FairValue

Checking model

Esc
  • EnvironmentOverviewWhat FairValue measures, what it found, where to go next.
  • EnvironmentTerminalExplore, compare and track players; market, league and model views.
  • EnvironmentStoriesValuations, head-to-heads and research findings, told as stories.
  • TerminalPlayer ExplorerSearch, filter and sort the player export.
  • TerminalComparison StudioTwo to four players on one scale.
  • TerminalWatchlistsNamed collections, kept in this browser.
  • TerminalMarket LandscapeHow observed value is distributed, and where the model drifts.

↑ ↓ to move · Enter to open · Esc to closeLoading player index…

Research story

The gap runs the other way

A model that never sees nationality still misses one group in one direction: non-UEFA players' market values sit about 10% above its estimates, relative to UEFA players in the same leagues (+0.093 log units). Without league and club, +0.164.

Built from Fairness audit outputs (reports/fairness, held-out predictions, league-stratified bootstrap intervals).7 chapters

+0.093

signed residual gap, non-UEFA minus UEFA, within leagues · Model A · 95% CI +0.063 to +0.124

The question

FairValue's models estimate a footballer's Transfermarkt market value from two years of club football: minutes, goals, assists, age, position, and for Model A the league and the club's strength. They never see nationality.

The audit asks something narrower than “is the model fair?”: do its errors fall differently on players from inside and outside UEFA?

Who is in the study

Of 50,149 players in the dataset, 5,820 meet the study's rules: a current valuation, a covered league and at least 450 club minutes in two years. 1,496 of them (25.7%) hold a non-UEFA first citizenship. It is a selected population, not the global market.

Equally accurate

On accuracy the groups are close. Model A misses UEFA players by 0.403 log units on average and non-UEFA players by 0.414. Within leagues the difference is +0.011, and its interval, −0.008 to +0.031, includes zero.

From every player in the dataset to the study population
  1. Every player in the dataset50,149
  2. With a current valuation19,364
  3. Valued while in a covered league7,378
  4. With 450+ club minutes in two years5,820

Players per selection stage

One direction

The difference is in direction. Relative to the model's estimates, non-UEFA players' observed values sit higher than UEFA players' in the same leagues: a signed gap of +0.093 log units, about 10%, with a 95% interval of +0.063 to +0.124.

The study expected the reverse: market values below those of comparable UEFA players.

Without league and club

Model B keeps only performance, age and position. Its gap is larger, +0.164 (+0.121 to +0.207). Adding league and club narrows the gap; it does not close it.

League by league

In the 8 leagues large enough to estimate on their own, the intervals are wide and 3 exclude zero. The other 6 contribute only to the pooled figure. The headline is an average across leagues, not a rule that holds in each of them.

What it does not mean

This is an observational pattern in one model's errors, on a selected population. The target is Transfermarkt's community valuation, which may itself be biased, so a gap measured against it is not evidence of discrimination by anyone, and nothing here claims the model meets or fails a fairness threshold.

Every number above, with its robustness checks, is in the Fairness Audit; the method is in Methods.

Signed residual gap, non-UEFA minus UEFA, within leagues · log units · 95% CI
0 · no gap
Model Awith league and club+0.093
Model Bperformance only+0.164

Right of zero: relative to the model's estimates, non-UEFA players' observed values sit higher than UEFA players' in the same leagues.