How the model actually did
The World Cup is over. 
Spain — the model's most likely champion since before the opening match — beat Argentina 1-0 in the final. This page is the full accounting: every one of the 104 matches was predicted before kickoff, logged, and never edited. Here's what the model got right, what it got wrong, and what it structurally couldn't see.
01Headline scorecard
Brier score measures probability accuracy: 0 is perfect, 0.25 is what you'd score by calling every match a coin flip. Lower is better. At 0.132, the model comfortably beat the baseline across all 104 matches — while publishing every number in advance.
02Calibration by confidence band
The real calibration test: when the model said 75%, did the favorite win about 75% of the time?
| Model said | Matches | Avg claim | Favorite won | Verdict |
|---|---|---|---|---|
| ≥90% | 5 | 91.8% | 60% | overconfident |
| 80–90% | 20 | 84.8% | 80% | calibrated |
| 70–80% | 26 | 76.0% | 77% | calibrated |
| 60–70% | 26 | 65.3% | 58% | overconfident |
| 50–60% | 27 | 55.9% | 52% | calibrated |
The middle held: the 80–90% and 70–80% bands landed almost exactly on their claims. The edges ran hot — 90%+ favorites won just 60% (small sample, n=5, but two of the tournament's ugliest upsets lived there), and the 60–70% band hit only 58%. Translated: when this model said "strong favorite," believe it; when it said "slight favorite," it should have said "toss-up" more often.
03Beyond match picks: the tournament-level claims
Match favorites are the easy test. The harder one: the day-one snapshot (2026-06-12) made a probability claim about every one of 48 teams — advance from the group, win the group, win the tournament. Scored against what actually happened:
All three metrics beat their base-rate baselines. The perfect 6-for-6 on the ≥90% claims is the flattering number; the honest one is at the bottom of the table — of the three teams the model wrote off at ≤30%, two advanced anyway (South Africa and Ghana). The model was better at spotting locks than at burying longshots.
04Biggest misses, owned
The five matches where the model was most confidently wrong. No excuses, no edited history — these are the same numbers that were published before each kickoff.
South Korea (78% favorite) lost→ South Africa won 1-0The tournament's biggest upset by the model's own numbers — and the start of a South Africa run the model never believed in (they entered the day at 23% to advance, and advanced).
Ecuador (77% favorite) lost→ Ivory Coast won 1-0Ecuador's +200 Elo re-seed in June made them a 77% favorite. Côte d'Ivoire won the actual football match 1-0.
Panama (74% favorite) lost→ Ghana won 1-0Same shape three days later: a one-goal loss to a team the Elo gap said should rarely win. Two of these in week one was the first hint the mid-bands ran hot.
Germany (66% favorite) lost→ Paraguay won on penaltiesLevel after 120 minutes, out on penalties. The model priced a winner; it has no shootout model at all — see the penalty problem below.
Australia (66% favorite) lost→ Egypt won 5-3An eight-goal extra-time chaos game. No probability model covers itself in glory here, including this one.
Honorable mention: England 6-4 France in the third-place match — the model had France at 57% and got a ten-goal absurdity instead. Bronze games obey no model.
05The penalty problem
Three knockout matches went to shootouts. The model's favorite lost all three:
- R32 · Germany (66%) lost to Paraguay on penalties
- R32 · Netherlands (65%) lost to Morocco on penalties
- R16 · Colombia (63%) lost to Switzerland on penalties
This is a structural limitation, not bad luck to be waved away: the knockout model prices a winner from the Elo gap, and a shootout is close to a coin flip that erases that gap entirely. A match that reaches penalties is one the model should have marked down toward 50% — it didn't, and it paid for it three times. The most-cited flaw of this model, and correctly so.
06Thirty-nine days of odds, drawn
Every day of the tournament, the model saved a full snapshot of every team's odds — 39 snapshots in all. These are the championship-odds curves for the final four, plus Norway as the chaos line.
View as table (weekly values)
| Date | ESP | ARG | FRA | ENG | NOR |
|---|---|---|---|---|---|
| Jun 12 | 22.6% | 14.8% | 12.6% | 7.0% | 2.2% |
| Jun 19 | 16.9% | 17.7% | 14.1% | 8.0% | 3.3% |
| Jun 26 | 19.3% | 19.6% | 13.8% | 7.0% | 2.5% |
| Jul 3 | 21.9% | 20.5% | 20.4% | 9.0% | 2.8% |
| Jul 10 | 25.9% | 25.5% | 27.8% | 12.3% | 4.3% |
| Jul 17 | 51.4% | 48.6% | 0.0% | 0.0% | 0.0% |
| Jul 20 | 100.0% | 0.0% | 0.0% | 0.0% | 0.0% |
- Spain started as the favorite (22.6%) and never gave up the top spot — the gold line dips through a nervy group stage, then climbs relentlessly to 100%.
- Norway is what an outsider run looks like in probability space: 2% for three weeks, a spike to 5% the morning after knocking out Brazil, then back to zero at the quarterfinal.
- Mexico (in the data, flat along the bottom, not drawn) never crossed 2% — the model never believed in the quinto partido, and the quinto partido never came.
07Model vs market, resolved
In June, the model's biggest disagreement with the betting market was Spain: the model said 23% to win it all, books implied ~17%. It also had Argentina (~15%) well above the market's ~9.5%, and France and England below their market prices. The final was Spain beating Argentina — the model's two "overpriced by us, underpriced by the books" teams, in that order. One tournament resolves nothing statistically, and a France final would have made the market look smart instead. But the one time this model and the market genuinely disagreed, the model's side of the argument lifted the trophy.
08What the model couldn't see
Every number on this page is computed from files published verbatim at /data: 104 pre-match predictions, 39 daily model snapshots, all scores. Re-run the math yourself.
This continues for the 2026–27 season
Same idea, new competitions: starting with the Premier League in August, every prediction published before kickoff, every miss kept on the record, calibration reported in the open — plus one short weekly email of hits, misses, and the week ahead. jetlagxi.com stays up permanently as the WC 2026 archive.
Leave your email and you'll get one launch announcement when the new site ships in August. Nothing else.
Methodology: Elo-seeded 10,000-run Monte Carlo, re-simulated as real scores landed. Per-match predictions logged to a public JSON file at first final whistle and never edited (104/104 matches). Daily odds snapshots 2026-06-12 → 2026-07-20. Final: Spain beat Argentina; England third, France fourth. Built with AI-assisted tooling, disclosed proudly.