Every match prediction on Kickoff Report comes from our own statistical model.
Here is the model architecture and the data behind it.
Team strength: Elo ratings
The backbone of the model is the Elo rating of each team: a strength score that moves after every match. For national teams we use the public World Football Elo Ratings. For clubs we keep our own Elo. It started from the last public ClubElo values, when that feed stopped in September 2026, and we update it from every result. The gap between two teams’ ratings translates into an expected goal difference for the fixture.
Recent form, as a fallback
When both teams have a rating, the rating gap alone sets the expected goals. In our tests that beat every mix with recent form. Form takes over only when a rating is missing: each side’s recent scoring and conceding rates, weighted toward the latest matches and normalised against the league’s average goals. A light head-to-head adjustment applies when the sides have met at least three times, using their five most recent meetings.
From expected goals to probabilities
The two expected-goal numbers feed a Poisson model with the Dixon-Coles correction — the standard academic adjustment that fixes plain Poisson’s tendency to underrate draws and low-scoring results. Out of that grid comes every number in the widget: win, draw and win percentages, expected goals and the most likely scores. We do not show both-teams-to-score or over 2.5 goals. Our model fixes the expected total of goals, so those numbers would only repeat the win probabilities. We checked our own both-teams-to-score number against 459 played matches. It did worse than a coin flip, so we removed it.
Most likely scores, honestly ranked
The widget lists the three most probable exact scorelines with their individual probabilities. One thing worth knowing: no exact score in football is ever very likely. Even the top scoreline rarely passes 16%, and in tight matches between well-organised sides it is usually 1-1 or 0-0 — that is real football, confirmed across thousands of historical tournament matches, not a quirk of our model. That is why the top scoreline and the most likely match result can name different outcomes: a draw can be the single most common exact score even when one side wins more often overall.
Knockout ties: who advances
In a single-leg knockout, a draw after 90 minutes is a real possible scoreline but not a possible result of the tie — someone goes through. So for knockout matches the widget shows a separate “To advance” figure: each side’s chance of winning in normal time, plus half of the draw probability, treating extra time and penalties as the coin flip the evidence says they roughly are. The win/draw/win percentages above it always describe the 90-minute result only. Our headline pick follows the same logic: when one side is more likely than not to win in normal time, we name their most likely winning score; otherwise we name the most likely 90-minute score and who advances.
Home advantage, honestly applied
Home advantage is worth roughly 65 Elo points — but only when a team genuinely plays at home. Neutral-venue matches (most World Cup fixtures, cup finals) get none; World Cup host nations playing in their own country keep it.
International tournaments use their own settings
The same rating gap gives a closer match at a World Cup than in club football. So for the World Cup, the Euros, the Copa América, the Africa Cup of Nations and the Asian Cup we use a separate set of coefficients. We fitted it on the final tournaments from 2006 to 2021 and tested it on those from 2022 to 2025. It is deliberately less confident than the club model.
What the model does not know
The model knows nothing about team news: injuries, suspensions, line-ups, weather or motivation. It works only from ratings, results and head-to-head history. That makes it consistent, but a star striker ruled out an hour before kickoff will not move it. Read our numbers as a calibrated starting point, not as a betting tip.
Backtested before it ships
Every version of this model is replayed against four seasons of historical results across Europe’s top five leagues, plus recent World Cups and Euros, and scored on calibration: when we say 60%, it should happen about 60% of the time. A new model version only replaces the old one if it measurably beats it on a held-out season it has never seen. The current version scores +0.0051 RPS above the bookmakers’ closing odds (0.2029 vs 0.1978) on the industry-standard ranked probability score. The absolute delta is small.
Predictions are statistical estimates for entertainment and editorial context — not betting advice.
Model Validation
These numbers come from our June 2026 backtest: calibrated on 2021/22–2024/25, then evaluated on the held-out 2025/26 season (N = 1,751 matches).
| Metric | Kickoff Report v2 | Bookmakers closing odds |
|---|---|---|
| RPS (Ranked Probability Score — lower is better) | 0.2029 | 0.1978 |
| Model | Elo-anchored Poisson + Dixon-Coles | — |
| Training | 2021/22 – 2024/25 (4 seasons) | — |
| Hold-out | 2025/26 season | — |
Our model scores +0.0051 RPS above bookmakers’ closing odds (0.2029 vs 0.1978) — a competitive result for a transparent, open methodology. The delta is small: well within measurement noise for a solo-built system, but the precise number to cite. Bookmakers benefit from odds shading and market depth that we do not replicate.
Live results. The numbers above come from a backtest on past seasons. Every prediction published since June 2026 is also checked against the real result. See the full record on Prediction Results and Accuracy.