Polymarket
Fully automatic LLM trading system for Polymarket prediction markets — archived in 2026-09, now paused.
Code2026Paused
- Three live accounts, one AI signal
- Scan to exit with zero human input
- Archived 2026-09, secrets removed

About this project
I built a fully automatic trading system for Polymarket prediction markets that runs with zero human input. Twice a day it scans the market, and an LLM researches the candidates on the web and returns structured picks. Hard-coded rules route them: only prices between 0.40 and 0.85 are bought with real money, the rest go to a paper account. It sizes orders with quarter-Kelly, monitors positions every 30 seconds, re-evaluates them automatically and exits on take-profit, stop-loss or a sell verdict.
The same recommendations fed three live-money accounts with different exit rules: the main account ran the full strategy, a baseline took every pick and sold only at −30%, and a low-price account bought only picks at 0.40 or below, with a −60% stop. It got there in stages: a browser-automation prototype (2026-04), the semi-automatic v3, the v4–v7 semi-auto main project, and finally the fully automatic AUTO v8.5.5.
I sealed the archive on 2026-09-20. Before uploading, I removed every private key, credential, browser session, AI-chat log and runtime log; by then the LLM keys were already disabled, because an upstream search-API change had sent costs out of control. The project is paused and no longer maintained. Trading prediction markets involves substantial risk of loss; nothing here is financial advice.
Version history
- 2026-09AUTO v8.5.5 three-account systemLatest version: the fully automatic pipeline driving three live accounts with three exit strategies.
- 2026-09Project summary report (PDF)Full timeline, final three-account results, incident log and lessons learned.
- 2026-07Weather-market side projectDaily max-temperature arbitrage on Polymarket, with its own wallet and strategy.
- Early docs and research PDFsTechnical docs for v1–v3, AI research PDFs from 2026-05 and curated project memory notes.
- 2026-05v4–v7 semi-auto main projectThe main stage, where the decision engine, re-evaluation system and dashboard matured.
- 2026-04v3 semi-auto initial versionPrototype: I placed the orders, Claude did the analysis, a bot ran tiered take-profit and stop-loss.
- 2026-04Browser-automation prototypeEarliest stage in the archive: browser automation driving Gemini Deep Research to find markets.
Full write-up on GitHub