live markets to teach LLMs actual research skills, not just answers. Markets keep getting harder as strategies decay, so unlike normal benchmarks, this one can't be gamed into uselessness.
Based on reporting by edotenv.com, Mzzzzz — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Rui and Michael spent their careers as quants before deciding the eval problem in AI training was basically the same problem they'd been living with in markets. Benchmarks get memorized, gamed, or saturated the moment enough labs train against them. Markets don't have that problem, because every profitable trading strategy erodes the moment enough people exploit it. That self-correcting decay is exactly the property their new startup, EdotEnv, is trying to bottle for LLM training.
The idea, launched this week on Hacker News, is to take real quant workflows — building predictive features, running backtests, constructing portfolios, adapting to shifting market regimes — and turn each stage into a verifiable RL task using actual historical market data rather than synthetic data. In one example task, a model gets clean data from time 0 to T, has to research and build a feature, backtest it on an earlier slice, then get scored on how it performs trading forward on unseen data. No LLM judge needed, no human grader in the loop. The market itself tells you if the strategy worked.
What they found running frontier models through these environments is more interesting than the pitch itself. Current SOTA models tend to skim across many shallow research ideas instead of digging deep into one; throwing more reasoning compute at the problem doesn't reliably help; and, perhaps most tellingly, the agents don't really grasp what trading is. When a model starts losing money, instead of adjusting its strategy, it just stops trading. That's not risk management, that's giving up — and it points to a real gap in how these models handle long-horizon, adaptive decision-making under real economic stakes rather than static text problems.
EdotEnv has open-sourced a sample feature-engineering task on GitHub and plans to sell continuously updated environments to AI labs and enterprises training agents on long-horizon planning, continual learning, or ML research skills generally. Notably, the founders are explicit that they're not building a stock-tip machine — the goal is teaching transferable research behavior, using quant trading as a uniquely well-instrumented, endlessly renewing training ground rather than the destination itself.
My take — AI-written commentary, not fact-checked reporting
Using markets as a live, self-refreshing benchmark is a genuinely clever hack around the eval-saturation problem everyone in AI training complains about, and it's refreshing to see founders lean into a niche they actually lived rather than bolting a chatbot onto finance. That said, the finding that models quit trading when they lose money is the real headline here — it's a blunt reminder that today's agents can sound sophisticated while having zero instinct for risk, adaptation, or consequence, which should worry anyone deploying them in domains where quitting isn't an option.
Read more about this at: edotenv.com