AAI
AAI (Automatic Arbitrage Intelligence) is a research instrument for how prediction markets set prices. Kalshi and Polymarket list the same events and disagree; AAI measures how and why. It cannot trade, by design.
IN BUILD
What it is
AAI treats each market as a separate observer of the same unknown probability. Each observer sees through its own fees, price steps, delays and settlement rules. The question is whether undoing those distortions recovers the true probability better than either market does alone, scored against markets that have already resolved.
It began as an automated arbitrage bot, which is where the name comes from. That design was removed after measurement showed its foundations were wrong in ways that would have lost money. The fee model was off by about seven times at even odds, a sign error made a guaranteed loss look like a guaranteed gain, and its headline metric could not come out bad, so it measured nothing. The old design is kept on a branch as history.
It cannot place an order. There is no exchange SDK, no wallet and no signing code. The cheapest way to make sure a research tool never trades is not to install the means.
What exists today
Private repository. Read-only clients for both markets, fee schedules built from the markets' own published parameters, and an append-only data store with content hashes. A collector runs at three speeds: a daily census of every open market, an hourly sweep of active ones, and a once-a-minute watch of 80 full order books. 177 tests. The research folders for hypotheses and runs are still empty: the data comes first, and no result is claimed.
Not done
No experiment has run yet, so nothing about price formation is claimed.
Polymarket publishes no history of bids and asks, so that record exists only from the hours the collector was running.
Next
Pre-register the first hypothesis and score it on markets resolved after the collection began.
Built with
Python DuckDB Parquet