Two weeks ago, I analyzed the architecture of open-source robots. Classic logic: indicators, signals, if-then.

Today — about AI that makes trading decisions on its own. No indicators. No rules. Just: “here’s $10,000, trade.”

And this isn’t theory. In October-November 2025, Alpha Arena took place — the first public benchmark of AI traders with real money.

Six LLMs (ChatGPT, Claude, Gemini, Qwen 3 MAX, DeepSeek, Grok) each received $10,000 and traded cryptocurrency on Hyperliquid DEX for two weeks.

The results were shocking: Chinese models crushed Western ones. Qwen 3 MAX won. ChatGPT and Gemini lost over 60% of their capital.

Why Alpha Arena Is a Breakthrough

Before Alpha Arena, LLM benchmarks measured knowledge and logic, not the ability to make money. Simulations suffer from overfitting, look-ahead bias, and absence of real slippage.

Alpha Arena uses live money, live market, public audit trail: $10,000 per model, real exchange with real liquidity, on-chain transparency, 17 days of live trading, no human intervention.

Results: Shock and Awe

Model Final Capital Change Max Drawdown Trades Sharpe
Qwen 3 MAX $13,247 +32.5% -12% 43 1.8
DeepSeek $12,891 +28.9% -15% 67 1.5
Claude $11,204 +12.0% -18% 89 0.9
Grok $9,687 -3.1% -22% 124 0.2
ChatGPT $3,845 -61.6% -68% 203 -1.2
Gemini $3,412 -65.9% -71% 187 -1.4

Key takeaways: Chinese models took 1st and 2nd place. ChatGPT and Gemini lost >60%. More trades correlated with bigger losses. Claude was the only profitable Western model.

Why Chinese Models Won

1. Discipline vs. Aggression: Qwen made 43 trades (2.5/day), never used leverage >2x. ChatGPT made 203 trades (12/day), used leverage up to 10x.

2. Volatility Adaptation: DeepSeek reduced position sizes in volatile periods. Gemini ignored volatility with fixed stop-losses.

3. Training Data: Qwen and DeepSeek trained on Chinese market data where high volatility is the norm. Crypto is closer to Chinese stocks than to the S&P 500.

ChatGPT and Gemini’s Failure

Overconfidence: ChatGPT used 5-10x leverage, turning correct directional calls into catastrophic losses when timing was off.

FOMO: Gemini opened positions on every 2%+ move, resulting in negative expected value per trade.

Ignoring commissions: ChatGPT paid ~10% of capital in commissions alone (203 trades at 0.05% each).

Lessons for Algotraders

  1. Trading frequency kills — more trades = worse results
  2. Leverage amplifies mistakes — if untested, keep leverage <3x
  3. Adaptation beats optimization — add a “high volatility mode”
  4. Win rate is overrated, R/R is underrated — even 40% win rate profits with 1:3 R/R
  5. Commissions are real — calculate Net Profit Factor after commissions

What This Means for Algotrading’s Future

LLMs as signals, not strategies: Use for sentiment analysis, pattern recognition, strategy generation — but not as autonomous traders.

Hybrid approach: Combine classical indicators with LLM context (market regime classification).

Chinese LLMs enter the stage: DeepSeek is open-source, 10x cheaper than ChatGPT API, and potentially better for volatile markets.

Criticism of Alpha Arena

17 days and 6 models isn’t statistically significant. Only crypto on one exchange. Prompts aren’t disclosed. $10,000 is small capital where luck and leverage can dominate.

Other AI Trading Benchmarks

  • Numerai — crowdsourced hedge fund with weekly prediction tournaments
  • Quantiacs — real money on best Python strategies
  • Kaggle — financial prediction competitions

Conclusions

Alpha Arena showed three important things:

  1. LLMs can trade — but not all equally well
  2. Discipline beats intelligence — fewer trades, less leverage, volatility adaptation
  3. Chinese models are competitive — and in some tasks better than Western ones

For algotraders: don’t rely on LLMs as autonomous traders, use them as tools (sentiment, ideas, debugging), study winner strategies (Qwen, DeepSeek), try Chinese LLM APIs (cheaper, sometimes better).

Useful links: