You commented TRADE. Here is the setup from the reel. TradingAgents is a free, open-source project. You give it a stock and a date. Four AI analysts read the fundamentals, sentiment, news and technicals. A bull researcher and a bear researcher argue it out. A trader, a risk team and a portfolio manager then reach a decision, and you get a report.
Read this first
It gives you a decision and a report. It does not place trades and it is not connected to a broker.
The authors say it is designed for research and is not financial, investment or trading advice. I'm not giving financial advice either.
The code is free. The model you run it on may not be. A cloud model like Claude bills you per run through its API. A local model through Ollama costs nothing.
Install
You need Python 3.11 or later. These are the README's own commands.
git clone https://github.com/TauricResearch/TradingAgents.git
cd TradingAgentsuv venv --python 3.13
source .venv/bin/activateuv pip install .Pick your model
Option A: Claude. Set your Anthropic key in the same terminal:
export ANTHROPIC_API_KEY=...Option B: free and local. Use Ollama. Pull a model with ollama pull <name>. In the CLI, pick "Custom model ID" for any model it does not list. No key needed.
Run your first analysis
tradingagentsA screen opens. Pick your ticker, analysis date and LLM provider. When it asks, save the report. The saved folder includes complete_report.html, the whole report as one page with its sections listed beside the text.
Read the bear
The README says the bull and bear researchers assess the analysts' findings in structured debates, balancing potential gains against risks. That debate starts once all the analyst reports are in.
Open the report and find the researcher debate. I have not checked the exact section names. Read the bear's side before the final decision. It is the argument most of us skip.
The catch
Run the same stock and date twice and the answers can differ. The README says this is expected for a tool built on language models, not a defect. Two causes:
Model sampling is not deterministic, even at a fixed temperature.
Live data moves. News and social sources reflect now, even for a past trade date.
You can lower the temperature, but the README says the current default models largely ignore it. So never trust one run.
Backtest it
One run gives one decision, which cannot tell you whether the system decides well. The backtest runs the same pipeline over a grid of tickers and dates, then scores the decisions whose holding window has since traded.
tradingagents backtest NVDA,AAPL --start 2026-06-01 --end 2026-08-01 --every 7Each cell is scored on realized alpha against the instrument's regional benchmark, grouped by rating. Alpha means the return beyond that benchmark. It is a score for the past, not a forecast.
Every cell is a full run. On a cloud model each one is paid, so keep the first grid small. The README also says backtest results are not guaranteed to match any published figure. Treat it as a research scaffold, not a strategy.
If you run a backtest, reply and tell me what it scored.
Mike
PS: I build lead-gen systems for local businesses at OptiMAX, optimax-ai.com.