Every prediction we publish for the 2026 World Cup goes up before kickoff, and every graded result — win, loss, or wrong pick — stays on the record. Through 20 graded matches, that record reads: a 15-5 pick record (75.0% hit rate), a model log loss of 0.722 against the market's 0.777 on the games where a line was posted, and — the number we're actually proud of — 14 flagged edge bets going 7-7 for +4.98 units, a 33.56% ROI.
An even bet record that's up nearly 5 units? It isn't a typo, and explaining why is the point of this post.
The model, briefly
The World Cup model isn't a black box tuned to this tournament. It's Elo ratings updated match-by-match, feeding a Poisson/Dixon-Coles goals model fit only on matches that happened before the tournament being predicted — the same leakage-free discipline every Predictium model runs on. No result from a tournament is ever used to help predict that same tournament.
We backtested this structure walk-forward across the last three World Cups — 2014, 2018, and 2022, 192 matches combined, none of them influencing the others' predictions. The combined numbers: 57.3% modal accuracy (correctly calling the most likely outcome — win, draw, or loss — across three-way soccer outcomes, where a coin flip isn't 50%, it's 33%), and a log loss of 0.986, about 7% lower than a baseline that just plays historical base rates. Full edition-by-edition numbers and the calibration tables are on the World Cup page.
That's the model earning its keep before this tournament ever kicked off. The 2026 numbers below are what happened once real money-shaped stakes were on the line.
The live 2026 record
Through 20 graded matches this tournament:
| Metric | Result |
|---|---|
| Pick record (modal 3-way call) | 15-5 (75.0%) |
| Advance-round picks | 15-5 |
| Model log loss | 0.722 |
| Market log loss (18 games with a line) | 0.777 |
| Flagged edge bets | 14 |
| Edge bet record | 7-7 |
| Edge units | +4.98u |
| Edge ROI | +33.56% |
The model beat the market's own implied probabilities on log loss — 0.722 versus 0.777 — which matters more than the pick record by itself, since log loss scores the whole probability distribution, not just which box we checked.
Why 7-7 is a winning record
Every flagged bet is sized to win 1 unit at the price posted when we flagged it — not to risk 1 unit. A favorite might have to risk 1.8 units to win 1; a longshot might only risk 0.3 units to win 1. That convention means win rate alone tells you almost nothing about profitability. An exactly-even 7-7 record still nets +4.98 units when the wins skew toward better prices than the losses did — which is the entire point of flagging edges by expected value instead of by confidence.
This is also why we publish the full graded ledger — every flagged bet, its price, and its outcome — instead of just the headline ROI. A number without the receipts isn't a track record, it's a claim.
Argentina's semifinal win, on the ledger
Our model had Argentina as the favorite to reach the final: 45.3% to beat England, against England's 27.1% and a 27.6% chance of a draw before extra time. Argentina won, 2-1, and it's graded: the modal pick was correct, the advance-round pick was correct, and the flagged edge bet — Argentina moneyline at +200, staked 0.5 units to win 1 — cashed. That's the extra unit in the numbers above, from 13 flagged bets at 6-7 and +3.98u to 14 flagged bets at 7-7 and +4.98u.
Argentina now plays Spain in the final on July 19. Whatever we flag there goes up before kickoff too, and it'll be graded the same way — win or lose.
That's the whole point of publishing before grading instead of after: the record only means something if it can't be edited after the fact.