Finding Edge in Prediction Markets
How I approach probability estimation and order placement on Polymarket using on-chain data and LLM-assisted event modeling.
Prediction markets are an unusual trading environment. The resolution is binary, the odds move fast, and most liquidity comes from retail punters reacting to news rather than base rates.
The core loop
My bot’s loop is simple: ingest the CLOB order book, fetch related on-chain data (resolution sources, oracle contracts), run a calibration pass via Claude to estimate implied probability drift, and place limit orders when the spread is fat enough.
The LLM step isn’t doing heavy lifting — it’s pattern-matching against historical resolution data and flagging obvious mispricing relative to comparable past events.
What actually generates edge
- Speed on newly-listed markets before arbitrageurs arrive
- Correct resolution-source reading (many traders misread the oracle contract)
- Staying out of illiquid markets with wide spreads and no natural counterparty
What doesn’t work
Sentiment signals. News-driven price moves happen before you can act. The edge is structural, not informational.