Bookmakers don’t publish true probabilities — they publish prices designed to make money. Every market carries a built-in margin (the “overround”, or vig), typically 4–8%, so the implied probabilities of all outcomes add up to more than 100%.

Implied probability  =  1 / decimal odds

To compare a price to fair value, that margin has to come out first.

Every bookmaker’s price contains its margin. Averaging the implied probability across 15+ books and normalising the result to exactly 100% removes it, which gives a margin-free reading of what the market as a whole thinks — and, more usefully, shows where the best available price on each outcome actually is.

Fair P(outcome)  =  mean(1 / odds)  ÷ Σ mean(1 / odds)

Two guards keep the arithmetic clean: a single book pricing an outcome wildly out of line (a “palpable error”) is discarded before averaging, and a market is only read when every outcome in it is priced.

What this is not. Until August 2026 this was our primary signal, published under the name “The Sniper” and justified by Kaunitz et al. (2017) — beating the bookmakers with their own numbers. Measured over 193 settled selections it returned a yield of −47.0% and closing-line value of −24.9%, and its Brier score came in at 0.3051 against the market’s own 0.3050. That last number is the whole story: the de-vigged consensus IS the market’s price, so it adds no information and cannot meaningfully disagree with the thing it was derived from. It is a best-price finder, which is a genuinely useful tool, and that is now the only job it does here. It no longer publishes selections.

And the one that publishes still does not set the price. Over 44,000 settled selections we measured what happens when the goals model and the sharp closing line disagree — and the answer is that the model loses, harder the louder it argues. The table below is that measurement, read live from the same place every figure on this site reads it from.

The bar a model has to clear to be turned back on is deliberately hard to fake: 300 settled selections, closing-line value above +0.5%, and a CLV z-score above 2. Yield is reported and does not count — at these sample sizes it is noise, and a perfectly efficient market produces a “significant” winning season 77% of the time.

Two models are being measured against that bar right now without publishing anything: a corners model and a cards model. Both discriminate better on historical results than the goals model does — the actual over-3.5-cards rate runs from 25.3% to 59.7% across the cards model’s deciles, in perfect order — but no bookmaker corner or card prices exist anywhere in 33 seasons of history, so neither can be backtested. They are being paper-traded against live prices instead, and they will publish when the numbers above say so and not before. Everything they log settles into the same public record on the performance page.

Your edge answers one question: for every £1 you put on this price, how much should you expect to win or lose over the long run? (The textbooks call it Expected Value.)

Edge (EV)   =  modelProb × decimal odds − 1
modelProb   =  (edge + 1) / decimal odds

Model gives 55% on a 2.10 price:
Edge = 0.55 × 2.10 − 1 = +0.155  →  +15.5% EV

A positive edge means the price is bigger than it should be. It is the same number everywhere on the site, because every page uses the same formula.

Implied %

The raw price read as a probability, margin included — 1 ÷ odds. A market’s three implied percentages always sum to more than 100.

Fair %

The same market with the vig removed, normalised to exactly 100. This is what the market really thinks, and it is the baseline an edge is measured against.

Model %

Our own probability for the outcome, produced independently of the price.

Edge

The EV, shown next to the probability-points gap between Model and Fair (e.g. +6.1pp vs fair) — the intuitive read of how far our number sits from the fair market price.

Raw edge alone is misleading — a +9% edge at 2.06 is a four-point disagreement with the market, and the same +9% at 12.0 is barely half a point. So the score is not built on the edge. It is built on the probability gap, measured against how wrong this model usually is:

p_market  =  the Fair % beside the pick     the price with the margin taken out
p_model   =  (1 + edge) / odds             what our edge at that price implies

z    =  (p_model − p_market) / σ
MES  =  100 × (1 − 0.35 ^ z)

σ is the model’s MEASURED probability error, not a chosen weight.

That makes a score readable as a sentence: 65 is a disagreement exactly one measured error bar wide, 88 is two. Nothing picks those numbers — invert the formula and every cutoff on the ladder falls out as a round multiple of σ. A model with no measured error bar gets no score at all rather than a flattering one.

The market side is the Fair % you can see — the price with the bookie’s margin taken out, which is the same number the ladder beside every pick draws. So the gap the score measures is the gap you are shown: one number doing one job, not two that agree when you are lucky.

There is one score. Until August 2026 the board also carried a Kelly-shaped rating, which is the right shape for deciding how much to stake and the wrong shape for deciding how sure we are — it contains no error bar, so the ladder’s rungs meant nothing when applied to it. Staking lives on its own axis now, in units, and the score answers one question only.

Match odds · 1X2

Home / Draw / Away, straight from the de-vigged consensus.

Goals · Over/Under & BTTS

A Dixon-Coles bivariate Poisson, anchored to the consensus 1X2 so goals and the result never disagree.

Data models · Corners & Cards

A team-stat Poisson, deliberately shrunk toward the market so sparse data can never publish an off-market edge.

One card per match merges all five markets; a market with no value stays collapsed, and value opens automatically.

Win rates over small samples are mostly noise. The honest measure of whether an edge was real is Closing Line Value — did our signal price beat the market’s final price at kick-off? Consistently beating the closing line is the single best predictor of long-term profit, and we track it on every settled signal.

CLV  =  ln( signal odds / closing odds )

Signalled 2.10, closed 1.90  →  ln(2.10/1.90) = +10.0%  beat the close

Every signal is settled against the real result and disclosed on the performance page — wins and losses.

Outlier filter

Palpable-error prices are dropped before they can create a fake edge.

Market shrink

Data-driven markets are pulled toward the market price, capping how far any signal can diverge.

Integrity check

Every engine cycle re-checks that all probabilities, edges and scores are internally consistent, and alerts on any violation.

No model is perfect. Team-strength and goals inputs update periodically and don’t yet capture late lineup changes, weather or motivation. The corners and cards models lean on patchy international stats, which is exactly why they’re anchored to the market. Treat every MaxEdge number as one input among many.