What Does the Confidence Number Mean? An Honest Answer
CheckLive prediction system · Signal score, not probability
The short answer
Every football prediction card on CheckLive carries a confidence readout. Here is a real card from our predictions hub, what each part of it actually means, and what the previous version of this page got wrong about it.
Example — a real card from the predictions hub
Kolos Kovalivka — Obolon Kyiv
- 1The pick and the market price. The outcome our model leans towards, next to the bookmaker odds at the time. The odds come from the market — they are not our number.
- 2The confidence readout — dots, word and percentage. All three say the same thing: how strongly the data leans towards the pick. Five green dots and “High” mean a strong lean; two dots and “Low” mean the match is close to a coin flip. The percentage is the strength of that lean — it is not a promise that the pick wins 79% of the time.
It says how strongly the data leans one way. Across most of the scale, predictions with a higher score have come in more often than those with a lower one — though not at its very top, which is one of the things we are fixing. Use it to decide which match to look at first.
Confidence 80 does not mean an 80% chance of winning. We checked the scale against real outcomes and it overstates. Confidence measures how strongly the data leans — not how often an outcome happens.
What this page used to say. An earlier version of this article claimed that “78% confidence means our AI genuinely calculates a 78% likelihood”, and advised readers to place larger stakes at 75%+. Both statements were false, and the staking advice should never have been published. They are gone.
Where the Confidence numbers come from
Two kinds of numbers appear around a prediction, and they are built differently.
Both kinds of numbers answer the same question — how strongly does the data lean — on slightly different ranges. Neither is a calibrated probability, and neither is a promised win rate.
One match carries several Confidence scores — one per signal
On a match page, the main pick is followed by additional signals. This block is from the same real match:
Example — “More match signals” from the same match page
- 3Each signal is scored by its own formula. The market type and the match profile adjust the score, which is why the numbers sit close together here. Read them as a ranking of the signals on this match — not as chances out of 100.
What we have actually measured
The team our composite rated higher won 48.1% of matches and avoided defeat in 68.5% — the full breakdown, including where the model is weak, is in our accuracy article. Our live signals returned −13.3% across 806 tested cases: we do not beat the bookmaker, and we do not claim to.
What happens next. We have already re-measured the scale against real outcomes, value by value, and the finding stands: the stated numbers overstate. The corrected scale is queued behind other work on the product. When it ships, we will publish the full before-and-after table on this page — including the numbers that look bad for us.
Questions people actually ask
No. The scale is not a calibrated probability, and reading it as one overstates reality. It tells you how strongly the data leans, which is useful for ranking predictions — nothing more.
No. CheckLive does not publish staking advice. An earlier version of this page did — that advice was wrong and has been removed.
Because they are built differently: the main-card percentage comes from the rating gap alone, while each additional signal blends more inputs. That is also why the recalibration exists — so that every number can eventually be read the same way.
Method. The scale was audited against graded past predictions in our own database. The finding — stated numbers overstate — is why a recalibration exists rather than a promise that everything is fine. The example card above is a real card from 28 August 2026, unedited. Everything on this page is measured on football; other sports have separate pipelines and are not covered by these numbers.
Last updated 28 August 2026. This page will be updated when the recalibrated scale is published.