1,900 EPL matches · Five seasons · Aggregate only
How much does the football odds de-vig method change 1X2 probabilities?
Direct answer: Shin and proportional margin removal differed by more than 0.50 percentage points on at least one outcome in 1,076 of 1,900 markets (56.6%). The mean largest gap was 0.66 pp; the 95th percentile was 1.41 pp.
Published by Football Proof AI · Published · football-1x2-margin-removal-benchmark/1.0.0 · Editorial technical note; not externally peer reviewed
Canonical publication record
Abstract
An aggregate-only comparison of proportional and Shin margin removal on closing-average 1X2 odds for 1,900 completed Premier League matches from 2021/22 through 2025/26, including outcome-level method gaps, paired proper scores, block-bootstrap intervals, source hashes and synthetic conformance vectors.
- Author and publisher
- Football Proof AI
- Technical report
football-1x2-margin-removal-benchmark/1.0.0- Published
- Last modified
- Release status
- Current release
- Review status
- Editorial technical note; not externally peer reviewed
- Version history
football-1x2-margin-removal-benchmark/1.0.0: Initial public release.
- Immutable artifacts
- empirical-1.0.0.json
sha256:2ae69ffb80fae20db7889a425634593b61ebf91115c0e9094d8c881091450b3b - football-1x2-margin-removal-benchmark-v1.csv
sha256:bdf5a8f7f6e380fc4cc3c8fb3054b99f8ff75dc5b8c23fe02884af684d0c151b - conformance-1.0.0.json
sha256:f2973c7573af0647055392d94cb40c5aa7e480ffc22c7fc87f32424ca91a40aa - empirical-1.0.0-manifest.json
sha256:f6159152fe67f77681fa1aef75bee18f7c6ea2d326b1c180580d9bca6cab8429
- empirical-1.0.0.json
- External release resources
First-party result
The method changes the probability shape more than the favourite label
Across this frozen sample, the two methods never changed which 1X2 outcome had the highest probability. They did change the displayed probability: 431 markets had at least one outcome move by more than one percentage point. That distinction matters when probabilities are compared with a model, calibrated, or converted to fair odds.
Where the gap grows
Stronger favourites produced the largest method sensitivity
The grouping uses the proportional no-vig favourite probability. It is descriptive, fixed before the table is displayed, and is not a rule for choosing either method.
| Favourite band | Markets | Mean largest gap | 95th-percentile gap | Above 0.50 pp |
|---|---|---|---|---|
| Below 45% | 597 | 0.24 pp | 0.41 pp | 0.2% |
| 45% to below 55% | 482 | 0.53 pp | 0.81 pp | 52.7% |
| 55% or higher | 821 | 1.05 pp | 1.56 pp | 100.0% |
Outcome-aware check
The paired score difference was tiny and its block-bootstrap interval crossed zero
Shin minus proportional class-averaged Brier score was -0.000033 and log loss was -0.000380 nats. Their deterministic weekly-block 95% intervals were [-0.000180, +0.000114] and [-0.001132, +0.000394]. This sample therefore does not justify calling either method a universal winner.
| Season | Markets | Proportional Brier | Shin Brier | Shin − proportional |
|---|---|---|---|---|
| 2021/22 | 380 | 0.184708 | 0.184598 | -0.000110 |
| 2022/23 | 380 | 0.190397 | 0.190357 | -0.000040 |
| 2023/24 | 380 | 0.175582 | 0.175275 | -0.000307 |
| 2024/25 | 380 | 0.191729 | 0.191802 | +0.000072 |
| 2025/26 | 380 | 0.202579 | 0.202799 | +0.000220 |
Margin-size check
Larger overrounds generally widened the method gap
Overround is sum(1 / decimal odds) − 1. It is not the same as the share of each stake retained by a bookmaker, and it is not a probability forecast.
| Overround band | Markets | Mean overround | Mean largest gap | Maximum gap |
|---|---|---|---|---|
| Below 4% | 864 | 3.81% | 0.60 pp | 1.68 pp |
| 4% to below 6% | 966 | 4.64% | 0.70 pp | 2.05 pp |
| 6% or higher | 70 | 6.25% | 0.85 pp | 2.39 pp |
Reproducible method
What was calculated—and what was deliberately not published
- 01
Freeze the inputs
Five declared Football-Data E0 files are bound by SHA-256 in the manifest. Each season contributes exactly 380 complete AvgCH, AvgCD and AvgCA odds rows.
- 02
Calculate both methods
Inverse decimal odds produce raw implied probabilities. Proportional normalization and a deterministic Shin bisection each return a complete distribution summing to one.
- 03
Compare the same markets
Outcome-level gaps, top labels, Brier scores and log loss use the same 1,900 completed matches. Weekly blocks preserve some temporal dependence in the descriptive interval.
- 04
Publish aggregates only
The release includes aggregate JSON and CSV, source URLs and hashes, but does not redistribute raw match or odds rows or claim a licence to do so.
Open evidence package
Download the exact aggregates, tests and provenance
The aggregate release is licensed under CC BY 4.0. Source files remain at their publisher; their URLs and hashes are included for reproducibility, not relicensing.
JSON v1
Headline, group results, formulas, source hashes, limitations and uncertainty.
CSV v1
Overall, season, overround and favourite-strength aggregate rows.
Conformance JSON
Four synthetic odds vectors with deterministic proportional and Shin outputs.
Manifest JSON
Artifact byte sizes, SHA-256 digests, generator and source-input identities.
Interpretation boundary
No-vig probability is a method output, not the true probability
A complete market is required because the margin is a property of the whole outcome set. Neither proportional nor Shin removal reveals a bookmaker's private view, later information, limits, liabilities or the actual chance of a result. Use the odds implied-probability calculator to inspect one complete market, and the model-vs-bookmaker benchmark for a paired historical model comparison.