player-value-lgbm

LightGBM model that forecasts the 6-month log-change in Transfermarkt market value for football players. Trained on the engineered features in player-value-features.

• Code: https://github.com/DanielRegaladoUMiami/player-value-ml
• Demo: https://huggingface.co/spaces/DanielRegaladoCardoso/player-value-ml
• Dataset: player-value-features

Performance (held-out test, ≥ 2024)

Model	Test MAE (log)	Test R²
Naive (y=0)	0.2145	−0.006
Ridge regression	0.2164	+0.130
AutoETS	0.2361	−0.198
AutoTheta	0.2740	−0.298
LightGBM	0.1970	+0.205

LightGBM wins by ~40% in R² over the best baseline and is the only non-trivial model to clearly beat the naive predictor.

Architecture

• LightGBM regression with L1 objective (MAE)
• 351 trees (early-stopped on validation)
• num_leaves=127, learning_rate=0.05, feature_fraction=0.8, bagging_fraction=0.8, lambda_l1=0.1, lambda_l2=0.1
• 84 features (92 raw, 8 dropped by the audit), 5 categorical handled natively (position, sub_position, foot, career_stage, country_of_citizenship)

Top features by gain

1. country_of_citizenship — the market prices passport over performance
2. as_of_year — captures market inflation
3. value_diff_1 — recent momentum
4. log_value — base level
5. age
6. as_of_month — transfer-window seasonality
7. value_diff_3
8. international_caps
9. age_minus_position_peak — distance from position-specific peak
10. value_lag_1

How to use

import lightgbm as lgb
import pandas as pd
from huggingface_hub import hf_hub_download

model_path = hf_hub_download("DanielRegaladoCardoso/player-value-lgbm",
                              "model.lgb")
booster = lgb.Booster(model_file=model_path)

# X must be a DataFrame whose categorical columns are pd.Categorical
# (see scripts/train_lgbm.py in the repo for the exact preprocessing)
preds = booster.predict(X)        # predicts log(value_T+6mo / value_T)

To convert a prediction back into euros:

predicted_value_eur = current_value_eur * np.exp(predicted_log_ratio)

Limitations

• The 6-month horizon is a median — Transfermarkt updates 2–3×/year irregularly. The model is trained on the irregular cadence; predictions are not calibrated for a fixed 180-day horizon.
• R² of +0.20 means ~80% of valuation-change variance is NOT explained by these features — markets are noisy.
• country_of_citizenship being the strongest feature reflects market bias the model captures (and does not endorse).
• Trained only on Transfermarkt's tracked competitions; lower-tier leagues are under-represented.

License

Apache 2.0.