Model accuracy
Predicting a table tennis match before it starts is hard. Published research and rating-based systems top out around 65–70% accuracy — upsets are a structural part of the sport, not a modelling failure. Any product claiming more than that pre-match is overclaiming.
So instead of claiming, we publish. Every prediction is logged before the match with its probability and interval, never revised once the result is known, and scored here against what actually happened.
- Resolved predictions
- 23
- Brier score
- 0.158
- Log loss
- 0.497
Matches completed with a logged pre-match prediction.
Mean squared error. Lower is better; 0.250 is a coin flip.
Penalises confident mistakes hardest. Coin flip: 0.693.
Reliability
Predictions grouped into ten bands by predicted win probability. A well-calibrated model’s observed bar matches its predicted bar in every band: when we say 70%, it should happen about 70% of the time.
| Band | Predictions | Mean predicted | Observed |
|---|---|---|---|
| 0%–10% | 0 | — | — |
| 10%–20% | 2 | 19% | 0% |
| 20%–30% | 2 | 27% | 0% |
| 30%–40% | 5 | 36% | 20% |
| 40%–50% | 3 | 47% | 0% |
| 50%–60% | 4 | 53% | 75% |
| 60%–70% | 3 | 63% | 67% |
| 70%–80% | 4 | 76% | 100% |
| 80%–90% | 0 | — | — |
| 90%–100% | 0 | — | — |
Statistics are computed live from every logged prediction with a known result. Predictions are created once, before the match, and never recomputed afterwards.