risk-aware-trading-agents / site

# Risk-Aware Trading Agents

> A benchmark for LLM trading agents, with news replayed by the minute it was published: three frozen US stock markets (the 2020 crash, the 2022 bear market and 2026), a public leaderboard (rta-bench-1) that scores each model blind and with names, apart before and after its training cutoff, and every result a JSON file that replays to the same numbers. The results are based on 65,863 headlines and SEC filings replayed by publication time; the headline data has license restrictions and is not distributed with the package. The open-source Python package (MIT license) also has a prompt leaderboard, a point-in-time research kit with a look-ahead check, risk-aware AI trading agents for US stocks that check the macro calendar, the VIX term structure and cross-verified news before they buy, an MCP server, and paper or live trading through 9 broker adapters. Nothing in it is investment advice.

- Install: `pip install risk-aware-trading-agents` (Python 3.10+), or run the dashboard with `docker run --rm -p 8501:8501 ghcr.io/jackson-cmd/risk-aware-trading-agents dashboard`.
- Run the benchmark: `riskaware research bench --llm openai:MODEL` runs the six tasks (2026, 2022 and 2020, each blind and named) with seeds 0, 1 and 2; `riskaware research submit RUN_ID --model NAME --cutoff YYYY-MM` writes the entry for a pull request.
- Leaderboard models so far (last update 2026-11-14, 67 entries): claude-haiku-4.5 (Anthropic, training cutoff 2025-07), claude-haiku-5.5 (Anthropic, training cutoff 2026-06), claude-sonnet-4.6 (Anthropic, training cutoff 2026-01), gemini-3.5-flash (Google, training cutoff 2025-01), llama3.2 (Meta, training cutoff 2023-12), gpt-5.6-luna (OpenAI, training cutoff 2026-02-16).
- Research kit: three frozen periods (2026 to date, the 2022 bear market, the 2020 crash), each with daily prices for 64 large-cap stocks and 25 ETFs from two years before, the VIX term structure and the macro calendar, with SHA-256 hashes. Every experiment runs a look-ahead check.
- MCP server: `uvx risk-aware-trading-agents mcp` (stdio, 22 tools). The tools are read-only or work on local paper accounts; none can send an order to a broker.
- Brokers: Alpaca (tested on a paper account), Public, Tradier, Interactive Brokers, Charles Schwab, tastytrade, Futu/moomoo, Tiger and Longbridge, plus a built-in paper account.

## Guides

- [LLM trading leaderboard (rta-bench-1)](https://jackson-cmd.github.io/risk-aware-trading-agents/leaderboard.html): LLMs trade 56 US large caps through 2026, the 2022 bear market and the 2020 crash on news replayed by publication time, blind and named, three seeds each, scored apart before and after each training cutoff; how each model makes money; memory probes; a prompt leaderboard; questions and answers; how to submit a model
- [LLM trading agent benchmarks compared](https://jackson-cmd.github.io/risk-aware-trading-agents/llm-trading-benchmarks.html): rta-bench-1 next to StockBench, CLQT, LiveTradeBench, AI-Trader, InvestorBench, FinSABER, BlindTrade and Agent Market Arena on news timing, frozen periods, blind and named runs, training-cutoff splits and replayable submissions
- [News backtesting without look-ahead bias](https://jackson-cmd.github.io/risk-aware-trading-agents/news-backtesting.html): point-in-time replay, a 16:00 cutoff, orders at the next open, and a check that the backtest never saw the future
- [Do news headlines predict stock returns? Ten news factors tested point in time](https://jackson-cmd.github.io/risk-aware-trading-agents/news-factors.html): rank IC, quintiles, decay and price controls for ten headline factors in 2020, 2022 and 2026, and how much a calendar-date news join inflates them
- [Point-in-time US stock market data for 2020, 2022 and 2026](https://jackson-cmd.github.io/risk-aware-trading-agents/point-in-time-market-data.html): daily prices, the VIX term structure and the macro calendar, with SHA-256 hashes, for reproducible backtests
- [Test LLM trading agents on a frozen market](https://jackson-cmd.github.io/risk-aware-trading-agents/llm-trading-agents.html): trade a frozen market one close at a time with Claude, GPT, Gemini or local models, with a blind mode against memorization

## Documentation

- [README](https://github.com/jackson-cmd/risk-aware-trading-agents#readme): quick start, features, configuration, FAQ
- [Leaderboard](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/leaderboard.md): the rta-bench-1 tasks and settings, training cutoffs, how entries are checked by replay, and how to submit one
- [Research kit](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/research.md): the three frozen markets, experiments, the look-ahead check, LLM trading sessions
- [Agents and the MCP server](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/agents.md): paper-trading agents and how to connect Claude, Codex, Gemini or Cursor
- [News history](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/news-history.md): point-in-time news for backtests and how to add your own archive
- [News factors](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/news-factors.md): `riskaware research factors`, the ten factors, the tests and the results
- [Custom strategies](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/custom-strategies.md): write a strategy class and backtest it
- [Strategy lab](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/strategy-lab.md): let an AI write a strategy from a description and test it safely
- [Brokers](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/docs/brokers.md): keys, gateways and paper accounts for each broker
- [Changelog](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/CHANGELOG.md)

## Optional

- [Everything above in one file](https://jackson-cmd.github.io/risk-aware-trading-agents/llms-full.txt): the benchmark, the data, how to run and submit a model, the comparison with other benchmarks, the current results and answers to common questions
- [Leaderboard on Hugging Face Spaces](https://huggingface.co/spaces/jacksonwangpro1/llm-trading-agent-leaderboard): the same tables, read from the results file below
- [Leaderboard results as JSON](https://jackson-cmd.github.io/risk-aware-trading-agents/assets/data/leaderboard.json)
- [Dataset on Hugging Face](https://huggingface.co/datasets/jacksonwangpro1/rta-bench-llm-trading-benchmark): the prices of the three periods as Parquet, with the original files
- [PyPI package](https://pypi.org/project/risk-aware-trading-agents/)
- [Docker image](https://github.com/jackson-cmd/risk-aware-trading-agents/pkgs/container/risk-aware-trading-agents)
- [README in Chinese](https://github.com/jackson-cmd/risk-aware-trading-agents/blob/main/README_zh.md)