The group stage of World Cup 2026 is done. 48 games played. Time to look at how the model performed — and more importantly, what happens when you only bet where you have an edge.

The raw numbers

The model uses FIFA Elo ratings to generate fair win/draw/loss probabilities for every fixture. Across all 48 group stage games, it correctly called the winner in 29 out of 48 matches — 60.4% accuracy.

If you had placed a flat £1 on the model's predicted winner in every game, you'd have staked £48 and ended up £4.07 down — a -8.5% ROI.

That might sound discouraging for a model that's right 60% of the time. But this is exactly why win rates alone don't tell the full story.

Why the favourite trap costs you money

The model predicted heavy favourites in a lot of games. Spain at 1.06, Germany at 1.04, France at 1.06, Brazil at 1.08. When you're right, you earn pennies. When you're wrong — Spain 0-0 Cape Verde Islands, England 0-0 Ghana — you lose your full stake.

Those two results alone cost £2 and returned nothing. The model had Spain at 84% and England at 83%, and both drew with teams ranked 50+ places below them. That's football. Short-priced favourites carry asymmetric risk that no model can fully escape.

Filter to value and the picture changes

A value bet is simple: our model thinks the outcome is more likely than the market does. Concretely, our fair odds are shorter than the bookmaker's price — meaning Bet365 is offering more than the bet is worth.

Applying that filter to the 48 group stage games left 12 value selections:

| Date | Match | Prediction | Result | Model | Market | Edge | P/L |

|------|-------|------------|--------|-------|--------|------|-----|

| 12 Jun | South Korea vs Czechia | HOME | ✓ | 2.11 | 2.70 | +27.7% | +1.70 |

| 13 Jun | USA vs Paraguay | HOME | ✓ | 1.77 | 2.05 | +15.6% | +1.05 |

| 14 Jun | Ivory Coast vs Ecuador | AWAY | ✗ | 2.28 | 2.50 | +9.9% | -1.00 |

| 16 Jun | Iran vs New Zealand | HOME | ✗ | 1.36 | 1.80 | +32.2% | -1.00 |

| 17 Jun | Argentina vs Algeria | HOME | ✓ | 1.43 | 1.50 | +5.3% | +0.50 |

| 17 Jun | Ghana vs Panama | AWAY | ✗ | 1.69 | 3.20 | +89.6% | -1.00 |

| 18 Jun | Switzerland vs Bosnia | HOME | ✓ | 1.51 | 1.55 | +2.8% | +0.55 |

| 19 Jun | Scotland vs Morocco | AWAY | ✓ | 1.53 | 1.70 | +11.0% | +0.70 |

| 20 Jun | Netherlands vs Sweden | HOME | ✓ | 1.55 | 1.70 | +9.6% | +0.70 |

| 21 Jun | Uruguay vs Cape Verde | HOME | ✗ | 1.43 | 1.44 | +0.7% | -1.00 |

| 22 Jun | New Zealand vs Egypt | AWAY | ✓ | 1.46 | 1.55 | +6.3% | +0.55 |

| 23 Jun | Norway vs Senegal | AWAY | ✗ | 1.91 | 3.10 | +62.1% | -1.00 |

7 from 12 winners (58.3%). £12 staked. +£0.75 profit. ROI: +6.3%.

What the numbers mean

Filtering to value doesn't just improve ROI mechanically — it changes which games you're betting on. You stop backing 1.06 shots that return almost nothing when correct, and start focusing on games where the market is mispricing the probability.

The two biggest edge calls — Ghana/Panama (+89.6%) and Norway/Senegal (+62.1%) — both went against the model. That's a £2 swing that hurts over a small sample. But edges that large don't appear often and over time they should resolve in your favour. Iran/New Zealand at +32.2% edge going against us is the same story — these are the bets you want to be making, even when they lose.

The 60% accuracy across all 48 games tells you the underlying model is solid. The +6.3% ROI on value selections tells you the edge is real when you apply the filter. The -8.5% on all games tells you that betting without a price edge — no matter how good your model is — will cost you in the long run.

The knockout stage starts now. Login to see the model's predictions and value ratings for every remaining fixture.