Every graded Tennis prediction, sorted by most recent. No cherry-picking; no filtered wins.
What these numbers describe
300 of these 300 predictions come from the live model; the remaining 0 are the walk-forward Elo floor.
The floor is Elo + Poisson with ratings frozen at each match — no news, no calibration, no market. It is a lower bound: the live model should beat it.
ECE(expected calibration error) is the count-weighted average gap between confidence and outcome — < 5% is well-calibrated, > 10% means the model's probabilities can't be trusted at face value yet. Independent of accuracy: a model can be 65% accurate and 25% mis-calibrated.
Above the dashed diagonal = under-confident on this sport; below = over-confident (the dangerous direction).
Sorted by sample count. Low Brier + high sample count = the model understands this league.