How the predictions are made, and how the record is kept.

Data collection

FutebolAI uses football-data.org for fixture identifiers, participating teams, competition codes, kickoff times and final scores. The site does not attribute advanced performance metrics, injury reports or suspension reports to that feed.

Statistical processing

Available fixture and competition fields are normalized into a consistent match record. Missing fields remain missing; they are not replaced with invented values.

AI assisted analysis

The publishing workflow uses automated analysis to produce one scoreline forecast and one result call. Outcome percentages shown on the Predictions page are derived from that published scoreline with a Poisson calculation and are labelled accordingly. They are not presented as independently calibrated probabilities.

Editorial review

Automated checks verify required fields, source match ID, kickoff timing, access rules and prohibited claims before publication. Sensitive or time-dependent claims require a named source and checked time. FutebolAI does not claim that a professional football editor manually reviews every forecast.

Pre kickoff publication and settlement

The forecast is written to the public record before kickoff and linked to its source match ID. After full time, the final score is fetched from football-data.org and reconciled automatically. Misses remain visible. Postponed, abandoned or unmatched fixtures are kept as void and excluded from accuracy denominators.

What the system does not claim

FutebolAI does not claim access to insider information, guaranteed outcomes, unavailable specialist data feeds or betting returns. Published analysis is for information and entertainment only.

How the accuracy numbers are calculated

The tracker is generated directly from the predictions table. No figure on this site is entered by hand.

• Correct result — the predicted outcome (home win, draw or away win) matched the final result.
• Exact score — the predicted scoreline matched the final score exactly. This is a stricter test than the result alone.
• Model calibration — last 10 — the ten most recently settled matches, plotted oldest to newest.
• Time windows — the last 7 days and last 30 days are sliced by kickoff time in UTC, so a match counts in the window it was actually played.
• Full history — the headline figures always use every settled match, never a hand-picked subset.

When a match cannot be settled

A prediction is written into the log before kickoff and is not edited once the match starts. If a fixture is postponed, abandoned, or its result cannot be matched against the results feed, it is marked void and excluded from the accuracy figures rather than being counted as a hit or a miss. Void entries stay visible in the log.

Timing and locking

Predictions are locked at publication. If fresh information arrives before kickoff, it is handled as a separate editorial note, not as a silent edit to the stored prediction.