Model

Model calibration, rolling closing line value, and scoring metrics.

Versionv1.01Samplen = 1868Range2026-07-15 – 2026-10-15Last logged2026-10-15

These are the results of my model that finds positive EV (expected value) odds, using Pinnacle as a sharp bookmaker and 14 “soft” UK bookmakers to compare. The model devigs the odds and compares them to see if there is an edge. To measure the success of this model we look at the CLV (closing line value) beat rate. If you have a better price than the last price the bookmaker publishes before kick off, then you have “beat” the CLV. A CLV beat rate of > 50% indicates an edge. Currently the model sits at a 62.2% CLV beat rate at n = 1868, at p = 0.000 — we’ve found an edge — p is at or below 0.01 and the beat rate is above 50%.

This page is a work in progress. The pipeline, the methodology, and this write-up will keep changing as more data comes in and as I find and fix issues with the approach.

Rolling closing line value

Calibration

Summary

Closing line value — n = 1868 (events with a finalized closing price)

Metric	Value
Beat-close rate	62.2% (1162/1868)
Mean CLV (± SE)	+8.05% (± 1.30%)
p-value (vs. 50% null)	0.0000

Settled bets — n = 1482

Metric	Value
Win rate	32.1% (475/1482)
Log loss	0.581
Brier score	0.198

Does the edge survive to kickoff?

The bot re-checks every logged bet’s EV against live odds up to kickoff. Two sticky flags come out of that: a bet is marked once its EV ever dips under 2% (the bar it had to clear to get logged at all), and again, separately, if it ever falls to near-zero — no realistic edge left. Once set, a flag stays set even if EV recovered later, so this measures whether the edge ever disappeared, not just where it ended up. Only re-checks under the corrected formula (n = 1034) count here — bets re-checked before that fix landed are excluded, since that version conflated EV decay with the soft book’s own price movement rather than Pinnacle’s.

CLV by decay bucket

Every corrected-formula bet falls into exactly one bucket: never flagged, dipped under the 2% EV barrier, or dipped further and confirmed value-gone by kickoff. The table below repeats the closing-line-value stats for each bucket separately, so the flag’s actual relationship to CLV is visible directly instead of inferred from a single filtered comparison.

Closing line value by decay bucket — corrected-formula bets only, n = 1034

Bucket	n	Beat-close rate	Mean CLV (± SE)	p-value
No flag	351	72.6% (255/351)	+8.41% (± 0.61%)	0.0000
Under 2% EV	181	84.0% (152/181)	+8.75% (± 0.64%)	0.0000
Value gone	502	62.4% (313/502)	+3.92% (± 0.42%)	0.0000