Project note · October 2026

Gecko: an AI forecaster for prediction markets

What started as “can an AI trade and make very high returns?” turned into a more useful experiment: find the one corner of the markets where a careful research agent has a genuine edge, and measure whether it does.

Why not just trade stocks or crypto?

Liquid markets like large-cap stocks and Bitcoin are priced by funds with enormous data and research budgets. Retail bots that appear to beat them are almost always overfit to past data, ignore trading costs, or survive only until the first large drawdown. Anything that can go up tenfold can also go to zero, and for most people it does. So the honest answer to “very high returns” is no, and the useful question became: where does a small team with good engineering and a language model actually have an edge?

Why prediction markets

On Kalshi and Polymarket, people trade contracts that pay one dollar if an event happens. Prices there are set mostly by retail opinion rather than quant desks, markets resolve on facts that can be researched, and the pools are small enough that big money ignores them. That plays directly to what an AI research agent does well: read the resolution rules literally, find dated primary sources, establish a base rate, and form a calibrated probability.

What the system does

Scan. Find liquid binary markets on Kalshi and Polymarket that resolve within a couple of months.
Research. For each one, Claude researches the question with web search and produces a probability and a confidence score, with evidence and sources.
Decide. Trade only when the model’s estimate differs from the market price by more than fees and spread. Size positions with fractional Kelly, capped at a few percent of the bankroll.
Guard. Hard risk limits sit in front of every order: a kill switch, a daily loss limit, concentration limits, and a cap on time to resolution.
Record. Everything runs in paper mode first, with simulated fills against the real order book and a ledger of forecasts, trades, and research costs.

How we’ll know if it works. Not by profit at first. The key metric is whether the model’s probabilities are better calibrated than the market’s own prices, measured with Brier scores across many resolved markets. If the model is not sharper than the market, no clever sizing fixes that.

What to expect

Likely	The AI agrees with the market most of the time and finds a handful of real mispricings around misread rules and stale prices.
Good result	A few genuine edges per month, sized sensibly, with a research cost that is clearly smaller than the edge.
Also a result	Finding out the model is not better than the crowd, cheaply, before any real money is involved.

Paper trading runs for weeks before real capital is considered. A market-neutral crypto funding-rate strategy is the planned low-risk base layer if steadier returns are wanted alongside the forecasting engine.