Accuracy
Predicted vs Actual Ks
3,502 pitcher-starts evaluated, self-updating daily
MAE
1.96
Ks off, on average
Bias
+0.24
neutral
Error Distribution
How far off are predictions, in Ks?
Green bar = perfect predictions · symmetric histogram = no bias · long tails = the model occasionally misses badly
| Rate n=3,502 | |
|---|---|
| Hit-rate bands · how often is the rounded prediction within X K? | |
| Exact (±0 K) | 15.8% |
| Within ±1 K | 45.7% |
| Within ±2 K | 69.6% |
| Within ±3 K | 84.3% |
| > 3 K off (catastrophic) | 15.7% |
| CI coverage · % of actual outcomes inside the predicted interval | |
| 80% CItarget 80% | 82.8% |
| 90% CItarget 90% | 90.8% |
| 95% CItarget 95% | 95.3% |
| 99% CItarget 99% | 99.3% |
Conformally calibrated intervals
The model is slightly conservative: nominal NB 80% catches 88.0% of actuals due to discrete-quantile overshoot. Pitcher pages use tighter calibrated percentiles that hit each target empirically.
80% CI
[10%, 90%]
[15.2%, 84.8%]
90% CI
[5%, 95%]
[8.4%, 91.6%]
95% CI
[2.5%, 97.5%]
[4.6%, 95.4%]
99% CI
[0.5%, 99.5%]
[1.2%, 98.8%]
MAE Over Time
Drift sentinel
Flat = consistent model · rising = retrain trigger
Scatter
One dot per pitcher-start
Dashed diagonal = perfect prediction. Hover any point to inspect.