athenara:~$ registry bench benchmarks/livetradebench

LiveTradeBench

A live forward-testing harness that runs LLM trading agents against streaming market data across US equities and Polymarket prediction markets, using a portfolio-allocation abstraction.

#benchmark #llm-agent #live-trading #portfolio-allocation #prediction-markets

added 2026-08-17 · PolyForm-Noncommercial-1.0.0 · external

$ git clone https://github.com/ulab-uiuc/live-trade-bench
collected 3 task environments
livetradebench::stock-trading
livetradebench::prediction-market-trading
livetradebench::portfolio-allocation

LiveTradeBench evaluates forward rather than backward: agents hold a cash-and-position portfolio and run repeated trading cycles against live data, which sidesteps the overfitting that backtests invite. Two design choices distinguish it. First, the agent interface is portfolio allocation across many assets, not a buy/sell call on one instrument. Second, it spans two market structures with different dynamics — US equities and Polymarket prediction markets — so a strategy that suits one is not automatically credited on the other. The data plumbing is named and real: a Yahoo Finance stock fetcher, a Polymarket CLOB fetcher, news from NewsAPI and Finnhub, and Reddit through PRAW. The repository is a self-hostable stack rather than a library alone, with a FastAPI backend, a frontend, and demo scripts. Three mock modes — mock agent, mock fetcher, and both — let the harness run offline with no LLM or market API keys, though real runs need model credentials.

The license is the constraint to read first. LiveTradeBench is released under the PolyForm Noncommercial License 1.0.0, which permits noncommercial use only and adds a competition clause withholding “the right to use this software in a way that competes with the licensor.” A separate LICENSE.COMMERCIAL file states that commercial use requires a license negotiated with the authors. GitHub’s license detector reports NOASSERTION for the repository; the PyPI metadata records the PolyForm identifier directly.

The accompanying work is an arXiv preprint with no venue (arXiv:2511.03628); the authors report evaluating 21 LLMs over 50-day live periods and conclude that high scores on general benchmarks do not translate into trading competence, with models displaying distinct risk preferences. Maintenance needs stating precisely: default-branch commits stop at November 2025, while pushes continued into February 2026 on side branches, and the PyPI package sits at 0.0.1. Branches and an example script exist for crypto and FX venues, but the README documents only equities and Polymarket — those are not shipped features. The project’s advertised leaderboard host, trade-bench.live, did not respond to connection attempts when this entry was verified; that is a single observation, not evidence the service has been retired.

trading [●●●··] moderate   ai [●●●●·] advanced   programming [●●●··] moderate   setup [●●●··] moderate

authors Haofei Yu, Fenghai Li, Jiaxuan You, ulab-uiuc (University of Illinois Urbana-Champaign)
origin external
license PolyForm-Noncommercial-1.0.0
markets multi-asset

athenara:~$