Signal-to-Ticket
Reads a company's SEC 8-K filing and turns it into a sized trade ticket, or kills the trade and names the rule it broke.
- When
- Jul 2026
- Role
- Solo
- Built with
- Python, SEC EDGAR, ChromaDB, sentence-transformers, Claude, GitHub Actions
- Code
- Public code
- Links
- GitHub
- By the numbers
- 11 commits of mine over 2 working days, Jul 3, 2026 to Jul 6, 2026
What I built
When a company files an 8-K, the stock can move 5 to 20% before a person finishes reading it. This agent classifies the event, finds the five most similar past events and how the stock reacted, sizes a position with half-Kelly scaled down for volatility, checks the idea against a fund mandate, checks the thesis against the latest 10-Q, and writes a ticket with a confidence score and citations.
What was hard
Not every rule needs a model. A restricted ticker is a set lookup, so it's checked instantly and a kill there never reaches the LLM. The model only handles judgment calls a lookup can't, like sector caps against a free-text thesis. Splitting those two gates made the agent cheaper and made each kill easier to explain.
What I learned
A demo pipeline isn't a strategy. The README says plainly that the analogue set was chosen for variety, not tested for survivorship bias, and lays out what a real backtest would take.
Results
Unit tests run in GitHub Actions on every push.