Predicting footballer market value, part 3: what drives value, according to SHAP

Part 3 of the series on my final-year major project. Part 2 covered the stacked ensemble and its results (R² 0.958, MAE ≈ €0.96M).

Why interpretability matters here

A model that predicts well but cannot explain itself is hard to trust and harder to learn from. SHAP (SHapley Additive exPlanations) attributes each prediction to individual features in a theoretically grounded way, so we can ask: what does the model believe makes a player valuable?

What the model learned

Ranking features by mean absolute SHAP value, three patterns stood out:

1. Scouting consensus dominates. The EA-attribute-derived prediction (ea_pred) is by far the most influential feature, followed by the ensemble agreement signals (pred_mean, pred_max). Expert-assessed ratings — potential and overall_rating — also rank highly on their own. The market, it turns out, prices players much closer to scouting judgement than to any single season's statistics.

2. Age is a curve, not a number. Alongside raw age features, the engineered abs_years_from_27 feature (distance from the traditional peak-value age) carries real weight — value rises towards a player's prime and decays past it, and the model uses that shape.

3. Wages leak the market's own opinion. Wage and wage-per-rating features rank above almost all on-pitch statistics. Clubs reveal their internal valuations through salaries, and the model picks that signal up. The highest-ranked pure performance feature, goal contributions per 90, appears only after all of these.

Honest limitations

• The target is Transfermarkt's crowd estimate, not realised transfer fees — the model learns the crowd's biases along with its wisdom.
• Coverage is the top five European leagues; transfer-market dynamics elsewhere may differ.
• Accuracy within ±20% for 56.5% of players still leaves the tails — wonderkids and superstars — systematically harder to price.

That last point is arguably the most interesting result: the players the market argues about most are exactly the ones observable data explains least.

This series summarises my final-year major project at Anglia Ruskin University. The notebooks, scraper and model outputs are all in the repository.