Polymarket prices are overconfident.
Correcting them pays, most near resolution.
How often prices come true
The model learns how far to move each price, separately for each horizon; the dots show what actually happened. Pick how far from resolution, and which kind of market.
Live decisions
Open markets right now: price, model, and the call after costs.
What we trained, and what it gave us
Ten models and rules, compared on validation only. The simplest model won; the money came from a rule, one bet per event.
Beats the market's own price out of time (Brier 0.1297 vs 0.1305). Shows the bias: near the end longshots are underpriced, favourites overpriced; a month out prices are fair.
An uncertainty band for every prediction; we bet only when its cautious end clears the costs.
Most bets lose (about 70%), so each stake stays small; max drawdown 0.22.
From +17¢ to about +30¢ per $1 after fees, with half the drawdown.
No gain over the simple model.
No better than the simple model.
−7.8¢ per $1 vs the simple model, walk-forward.
Domain terms overfit.
−10¢ per $1: it assumes overpriced longshots; Polymarket's are underpriced.
Model errors were 99% correlated, so averaging added nothing.
No extra profit, slightly smaller drawdown (0.20 vs 0.23).
Worse than the market (Brier 0.1471 vs 0.1399), −16.5¢ per $1.
Checks: walk-forward month by month (positive in 15 of 17), a placebo where prices are assumed right (−8¢), deflated Sharpe over the 12 rules tried. Formulas: cheat sheet (PDF).