Football Model v2: Rebuilt for International Football and World Cup 2026

Every closed season we go back to the data and ask an honest question: what is the model still getting wrong? This year, the answer was clear. Our football model was trained almost entirely on league football — and when it saw international fixtures and knockout rounds, it made a systematic mistake. We fixed it, retrained the model, and shipped football v2. It is now live ahead of World Cup 2026 and the new club season.

The draw problem

In league football, draws are common. Teams play for points, defend leads, and settle for a share when a win looks unlikely. The model learned that pattern well. But international knockout football is different: draws resolve into extra time and penalties, which means the tactical incentives and the true distribution of results shift. A model that never saw international data did not know that.

The root cause was a data gap, not a modelling flaw. The fix was to close the gap: we added approximately 1,275 international matches drawn from World Cup, Euros, Copa América, AFCON, and Asian Cup tournaments between 2010 and 2024 to the existing training set of ~105,000 league fixtures. The model now knows what it is looking at when it sees a knockout tie.

Tournament-context awareness

Beyond the raw results data, v2 gained several features that help it understand the specific context of a tournament match — things a league model simply cannot infer:

• Neutral venue signal — at a World Cup, neither team has a home crowd. The usual home-advantage factor is switched off.
• Knockout-round depth — a Round of 16 match, a quarter-final, and a final all carry different psychological and tactical weight. The model now encodes where in the bracket the match sits.
• Group-stage standings — a team that has already qualified plays differently to a team needing a result. Group-stage context feeds directly into the prediction.
• National-team strength ratings — every World Cup 2026 contender starts with a sensible Elo rating bootstrapped from over a decade of international results, not a default or an arbitrary carry-over from club football.

We also pruned one feature that carried zero predictive signal in the retrain, keeping the model lean. A feature that does not predict outcomes is noise that can only hurt calibration over time.

Calibrated to the sharpest market

The bar we set for ourselves is not beating the market — it is matching the calibration of the sharpest market, then exploiting soft bookmakers who price a little off. If you understand closing line value, you know why this matters: a model whose probabilities are as accurate as Pinnacle's will consistently identify when a soft book is offering odds above the fair price. That gap is the measurable, repeatable edge.

Football v2 reaches that bar. The calibrated Brier score on our core markets is approximately 0.177 — in line with what Pinnacle's own implied probabilities produce. For international and knockout football specifically, the v1 model never got close to this; the draw-prediction error alone was enough to distort every EV calculation involving international fixtures. You can see the live performance numbers on our model page.

We are not claiming certainty or guaranteed profit. What we are claiming is that the model's probability estimates are now well-calibrated enough that when a soft bookmaker prices a match differently from our model, that difference is more likely to be a real mispricing than a product of our own model error. That is the honest framing — and it is the only framing that holds up over hundreds of bets.

Ready for World Cup 2026 and the new season

The timing was deliberate. World Cup 2026 is the largest single betting event of the year — dozens of knockout matches, national teams from every confederation, neutral venues throughout. A league-only model would have been flying blind through all of it. Football v2 is not.

The same model carries forward into the new club season. The international data does not dilute the league calibration — it adds context the model was missing. We will continue tracking CLV match-by-match and will share findings as the season progresses. Past performance is not a guarantee of future results; bet responsibly, 18+ only.