A comprehensive look at 184 football matches analyzed between December 28-30, 2025. This period represents the early system performance — before the DIAMOND tier classification was implemented in early 2026.
Overview
During December 28-30, the system analyzed 184 matches from multiple leagues worldwide. The overall 1X2 accuracy stood at 44.6% — which was the baseline performance of our early ensemble model without the confidence filtering that DIAMOND now provides.
Key takeaway: raw ensemble predictions without quality filtering produced mediocre results. This was the catalyst for developing the DIAMOND tier system, which now filters predictions to only include matches where model agreement, profit model, and confidence thresholds all align.
Day-by-Day Results
December 28, 2025
65 matches analyzed with 42% accuracy. Heavy schedule with many lower-league matches where our models had limited training data. Value bets performed better than 1X2 picks, with 3 successful value bet detections.
December 29, 2025
58 matches analyzed with 46% accuracy. Better performance on top-5 league matches (Premier League, La Liga). The ensemble model showed stronger agreement when analyzing teams with richer xG data.
December 30, 2025
61 matches analyzed with 45.9% accuracy. Mixed results — good predictions on home favorites but weak on draw and away picks. 5 value bets identified, of which 3 were successful.
Key Insights
Why 44.6% was not enough
A 44.6% accuracy on 1X2 picks means the system was barely above random (33.3%). The problem was no quality filtering — every match got a prediction regardless of model confidence. This led to three major improvements:
- DIAMOND tier — Only predict when ensemble confidence is high (≥60%), model agreement is strong (≥75%), and the profit model confirms value
- XGGe/XGG2e models — New XGBoost classifiers with 73-87 features replaced the basic ensemble voting
- Profit Model gate — Every prediction now requires positive expected value from the meta-model
Value Betting Performance
Despite the low 1X2 accuracy, the value betting detection showed promise:
- 8 out of 14 value bets hit (57.1% strike rate)
- Average odds on successful value bets: 2.35
- Positive ROI on value bets: +8.3%
This early signal confirmed that our probability model was capable of finding true value — it just needed better filtering to separate the good predictions from the noise.
What Changed After This Period
These December results directly led to the overhaul of our prediction pipeline in January 2026:
- Implemented DIAMOND tier classification — filtering from ~184 matches/day to ~30-50 high-confidence picks
- Added XGGe model (73 features) and XGG2e model (87 features) as primary classifiers
- Built the Profit Meta-Model to gate all predictions by expected value
- Introduced AKO scoring for ranking prediction quality
Result: DIAMOND predictions now achieve 75-80% accuracy on selected matches, compared to the 44.6% we saw when predicting everything.
Conclusions
December 28-30, 2025 was a turning point for ROWIN88. The data proved that:
- Raw ensemble predictions on all matches are insufficient for profitable betting
- Value bet detection was already promising and worth developing further
- Quality filtering (what became DIAMOND) was essential for delivering reliable predictions
See today's DIAMOND predictions
Since implementing DIAMOND, our accuracy has been consistently above 75%.
Check predictions →