@aifootball26
How every number is sourced

Methodology

AI Football's pitch is simple: every claim on the screen is backed by a number from a citable source. This page lists those sources, the formulas we publish, and what the model cannot see.

Primary data sources

Stat-freeze dates

Sports data is alive — a stat-freeze date tells you what the model could see when the video shipped. Every published deep dive declares its freeze date in development.md.

Published composite formulas

Best XI from Pure Data (0007)

Each position scored on a 100-point composite weighted as follows:

GK = 40% PSxG±/90 + 30% save% + 20% clean-sheet rate + 10% clearances/distribution
CB = 30% Tkl+Int/90 + 25% aerial duel% + 25% progressive passes/90 + 20% pass% under pressure
FB = 25% dribbles + 20% progressive carries + 25% crosses/key passes + 20% tackles + 10% F3 recoveries
CM = 30% progressive carries + 25% chance creation + 20% pass% under pressure + 15% tackles + 10% xG buildup
W = 30% non-penalty xG + 25% G+A/90 + 25% shot-creating actions + 20% successful dribbles
ST = 40% npxG/90 + 30% G+A/90 + 20% shot conversion% + 10% aerial duels

Eligibility: in a 2026 squad, 1,500+ club minutes in 2025-26, 5+ international appearances, not retired internationally.

Group strength claims

Where a group description on /groups/ says "widest rank spread", "highest combined rank", or "toughest by average rank", the underlying formula is:

RANK GAP = max(FIFA rank, 4 teams) − min(FIFA rank, 4 teams)
RANK SUM = sum of all four FIFA ranks (lower = stronger group on average)
AVG ELO = arithmetic mean of the four teams' Elo ratings (higher = tougher)

Each card declares which metric drives its label. April 2026 FIFA rankings are the basis for every published Group X-Ray.

R32 prediction model (Round-of-16 forecast)

A simple Elo-based win-probability model with home-advantage adjustments for matches in the three host countries. Methodology is open — Elo input, +30 for hosts in their own city, +15 for hosts in any host country.

What the model cannot see

Every published video discloses its biases on-screen. Sitewide, the recurring caveats are:

Errors and corrections

If a published number is wrong or out of date, file it in our bug tracker (repo issues) or email contact@aiftbl.com. We track corrections in the per-video development.md.

Last updated: 18 May 2026