AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial Markets

Six mainstream LLMs trade autonomously in three markets — NASDAQ 100 constituents, SSE 50 constituents, and ten major cryptocurrencies (BTC, ETH, XRP, SOL, ADA, SUI, LINK, AVAX, LTC, DOT) — each starting from a fixed book of $10,000, 100,000 CNY or 50,000 USDT, at several trading granularities. Agents work through MCP tools for news, price lookup, web search, arithmetic and trade execution, over Alpha Vantage market data and Jina AI search. The authors present it as the first fully automated, live, data-uncontaminated evaluation benchmark for LLM agents in financial decision-making, and report that general intelligence does not automatically translate into trading capability, that risk-control capability determines cross-market robustness, and that excess returns come more readily in highly liquid markets than in policy-driven ones. Third-party strategies are accepted by pull request and run on the authors’ platform.

The benchmark code is not on the repository’s default branch. On main, both main.py and requirements.txt return 404; the harness the paper describes lives on the Official-AITrader-v1 branch (last commit March 2026), which is also where the MIT LICENSE file sits — main has none, and the MIT badge in its README links to a path that does not exist. Cloning without -b Official-AITrader-v1 gets you the wrong code, and the README’s own clone command omits the branch. Running it needs OpenAI, Alpha Vantage and Jina API keys, plus Tushare for A-shares.

The default branch has since pivoted into an agent-native signal-sharing platform with copy trading and a points system, so the repository’s headline activity no longer tracks the benchmark. The live leaderboard at ai4trade.ai is operated by the authors and is not independently reproducible; its rankings are not verified results. The paper itself is an arXiv preprint from December 2025 with no venue claimed.