1. How the Models Work
Modern football prediction reduces match outcomes to a probability problem. The main approaches:
- Poisson models — estimate goal distributions from attack/defense strength;
- ELO-style ratings — dynamically updated team strength scores;
- Machine learning — classifiers trained on dozens of features: shots, possession, xG, lineups, schedule congestion;
- Market models — reverse-engineer consensus probability from odds movement (see our Kelly Index guide).
Our three models take different routes: the Kelly model reads market information, the 14-Match model is feature-based machine learning, and the Sanmuban model uses multi-factor composite scoring.
2. Reading Probability Outputs
A model outputs a distribution, not an answer:
Home 58% / Draw 24% / Away 18%
This means about 58% of matches under comparable conditions end in a home win. Confidence reflects how certain the model is — clearer feature signals and richer historical samples yield higher confidence.
Practical rules:
- Treat probability > 65% with confidence > 70% as priority matches;
- Independent models agreeing on one outcome compounds reliability;
- Sharp disagreement between models ("model conflict") is a classic upset zone.
3. Three Truths About Hit Rates
- Single-day hit rates are meaningless — every model streaks both ways; judge over 100+ matches.
- Structure matters — a 58% overall rate with 70% on high-confidence picks and 45% on low-confidence picks indicates a well-calibrated model.
- Hit rate ≠ profit — always backing favorites yields a high hit rate and long-run losses; a model's real value is finding outcomes the market underprices.
Our Dashboard publishes every prediction from all three models over the past 30 days — including the misses. No cherry-picking.
4. What Models Cannot Know
Dressing-room conflicts, late injuries, weather, unusual motivation. These are unquantifiable or arrive after the data cutoff. Therefore:
- Use AI predictions as a frame of reference, not an oracle;
- Respect the confidence score more than you crave the result;
- Long horizon, diversified, small stakes — the precondition for using any prediction tool.
Disclaimer
Educational content only, not betting advice. Please gamble responsibly.