Methodology
Every number on GoalMath comes from the model documented below. We publish it in full — including the calibration record and the failure modes — because a prediction site that won't show its misses can't be trusted.
The model
Expected goals for each team come from attack/defense strengths relative to league averages:
λA = away attack × home defense × league avg away goals
Scoreline probabilities follow the Dixon-Coles (1997) distribution: independent Poisson marginals with a low-score dependence correction, ρ = −0.13:
We compute the full grid (0–0 through 12–12), normalize, and aggregate to 1X2, over/under 2.5, BTTS, clean sheets and most-likely scores. Full derivation in our explainer.
League parameters
| League | Avg home goals | Avg away goals | ρ |
|---|---|---|---|
| England · Premier League | 1.53 | 1.22 | −0.13 |
| Spain · La Liga | 1.57 | 1.11 | −0.13 |
| Italy · Serie A | 1.28 | 1.15 | −0.13 |
| Germany · Bundesliga | 1.78 | 1.46 | −0.13 |
| England · Championship | 1.40 | 1.20 | −0.13 |
Averages from recent completed seasons; ρ fixed at the literature standard. We re-check these each season.
Calibration record — where the model fails
Backtesting the model taught us more from the failures than the successes:
- Model-vs-market divergence is not an edge. When our numbers disagreed with bookmaker odds, the market was usually right — divergence signals missing information in the model (injuries, motivation, lineups), not a mispriced market. We treat large divergences as a warning to check the news, never as a bet signal.
- Promoted teams break the priors. A team new to a league has no meaningful history at that level; strength estimates from a lower division don't transfer. Treat these matches as low-confidence.
- Slow-moving averages lag regime changes. When a league's scoring environment shifts (e.g., a +19% goals season), season-long averages underreact for weeks. We monitor this and say so when it happens.
- Small samples, wide error bars. With fewer than ~10 games of team data, strength estimates are mostly prior, not signal. The predictor is honest about inputs; garbage in, garbage out.
What we don't do
- No "guaranteed tips", no locks, no fixed matches — anyone selling those is lying.
- No affiliate links. GoalMath currently earns nothing from bookmakers or anyone else.
- No personal data collection beyond what's needed to run the site.