In one sentence
For each match requested, Prediction AI first has recent facts about the two teams researched on the web, then asks an artificial intelligence model to estimate probabilities from those facts. The result is an estimate, shown with its assumptions and sources, never a certainty.
Step 1: the input
The user enters two teams (the home side and the away side), optionally the match date, and picks a language. If the real fixture is the other way round from what was entered (the team listed as home actually plays away), the research detects it and the analysis is corrected accordingly.
Step 2: researching facts on the web
An AI model with web access (the Perplexity service, with a configurable model) runs several searches in parallel, with instructions not to invent anything and to leave blank whatever cannot be found:
- The match itself: competition, date, kick-off time, venue, and which team is really at home.
- Each team: its last five official matches played, with exact scores, its league position and its absences (injuries, suspensions).
- Head-to-head: up to the last five meetings, plus a few facts of context (what is at stake, fixture congestion).
- Key players: for each side, the top scorers and assist providers of the season and whether they are available.
- The referee: searched once the real fixture is identified, and only if the match is close, because the referee is announced only a few days beforehand. If one has been appointed, their averages of yellow cards, red cards and penalties per match are looked up.
Step 3: calculations done by the program
Some indicators are not estimated by the AI but calculated from the scores found: the run of results over the last five matches (win, draw, loss), average goals scored and conceded, the number of clean sheets and the number of matches in which both teams scored.
Step 4: estimation by the AI model
The model receives a fact file (results, league position, absences, head-to-head, key players, referee) and may supplement it with a search if recent information is missing. It produces:
- probabilities: home win, draw, away win (adding up to 100%), over or under 2.5 goals, both teams to score, and the most likely score;
- a reading of the match: form, key factors, risks and the assumptions made;
- three forecasts at different confidence levels: a “cautious” forecast (probability of at least 70%), a “main” one (between 50% and 70%) and a “bold” one (between 25% and 50%). Their probabilities must follow from the estimated probabilities: a combined scenario is always less likely than each of its parts;
- two to four likely scorers among the available players, with a probability and a justification;
- the probability of more than 4.5 cards in the match, based on the referee's averages where known;
- a data quality indicator (high, medium or low), which the model assigns according to whether form and absences are well documented.
The model is given guard rails: a home advantage of roughly 8 to 12 points, no more than 75% on a 1X2 outcome unless the gap in quality is very large, and a low generation temperature to limit variability. A check verifies that the “bold” forecast really is bold; if not, a single re-run is made to replace it.
Step 5: the display
The analysis appears with the facts used, the assumptions and the sites consulted by the web search (their domain names are shown). The latest analyses are kept on the user's own device so they can be read again. Team logos come from a separate public database and are used for display only.
The limits, plainly
- Data from web searches. It can be incomplete, out of date or wrong, and its quality depends on the sites consulted. The tool does not rely on a database of shots or expected goals (xG).
- An estimate by an AI model, not a validated statistical model. The probabilities are produced by a language model following instructions, not by a model fitted and tested on a history of matches.
- No published calibration. We do not publish any measure of how good the probabilities are over time. We therefore cannot claim that they are calibrated, and we quote no success rate.
- Numbers even when the facts are partial. The tool is designed to always give an estimate; it is the data quality indicator and the list of assumptions that flag when a result is fragile.
- A result that can vary. Two requests for the same match can give slightly different numbers, and an analysis reflects the information available when it was requested (including last-minute line-ups and absences).
- Match identification that relies on search. Always check the date, competition and teams shown.
- Unpredictable events (an early injury, a red card, a refereeing error) cannot be anticipated.
How to read the result
Look at the date of the analysis, the sources and the data quality indicator. Compare the probabilities with one another rather than fixating on one number. A probability of 60% means that in about four cases out of ten the scenario does not happen. Our guides on reading a 1X2 probability and on expected goals explain these ideas in more detail.
About Prediction AI
Prediction AI is an AI-powered predictive sports analysis platform: for a football match it presents estimated probabilities together with the facts behind them and their limits. Take a look at the Prediction AI plans. An estimate is an estimate, never a certainty.
Further reading
- Expected Goals (xG), Over/Under 2.5 Goals and Both Teams to Score: Understanding the Indicators
- The Factors That Influence a Football Match and How to Weigh Them Cautiously
- How to Read a Win, Draw or Loss Probability (1X2)