Methodology

How our analyses are really made: data, model, limits

The principle: original, verifiable, immutable work

Every AI Analysis is an original piece of work produced by our system for a specific match, before kick-off. We do not redistribute third-party predictions and we do not copy content: the raw data (results, standings, odds) comes from our providers, but the reading, the weights, the probabilities and the reasoning you read on the page are generated by our system and signed by the editorial team that designs and supervises it.

Two rules make the work verifiable: every analysis is immutable after publication (we never rewrite a prediction after the fact) and every prediction is automatically compared with the real result at full time, feeding a public accuracy dashboard that also includes the errors.

The input data

For every match the system automatically collects from the database: the current table for both teams (position, points, matches played, form, goals scored and conceded) and the size of the tournament to put it in context; for international cups also the domestic-league table of each club; the last 5 results of each team with venue and score; the last 10 head-to-head meetings in any competition; the pre-match consensus odds for 1X2, Over/Under 1.5 and 2.5 and both teams to score; the squad actually available (excluding injured and suspended players, with injuries synchronised close to the match) and the list of known absentees; the exact stage of the competition (matchday, knockout round); the recent news of the last three days mentioning the two teams, used to catch game-changing factors.

Sports data comes from professional providers; news from our editorial system. No personal user data enters the process.

How the probabilities are formed

The core of the system is a large language model (Anthropic Claude) to which we pass the data context described above in structured form. The model does NOT draw on its prior knowledge for facts (regulations, tournament stages, the state of clubs): it is bound to use only the data provided, precisely to avoid outdated information. For every market (full-time result 1X2, Over/Under 2.5, both teams to score, double chance) the model produces an indicated outcome, a confidence level as a percentage and explicit reasoning that explains which factors weighed; it also indicates likely scorers choosing ONLY among the available players in the squad.

Market odds act as a thermometer of consensus: when our statistical reading and the market diverge, the reasoning must explain why. We are not a parametric model with fixed coefficients: the weighting of form, head-to-head, home advantage and absences is contextual, and the reasoning published in every analysis IS the documentation of that weighting.

One market at a time: 1X2, Over/Under, both teams to score

Each market answers a different question and uses the data differently. For 1X2 the main weights are the gap in standings and form between the two teams and home advantage. For Over/Under 2.5 what counts are both sides' scoring and defensive averages (goals scored and conceded per match) and the attacking tendency shown in recent games. For both teams to score the question is whether BOTH score: here the combination of one side's attack and the other's defensive solidity, in both directions, weighs most. Double chance derives from 1X2 as cover for the two most likely outcomes. Likely scorers, finally, are chosen exclusively among the players available in the squad, with an indicative probability for each.

The weighted factors

In order of typical impact: position and points in the table (with the tournament context), recent form (last 5 results with the quality of opponents), home advantage, head-to-head, market odds as external consensus, squad availability (heavy absences shift the probabilities), stage of the competition (a knockout round changes the incentives compared with a league matchday), relevant recent news. The relative weight varies match by match and is always made explicit in the reasoning.

How to read a probability

"Home win at 62%" means: in a hundred matches with this exact context, we expect the home team to win about 62 of them, and NOT to win 38. A 62% outcome that does not happen is not a model error: it is within expectations almost four times out of ten. Error is measured only over the long run, comparing real frequencies with declared probabilities: that is exactly what our accuracy dashboard does, updated automatically at every verified match. For more, there is the dedicated guide on how to read probabilities.

The declared limits

What the model does NOT do and CANNOT predict: incidents (a red card in the 5th minute, a doubtful penalty, an injury in the warm-up); non-public internal motivations (rotation decided during the week, dressing-room tensions) beyond what emerges from the news of the last three days; matches with scarce data (minor leagues without odds or with few rounds played produce less reliable estimates); last-minute changes after the analysis is generated, which is frozen at publication and not updated along the way. Injuries depend on the timeliness of sources and may be incomplete.

Furthermore: the confidence level is the probability the model estimates for that outcome: a 70% confidence means that, over many matches with the same confidence, we expect to get about 7 out of 10 right. It is not a guarantee on the single match; how reliable the estimate is gets checked afterwards on the accuracy page, where we publish all outcomes without selection.

Update frequency and verification

Analyses are generated before kick-off (by the editorial team for the main matches, or on users' request) and carry their publication date and time. After the final whistle an automatic process compares every prediction with the real result and records the outcome in the database: no human intervention can select or change the counts. The aggregate accuracy statistics, by market, by league and by month, update accordingly and are public.

Guides and insights

Start free
Today
Competitions
Favorites