ypolatog/player-value-lgbm-core

Player value forecaster: LightGBM core model (lgbm-core-v1)

Category: trained from scratch. The primary model of CMU 24-679 Project 1 (Yunus Polatoglu and Jack Stevens), and the model behind the live app.

• App: https://huggingface.co/spaces/jackstev/player-value-forecaster
• Dataset: https://huggingface.co/datasets/ypolatog/football-contracts-500
• Second model (off-the-shelf): https://huggingface.co/ypolatog/player-value-chronos-bolt
• Code: https://github.com/jackstev-edu/player-value-forecaster
• End-to-end notebook: https://colab.research.google.com/github/jackstev-edu/player-value-forecaster/blob/main/notebooks/02_end_to_end.ipynb

What it does

Given a football player on a 1 July snapshot, it forecasts how their Transfermarkt market value will change over the next season: a median and a 10th to 90th percentile range of log(value in one year / value today). The app turns those into euros.

Inputs (4 features): age (years, decimal), position (Attack, Midfield, Defender, Goalkeeper), value_now (EUR), value_change_12m (log change in value over the past year).

Outputs: q10, q50, q90 as log ratios. Forecast in euros = value_now * exp(q).

Why this model

• Small tabular data. 1,000 rows of mixed numeric and categorical columns, where gradient boosting is the standard strong choice. Shallow trees (15 leaves), slow learning (0.03) and early stopping on held-out players keep it from memorising 800 rows.
• Ranges, not just points. One LightGBM per quantile (0.1, 0.5, 0.9), then split-conformal calibration (CQR) on held-out players, so the 10 to 90% band really holds about 80%.
• Four features, not 49. The full model had 49 features (contract, minutes, club and league strength, peer rank, injuries, transfers). An ablation with paired bootstrap intervals found no feature group beyond age, position and value that measurably helps (decision #47 in the repo), so the simplest model shipped. It also needs no contract or panel row, so it can score every league player.

Training data

The football-contracts-500 dataset: 500 contracts collected by hand (the 500 most valuable players, each at one past season) plus 500 augmented rows that reuse each contract for its other seasons. Values come from Transfermarkt via the CC0 Kaggle dataset player-scores.

Fitting: all 1,000 samples, with 15% of players held out for early stopping and band calibration. Seed 42. Parameters and the band shift are in metrics.json.

Evaluation

Five-fold cross-validation grouped by player, so no player is in both training and test. Only the 500 hand-collected rows are scored; augmented rows are used for training only.

Model	log MAE	80% band coverage
No change (predict 0)	0.959	–
Age × position median curve	0.744	–
Ridge linear baseline	0.602	–
LightGBM core (this model)	0.608	0.788
LightGBM core + Chronos forecast as features	0.609	0.802
Chronos-Bolt zero-shot	1.327	0.405

A log MAE of 0.61 is a typical miss of about ×1.8. The linear baseline ties it (difference −0.006, 95% interval −0.028 to +0.016); LightGBM is kept for its calibrated band.

Walk-forward backtest on five named players (Mbappé, Gavi, Anton, Schrijvers, Salah, picked by a stated rule from 2024 data, not by outcome): log MAE 0.50, 6 of 10 actuals inside the range.

Ordinary league players (2,000 panel rows anchored in 2023 and 2024, none of them training players), the population the app serves:

Model	log MAE	predicts a rise	80% band coverage
No change	0.457	0%	–
LightGBM core (this model)	1.110	97%	0.401
Chronos-Bolt zero-shot	0.560	28%	0.709
Actual outcome	–	28%	–

Intended use and limits

Intended for coursework and illustration: how a forecast with an honest range can be built from a small hand-collected dataset. Not for transfer, contract, betting or any decision about a real person.

• Strong upward bias outside its training population. The 500 players were chosen because they are among the most valuable today, so the cheap ones in training are future stars (selection on the outcome). On ordinary league players it predicts a rise for 97% when 28% rose, and its band covers 40% instead of 80%. It suits young players on a breakout path; it is not a general league forecast. The app states this on the page.
• Training covered ages 16 to 31 and values from about EUR 0.75M (10th percentile) upward. Older or cheaper players are extrapolations.
• One season ahead only. Two and three seasons were measured and held back.
• Blind to news. Injuries, transfers and contract changes after the snapshot are invisible.

Use it

# pip install lightgbm pandas huggingface_hub git+https://github.com/jackstev-edu/player-value-forecaster
import numpy as np, pandas as pd
from huggingface_hub import snapshot_download
from pvf.models.gbm import QuantileGBM

model = QuantileGBM.load(snapshot_download("ypolatog/player-value-lgbm-core"))
X = pd.DataFrame({"age": [21.5], "position": ["Attack"], "value_now": [30e6],
                  "value_change_12m": [0.4]})
q = model.predict(X)[["q10", "q50", "q90"]]   # log ratios
print((30e6 * np.exp(q)).round(-5))             # euros one season ahead

Files: h1_q10.txt, h1_q50.txt, h1_q90.txt (LightGBM boosters), meta.json (features, category levels, conformal band shift), metrics.json (every number on this card).

Ethics

Players are public figures and the values are published Transfermarkt estimates, but they are real people, and a forecast of a named person's worth can be wrong in ways that matter to them. Keep outputs inside coursework. Built with help from Claude (Anthropic); every modelling decision and its reason is logged in the repo's docs/decisions.md and docs/genai_log.md.