Methodology

1 · The pipeline

Every five minutes, the scanner pulls active sports markets from Polymarket, prices each fixture with the model, and compares. A signal fires only when every gate passes:

• A liquidity floor — thin books produce phantom edges
• A bounded edge band — both a minimum and a maximum; see section 3
• One direction only — the side the data supports; see section 4
• One signal per market — no re-firing while a position is open

The exact thresholds are part of the product and ship in the full model card with Pro.

Settlement is read from the chain. When a market resolves through the UMA oracle, the settler marks the signal won or lost against the on-chain outcome. We never grade our own results.

2 · The model

Two components, both deliberately boring:

• Elo ratings seeded from international results and FIFA rankings give a win/draw/win probability for every fixture, with a draw factor calibrated to international football (draws happen roughly 26% of the time — markets often price them much lower).
• A Poisson goal model estimates expected goals per team, which prices over/under markets directly.

No machine learning, no news feeds, no sentiment. Simple models are easier to calibrate, and calibration is the entire game: a model that says 22% must be right about 22% of the time in the long run.

3 · Bounded edges — a conservatism policy

Edge is the gap between the model's probability and the Polymarket mid-price. Signals fire only inside a narrow band of modest disagreement — below the floor there's nothing worth trading, and above the cap we assume the model is wrong, not the market.

That cap is the least intuitive part of the system: when a model disagrees with a liquid market by a lot, the likeliest explanation is model miscalibration at extreme prices, so we stand aside. Be clear about what our sample can and cannot prove — with a few dozen settled picks no sub-slice is statistically significant, which is exactly why the band is set by policy rather than curve-fit to early results.

The band's exact bounds and per-band settled results ship in the full model card with Pro.

4 · Model fine-tuning — what we cut, and why

• Draw fades. Betting against overpriced draws won more often than it lost and still lost money — short-priced fades must win at a very high rate to profit, and these didn't. Removed from the live strategy.
• Extreme edges. Capped as policy — see section 3. When the model and a liquid market disagree strongly, we assume the model is wrong until proven otherwise.
• Sportsbook-vs-Polymarket arbitrage. We ran a strategy comparing devigged Bet365 odds to Polymarket for a week. Bet365 and Polymarket agree within noise — no persistent edge to harvest. Shelved, and we tell you so instead of pretending it works.
• Our own inflated numbers. In July 2026 we found a duplication bug that had counted one winning market thirteen times in our published record. We pruned 70 duplicate rows and restated everything on this site downward — from "94 settled, +30% ROI" to "24 picks, +9% ROI". Publishing the correction is the product.

5 · Limitations

• The published record is simulated at flat stakes against signal-time mid-prices. It does not model slippage or fill quality; live results at size will differ.
• The sample is 24 picks from one tournament. The method extends to any competition Polymarket lists with liquidity, but each new competition resets the burden of proof.
• Longshot strategies draw down hard between wins. Roughly three of four signals lose; the math works over dozens of signals, not three.