On 23 July we froze the rules of a public experiment: take every value signal our site publishes, check it against the live Betfair price every half hour, and record a dry-run bet whenever the exchange price beat our fair probability by at least 2.5% after commission. Half-Kelly stakes on a simulated bankroll, settled at real results, every decision on a public ledger. The file was hashed and timestamped so nobody, us included, could move the goalposts.
Two months and 262 settled bets later, we are closing it. Not because it failed its gate. Because of what the ledger showed once we looked at it properly.
| Value | |
|---|---|
| Settled bets | 262 |
| Win rate / break-even at average odds 2.50 | 43.9% / 40.0% |
| ROI, stake-weighted (half-Kelly) | +3.4% |
| ROI, flat one unit per bet | +0.8% |
| Closing line value vs Pinnacle close | +2.5%, 61% of bets positive |
| CLV as originally published | +6.5%, 83% of bets positive |
The last two rows are the uncomfortable part. Until this week the ledger compared the Betfair back price with Pinnacle's raw closing odds. Pinnacle's raw odds include its margin — about 4.95% overround on these matches, roughly 1.7 points per selection. An exchange price carries no margin. So an exchange price will "beat" a raw bookmaker price nearly by construction, and our published CLV was flattered by about four points.
The pre-registered rule said "same-book CLV versus the Pinnacle close". Same-book has to mean the same kind of price, which means the devigged close, using the same power devig every other page on this site uses. Recomputed that way, every bet's CLV on the ledger has been corrected, and the average is +2.5%, not +6.5%. This is the second measurement error we have found and disclosed this month; the first used in-play prices as closing prices. We would rather publish these than have someone else find them.
The overall figure hides a split that is too clean to be noise.
| Segment | Bets | ROI |
|---|---|---|
| Fair price stale (benchmark reading older than 2 h) | 77 | −14.8% |
| Bet placed 3 to 8 hours before kick-off | 59 | −21.0% |
| Betfair liquidity under 500 | 115 | −7.0% |
| Clean core: fresh fair, under 3 h, liquidity ≥ 500 | 119 | +15.3% weighted, +2.4% flat |
| Everything outside the core | 143 | −6.3% |
Each of the three losing segments has a mechanism, not just a bad run.
Stale fair. Our fair probability is derived from the sharp market. When the latest live reading of that market was older than two hours, the code fell back to the fair price logged when the signal was first detected — sometimes days earlier. An "edge" measured against a days-old price is mostly the line having moved since, not value. The signature is visible: these bets showed the highest CLV and the worst results.
Early execution. We charged a 2% extra edge for bets placed more than 2.5 hours out. It was not enough. The fair price simply degrades with distance from kick-off; lineups, news and money arrive later.
Thin liquidity. The specification itself warned that displayed Betfair prices at low liquidity "may not be fillable at size" and would flatter results. It did. At liquidity under 500, CLV looked better and results were worse, which is exactly what a price that is not really there looks like.
The much-discussed "edge paradox" — bets with the largest claimed edge, 8 to 12%, lost the most at −34.9% — is the same finding from a different angle. The largest edges were disproportionately stale, early or thin. A 10% edge at Betfair is a data artefact.
The clean core is +15.3% stake-weighted, which sounds like a strategy. At flat stakes it is +2.4% on 119 bets, inside an interval that comfortably includes zero. The difference is 17 large half-Kelly bets that happened to win. Kelly sizing multiplies whatever your edge estimate says, and where the estimate was inflated (the three segments above) it multiplied the losses too. Weighted ROI on a small sample tells you who got lucky, not who has an edge.
So the honest summary of v1 is the one we wrote before it started: our own simulations put value-signal edge at the exchange near zero, and 262 bets did not contradict that.
Take v1's ledger, apply the three filters, publish +15%. That is exactly the after-the-fact selection we spend most of this site criticising in tipsters, and pre-registration exists to stop us doing it. v1 stands as it ran, with its corrected CLV, on a ledger nobody can edit.
Instead, the three lessons are now rules in a separately pre-registered successor, TIPS-EXP-003: fresh fair only or skip, execution only within three hours of kick-off, liquidity of at least 500, ROI reported flat as well as weighted, CLV against the devigged close. Its gate is stricter than v1's: 150 or more bets with flat-stake ROI above zero and an interval excluding zero. Its stated prior is "roughly break-even, possibly slightly positive". It started today, with its own hash and timestamp, and its ledger is live here.
If v2 also lands at zero, that will be the result, and it will be published the same way.
Method notes: 262 bets recorded between 18 July and 21 September 2026 under the frozen rules of TIPS-EXP-002; full autopsy document with SHA-256 and OpenTimestamps anchor on the public record repository. CLV = Betfair back price × devigged Pinnacle closing probability − 1. Flat ROI = mean of (odds − 1) × 0.98 on wins and −1 on losses. Segment ROIs are stake-weighted unless stated.