# AI trading in real markets Source: https://nof1.ai HomeAlpha Arena Nof1 is an AI research lab building new models for financial markets. We believe markets are the next great domain for frontier AI research. They are open-ended, deeply consequential, and highly measurable. Why markets The current generation of LLMs are shockingly bad at navigating financial markets. They’ve read vast amounts of financial text, but struggle with even the simplest investing tasks, showing minimal signs of improvement between releases. Part of the problem is that markets are non-stationary. Things that worked well in the past most likely won’t work in the future. This does not fit into the current training paradigms used by frontier AI labs, despite working for domains like coding, math, and law. To train frontier models for markets, we need new techniques that can handle noisy, constantly changing environments. It’s in these environments where the most interesting, economically relevant problems live. New capabilities At the same time, the field of AI research is running out of good ways to measure AI progress. New benchmarks get quickly memorized and saturated, specifically because they don’t adapt alongside the models. This all leads to an interesting question: what if markets are a solution? What if we can use them as both an evolving benchmark and a training environment for new capabilities? This would directly address one of the biggest failure modes of modern LLMs: adaptability. Current models are overly static and backwards-looking, struggling to update or think ahead. Markets provide a natural setting for developing this capability, exposing weaknesses that static benchmarks often miss. They also provide something increasingly valuable for training AI systems: objective feedback. Earnings beat consensus and management raises full-year guidance A surprise CPI print Quarter-end rebalancing flows push against the rally Price breaks above its twelve-month high on rising volume Fed decides to raise interest rates The regime shifts to risk-off Earnings beat consensus and management raises full-year guidance A surprise CPI print Quarter-end rebalancing flows push against the rally Price breaks above its twelve-month high on rising volume Fed decides to raise interest rates The regime shifts to risk-off Our models We’re building new models for financial markets, alongside the environments and benchmarks needed to train and evaluate them. Our training objective is general market intelligence, something we’re constantly improving our measure of internally. We’ve identified a proto scaling law for financial intelligence, with signs of skill-transfer to other domains. Upon release, our models will be able to perform well-calibrated research, analyze large time-series datasets, do complex feature engineering, write market-aware code, make forecasts, and improve based on the outcomes of their decisions. An adaptive reasoner that can be used in any harness by hedge funds, individuals, aggregators, brokerages, and financial institutions. Forecasting Risk management Portfolio management Systematic trading Backtesting & simulation Execution Team Our team’s work spans across open-ended learning, world modeling, and large-scale RL, with alumni from DeepMind, Two Sigma, Citadel, Bridgewater, and Netflix. Research is led by Julian Togelius, the founder of NYU’s Open-ended Learning Lab, with 20+ years in frontier AI research. We’re backed by Village Global and leading AI researchers & founders including Tim Rocktäschel (DeepMind), Jack Parker-Holder (DeepMind), Oliver Cameron (Odyssey ML), alongside angels from Jane Street. ## Comments **thatalexpalmer.eth**: yeah this is very cool. curious how many people would entrust models with their money **timdaub.eth**: I actually did entrust some models with my money. Eg I once asked how to get the most exposure to LLMs by buying public stocks and based on the research I bought various tickers. But yeah I never had the model fully autonomously do that. I imagine it‘d be highly initial-prompt-dependent too, plus the context window eventually is fully **thatalexpalmer.eth**: What was the end result for you? Did it match expectations or no? **timdaub.eth**: So far, so good. I bought Google (up 44%) and Tesla (up 65%) and that's actually all. Must have done this like in the last 6 months. All other options had less than ten cents of the invested dollar going towards investing in LLMs actually. E.g. it ranked MS at 4 cents on the dollar, I think. Would be interesting to revisit and see whether the data is now different. **thatalexpalmer.eth**: Gotcha so the usual suspects. Very cool though. Have you seen this repo? https://github.com/virattt/ai-hedge-fund