In short: aiALYSIS Live publishes pre-match and live AI analyses. Every analysis carries a confidence score, and only analyses above a confidence threshold are published. Each one is resolved automatically against the final match data and counted as correct or wrong, wrong results are never hidden, and all results are public on the accuracy page. This page also shows how well the confidence score holds up in practice.
Five steps: from data to result
- DataMatches, scores, events and in-match statistics arrive automatically from licensed live-data providers: live data over a streaming connection within seconds; fixtures, standings and squads several times a day. No figure is entered by hand.
- ModelBefore kick-off, selection combines the pre-match probability expectations of independent sources with a correction learned from our own historical results; a machine-learning model computes goal, corner and card projections, and a language model writes the preview text based on team form and head-to-head history. During the match, from the 5th minute, the model keeps re-evaluating the score, minute and in-match statistics (the scan runs every minute; at busy times each match comes up every few minutes).
- Confidence scoreEach analysis gets a probability estimate from 0 to 100. Only one pick is made per match and category group, and two conflicting directions (over and under) in the same category are never published together.
- PublicationOnly analyses above the confidence threshold are published (pre-match 62%, live 65%; absolute floor 60%). The publication time of each pre-match analysis is recorded and shown on the match page as a “Published” timestamp.
- Result checkWhen the match ends, every analysis is automatically marked correct or wrong against the final data. Analyses for postponed or abandoned matches, or where the required statistic was never published, are not counted either way. The count is the same as the in-app AI Performance screen and the accuracy page.
What does the confidence score mean?
The confidence score is the model's estimate of the probability that the analysis turns out correct. 70% confidence means that, over the long run, about 7 in 10 similar analyses are expected to be correct, and 3 in 10 are not. It is not a certainty for any single analysis.
A confidence score is only meaningful if it is calibrated: analyses labelled 65% should actually be correct about 65% of the time. To check this we group every published analysis by the confidence we showed and compare it with the actual accuracy. The tables below are recomputed from the live database every time this page is built.
Live analyses have an extra safeguard: the confidence shown on screen cannot exceed the actual accuracy of the same category and confidence band over the last 14 days (when at least 30 results exist). That is why a live confidence shown on screen can sit below the publication threshold; the publication decision uses the raw confidence.
Calibration: confidence shown × actual accuracy
System analyses only, resolved ones only (correct/wrong), window by match date. Each cell shows the actual accuracy and, below it, the number of resolved analyses; cells with fewer than 30 results show “—”. Counted at 3 November 2026 at 17:39 UTC.
Last 90 days
| Confidence shown | Pre-match | Live | All |
|---|---|---|---|
| 60–64% | 63.8%6,624 | 64.6%3,212 | 64.1%9,836 |
| 65–69% | 66.4%10,689 | 64.5%10,532 | 65.5%21,221 |
| 70–74% | 69.4%2,894 | 66.8%6,095 | 67.7%8,989 |
| 75–79% | 59.1%1,344 | 59.9%1,485 | 59.5%2,829 |
| 80%+ | 58.7%2,584 | 59.9%3,657 | 59.4%6,241 |
| All | 64.8%24,135 | 64.2%24,985 | 64.5%49,120 |
Bands that landed clearly below the confidence shown in this window: 75–79%, 80%+.
Current method (since 6 October 2026)
| Confidence shown | Pre-match | Live | All |
|---|---|---|---|
| 60–64% | 64.0%6,234 | 65.6%2,912 | 64.5%9,146 |
| 65–69% | 66.5%10,246 | 66.8%7,286 | 66.6%17,532 |
| 70–74% | 70.9%2,603 | 70.2%2,986 | 70.5%5,589 |
| All | 66.3%19,083 | 67.3%13,214 | 66.7%32,297 |
In this window every band landed within a few points of the confidence shown.
Two changes were made on 6 October 2026: pre-match selection moved to the method described above, and live analyses got an actual-accuracy cap on the confidence shown. Before that date, analyses showing 75% confidence or more in particular were correct less often than the confidence shown; the 90-day table still includes that period.
The 60% rule and closed categories
Absolute floor. No analysis is published with a confidence below 60%. The floor cannot be lowered by configuration and is checked a second time when the analysis is saved.
Current thresholds. Pre-match 62%, live 65%.
Category rule. A category that measures below 60% and cannot be lifted by adjusting the threshold is removed from publication.
Pre-match categories currently closed. Double chance, Both teams to score (measured below 60%; closed since 8 October 2026). For live corners only the “over” direction is published.
Pre-match vs live analysis
| Pre-match | Live | |
|---|---|---|
| When | Hours or days before kick-off; median about 2 days (October 2026 measurement) | During the match, from the 5th minute; each match is re-evaluated every few minutes |
| Input | Pre-match probability expectations of independent sources, our own result history, team form and head-to-head history | Live score, minute and in-match statistics |
| Categories | Total goals, first- and second-half goals, home / away team goals, corners, cards, match result | Total goals, first-half goals, corners |
| Publication threshold | 62% | 65% (raw confidence) |
| Confidence shown | The model's probability | Capped by the actual accuracy of the last 14 days |
| Updates | Published once; a match with no pick is re-evaluated on match day if new data arrives | Continuous; each analysis carries the minute it was written |
Timestamps and record integrity
Publication time. The first publication time of every pre-match analysis is recorded and shown on the match page as “Published: date time · X hours before kick-off”.
No additions after kick-off. A pre-match analysis cannot be written for a match that has started; this is a database rule (since 26 September 2026). Before the rule existed, 381 pre-match records were saved after kick-off and only 5 of them entered the count.
Result corrections are disclosed. Results are set automatically from the final data. If the data provider later corrects a statistic (for example the number of cards or corners), the result is recalculated and the original result is kept on record. Over the last 90 days this happened to 429 of 49,120 resolved analyses (0.9%): 211 from wrong to correct, 218 from correct to wrong.
Limitations
- Analyses are statistical estimates, not certainties. Even high confidence can be wrong in a single match.
- Data is sparser in lower divisions and less-followed competitions: without pre-match probability data there is no pre-match analysis, and statistics such as corners and cards are not published in every league.
- Out of scope: club friendlies, youth and age-group matches (U16–U23), amateur and regional lower-tier leagues, reserve leagues and some regional tournaments. Pre-match analysis is also not produced for women's leagues, international friendlies, academy and development tournaments and some continental cups (Copa Libertadores, Copa Sudamericana), where measured accuracy was low or samples were too small. Live analysis continues in women's leagues.
- Squad, injury or weather news that arrives after a pre-match analysis is published is not reflected in it.
- Accuracy can move by a few points in small categories or short windows; the 30-day and 90-day windows are the meaningful ones.
Why free and ad-free?
aiALYSIS Live is currently completely free and ad-free: there are no subscriptions, in-app purchases, paid features or third-party ad networks. Analyses are not written to steer anyone towards a payment, and the accuracy count is public. Community points are virtual and have no monetary value.