Alpha Arena (Nof1)
Reading research lab Nof1's AI trading experiment as structure rather than story — the real variable that split the results of six models starting from the same capital and the same market
Published 2026.07.01 · Updated 2026.07.07
Anyone searching for this case is really asking one thing: "If I hand trading over to an AI, does it make money?" Alpha Arena, launched by research lab Nof1 in October 2025, is one of the rare experiments that answers that question with measured data. Six AI models were each given about $10,000 and set loose to trade Hyperliquid perpetual futures with no human intervention — and the final results split from roughly +22% down to −63%. Same capital, same market, same data. This piece breaks down where that spread came from as structure — and the answer lay not in the models' intelligence but in the rules for handling losses.
- A public experiment Nof1 launched in October 2025 — six AI models were each allotted about $10,000, and the AI decided every entry, exit, and position size autonomously on Hyperliquid perpetual futures
- DeepSeek (DeepSeek Chat V3.1) led early, but final returns split from roughly +22% to −63% (historical, per public reporting, and not a typical result)
- The variable behind the spread wasn't predictive accuracy but how fast and how small the losses were when a model was wrong — a matter of stop rules and position sizing
- The fact that even an emotionless AI printed −63% shows that trading success isn't solved by removing psychology alone; it hinges on the design of the risk rules
Same Money, Same Market, Different Brains — The Experiment's Design
Alpha Arena is an AI trading experiment run by research lab Nof1. It isn't a contest promoting some particular bot or product — it's closer to a public benchmark observing 'how a language model actually makes decisions with real money in a real market.' To control the variables, the conditions were kept as simple as possible — identical capital, identical arena, zero human intervention. Thanks to that control, the differences in outcome can be read not as 'luck' or 'unequal conditions' but as differences in each model's decision-making structure.
- Enter six different AI models
- Give each model the same real capital — roughly $10,000
- Unify the trading arena as Hyperliquid, a crypto perpetual futures exchange
- Allow no human involvement whatsoever in entry, exit, or position-sizing decisions
- Compare performance by return over a fixed period
Started October 2025 · 6 AI entrants · about $10,000 in seed capital per model · arena: Hyperliquid perpetual futures · no human intervention. Because everyone started from the same conditions, the entire spread in final results came from trading decisions alone.
The Scorecard — An Early Leader and the Final Spread
Per public reporting, DeepSeek (DeepSeek Chat V3.1) led early in the experiment. Here we have to flag the first trap — 'the leader at a given moment' carries less information the shorter the sample. A few days' ranking can be nothing more than the market regime happening to line up with that model's disposition, and in fact the standings kept shifting afterward. It's exactly the same structure as why you shouldn't judge a trader's performance by 'last week's return' alone.
Final results ranged from roughly +22% to −63%. That came out of the same $10,000, the same market, and the same price data. These figures are historical and based on public reporting, and are not a typical result that anyone can reproduce.
What separated the standings wasn't 'which model is smarter' but 'how small the losses were when it was wrong' — an interpretation raised again and again in discussions of the public results.
Why It Split Between +22% and −63% — The Structure of the Spread
Looking at the exact same market data, some models protected their assets while others lost more than half. It's hard to argue that prediction was the divider — all six models had stretches where they got the direction right and stretches where they got it wrong. The difference showed up in how they handled the wrong stretches. The models that cut losing positions small still had capital left for the next opportunity, while those that bet big and stopped out late slid into an unrecoverable zone after just one or two failures.
The reason it becomes unrecoverable is arithmetic. An account down −63% needs roughly +170% just to get back to breakeven. Losses pile up linearly, but the return required to recover grows non-linearly — and that asymmetry is the mathematical basis for the principle covered in risk/reward: that not losing big in a single shot comes before any strategy.
Perpetual futures are an instrument that carries leverage, so a judgment error gets amplified into account loss and, in the worst case, ends in liquidation. The exact sequence in which liquidation actually unfolds is covered in the liquidation mechanics piece. AI or human, no one gets to dodge this structure itself.
What separated the top from the bottom wasn't predictive power but risk management — position sizing that limits the scar a single failure leaves on the account, and entry and exit principles that cut losses by rule instead of by emotion. This experiment showed, under controlled conditions, that those two things decide the outcome.
That Even an Emotionless AI Printed −63%
The part of this experiment most worth chewing on is elsewhere. The causes usually blamed for trading failure — fear, greed, revenge trading — are all human emotions. And yet an emotionless AI still produced −63%. This confirms, in inverted form, the proposition covered in the trading psychology piece: removing psychology is a necessary condition, not a sufficient one, and if the rules themselves — loss limits, invalidation criteria — aren't designed in, the account collapses even without emotion.
Conversely, the structure behind the side that held onto +22% reads through the same frame. The interpretation that fits the public results is that it came not from brilliant prediction but from systematically holding down 'the cost of being wrong.' The basic principle of risk management — that as long as capital remains there's a next chance — was, in effect, observed again in a controlled experiment whose sample was AI rather than humans.
The Limits of Reading This Experiment
Overreading has to be guarded against. First, the sample is one season, one market regime. Put the same models into a different regime and the standings could flip; the very fact that the early leader wasn't the final leader shows how unstable the rankings are. Second, 'risk management was the divider' is itself an interpretation of the public results, not proof that any particular model is superior. Third, if the experiment's conditions — capital size, duration, instruments — change, the distribution of results can change too — this is not a result whose reproducibility has been verified.
This case isn't a piece recommending any particular AI or automated trading. Perpetual futures are an instrument that carries leverage and liquidation risk no matter the tool, and automation only reduces the psychology problem — it doesn't fill in for the absence of risk rules. What the experiment showed isn't 'AI makes money' but the structure that 'an account without rules collapses no matter who's at the wheel.'
What Alpha Arena left behind is a counterexample under controlled conditions. If performance split from +22% to −63% on the same capital and the same market, then the variable is not intelligence but the rules that cut losses and position sizing — and even an emotionless AI collapsed without them. That's why Whale Story reads this case not through the lens of 'let's hand it over and copy them' but of 'what keeps an account alive.' To carry this structure into your own trading, we suggest the order of first learning the arithmetic of loss asymmetry through risk/reward, then confirming in trading psychology the proposition that rule design comes before emotional control.
FAQ
What exactly is Alpha Arena?
It's an AI trading experiment that research lab Nof1 launched in October 2025. Six AI models were each allotted about $10,000 and set to autonomously trade crypto perpetual futures on Hyperliquid with no human intervention, after which their returns were compared. It's less an ad for a particular product than a public benchmark run for observation.
Does AI trade better than humans?
This experiment alone can't settle that. Final returns split from about +22% to −63% (historical, per public reporting, and not a typical result), and even the same AI produced opposite outcomes depending on its risk rules. Being an AI doesn't make the loss risk of a leveraged instrument disappear.
Why did the six models' results differ?
The interpretation raised again and again in the public results is a difference in loss handling, not predictive power. The models that cut small when wrong protected their capital, while those that bet big and stopped out late were pushed into a zone that's hard to recover from. The asymmetry that a −63% account needs roughly +170% to recover its principal is the heart of that structure.
So should I just hand my trading over to an AI?
This article isn't content recommending automated trading or any particular AI. Automation only reduces emotional interference; without risk rules like loss limits and invalidation criteria, the account can collapse just as it did in this experiment. The judgment and the responsibility are yours, and we suggest learning the basics of risk/reward and risk management first.