World Cup 2026 Review: What Our Model and Value Bets Actually Did

The World Cup is over, so here is the honest accounting — not a highlight reel, but the full data debrief of how our World Cup 2026 model and value bets actually performed across all 104 matches. The headline: 89 settled value bets, 42 wins and 40 losses, for +8.45 units and a +3.7% yield, with average closing-line value of +3.36pp. That lands inside the realistic 3–8% range we tell people to expect, and the CLV is positive — the market kept shortening the prices we took. But the tournament also exposed two things clearly: the handicap and knockout books carried us, the Unders bled, and the model's one real blind spot was draws. This is all of it, misses included, straight from the database that our public track record is built on.

The book: +8.45 units across 89 value bets

Convention first, because it governs every number here: voids are counted separately — a void returns the stake in full and is never a loss — so the 42–40 record and 51.2% win rate are over settled wins and losses only, with 7 voids on top. Units are fractional-Kelly stakes recomputed from the database (1 unit = 1% of bankroll), which is why they reconcile exactly to the public track record rather than a flat-stake tally. Against 225.8 units staked, the tournament returned +8.45 units.

The shape of that profit matters more than the total, because it splits cleanly by market:

• Asian handicaps — the engine. 32 bets, 16–16, +13.34 units (+15% yield, CLV +2.3pp). A coin-flip win rate that still made money, because the winners were priced generously and the draw-no-bet lines around evens did their job.
• Match winners (1X2) — small book, sharp read. 11 bets, 10–1, +11.67 units (+27% yield). Tiny sample, helped by variance, but the model picked its spots well — it only fired a 1X2 bet when it genuinely disagreed with a soft price, and it was right ten times out of eleven.
• Totals (over/under) — the drag. 46 bets, 16–30, −16.55 units (−18% yield) — and this is despite the totals book closing at +4.8pp CLV. That gap between a positive close and a negative return is the single most important lesson of the tournament, and we come back to it below.

There is also a clear arc across the tournament. The group phase was a net loser: 60 settled bets for −4.84 units. The knockout rounds turned it around completely — 29 settled bets, 15–11, for +13.29 units at a +20% yield. A 29-bet sample is far too small to call a durable knockout edge, but the direction is honest: our worst stretch was the group-stage goal markets, and the book climbed back into profit once the bracket began.

Where it bled: the Unders, and why beating the close did not save them

Almost the entire totals loss lives on one side of the market. Split the 46 over/under bets by direction and the Overs were roughly break-even — 25 bets, 10–15, +1.35 units — while the Unders were a bloodbath: 21 bets, 6–15, −17.90 units. World Cup 2026 was a high-scoring tournament, and every time our international totals model leaned Under, the goals kept coming.

Here is the part that separates a real edge from a mirage. Several of those losing Unders closed with strongly positive CLV — Senegal vs Iraq Under 2.5 lost but closed +28pp to the good; Netherlands vs Sweden Under 2.5 lost at +22.8pp CLV. In other words, we consistently beat the closing line and still lost money on that market. This is exactly why we do not treat closing-line value as a universal profit metric. For moneyline-type markets — match winners, draw-no-bet — CLV is a reliable long-run signal. For totals, it is not: the closing line can move our way and the tickets can still be underwater, because the mispricings that generate CLV on a goals line are not the same as the ones that win the bet. On totals we let realized return and calibration lead, and CLV only corroborate — and the tournament proved the point in red ink.

The model scorecard: sharp on winners, blind to draws

Across the 104 matches the model's average 1X2 Brier score was 0.56. That beats a no-skill baseline (uniform 1/3–1/3–1/3 scores 0.67), but it is well above the model's long-run club-football figure of about 0.18 — a reminder that international tournaments are the single hardest thing we price. National teams meet rarely, form and lineup features are sparse, and the signal collapses toward Elo. On the clear favourites the model was excellent; the damage was concentrated in one place.

Draws. The model predicted a draw in only 13.6% of matches. The tournament actually drew 27.9% of the time — more than double. It was not a knockout artefact either: even the knockout rounds finished level after 90 or 120 minutes in 9 of 32 games (28%). Under-predicting draws by that much is the model's clearest weakness of the tournament, and it is the honest reason the Brier looks the way it does — a missed draw is punished heavily, and we missed a lot of them.

We traced it to the calibration layer, not the raw model. Our international predictions were being run through a calibration step tuned mostly on draw-light club football, which systematically shrank the draw — on some knockout ties almost to zero. The raw model, it turns out, reads international draws well; the calibration was correcting away a signal that was already right. We shipped the fix late in the tournament — international matches now bypass that club-tuned step — so most of the settled bets above still ran on the old numbers, but re-scored on the full World Cup the corrected model predicts draws at 27%, in line with the 27.9% that actually landed, and its match Brier improves from 0.56 to 0.48. The lesson we carry into the club season: calibration tuned on one domain can quietly break another.

Standout calls, misses included

A representative slice of the settled book — the biggest winners and the biggest misses, never cherry-picked:

• Algeria vs Austria, Over 2.5 @ 2.39 (+16% EV). WON ✓ — +4.09u. The single best ticket of the tournament.
• USA vs Paraguay, Over 2.5 @ 2.42 (+15% EV). WON ✓ — +3.82u.
• Croatia vs Ghana, Over 2.5 @ 2.38 (+15% EV). WON ✓ — +3.71u.
• England draw-no-bet vs Mexico (last 16) @ 1.817 (+12% EV). WON ✓ — +2.91u. England edged it 3–2.
• Spain draw-no-bet vs France (semi-final) @ 2.337 (+11% EV). WON ✓ — +2.82u. Spain won 2–0 and reached the final.
• Morocco to beat Canada (last 16) @ 1.80 (+11% EV). WON ✓ — +2.69u. Morocco won 3–0.
• Uruguay to beat Saudi Arabia (group) @ 1.50 (+8% EV). LOST ✗ — −4.20u. Our worst single call of the tournament.
• Türkiye vs Paraguay, Over 2.5 @ 2.30 (+17% EV). LOST ✗ — −3.36u.
• Senegal vs Iraq, Under 2.5 @ 2.55 (+18% EV). LOST ✗ — −2.90u, closing +28pp CLV. The perfect illustration that beating the close does not save a totals ticket.
• Netherlands vs Sweden, Under 2.5 @ 2.48 (+17% EV). LOST ✗ — −2.86u (CLV +22.8pp).

How the tournament finished

For the record, our system logged all 104 matches. Spain were the champions, beating Argentina 1–0 after extra time in a final that was still goalless at 90 minutes — the model and the market had made Argentina a slight favourite, and it stayed scoreless until Spain's extra-time winner. Spain had reached the final by beating France 2–0 in the semis; Argentina came through 2–1 against England, who took third place in a wild 6–4 play-off against France. Those are the results our bets settled against (on the 90-minute lines) and the model was scored on; they are what this debrief measures.

What we carry into the club season

Three things travel out of this tournament. The draw-calibration fix and the World-Cup-informed retrain go in before the domestic season ramps up. The totals-versus-CLV lesson gets reinforced — on goals markets we trust the realized number, not the flattering close. And the parts that worked — disciplined 1X2 selection and the handicap book — carry straight over into the leagues, where the model is at its strongest and the features it lacked for national teams are all present. Follow the full, honest record on the track record page, revisit the tournament on the World Cup 2026 hub, and see how the group stage played out in the Matchday 14 recap.