Expected Value Explained: Why a High Hit Rate Is Not Enough
Why This Concept Is Worth an Article
Expected value is the single idea that separates people who evaluate football predictions from people who collect them. It is also the idea most often reduced to a slogan — "positive EV good, negative EV bad" — which is true, useless, and hides everything that makes EV worth learning.
The mechanics are simple enough to do on paper. The consequences are not intuitive at all, and the least intuitive consequence is the one this article is built around: a high hit rate and a method worth following are different things, and they routinely point in opposite directions.
This is a teaching article. All numbers below are invented, neutral examples chosen to make the arithmetic legible. None of them describe a real fixture or a real result.
Step One: A Price Is a Probability Quote
Every decimal price implies a probability, and the conversion is one division: 1 divided by the price.
- A price of 2.00 implies 50%
- A price of 2.50 implies 40%
- A price of 1.30 implies about 77%
That implied probability is the market's estimate, inflated slightly by the margin built into the price. Stripping that margin out and reading a full market as a probability distribution is walked through step by step in the guide to football odds, which is the prerequisite to this article rather than a repeat of it.
Step Two: EV Is a Disagreement, Measured
EV exists only when two things are present: an estimate and a price. The estimate alone has no EV, and the price alone has no EV.
Take a neutral example. Some outcome is quoted at 2.50, implying 40%. A model estimates the same outcome at 46%. To turn that six-point gap into a single comparable number, score each published call: a correct call scores the price minus one, an incorrect call scores minus one.
- Correct: 2.50 − 1 = +1.50, which happens 46% of the time
- Incorrect: −1.00, which happens 54% of the time
- Expected score per call: 0.46 × 1.50 − 0.54 × 1.00 = 0.69 − 0.54 = +0.15
Now change one thing. The estimate stays at 46%; the price drops to 2.00.
- 0.46 × 1.00 − 0.54 × 1.00 = −0.08
Identical read of the match, opposite conclusion. EV is a property of an estimate and a price together, which is why a signal card has to be anchored to the price at the moment of publication, and why the same card read an hour later may describe a situation that no longer exists.
Step Three: Why a High Hit Rate Is Not Enough
Here is the part that surprises people. Three neutral examples, each publishing a hundred calls:
| Example | Average price | Hit rate | Expected score per call | Over 100 calls |
|---|---|---|---|---|
| A | 1.30 | 70% | −0.09 | −9 |
| B | 3.00 | 40% | +0.20 | +20 |
| C | 2.00 | 50% | 0.00 | 0 |
Example A gets seven calls in ten correct. Example B gets four. A's hit rate is close to double B's, and A's long-run direction is negative while B's is positive. The arithmetic for A: 0.70 × 0.30 − 0.30 × 1.00 = −0.09. Seventy percent sounds like mastery, but at a price of 1.30 the market was already saying the outcome happens about 77% of the time — so 70% is underperformance dressed up as a strong number.
That is the whole reason a screenshot of a hit rate tells you nothing. A hit rate quoted without the prices those calls were published at is not an incomplete statistic. It is an uninterpretable one.
Step Four: Variance, and Why Twenty Calls Show Nothing
Example B above has a genuinely positive expected score, and it is still wrong six times in ten. Five consecutive misses happen with probability 0.6 to the fifth power — about 7.8%, which is to say roughly once in every thirteen sequences of five. Long stretches of nothing but misses are a normal output of a method that works.
The practical consequence: twenty calls, or fifty, cannot separate skill from variance. A reader watching a method with a real edge will frequently see a losing month, and a reader watching a method with no edge at all will frequently see a winning one. This is exactly why we do not publish a headline accuracy figure anywhere on this site, and why we point verification at the full record instead — the reasoning is set out in is AI football prediction accurate.
Step Five: How to Read EV on a Prediction Signal
Four rules, each of which prevents a common mistake.
- 1.EV is not a ranking key. A large EV number usually means the model and the market disagree strongly, and strong disagreement is as often a symptom of thin data or an unusual market as it is a sign of opportunity.
- 2.EV is timestamped. It was computed against one price at one moment. In fast-moving markets that moment is short, which is why the publication time matters as much as the number.
- 3.A negative-EV candidate should never reach you. If a service publishes calls at prices where its own stated probability makes the score negative, it is publishing on a schedule rather than on a signal.
- 4.EV cannot verify itself. Every EV figure contains a model probability, and every model probability carries error. If the model says 46% and the truth is 38%, a positive-looking EV is negative in reality. Only calibration, tested over a long record, can check that — the arithmetic cannot.
The field-by-field version of this, on an actual signal card, is in how to read a signal card. The settled outcomes it all resolves against sit on our public prediction record.
FAQ
Is positive EV the same as expecting to be right? No, and the confusion is the source of most complaints about prediction services. A positive expected score can coexist with being wrong most of the time, as Example B shows. EV describes a distribution over many calls. It says nothing about the next one.
Why not just report EV as a percentage? Both forms exist and they are not interchangeable. The percentage-point gap between the model's probability and the implied probability (six points, in the example above) measures the disagreement. The expected score per call (+0.15) weights that disagreement by the price. Quoting one and calling it the other is a common sleight of hand.
Can I compute EV myself from a published call? You can compute the implied probability with one division, and you can compute the expected score if the service publishes its own probability estimate. If it does not publish a probability, EV is not recoverable from the outside — and a service that reports EV while hiding the probability behind it is asking you to accept the more derived number and skip the more basic one.