PROVEN, NOT PROMISED

How accurate is Goalmetry, really?

Anyone can claim they're good. So here are the actual numbers — measured the only honest way: every prediction made using only matches played before that game, with the same model that runs in production. No cherry-picking. We show where we win and where we don't.

Predictions scored

376,004

out-of-sample

Skill vs no-skill

+19.2%

Brier skill score

Brier score

0.199

baseline 0.246

Calibration error

0.5%

avg predicted vs real

Performance by market

Brier score is the average squared error of our probabilities — lower is better. The baseline is a model that always guesses the long-run average. The gap is our skill.

MarketPredictionsBrierBaselineSkillCalib. errorConfident calls
Goals (Over/Under)112,0040.1910.211+9.2%0.8%78.0% righton 75,662 calls ≥65%
Shots on target60,0000.1850.210+12.1%1.4%79.1% righton 40,525 calls ≥65%
Cards48,0000.1990.250+20.5%1.5%76.0% righton 15,500 calls ≥65%
Goalkeeper saves72,0000.2070.234+11.2%0.7%70.6% righton 2,841 calls ≥65%
Corners48,0000.2270.246+7.6%1.8%72.9% righton 14,610 calls ≥65%
Total shots36,0000.1930.230+16.1%1.1%69.4% righton 2,667 calls ≥65%

"Confident calls" = how often the outcome happened when our model put it at 65% or higher. Honest confidence means this tracks the probability, not 100%.

Are our probabilities honest? (Goals markets)

We grouped every goals prediction by the probability we gave it, then checked how often it actually happened. A trustworthy model's bars line up: when we say 70%, it lands about 70% of the time.

20-30%
27.0%39.0%
30-40%
36.6%38.8%
40-50%
45.7%46.0%
50-60%
55.0%54.3%
60-70%
65.5%66.6%
70-80%
75.0%74.7%
80-90%
84.7%84.0%
90-100%
92.3%90.5%

what actually happened · what we predicted. When they line up, our probabilities are honest.

How we measured this

Walk-forward out-of-sample backtest: every prediction uses only matches played before that fixture, with the same model and parameters that run in production. Brier score = mean squared error of our probabilities (lower is better); the baseline is a no-skill model that always predicts the historical base rate.

Report generated on 20 June 2026 from 376,004 out-of-sample predictions. Numbers are recomputed from live historical data, never hand-entered. Goalmetry is data-driven analysis, not betting tips — markets are unpredictable and no model is right every time. 18+.

See today's edges free for 3 days