Project

Polymarket Insider

Watches every large trade on Polymarket and flags the statistically unusual ones a few minutes after they happen.

What it does

Polls Polymarket's public trades API every 60 seconds for every trade of $500 or more, and backfills about six hours of history on startup.
Keeps running statistics per wallet and scores each trade from 0 to 100: a large first trade, a size spike far above the wallet's usual bets, and an Isolation Forest.
Flags a trade at 60 or more, so a rule and the model have to agree, and sends it to a Discord webhook and a small FastAPI endpoint.

How it works

An asyncio feed asks the API only for large trades, so consecutive pages overlap and nothing slips through its five-minute cache. Each wallet's mean and variance are updated with Welford's online algorithm, and every trade gets a z-score against the wallet's previous average. Two rules and an unsupervised Isolation Forest, retrained every hour on the latest 50,000 trades, add up the score.

The honest bit

A flag means "statistically unusual", not insider trading: nobody publishes confirmed insider trades, so precision and recall cannot be measured. The one recorded run lasted 26 minutes and flagged nothing; its top scores were all sports bets, most of them near-certain, which is the obvious next filter. The thresholds are judgement calls, not tuned values.

Python
asyncio
httpx
scikit-learn
SQLAlchemy
SQLite / PostgreSQL
FastAPI