How it works
Every prediction comes from a Dixon-Coles goal model: separate attack and defence coefficients per team, fit by maximum likelihood on 49,000+ internationals since 1872, with a low-score correlation correction and squad market value as a prior. From the fitted rates it builds a full scoreline distribution, then reads off win/draw/loss probabilities and the modal score. Two calibration steps follow: a recent-form nudge, and a draw transfer that shifts surplus draw probability onto the favoured side, both checked out of sample.
The model trains on 49,000+ international matches stretching back to the first official fixture in 1872, a dataset maintained by martj42 and refreshed daily. Squad strength comes from Transfermarkt's market valuations, covering 182 national teams from 2006 onward, which feed the value prior.
Every figure here comes from World Cups the model never trained on, and each new idea is gated the same way, tested leave-one-tournament-out before it ships. Several intuitive ones (rest-day edges, dynamic ratings, lineup-weighted squad value) were tried and dropped. They stay documented in the repository, which is what lets the rest be trusted.
Each shipped part had to earn its place out of sample. Adding them one at a time steps the held-out Brier down this chain, from a random guess to the production model.
The line steps down from a random guess (0.667) to the shipped model (0.5441); the figure under each step is the out-of-sample gain it bought. The axis is zoomed so the small steps are visible. One caveat on the first step: the value-prior weights were tuned on 2006 to 2018, so four of the five held-out tournaments overlap that window there, and only 2022 is fully out of sample (0.6123 to 0.5907).
The one data edge that shipped was squad market value. Here are its cleanest numbers on one scale: the grey dot is the held-out Brier without the prior, the blue dot with it, and lower (left) is better. Both slices move left, 2022 the most.
The raw goal engines are nearly tied. On the same 320 held-out matches they land within 0.003 of each other, so the edge is the gated calibration chain, not the engine. The shipped model sits well left; a random guess sits far right.
These four are the raw engines scored head to head. The shipped 0.5441 is reached by the calibration chain above, measured from its own 0.5717 Dixon-Coles baseline, so the engine table and that baseline are separate configurations rather than one comparison. Every value here is multiclass Brier over the same 320 matches.
Championship odds
For each side still in, the chance of reaching the semi-finals, the final, and winning the title. The nested bar darkens from the semis (lightest) through the final to the title (solid).
These come from 10,000 Monte-Carlo runs of the bracket. The simulator keeps it simple: neutral venues, no draws, and the knockout paired in schedule order. Read them as the model's view, not a hard forecast.
Title odds over time
The chart starts on 13 June, the tracker's first recorded day, so any earlier moves are not shown. Switch to all teams to add the knocked-out sides as thin grey lines that stop on the last day their title chance was above zero.
No title-odds history yet. It builds up as the daily tracker run records each day's simulation.
Live record
| Date | Match | Pick | H/D/A | Pred | Result | Δ | Hit |
|---|
Upcoming predictions
| Date | Match | Pick | Win | Draw | Loss | Pred |
|---|
Why it never picks a draw
The model's most visible quirk, owned up front: it never makes a draw its single most likely call. A draw only tops the list when two sides sit almost dead level, and the model's draw probability never climbs that high.
predicted draw probability across the 320 held-out matches (0.05 bins)
The favourite-directed draw-transfer pulled the mean predicted draw from 0.283 to 0.214, against a 0.225 base rate. It shifts probability onto the favoured side, so the model sharpens its draw estimates without ever picking a draw. Bookmakers share the same ceiling.
Where does that ceiling come from? This is the empirical curve behind it: the observed draw rate as two teams pull apart in rating, straight from the data. Dots are the observed rate per 50-point gap bin, sized by how many matches fall in each; the blue line is the smoothed table the model actually carries.
One label to keep it honest: this curve belongs to the Elo-side draw model, which drives the baseline and the title simulator. The production Dixon-Coles model derives its draws from the scoreline distribution instead, so read this as the mechanism behind the 30% ceiling, not the production draw engine.
Past World Cups results
Validated on five World Cups it never saw during training. For each tournament the model trains only on matches played before it, then predicts every match in it, the same information regime as predicting live.
Combined across all 320 matches: 59.4% accuracy, multiclass Brier 0.544 (uniform guess = 0.667), beating both the baseline and a coin flip. The model reads the earlier tournaments best; the tighter recent fields are harder.
A probability model should be right about as often as it claims. Each point groups the 320 picks by how confident the model was and plots that against how often it actually won. Well-calibrated points sit on the dashed diagonal.
The points sit close to the diagonal, so the model wins about as often as its confidence implies. The top bin is thin, only 18 matches above 80 percent, and the small dip at 65 percent is sampling noise rather than miscalibration. Murphy splits the Brier into the three parts on the right: the value prior and recent form bought resolution, the draw-transfer bought reliability.
The 320 matches
Held out leave-one-tournament-out: each World Cup is predicted by a model trained only on matches played before it, the same information regime as a live prediction. This is the full evidence base behind the headline, one row per match.
| Date | Match | Pick | H/D/A | Result | Hit |
|---|
Tested and rejected
Every idea here was built, measured leave-one-tournament-out, and rejected. The column on the right plots what each one did to the held-out Brier, against a centre line at zero (no change): lower Brier sits to the left, and the bar shows how sure the test is. An idea only ships if its whole 95% interval clears zero on the better side. None of these managed it.
Negative results with their intervals are the point: they are why the three changes that survived (the value prior, the recent-form nudge and the draw-transfer) can be trusted. The one highlighted row, form λ = 0.60, is the near miss. It sits left of the line, but 2022 regresses and its interval grazes zero, so it was left on the table: rejected, not pending.
Against the closing line
The hard benchmark for any predictor is a bookmaker's closing line, the sharpest probability on the market. The model's held-out predictions were scored against real de-vigged closing odds from two past World Cups, on the same matches. Across 99 games it comes out level with the line: a shade ahead after calibration, but inside the margin of error, so parity rather than a proven edge.
| World Cups | Matches | Model Brier | Market Brier |
|---|---|---|---|
| 2018, 2022 | 99 | 0.564 | 0.568 |
The margin sits inside one standard error, so the honest read is parity with the sharpest line rather than a beat. The scoreboard below carries the same method into 2026, though the 2026 benchmark is a cross-book average rather than a single book's line: for every match where we captured a market price, it scores the model's probabilities against that market consensus on the result.
How the model reached parity, in three moves.
The gap sat on moderate favourites, the biggest and hardest band. Held-out hit rate, split by how strongly the model favoured a side:
Model versus the market consensus, live.
| Date | Match | Model | Market | Sharper |
|---|
These are the 2026 matches where we captured a market price. The market figure is the de-vigged average across many bookmakers, not a single book's closing line. Because it pools many books it carries almost no margin (on nearly half of these matches the raw prices sum to under 100 percent), which makes it sharper than any one bookmaker would quote, so the model trailing it here is expected and is not the same as trailing a single book. The calibrated comparison is the rigorous test above: on the 99 matches from 2018 and 2022 with real single-book closing lines, the model was level (0.5643 against 0.5684). This 2026 sample is small and noisy, so read the running gap, not a verdict.