An autonomous event-driven trading research system that uses LLMs to analyze stocks and run paper trading strategies. It's a personal project — not a product, not a service, not financial advice.
There are two main pieces:
The core pipeline simulates a small trading firm. Six AI agents — a fundamentals analyst, sentiment analyst, news analyst, technical analyst, options analyst, and a pair of bull/bear researchers — each look at a stock from their angle. They debate, then a trader agent makes a call. A risk manager and portfolio manager review it before anything happens. You give it a ticker and a date, it gives you a buy/hold/sell decision with reasoning.
The autoresearch system is the bigger experiment. It runs 12 event-driven strategies every day across 16 paper portfolios (4 time horizons × 4 portfolio sizes), tracks what works, and learns from the results. Each strategy looks at a specific kind of market signal — not price charts, but things like SEC filings, insider trades, and congressional trading disclosures. A portfolio committee (also LLM-powered) synthesizes the signals and sizes positions.
Each one watches for a different kind of event and generates trade signals:
- Earnings calls — clustering around earnings dates and estimate revisions
- Insider activity — Form 4 filings (when executives buy or sell their own stock)
- Filing analysis — anomalies in 10-K and 10-Q filings
- Regulatory pipeline — FDA approvals, FCC licenses, other regulatory signals
- Supply chain — stress indicators across supplier/customer networks
- Litigation — SEC enforcement actions and major lawsuits
- Congressional trades — stock trades disclosed by members of Congress
- Government contracts — federal contract awards (USASpending data)
- State economics — FRED macroeconomic indicators by region
- Weather/agriculture — NOAA weather anomalies, USDA crop conditions, drought severity
- Commodity macro — CFTC COT positioning extremes, futures curves, macro regime alignment
- Quantum readiness — post-quantum cryptography migration signals from SEC filings and news, regime-switching across PQC vendor/crypto-exposed/quantum hardware baskets
Data comes from about a dozen sources: yfinance, Finnhub, SEC EDGAR, OpenBB, FRED, NOAA, USDA, US Drought Monitor, Capitol Trades, CourtListener, Regulations.gov, and USASpending.
Daily cron job on a MacBook Air (16GB M4). The generation management system lets me A/B test different code versions in parallel using git worktrees — each generation gets its own frozen copy of the code and independent state. LLM costs run about $0.03/day per generation using Claude Sonnet and Haiku.
The 16 portfolios vary in size ($5k to $100k) and time horizon (30 days to 1 year). Bigger portfolios unlock more instruments: $10k+ can write covered calls, $50k+ can short stocks with margin and borrow cost gates.
git clone <this repo>
cd TradingAgents
pip install . # or pip install -e . for development
cp .env.example .env # add your API keysYou'll need API keys for at least one LLM provider (OpenAI, Anthropic, Google, xAI, or OpenRouter) and Alpha Vantage for market data. See .env.example for the full list.
# Interactive CLI — analyze a single ticker
tradingagents
# Daily automation — run all active generations
python scripts/run_generations.py run-daily
# Start a new generation (A/B test a code change)
python scripts/run_generations.py start "description of what changed"
# Dashboard
python -m streamlit run tradingagents/dashboard/app.pyDocker works too:
docker compose run --rm tradingagentsThis started as a fork of TauricResearch/TradingAgents, an open-source multi-agent trading framework from this paper. The core pipeline architecture comes from their work. Everything else — the autoresearch system, the strategies, the generation management, the portfolio committee, the paper trading infrastructure — was built on top.
The original TauricResearch core pipeline code is Apache 2.0 (see LICENSE-APACHE). All other code is proprietary (see LICENSE and NOTICE).
Not financial advice. Not investment advice. Not trading advice.