// model calibration · graded picks only
How accurate are the predictions?
Every finished pick stays on the record, right or wrong. We group those picks by the win chance the model gave its predicted winner, then compare that chance with how often the pick actually won.
A plain reading: a 70% call should miss about 3 in 10. If the 70–75% bucket has won roughly seven times in ten over a large sample, the model is well calibrated there. Early-season buckets with few picks will jump around.
Overall · 203 of 324 (63%)
Sample size: 324 graded picks across NBA, NFL, NHL and the Premier League. Expected correct from the model's own figures: about 201.3.
| Model chance | Picks | Expected win rate | Actual win rate |
|---|---|---|---|
| 50–55% | 74 | 52.5% | 50.0% (37–37) |
| 55–60% | 63 | 57.2% | 60.3% (38–25) |
| 60–65% | 58 | 62.4% | 60.3% (35–23) |
| 65–70% | 38 | 67.4% | 65.8% (25–13) |
| 70–75% | 28 | 72.5% | 78.6% (22–6) |
| 75–80% | 24 | 77.2% | 79.2% (19–5) |
| 80%+ | 22 | 83.9% | 100.0% (22–0) |
Accuracy score (average squared error): 0.212.
NBA · 95 of 139 (68%)
Sample size: 139 graded picks. The model's own figures said to expect about 91.5 correct, so it is running 3.5 ahead of what it forecast.
| Model chance | Picks | Expected win rate | Actual win rate |
|---|---|---|---|
| 50–55% | 22 | 52.5% | 59.1% (13–9) |
| 55–60% | 22 | 57.3% | 63.6% (14–8) |
| 60–65% | 30 | 62.6% | 53.3% (16–14) |
| 65–70% | 21 | 67.6% | 61.9% (13–8) |
| 70–75% | 14 | 72.7% | 71.4% (10–4) |
| 75–80% | 16 | 77.1% | 93.8% (15–1) |
| 80%+ | 14 | 84.9% | 100.0% (14–0) |
Accuracy score (average squared error, lower is better): 0.200 · coin-flip baseline 0.250. Full NBA record →
NFL · 25 of 42 (60%)
Sample size: 42 graded picks. The model's own figures said to expect about 27.8 correct, so it is running 2.8 behind what it forecast.
| Model chance | Picks | Expected win rate | Actual win rate |
|---|---|---|---|
| 50–55% | 8 | 52.7% | 25.0% (2–6) |
| 55–60% | 6 | 57.2% | 50.0% (3–3) |
| 60–65% | 5 | 62.6% | 60.0% (3–2) |
| 65–70% | 7 | 67.4% | 71.4% (5–2) |
| 70–75% | 6 | 71.8% | 83.3% (5–1) |
| 75–80% | 4 | 77.3% | 25.0% (1–3) |
| 80%+ | 6 | 82.0% | 100.0% (6–0) |
Accuracy score (average squared error, lower is better): 0.217 · coin-flip baseline 0.250. Full NFL record →
NHL · 77 of 123 (63%)
Sample size: 123 graded picks. The model's own figures said to expect about 73.0 correct, so it is running 4.0 ahead of what it forecast.
| Model chance | Picks | Expected win rate | Actual win rate |
|---|---|---|---|
| 50–55% | 42 | 52.4% | 52.4% (22–20) |
| 55–60% | 35 | 57.1% | 60.0% (21–14) |
| 60–65% | 22 | 62.3% | 68.2% (15–7) |
| 65–70% | 10 | 67.1% | 70.0% (7–3) |
| 70–75% | 8 | 72.5% | 87.5% (7–1) |
| 75–80% | 4 | 77.3% | 75.0% (3–1) |
| 80%+ | 2 | 83.1% | 100.0% (2–0) |
Accuracy score (average squared error, lower is better): 0.223 · coin-flip baseline 0.250. Full NHL record →
Premier League · 6 of 20 (30%)
Sample size: 20 graded picks. The model's own figures said to expect about 9.0 correct, so it is running 3.0 behind what it forecast.
| Model chance | Picks | Expected win rate | Actual win rate |
|---|---|---|---|
| 50–55% | 2 | 51.6% | 0.0% (0–2) |
| 55–60% | 0 | — | — |
| 60–65% | 1 | 60.6% | 100.0% (1–0) |
| 65–70% | 0 | — | — |
| 70–75% | 0 | — | — |
| 75–80% | 0 | — | — |
| 80%+ | 0 | — | — |
Accuracy score (average squared error, lower is better): 0.218 · coin-flip baseline 0.250. Full Premier League record →
Probabilities come from ProbaLab's own model. They are for entertainment, not advice. Figures match the sport record pages for the same graded set.