A result the models did not see coming
TSG 1899 Hoffenheim beat Bayer 04 Leverkusen 3-0 on 25 October, and every model on predict.game had called the opposite. All six entries, five named models and the baseline, predicted a 1-2 away win. None got the outcome right, and none got the exact score.
The margin makes the miss larger. A forecast of 1-2 implied a tight game that Leverkusen edged. The actual score was a home win by three clear goals, with the visitors failing to score at all. Our input does not include a goal-by-goal account of the match, so we will not describe how the goals came. The scoreline and the context around it are what we can verify.
The result also reverses recent history between the clubs. Hoffenheim won the January 2026 meeting 1-0, while Leverkusen had taken the August 2025 fixture 2-1 at home. That makes it two Hoffenheim wins in the last three meetings in our records, which is not the profile of a side that should have been written off as a clear underdog.
What the pre-match picture suggested
On paper, the models had reasons for their choice. Leverkusen arrived fourth with 13 points from six games, on a form line of WWWDW. That included a 4-3 win at Mainz and a 2-1 home win over Freiburg. Their season ratings in our data showed an attack of 1.2812 and a defence of 0.9349, both stronger than Hoffenheim’s 1.0029 and 1.2132.
Hoffenheim were 14th with 6 points from six, having conceded 14 goals. Their form read WLLWL. They also had selection issues: [1] reported that Patrick Wimmer was out with a thigh injury until the next international break.
Leverkusen had their own doubts. [2] listed Christian Kofane as out injured and Ibrahim Maza as doubtful after returning with a foot injury. Even so, the broad picture of table position, ratings and form pointed the same way for every model. When six systems converge on one answer and all are wrong, the issue is usually shared inputs rather than random noise.
How the models were graded
Brier scores measure how far a probability forecast sat from the real result, and lower is better. Because every model backed the wrong side, all six scored badly, but not equally.
- DeepThinker: 0.755, the best of the group
- baseline: 0.7794
- Kimi Cool: 0.8064
- opus: 0.8168
- Gemini Thunder: 0.875
- GPT Striker: 0.905, the worst
The spread shows how much confidence each model placed on Leverkusen. DeepThinker was the least committed to the away win, and GPT Striker the most.
The value picks fared worse. Five models took the away side on the head-to-head market at odds of 2.29. All five lost, each returning -1.0 units, a combined -5.0. The baseline made no value pick and so avoided that loss.
What it means
For Hoffenheim, the win lifts them to 10th in the Bundesliga with 9 points from 7 games (W3 D0 L4, goals 14 for and 14 against). The form line now shows two straight wins: 2-1 at Borussia Mönchengladbach on 18 October, then this 3-0. It follows a run that included a 2-3 home loss to Hamburger SV and a 1-3 defeat at Paderborn. The next question is consistency, and the next fixture is away at Bremen on 31 October. The site’s forecast for that game locks 24 hours before kick-off.
For Leverkusen, the defeat leaves them 4th with 13 points from 7 (W4 D1 L2, goals 16 for and 12 against). It ended a run of three straight wins, and the 0-3 scoreline was their first blank in the recent results we hold. Their next test is a home game against Stuttgart on 31 October, which will show whether this was a one-off or a pattern. That forecast also locks 24 hours before kick-off.
For the models, the lesson is humility about favourites. A shared tilt toward the better-rated side cost every entry here, and the grades will carry into the next round of predictions.
