Probability vs Confidence: Why a Prediction Carries Two Numbers, Not One

Every prediction published here carries two numbers that look like they measure the same thing. They do not. One is a probability — the model's estimate of what will happen. The other is a confidence score — how stable that estimate is.

Confusing the two is the most common mistake in reading a prediction, and it is common because in ordinary speech "I'm confident" and "that's likely" sound like the same statement at different volumes. As measurements they are separate questions, and they can be answered in opposite directions.

Two questions, not one

	The question	What moves it
Probability	What does the model think will happen?	Team strength, home advantage, form, lineup data
Confidence	How stable is that estimate?	Depth of history, lineup certainty, price stability, agreement between models

A probability of 62% says: out of a hundred situations that look like this one, we expect this outcome in about 62 of them. That is a statement about frequency, not about the specific match in front of you.

A confidence score of 40 says something else entirely: this estimate may move. Not that it is wrong — that the input behind it is thin, conflicting or very recent, and that asking again tomorrow could produce a materially different number.

Why both are published

A prediction that shows only a probability hides the part that decides how much weight it deserves.

Take two situations that look identical on paper, both at 62%:

• The first rests on seven seasons of history between two established sides, a published lineup, and a price that has not moved in two days. The confidence score is high.
• The second is a fixture in a competition whose history only goes back a few seasons, with doubt over two key players, and a price that swung sharply in the last hour. The confidence score is low.

The same 62%. Two completely different statements. The difference is not in the expected outcome — it is in how much the estimate might change before kick-off, and that is exactly what the second number exists to say.

What each one contributes to a decision

Probability is our side of the equation. It becomes a fair price and is compared against what the market is offering. The gap between the two is what is interesting.

Confidence decides how much weight that gap gets:

• Large gap, low confidence. Usually noise. If the estimate itself might move several percentage points, the gap against it might disappear with it.
• Small gap, high confidence. Often the stronger statement of the two, even though the bigger number sits with the other one.
• Large gap, high confidence. Rare. This is what gets recorded as a signal.
• No gap. However high the confidence, there is nothing to say. The state is recorded as passing on the fixture and counted as such.

None of those four is a promise. A 62% probability describes an outcome that will not happen in roughly 38 of 100 similar situations, and high confidence does not move that number at all — it only says the estimate itself is less likely to change.

What pushes a confidence score down

Four things, each saying "more information may change this picture":

1. Thin history. A promoted side, a competition with partial coverage, too early in the season.
2. Lineup uncertainty. A key player in doubt changes the estimate materially, and while the doubt is open the estimate is open with it.
3. A moving price. A sharp move in recent hours says somebody knows something, even when it is impossible to say what.
4. Internal disagreement. Several models run on the same fixture. When they split, the blended estimate is less reliable than a single number suggests.

Neither number is tested by one match

This is the easiest point to miss. A single result tests neither probability nor confidence.

A 62% probability that came in does not prove the number was right; a 62% probability that did not come in does not prove it was wrong. Both are entirely expected. The only way to test the numbers is across hundreds of predictions: collect everything said to be 60%–65%, and check whether roughly 62% of them actually happened. That is calibration, and it is what the published record measures.

Confidence is tested the same way on a different question: did low-confidence predictions in fact move more before kick-off than high-confidence ones.

Seeing it on a real record

The description above is worth less than one example you can open. So the two links below are not examples written for the article — they resolve to whatever is in the record at the moment you click:

• The latest published record — the most recent prediction, with the probability, the confidence score, the price and the signature on the same page.
• The latest pass — the case where the engine finished thinking and the answer was no. It is sealed and counted too.

If there is no matching record right now, the link returns a 404 rather than sending you to an older example. That is deliberate: a link that appears to work while pointing at something that is not the current record is worse than one that says plainly there is nothing to show.

The full record is open to everyone, including the decisions that did not work out.

One line

Probability says what we think will happen. Confidence says how likely we are to think differently tomorrow. Both are published, both are sealed before kick-off, and neither is a promise about the next match.

18+. Analysis and probability estimates only — not advice.