Reading win probabilities & calibration
A win probability is only useful if it is honest. Here is how to read ours — and how to tell a calibrated model from a confident-sounding one.
A percentage is a long-run frequency
When we say a player has a 58% chance, we mean: across many matches with that same rating gap, that side wins about 58 out of 100. It is not a guarantee for one match — table tennis is high-variance and the underdog wins often.
Why so many predictions sit at 52–58%
In the round-the-clock leagues the players are closely matched, so honest probabilities cluster near a coin-flip. That is a feature, not a weakness. A model that constantly showed 75–80% in these fields would simply be lying.
More lopsided fields — the Challenger Series, the women's events — genuinely produce higher probabilities, and the model shows them.
How to read a calibration chart
On every league page we plot predicted win rate against actual win rate. If the dots sit on the diagonal, the model is calibrated: its 60% really happens 60% of the time. Dots above the line mean it was too cautious; below, over-confident.
We grade every prediction walk-forward — using only information available before the match — so the accuracy you see has no hindsight baked in.
Brier score in one sentence
The Brier score measures how close probabilities are to outcomes (lower is better); 0.25 is a coin-flip, and a good model beats that while staying honest. It is the single number we trust most to judge prediction quality.