Principle 1
Completed fixture data is imported locally from controlled Sportmonks endpoints.
Methodology
How Football AI evaluates matches
This page explains the research process in accessible language without code-level detail.
Core principles
The portal separates historical evidence, current fixtures, AI reports and public publication status.
Principle 2
Pre-match statistics are generated using only information available before kickoff.
Principle 3
Backtests are chronological, so later fixtures cannot influence earlier predictions.
Principle 4
Random splitting was rejected because football model evaluation must respect time.
Principle 5
V2 and ML experiments were not promoted because evidence was not strong enough.
Prediction workflow
A match page should not look public-ready until the analysis has passed the review pipeline.
Step 1
Import fixtures
Supported leagues are pulled into the local lifecycle queue before any public page is treated as an analysis.
Step 2
Build model inputs
Team history, goal context and market-style signals are separated from future results.
Step 3
Generate draft
The numerical model creates probabilities and confidence context, then AI text explains the evidence.
Step 4
Validate before public
Drafts stay private until quality checks pass. Public pages avoid odds, staking advice and guaranteed language.
Evidence layers
These layers are deliberately kept separate so visitors can understand what is historical, pending or public.
Historical baseline
Visible
Completed seasons are used for backtesting and team context. They are not presented as live results.
Current lifecycle
Visible
Upcoming fixtures move through import, AI review, validation and public publication.
Settled records
Visible
Accuracy is shown only after a public prediction reaches full-time and can be checked.
Known limitations
Visible
Draw handling, confidence calibration and league variation remain visible instead of hidden.