A machine-learned model prices every open market we scan from what the price alone misses: how such prices have really resolved, where profitable wallets put their money, their track record and the last day's move. Where its probability differs from the market's you get the gross edge, the net edge after the spread, Polymarket's fee and slippage, a confidence score and the model's reasons. Open a row to see them.
Fitted on markets that closed before Oct 17 and tested on the 6,468 that closed after, which it never saw. Its probabilities scored a Brier of 0.1698 against 0.1705 for the price alone (0.4% better; lower is better). The edge is small and real on average, not a sure thing on any one market.
| Edge called | Sides | Model said | Came true |
|---|---|---|---|
| 8+ pts | 17 | +10.3 pts | +39.9 pts |
| 5-8 pts | 270 | +5.5 pts | +8.7 pts |
| 3-5 pts | 2,056 | +3.9 pts | +2.2 pts |
| 1-3 pts | 2,955 | +2.0 pts | +1.6 pts |
“Came true”: how often those sides won, minus the price paid. Before costs.
| Time to close | Sides | Model said | Came true |
|---|---|---|---|
| up to 3 h | 2,665 | +3.4 pts | +3.3 pts |
| up to 12 h | 1,755 | +2.6 pts | +0.9 pts |
| up to 48 h | 606 | +2.3 pts | +1.1 pts |
| over 48 h | 272 | +1.9 pts | +5.0 pts |
Where its calls have come true less, its confidence is lower.
The model is a logistic regression retrained every six hours on every binary market the scanner watched resolve since Sep 23 (21,558 snapshots taken 1 hour to 7 days before the close). Prices are the live order book's midpoint where there is one; costs are worked out by buying the chosen stake from the live asks, with Polymarket's taker fee. Edges come and go in minutes and the record is short, mostly sports. Not financial advice.