Methodology
Every probability on a match card is read from a single scoreline grid, built from six seasons of data and checked forty ways before it is sent.
Probabilities are the model's estimates, not recommendations. Sample shown with illustrative figures.
A card as it arrives by email. Each selection carries a probability, a tier, and one line of evidence.
01 — The card
Each match gets one card. The engine prices around 47 markets on every fixture. Roughly 14 make the card: match result, both teams to score, goals from 0.5 to 5.5, corners from 6.5 to 15.5, cards from 0.5 to 4.5, total shots, and shots on target. Every line is priced on both sides. A market without enough evidence behind it is withheld.
The tiers are fixed probability bands. Nothing below 35% is published. A tier describes likelihood. It is not a promise.
Each band is the probability range a tier covers; the tick marks its floor.
02 — The pipeline
Confirm the match itself: teams, competition, venue, referee, kick-off.
Six seasons of match records, including the current one. Shot-quality data, club strength ratings, referee history, injury news.
Each side receives an expected-goals figure — the chances a team creates, weighted by how often chances of that kind are scored. Figures are adjusted for division strength, so a promoted club is not rated on numbers earned in an easier league.
A scoreline model turns the two figures into a grid: every possible final score and its probability. Every market on the card is read from this one grid, so no two numbers can contradict each other.
The model is recalibrated against settled results as they arrive. The evidence behind each market is scored, and under-supported markets are withheld. What remains is assigned a tier.
Forty automated checks run against the finished card. A single failure and the card is not sent.
03 — A worked example
Newcastle at home to Wolves. The model puts Newcastle on 1.8 expected goals and Wolves on 0.9.
The grid turns those two figures into a probability for every scoreline. The single most likely score is 1-0, at 12%. 0-0 sits at 7%. Even the most likely score is unlikely, which is why the engine never predicts one. It answers market questions by adding cells.
Newcastle to win. Add every cell where Newcastle score more than Wolves: 59%.
Over 2.5 goals. Add every cell containing three goals or more: 51%.
Both teams to score. Add every cell where neither side is on zero: 50%.
Tiers follow from the same grid. Under 4.5 goals sums to 86% and publishes as ANCHOR. Over 1.5 goals sums to 75%, STRONG. Over 2.5, at 51%, is PROBABLE.
Darker cells are likelier scorelines. The dashed outline groups every scoreline that clears 2.5 goals; those cells sum to 51%.
Two-leg combination
Newcastle v Wolves · same grid
The legs overlap: a 3-1 win counts toward both. The grid adds only the cells where both are true, so the combination is priced on its real joint probability. A combination whose legs cannot all be true sums to zero and is refused.
04 — The record
Calibration is the test of whether stated probabilities match reality: outcomes given 80% should happen about 80% of the time. The engine is measured on every pick it publishes.
Backtest · Premier League 2025-26 · all 5,356 published picks, settled. Each dot is a band of stated probabilities plotted against how often those selections landed.
By stated probability
| Stated | What happened | Sample |
|---|---|---|
| 70–80% | 73.7% | 1,200+ picks |
| 80–90% | 85.5% | 1,200+ picks |
| 90–100% | 94.0% | 1,200+ picks |
By tier
| Tier | Strike rate | Picks |
|---|---|---|
| ANCHOR | 91.9% | 2,848 |
| STRONG | 77.4% | 2,269 |
| PROBABLE | 59.0% | 239 |
These figures describe how the engine has performed, not how it will.
05 — Limits
It does not beat bookmakers. Measured against the sharpest closing prices, the engine's numbers are slightly worse. A four-season simulation that staked one flat unit on every published pick lost 3.2%. The engine exists to describe matches accurately, not to promise profit.
A strike rate is not a return. Over 0.5 goals lands about 94% of the time and pays about 1/12. High percentages and high payouts rarely coincide.
It does not price player markets. Goalscorer and other player markets are not offered.
It does not recommend bets. The engine states probabilities. Decisions belong to the reader.
It does not curate its record. Settlement runs daily, and every published figure comes from settled matches. Picks that cannot be verified are counted, never dropped. A win rate never appears without the number of picks behind it.