Spain are world champions. La Roja lifted their second World Cup title on July 19 at MetLife Stadium, beating Argentina 1-0 after extra time in a final that was goalless after 90 minutes. A historic tournament — 48 teams, 104 matches — and our AI tracked 102 of them from the opening kickoff on June 11.
ProbWin ran no official soccer picks during the World Cup. European leagues were on summer break, and our pipeline switched to analysis mode. But the model kept running: generating predictions, logging results, and building a dataset. Here's what 102 matches and 421 predictions across five markets actually tell us — what worked, what failed, and what we're fixing before 2030.
Key takeaways - Our model analyzed 102 of 104 World Cup matches (421 predictions in shadow mode) - Strongest signal: over 2.5 goals at 55.9% accuracy across 68 matches (+9.5u simulated) - The draw market surprised us: 28% win rate on the surface, but +23u simulated thanks to consistently high draw odds - The final: Spain 1-0 AET Argentina — 0-0 at 90 min, our model gave the draw a 33% chance
Why we ran in analysis mode, not official picks
Our soccer model is calibrated on the top five European leagues: Premier League, Liga, Bundesliga, Ligue 1, Serie A. These leagues stop in summer. Facing a 48-team World Cup that included Haiti, New Zealand, and Panama — teams with thin historical data — our confidence intervals widened significantly.
The call to stay in shadow mode was deliberate: stress-test the model outside its comfort zone without putting subscriber capital on predictions we hadn't validated for international tournaments. Every prediction was archived and auto-settled. This recap is built on those 421 records.
The strongest signal: over 2.5 goals
This is the clearest lesson of the tournament. Across the 68 matches where our model predicted over 2.5 goals, it was right 38 times — 55.9% accuracy, +9.5 simulated units.
| Phase | Picks | Win rate | Simulated PnL |
|---|---|---|---|
| Group stage | 56 | 55.4% | +7.66u |
| Round of 16 | 7 | 71.4% | +2.91u |
| Quarter-finals | 3 | 33.3% | -0.93u |
| Semis + Final | 2 | 50.0% | -0.16u |
| Total | 68 | 55.9% | +9.48u |
The group stage is the most readable phase: mid-tier nations don't sit back and absorb, favorites score early, and the ultra-cautious knockout mentality hasn't kicked in yet. The Round of 16 went even better — 71.4% on seven matches. And it produced some wild games: France 4-6 England in the third-place match, Argentina 3-1 Switzerland AET.
The signal collapses in the QF and beyond. With only elite teams left, game management takes over and scoring drops. Our model saw this coming — it only generated two over 2.5 predictions in the semis and final, down from 56 in the group stage.
Empirical rule confirmed by WC2026:
✅ Group stage (win-or-bust matches) → over 2.5 reliable
✅ Round of 16 (strong favorites pressing) → over 2.5 exploitable
⚠️ Quarter-finals and beyond → signal fades, under 2.5 becomes plausible
To understand how our model builds its goal distribution match by match, see our soccer exact score predictor guide.
The hidden value: draw odds were mispriced
28.4% accuracy across 102 matches — the draw prediction looks poor at first glance. In simulation it made +23 units. The math explains why.
Bookmakers in international tournaments systematically undervalue the draw when a "known" nation (France, England, Argentina) is playing. Our xG-based model assigned more draw probability than the market did — creating an edge on odds that typically land between 3.00 and 3.80 for draws in World Cup group games.
| Phase | Draw picks | Win rate | Draw PnL |
|---|---|---|---|
| Group stage (n=72) | 72 | 27.8% | +16.83u |
| Round of 16 (n=16) | 16 | 31.3% | +6.77u |
| Quarter-finals (n=7) | 7 | 14.3% | -3.91u |
| Semis + Final (n=7) | 7 | 42.9% | +3.35u |
In the group stage, evenly matched nations often trade early goals and sit on 1-1 — the draw is a real outcome that the market underprices. The point: a 28% win rate can be profitable if average odds exceed 3.57 (= 1/0.28). This is the core of expected value betting — win rate alone doesn't determine profitability.
Rejoignez ProbWin gratuitement
Accédez à nos prédictions IA pour MLB, NHL et NFL. Inscription en 30 secondes.
The market to avoid: away win (1x2)
Across all five markets, the 1x2 away win prediction lost -45.46 simulated units on 102 matches — by far the worst result of the tournament.
| Market | Picks | Win rate | Simulated PnL |
|---|---|---|---|
| Over 2.5 goals | 68 | 55.9% | +9.48u |
| Draw (1x2) | 102 | 28.4% | +23.04u |
| Home win (1x2) | 102 | 44.1% | -16.87u |
| Under 2.5 goals | 47 | 42.6% | -9.33u |
| Away win (1x2) | 102 | 27.5% | -45.46u |
The root cause: home advantage bias. Our model is calibrated on domestic leagues where the home team has a genuine 5-10% edge. We applied that same parameter to a World Cup played on neutral ground in the USA, Canada, and Mexico. On neutral venues, there is no home team — but our model still treated the designated "home" side as having a structural advantage, which distorted every 1x2 prediction.
Bias identified — WC2026:
❌ Home advantage parameter applied to neutral-venue matches
→ inflated probability for "home" sides
→ systematic underestimation of the "away" win
Planned fix for WC2030:
→ neutral_venue flag: home_weight = 0
→ host-nation proximity correction for crowd support bias
Key matches: what the AI saw
Argentina in the semi-final: the clearest edge of the knockout stage
Before the England vs Argentina semi-final (July 15), our model assigned 49% probability to Argentina — against 31.6% implied by bookmaker odds. A raw edge of +17.4 percentage points, the largest we identified across the knockout rounds.
England vs. Argentina — SF, July 15, 2026
Our model : Argentina 49% | Draw 28% | England 23%
Bookmakers : Argentina 31.6% | Draw 29.3% | England 39.1%
Result : Argentina 2-1 England ✓
Argentina had posted the highest xG across their previous five matches in our pipeline. Bookmakers overweighted England — partly recency bias from their tournament form, partly the public money behind the English. Simulated pick: 1x2 away (Argentina) WIN.
Spain: three knockout matches, three correct calls
Spain's run to the title was the most consistently modeled trajectory of the tournament:
- QF Spain 2-1 Belgium (July 9): 1x2 home WIN ✓
- SF France 0-2 Spain (July 14): 1x2 away (Spain) WIN ✓
- Final Spain 1-0 AET Argentina (July 19): 1x2 draw WIN at 90 min ✓ (0-0 after regulation, Spain scored in extra time)
Spain had the best xG ratio (offense/defense) of the final four. That's exactly what our Dixon-Coles model weights most heavily. For the final: 0-0 at 90 minutes, a draw at exactly the 33% our model gave it. The winning goal came in extra time.
Frequently asked questions
Why no official picks during the 2026 World Cup? Our model is built on European domestic league data. Those leagues pause in summer. With 48 teams — several with thin historical data — our uncertainty intervals were too wide to justify official picks. We ran the full pipeline in shadow mode to stress-test the model and collect data for future tournament calibration. Caution over capital exposure.
Is over 2.5 goals actually a reliable World Cup market? Across 68 WC2026 matches, our over 2.5 signal hit 55.9% (+9.5u simulated). The signal holds in the group stage and Round of 16. From the quarter-finals onward, elite teams lock down defensively and the under 2.5 becomes more viable. Never bet over mechanically through to the final.
How do you explain a 28% draw win rate being profitable? A 28% win rate is profitable if average draw odds exceed 3.57 (= 1/0.28). In WC group games, draws typically paid 3.00-3.80. Our model consistently assigned a higher draw probability than the market implied — creating positive expected value even at a low hit rate. Win rate and profitability are not the same thing.
What is the main model fix before the 2030 World Cup?
Neutralizing the home advantage on neutral-venue matches. In domestic leagues, home advantage is real and well-calibrated at around +7%. In international tournaments on neutral ground, that coefficient becomes a systematic bias. The 2030 model version will include a neutral_venue flag that sets home weight to zero, plus a host-proximity correction for the local crowd factor.
Next step
The 2026 World Cup is over. European leagues restart in August — our natural territory, where the model is fully calibrated and official picks resume from matchday one of the Premier League.
For the full methodology behind our daily soccer predictions, read our guide to AI soccer predictions and how our free AI predictions work. To revisit the semi-finals that set up this memorable final, our WC2026 semi-final breakdown covers France, Spain, Argentina, and England in detail. And our original pre-tournament preview — AI predictions for all 104 matches — is worth comparing against these results.
