Expected Value and Closing Line Value, Without the Hand-Waving

"Positive EV" has become shorthand for seriousness in this category. The awkward fact is that the calculation takes five seconds, and the only thing that determines whether it means anything — the probability going into it — takes seasons to build and is never exact.

The calculation

Per unit staked:

EV = (probability × decimal odds) − 1

Probability	Odds	Calculation	EV per unit
45%	2.40	0.45 × 2.40 − 1	+0.080
45%	2.20	0.45 × 2.20 − 1	−0.010
55%	1.90	0.55 × 1.90 − 1	+0.045
55%	1.75	0.55 × 1.75 − 1	−0.038
30%	3.60	0.30 × 3.60 − 1	+0.080

Look at the pairs. Identical probability, a tenth of a point difference in odds, and EV flips from positive to negative. The price is not a detail — it is half the calculation.

Where it breaks: the probability

Three ways people obtain a probability, in ascending order of trustworthiness:

1. A feeling converted into a number. "I make them 60%." That is not a probability, it is confidence with a label attached. It cannot be checked and it cannot be improved.

2. A model never tested on unseen data. There is arithmetic now, but if the parameters were chosen on the same data the accuracy is reported from, the number is systematically inflated.

3. A model walked forward in time and then calibrated. This is the minimum bar. Weights for a given test season are fitted only on seasons before it, and the output is re-mapped so a stated 60% occurs around 60% of the time.

Even at level three the probability is not exact. It is only exact enough for its error to be estimable — and that is what makes it possible to know when not to act.

How small an error erases the edge

Worth computing once and remembering forever.

EV per unit at odds O and probability p is p·O − 1. The derivative with respect to p is simply O. In words: each probability point is worth O/100 units of EV.

Odds	EV destroyed per point of error
1.50	0.015
2.00	0.020
2.40	0.024
3.50	0.035
6.00	0.060

Now the troubling combination: at 2.40, an EV of 0.08 is wiped out entirely by an error of 3.3 points. Three points is a small error for a football model — comfortably inside the error bar of an estimate on a fixture with mediocre data.

Conclusion: a small positive EV built on an unstable estimate is not an edge. It is the appearance of an edge, created by displaying the number to three decimal places.

Why this forces selectivity

If a three-point error erases an 8% edge, then the bar for action cannot be "EV is positive". It has to be "EV is positive enough to survive the error bar here".

And the error bar is not constant. It widens when lineups are unconfirmed and kickoff is close, when data quality is low, when ensemble components disagree, and when the price is moving quickly between snapshots.

That is exactly the definition of the uncertainty budget: the discrepancy required before a fixture is classified as actionable, growing with each of those factors. A raw three-point gap can be actionable on one match and a PASS on another, correctly.

Four common traps

Computing EV against a raw price. The displayed price contains margin. Comparing a model probability to a raw implied probability produces systematically imaginary negative EV, then imaginary positive EV the moment you find a slightly better price. Devig first.

Hunting EV where margin is highest. High margin signals that the book is uncertain. In marginal leagues that usually coincides with our data being thin too. Two uncertain parties generate many apparent gaps and very little truth.

Accumulating many small-EV decisions. The logic sounds right — more and more small positive expectation. In practice, if your probabilities are biased in one direction, volume does not average the bias out, it multiplies it. Volume increases confidence in the outcome only if the estimates themselves are unbiased.

Confusing EV with outcome. Positive EV that lost was not a mistake. Negative EV that won was not skill. This is the hardest one to internalise, and it is why decision quality is graded entirely separately from results.

What you can actually measure after the fact

You cannot measure EV retrospectively. You can measure three other things, all more useful than they sound:

Metric	What it tells you
Calibration	Did the 60% bucket land near 60% across a sample?
Brier score	How far the stated probabilities sat from the actual outcomes
Closing line value	Was the price taken better than the market's final price?

The third is the closest available answer to "was this a good decision", and it is measured without reference to the result.

Closing line value, properly

The market moves from open to kickoff because information arrives: lineups, injuries, money, weather. The closing price is the market's summarising estimate after all available information has entered.

CLV asks one question: was the price I took better or worse than that?

Price taken	Closing price	CLV	Meaning
2.20	2.00	Positive	You were in before the market moved your way
2.20	2.20	Zero	Neither early nor late
2.20	2.40	Negative	The market moved against you after you entered

Why it isolates decision from luck

A single match result contains an enormous amount of luck: a corner, a deflection, a refereeing call. Two people making the identical decision get different results if the match is played twice.

The closing price contains no ninety-minute luck, because it is set before the match starts. It contains only information. Comparing against it isolates the question: did I see something before the market did?

Three ways to measure it

Raw odds difference is the tempting one and the worst, because 0.20 means something entirely different at 1.30 than at 6.00. Two better options:

Difference in percentage points. Convert both prices to implied probabilities, devig both, and take the difference. Most stable across markets.

Relative ratio. Price taken divided by closing price. 2.20 / 2.00 = 1.10, i.e. 10% above the close. Convenient for averaging across many decisions.

Critical in both: devig both sides. Margins often tighten toward kickoff, and comparing raw prices will attribute negative CLV to you purely because of a margin change.

What CLV does not tell you

It does not guarantee profit. A long run of positive CLV can accompany a losing period. CLV measures that you were early relative to market movement, not that the movement was toward the truth.

It is weak in thin markets. In a market with little money, the closing price is not a wise collective estimate — sometimes it is just the last price standing.

It does not fix a biased estimate. You can consistently beat the close on consistently wrong selections if a particular bias is shared by you and part of the market.

It is not always available. Suspended markets and missing snapshots mean the metric simply does not exist for that fixture, and that is better than a filled-in guess.

Where this leaves us

CLV feeds into the decision-quality score as one component of five, alongside decision-state discipline, price discipline, information freshness and acknowledged uncertainty. It carries weight only where a closing price exists; where it does not, the weight is redistributed rather than invented.

And the honest current status: CLV is not measured in the public record yet. The field is shown as not tracked, and the performance-pledge infrastructure that exists in the codebase is disabled and stays disabled until there are enough real settled predictions to stand it on. A promise made before the data exists is marketing, not measurement.

---

18+. WinPIQ is an analysis tool, not advice and not a promise. Betting can be addictive and money can be lost. Only stake what you can afford to lose, and if betting stops being entertainment, seek help. WinPIQ is not affiliated with Winner or the Israeli Council for the Regulation of Sports Betting.