Findings / Market calibration
Is the published market calibration consistent? A favorite-longshot audit
Proportional devig of the two-way closing price -> implied probability; favorite = max(q, 1-q); bucketed by favorite implied probability, each bucket reports n, mean implied, realized favorite win-rate, the gap, and a 95% Wilson interval. This exhibit grades the closing price, not our model. These historical comparisons do not establish market efficiency or future calibration.
The published bounds are row-count Wilson calculations using z=1.96. Distinct match or game identity and independence of retained records have not been verified, so 95% coverage is not established. This is a limitation of the available evidence, not a finding of duplicate or dependent records. No adjustment for possible dependence is included; the original rates and bounds are preserved.
Wilson interval method (NIST)Tennis -- match winner
Pinnacle (close) — match winner — source records n=33,713
Published bins cover 33,713 of 33,713 source records.
| Implied bucket | n | Mean implied | Realized fav win | Gap | Nominal 95% Wilson |
|---|---|---|---|---|---|
| [0.5, 0.6) | 9,664 | 0.5532 | 0.5506 | -0.0026 | (0.5407, 0.5605) |
| [0.6, 0.7) | 9,744 | 0.6483 | 0.6495 | +0.0013 | (0.64, 0.6589) |
| [0.7, 0.8) | 7,570 | 0.7474 | 0.7563 | +0.0089 | (0.7465, 0.7658) |
| [0.8, 0.9) | 4,885 | 0.8452 | 0.8583 | +0.0131 | (0.8483, 0.8678) |
| [0.9, 1.001) | 1,850 | 0.9321 | 0.9497 | +0.0176 | (0.9388, 0.9588) |
Scroll horizontally for all columns.
Published interpretation — mild monotone favorite-longshot bias.
MLB -- moneyline
closing line — moneyline — source records n=27,983
Published bins cover 27,976 of 27,983 source records. 7 source records are outside the published bins. No bucket-level rates or intervals are published for those records.
| Implied bucket | n | Mean implied | Realized fav win | Gap | Nominal 95% Wilson |
|---|---|---|---|---|---|
| [0.5, 0.55) | 10,685 | 0.5226 | 0.5258 | +0.0032 | (0.5163, 0.5352) |
| [0.55, 0.6) | 8,497 | 0.5731 | 0.5786 | +0.0055 | (0.568, 0.589) |
| [0.6, 0.65) | 5,800 | 0.6233 | 0.6155 | -0.0078 | (0.6029, 0.628) |
| [0.65, 0.8) | 2,994 | 0.687 | 0.7031 | +0.0161 | (0.6865, 0.7192) |
Scroll horizontally for all columns.
Published interpretation — essentially efficient -- no systematic bias.
Observation window
committed odds corpora (decade-plus per sport), not a rolling cache; pregame closing prices.
Confounds / notes
- proportional devig shifts absolute probability levels vs. other devig methods
- closing snapshot only, not a continuous feed
- favorite is defined by closing price, not by any external ranking
- bucket boundaries are declared upfront, not tuned to the data
- this grades the closing price's own calibration, not our model -- nothing here is a claim about beating it
This is also a calibration exhibit, alongside reliability (which checks our own probabilities, not the market's) and league parity (a different descriptive read on competitive balance).