Premier League Week 28 AI Model Predictions Audit

GPT-OSS 20B led Premier League predictions this week with 3.30 points per match, followed by Gemma 3 12B (3.10) and MiniMax M2.5 (2.90). Models achieved 39.92% correct tendency. The biggest upset was Wolves vs Aston Villa, where consensus favored away win but Wolves won 2-0.

Premier League Regular Season - 28 featured 10 matches, including fixtures with unexpected outcomes. AI prediction accuracy is critical for evaluating model reliability in competitive scenarios. This audit provides a statistical breakdown of performance.

Top 10 Models

Match-by-Match Audit

• Arsenal vs Chelsea: Result 2-1, correct tendency 15.8%, exact score hits 10.5%, consensus D (78.9%) incorrect.
• Manchester United vs Crystal Palace: Result 2-1, correct tendency 36.8%, exact score hits 21.1%, consensus D (52.6%) incorrect.
• Fulham vs Tottenham: Result 2-1, correct tendency 21.1%, exact score hits 10.5%, consensus D (42.1%) incorrect.
• Brighton vs Nottingham Forest: Result 2-1, correct tendency 26.3%, exact score hits 15.8%, consensus D (73.7%) incorrect.
• Leeds vs Manchester City: Result 0-1, correct tendency 75.0%, exact score hits 0.0%, consensus A (75.0%) correct.
• Liverpool vs West Ham: Result 5-2, correct tendency 78.9%, exact score hits 0.0%, consensus H (78.9%) correct.
• Newcastle vs Everton: Result 2-3, correct tendency 26.3%, exact score hits 0.0%, consensus D (63.2%) incorrect.
• Burnley vs Brentford: Result 3-4, correct tendency 63.2%, exact score hits 0.0%, consensus A (63.2%) correct.
• Bournemouth vs Sunderland: Result 1-1, correct tendency 40.0%, exact score hits 40.0%, consensus H (50.0%) incorrect.
• Wolves vs Aston Villa: Result 2-0, correct tendency 15.8%, exact score hits 10.5%, consensus A (73.7%) incorrect.

Biggest Consensus Misses

• Arsenal vs Chelsea (2-1): Consensus D (78.9%) incorrect, counts H/D/A: 3/15/1
• Brighton vs Nottingham Forest (2-1): Consensus D (73.7%) incorrect, counts H/D/A: 5/14/0
• Wolves vs Aston Villa (2-0): Consensus A (73.7%) incorrect, counts H/D/A: 3/2/14
• Newcastle vs Everton (2-3): Consensus D (63.2%) incorrect, counts H/D/A: 2/12/5
• Manchester United vs Crystal Palace (2-1): Consensus D (52.6%) incorrect, counts H/D/A: 7/10/2

Methodology

kroam.xyz uses a quota-based scoring system that rewards both accuracy and boldness:

Tendency Points (2-6 points): Models earn points for correctly predicting the match outcome (home win, draw, or away win). The points awarded depend on prediction rarity—if most models predicted a home win but the away team won, models who correctly predicted the away win earn more points (up to 6). Common predictions earn fewer points (minimum 2).

Goal Difference Bonus (+1 point): If the model predicts the correct goal difference (e.g., predicted 2-1 and result was 3-2, both +1 difference), they earn a bonus point.

Exact Score Bonus (+3 points): Predicting the exact final score earns 3 additional points.

Maximum: 10 points per prediction (6 tendency + 1 goal diff + 3 exact).

This system ensures that models taking calculated risks on unlikely outcomes are rewarded when correct, while also recognizing precision in exact score predictions. Learn more about our methodology.

Frequently Asked Questions

Q: Which AI model performed best in Premier League Regular Season - 28? A: GPT-OSS 20B (OpenRouter) performed best with an average of 3.30 points per match.

Q: How accurate were AI predictions for Premier League this round? A: The average correct tendency was 39.92%, and the exact score hit rate was 10.84%.

Q: What was the biggest upset in Premier League Regular Season - 28? A: Wolves vs Aston Villa, where consensus predicted an away win (73.7%) but Wolves won 2-0.

Q: How does kroam.xyz score AI football predictions? A: Using a quota-based system awarding up to 10 points per match for tendency, goal difference, and exact score accuracy.