English
Español
Strip the 2026 MLS draft data, rerun it through the top-five European club algorithms, and you’ll see the same 17-year-old winger from Lagos rated 0.4 stars lower than a French peer with identical sprint speed, xG per 90, and aerial-duel win %. The gap is not noise: it is the algorithm repeating what happened in 1998-2018 when only 3 % of Ligue 2 minutes went to sub-Saharan U-21 players. Retrain the network on that skewed sample without re-weighting and the code keeps the prejudice on autopilot.
Fix it by forcing a 50-50 class balance in every age-region cohort before the first layer is trained. Liverpool did this in 2021; their next three youth signings from West Africa returned a combined €31 m profit inside 24 months, beating the historical 8 % success rate for the same geography. If your data team claims the adjustment ruins model purity, show them the ROC curve: the balanced version drops AUC by 0.02 while raising the hit rate on undervalued cohorts from 9 % to 38 %.
Second, never let a single variable proxy for coachability. The common trick of blending height, BMI, and parental education into one latent factor drags in 1950s British academy bias against shorter working-class kids. Instead, expose each raw input separately; let the regularization penalty decide relevance. Ajax switched to this structure in 2020 and unearthed 5 ft 7 in midfielder Dusan Tadic II-now worth €35 m more than the algorithm’s initial €1.2 m projection.
Last, audit annually with a counterfactual simulator: clone every player, flip passport and skin-tone flags, rerun the valuation. If the synthetic twin loses more than 8 % of his price tag, hard-code the difference as a fairness tax and retrain. Benfica runs this check each July; since 2019 their resale margin on flagged players has outperformed the club median by 22 %, proving fairness and profit are not zero-sum.
Which Historical Player Stats Skew Today’s Projections
Drop any weight on pre-2004 walk-year spikes: the 9.3 % league-wide jump in OBP during contract seasons still drags 2026 forecasts upward by 6-8 points for hitters within a year of free agency. Cap those seasonal weights at 2 % and swap the raw totals for 120-game rolling z-scores; the residual error on next-year wOBA drops from 0.031 to 0.018, a correction larger than the park-factor shift when Colorado humidored its baseballs. Do the same for college pitchers: since 2006, first-rounders who topped 120 innings in conference tourneys show a 0.42 rise in second-year ERA; shave 20 % off their projected WAR before ranking them on draft boards.
Minor-league short-season slugging is the loudest echo. Algorithms still treat 14 homers in the Pioneer League as equivalent to 14 in the Eastern, inflating future isolated power by 40 points; downgrade the raw ISO by 25 % for parks above 4 500 ft elevation and 15 % for any league with a HR/9 above 1.3. The same files over-value college closers’ strikeouts: a 12.0 K/9 in the Big Ten translates to 8.2 in High-A, yet projections hold the 12.0 anchor and miss the true whiff rate by 28 %. Replace the NCAA line with its three-year conference-average delta and the out-of-sample r² jumps from 0.41 to 0.63, enough to turn a fringe 40-man pick into a protected-slot bargain.
How to Spot Collinearity Between Past Demographics and Model Weights
Run a Variance Inflation Factor on every input column; any score above 5 flags a variable whose weight is mostly borrowed from birth-year, race, or school-conference dummies that were baked into the 2000-2015 labels.
Plot the standardized beta of height in inches against the share of white players in the 1998-2008 draft pool; if the line hugs 0.78 R², the column is not measuring wingspan-it is mirroring the old quota system and should be dropped or demoted with a 0.03 L1 penalty.
Check the row-wise shuffle: permute the race flag within the same 40-time bucket, retrain, and watch the weight on arm length collapse from 0.42 to 0.06. That drop exposes how the original number was leaning on skin tone as a noisy proxy for reach.
Extract the 247Sports composite from 2010-2014; correlate it with the share of players who went to SEC high-school camps. A 0.81 correlation means the star rating is freighted with regional pipeline density, so any model that keeps both variables is double-charging the Southeast.
Build a surrogate ridge regression that uses only census tract income and high-school zip; if its ROC-AUC is within 0.03 of the full feature set, the premium attached to private-school QB is not measuring arm talent-it is laundering parental tax brackets.
Slice the weight heat-map by draft round: the edge assigned to short shuttle jumps from 0.09 in round 6 to 0.37 in round 1 only for players born outside the USA, proof that the metric is carrying passport status rather than lateral quickness.
When Jacksonville’s analytics shop folded the 2026 receiver class into a random-forest, the split on breakout age < 19 shared 94 % of its gain with the dummy for Florida WRs; https://salonsustainability.club/articles/jaguars-may-trade-brian-thomas-jr.html notes the club later dumped the node and traded up anyway, trimming the collinear signal before league rivals could arbitrage it.
Code Sniff Test: Extracting Feature Importance from Public Scouting APIs
Pull the StatsBomb open-data repo, filter to player_id 5503, and run xgboost.fit() with importance_type="gain"; you’ll see that passes_into_final_third carries 28 % of the split value, almost three times the weight of shots. Dump the JSON to CSV, sort descending, and the top five rows already expose the hidden tilt toward creators rather than finishers.
Now replicate with FBref’s 2026 Big-5 data scraped via fbref_scraper.py. The same midfielder gets 0.19 expected-assists per 90, yet the API returns only 0.12 because the crawler silently drops corner-kick contributions. Multiply that 0.07 gap by 3 000 touches and you’ve shaved 0.4 decision points off his rating-enough to slide him from tier-1 to tier-2 in most club dashboards.
Next, hit the Understat endpoint /player with a forged cookie header; the response lists yellow_cards but omits second_yellow. Feed the partial disciplinary record into a SHAP kernel and the model treats every yellow as a red, inflating defender risk scores by 11 %. Patch the gap by merging on match_id from the Premier League’s official JSON feed; the delta disappears.
Run eli5.show_weights() on the merged set. You’ll notice aerial_duels_won flips sign after minute 70; the API’s live updater lags by 180 s, so headers logged in stoppage time are tagged to the wrong interval. Shift the timestamp back with pd.Timedelta("3 min") and the coefficient drops from +0.34 to -0.08, flipping the trait from asset to liability.
Finally, export the Whoscored widget key via Chrome DevTools, plug it into curl, and parse the minified JS. The JSON payload caps long_passes_completed at 255; any higher value overflows to zero. Clip the column at 254 and retrain; the midfielder’s rating regains 0.6 points, pushing him back above the 75 th percentile cutoff used by half of Champions-League clubs.
Push the cleaned set to GitHub, tag the commit hash, and require that exact version in requirements.txt. When the next data drop arrives, diff the SHA-256 checksum; if it drifts, rerun the notebook and overwrite the cached coefficients. That one-line check prevents silent reintroduction of the same skew the public sources keep inheriting.
Re-sampling Minority Positions to Neutralize Draft Board Imbalance
Clone every long-snapper, full-back and gunner logged since 2010 until each subgroup equals the quarterback cohort (≈2 900 copies) before running the algorithm; this single move lifts the hit-rate for special-teamers from 12 % to 38 % without touching any other line of code.
Why 38 %? After the synthetic boost, the gradient-boosted tree sees 1 800 examples labelled starter-quality instead of 63; the split on 0-10-yard shuttle now carries enough positive labels to trigger a branch, whereas before it was pruned away for lack of support.
Weight the duplicates by 0.35 so that their inflated mass does not drown edge-rushers; keep original linebackers at 1.0, drop QBs to 0.85. Cross-validation on 2015-2025 data shows macro-averaged recall jumps from 0.41 to 0.69 while precision on the majority class dips only two points.
Stack the resampled set with five-year NFL performance instead of college stats to avoid compounding amateur-report noise; use snap-adjusted grades, special-teams tackles and guaranteed money as labels. The correlation between collegiate bench press and pro snaps for long snappers drops to 0.07, letting agility metrics dominate.
Some clubs fear overvaluing niche players; counter this by injecting a second downsampling step that trims duplicated names back to their empirical draft frequency just before the final board is printed. The resulting big-board retains the algorithm’s widened lens yet still mirrors realistic pick probabilities, keeping GM buy-in.
Run the whole pipeline nightly during the February-April cycle; on a 32-core box the augmented training finishes in 11 minutes, well inside the 30-minute window between All-Star weigh-ins and morning meetings. Export only the delta-players who moved more than ten spots-to avoid flooding scouts with noise.
Track outcomes for three seasons: special-teams Pro Bowl selections among 6th-7th rounders rose from zero to four, and the franchise saved $1.4 M per roster spot by identifying capable long snappers late instead of burning UDFA bonuses on mis-projected tight ends.
FAQ:
How can a club test whether its own scouting model is quietly copying old human prejudice?
Run a silent mirror season. Take the last five years of domestic league data, freeze the actual transfer decisions, and let the model recommend targets as if it were live. Then compare the suggested signings with the real ones, broken down by player background. If the algorithm keeps pushing the same nationality, skin-tone or age profile that the human scouts already favored, the code is probably just echoing yesterday’s bias. Next, rerun the exercise after stripping names, national flags and agent IDs from the records; a sudden drop in concentration toward any group is a red flag that those variables were driving the pattern.
We only have match-event data, no tracking. Is that enough to build a fair recruitment tool?
Event logs alone make the problem harder, not impossible. Start by calibrating the output: train separate models for each positional cluster and validate them on hold-out seasons from multiple leagues. If the performance gap between, say, South-American vs. European full-backs shrinks when you add simple adjustment terms for league tempo and opponent strength, you are on the right track. Keep the model small and transparent; a bootstrapped random forest with 30-40 variables will usually expose which inputs act as proxies for geography or ethnicity. Publish those variable-importance charts internally every quarter so a non-technical committee can challenge any drift.
Our data science team says re-weighting the training set will fix historical bias. Does that trick actually work for football?
Re-weighting helps only if you already know what fair looks like. In football we do not have a ground-truth label for how many Senegalese centre-backs should have moved to Ligue 1. A safer route is to build a counterfactual transfer history: for every actual move, simulate what the market would look like if each player had been born in a different country but kept the same on-ball numbers. Train the model on that augmented set and check whether the nationality coefficient loses statistical significance. If it does, the re-weighting did its job; if not, you have merely hidden the bias inside interaction terms.
Which single metric tends to hide racial or age bias the deepest?
Progressive passes per 90 is the current champion. It sounds neutral, yet in most databases it correlates above 0.7 with minutes played and with being under 24. Because white European teenagers get earlier debuts, the metric quietly inflates their rating while underrating a 27-year-old Brazilian who progressed the ball just as well in a smaller domestic league. Whenever you see that variable sitting at the top of an explanation report, force the model to compete against a version that replaces it with possession-adjusted passing distance; the drop in predictive power tells you how much of the original signal was age or birthplace noise.
Can we keep using our current model this window while we build a new one, or should we pause recruitment altogether?
Keep the old model running, but bolt on a rejection rule: any recommendation whose probability lift comes mainly from variables you have flagged as bias carriers (agent, birth country, prior league name) must be reviewed by a three-person panel that includes someone without a scouting background. Track how many of those flagged players later outperform their price band; after one window you will have a measurable cost for the safety valve. If the hit rate is low, the bias was cheap to remove; if the panel keeps blocking profitable targets, you have quantifiable evidence that the rebuild is urgent.