AI-driven trading research bot

Does reasoning beat indicators on long holds?

Not a product. An experiment with a falsifiable question: on positions held for days rather than seconds, does a language model reading the same data as a rules engine make better decisions than the rules engine?

Run on paper, with real prices, and a hard daily loss cap.

The model proposes, the risk engine decides size.

Claude reads market context and produces a thesis with a confidence number. The risk engine, which is ordinary deterministic code, turns confidence into a position size bounded by available balance and a daily loss cap. The model never sizes its own bets.

Every decision is logged with its inputs, so a bad trade can be replayed and argued about later.

Scheduler, exchange, ledger, dashboard.

Python 3.12. ccxt for exchange access, pandas-ta for indicators, Anthropic Claude Opus for the reasoning step, SQLAlchemy over aiosqlite as the ledger, APScheduler driving the loop, Streamlit for the dashboard.

Promising, and still an experiment.

One week on $100 of paper is a signal, not a result. It is reported here as what it is: an open experiment with its logs intact.