Receipts — is the model calibrated?
Calibration means the probabilities are honest: when we say 60%, it should happen about 60% of the time. Below is our live record — every settled prediction, bucketed by what we predicted vs. what actually happened. Dots on the dashed line = perfectly calibrated. Nobody selling picks shows you this; it's the only score that can't be cherry-picked.
Regenerated monthly from the full prediction record · last run 2026-07-10 · window 2026-04-19 → 2026-07-08
ATP — the validated scope (1234 settled predictions)
What the plot says, in plain English
- In the 90–100% bucket we said 93.8% and it happened 81.3% (n=64) — overconfident there: those picks win less often than we say.
- In the 0–10% bucket we said 6.7% and it happened 15.2% (n=33) — underconfident there: those picks win more often than we say.
- In the 40–50% bucket we said 45.2% and it happened 37.3% (n=225) — overconfident there: those picks win less often than we say.
Dots above the diagonal = we were underconfident; below = overconfident. Bars along the bottom show each bucket's sample count.
| Bucket | n | Mean predicted | Observed | Gap (obs − pred) |
|---|---|---|---|---|
| 0–10% | 33 | 6.7% | 15.2% | +8.5% |
| 10–20% | 52 | 14.2% | 7.7% | −6.5% |
| 20–30% | 129 | 25.6% | 20.9% | −4.7% |
| 30–40% | 113 | 35.7% | 34.5% | −1.2% |
| 40–50% | 225 | 45.2% | 37.3% | −7.8% |
| 50–60% | 260 | 55.1% | 48.5% | −6.6% |
| 60–70% | 137 | 67.5% | 62.8% | −4.7% |
| 70–80% | 136 | 74.9% | 69.1% | −5.8% |
| 80–90% | 85 | 83.3% | 85.9% | +2.6% |
| 90–100% | 64 | 93.8% | 81.3% | −12.5% |
WTA — shown for transparency (1082 settled predictions)
These come from applying the ATP-trained model cross-tour. WTA rows never carry the validated filter badge; this table exists so you can see how the model behaves outside its validated scope, not to sell it.
| Bucket | n | Mean predicted | Observed | Gap (obs − pred) |
|---|---|---|---|---|
| 0–10% | 42 | 7.5% | 21.4% | +13.9% |
| 10–20% | 77 | 13.8% | 13.0% | −0.8% |
| 20–30% | 102 | 25.0% | 29.4% | +4.4% |
| 30–40% | 83 | 34.6% | 25.3% | −9.3% |
| 40–50% | 166 | 45.6% | 44.6% | −1.0% |
| 50–60% | 229 | 54.3% | 53.3% | −1.0% |
| 60–70% | 132 | 67.0% | 72.0% | +5.0% |
| 70–80% | 99 | 73.9% | 64.6% | −9.3% |
| 80–90% | 96 | 84.2% | 87.5% | +3.3% |
| 90–100% | 56 | 93.6% | 96.4% | +2.8% |
Check our work
Don't trust the rendering — pull the raw rows. The full ledger CSV is the stored fields row-for-row (no derived or renamed values), so you can recompute the aggregate CLV and every per-pick grade yourself; the formula is printed on the ledger. The calibration numbers above are the monthly job's stored output, rendered untouched — the bucket table IS the raw data.
What calibration means (and doesn't)
A calibrated model is not a profitable model — Pinnacle's closing line is calibrated too, and you can't beat it by agreeing with it. Calibration is table stakes: it makes our probabilities a usable fair line. We calibrate per surface with isotonic regression because raw model scores drift differently on clay, grass and hard courts. The product is what we do with the fair line: finding recreational books that disagree with it by 5–12% — the one band a six-year walk-forward validated. The full reasoning is in the methodology essay.
2026-07: We tested removing a calibration step — and kept it
Our research sprint found a calibration layer in our model that slightly hurts raw predictive accuracy. Before removing it, we re-checked whether our pick filter still worked without it. It doesn't: the picks that make our strategy profitable come disproportionately from exactly that layer's adjustments. Accuracy on paper and betting value are not the same thing — so the model stays as it is, and both experiments are in the research log.
Full numbers for the technical reader: the calibration step we kept.
The kill-gate, pre-committed
If the ledger's aggregate CLV on filtered picks is ≤ 0 over any rolling 6-month window with n ≥ 100 picks, the 5–12% edge claim comes off the site and the positioning reverts to pure process/education.
That sentence is the deal. The public ledger is the meter; nobody gets to move the goalposts after the fact.
Not betting or financial advice. 21+ where legal. About Baseline.