What the model got right last time.

77% of 39 matches is meaningfully better than chance ( 50% ) and better than what a random forecaster (Brier ≈ 0.25) would manage. It's in the same range as published academic models on comparable datasets.

The headline winner accuracy is the same as a naive baseline that always picks the higher-Elo side (77%). That's an honest finding: the Dixon-Coles addition doesn't move the picks much because both methods share the same Elo input. Where Dixon-Coles does help is in confidence calibration — the probability assigned to the winner — which shows up in the Brier and log-loss scores rather than the binary pick.

The misses are worth noting. The model failed to predict major upsets — Morocco beating Spain and Portugal at WC 2022, Switzerland beating Italy at Euro 2024, Uruguay beating Brazil at Copa 2024. Pre-tournament Elo simply doesn't capture the in-tournament form of a side hitting their stride. This is a known limitation, not a fixable one without live in-tournament updates.

Sample size warning. 39 matches is not enough to draw strong statistical conclusions about model calibration. The 95% confidence interval on a 77% accuracy estimate from 39 trials is roughly ±13 percentage points. Three tournaments' worth of knockout matches is illustrative, not definitive.