Skip to content
← All projects
Data Science CompetitionDeployed - Top 10% (GCI World 2026)

NFL Draft Prediction - Gradient-Boosting Ensemble

A LightGBM + CatBoost ensemble predicting whether a college athlete gets drafted into the NFL from Combine data, served as a live serverless API with SHAP explanations. Placed in the top 10% worldwide at GCI World 2026 (University of Tokyo, Matsuo Lab).

Live APICode

0.829

OOF ROC AUC

Top 10%

GCI World 2026

50

Optuna trials

2,781

Train rows

Overview

A competition-grade tabular ML project that predicts NFL Draft outcomes using careful validation, feature engineering, and gradient-boosting ensembles - then packaged, tested, explained, and deployed as a live service.

The core analytical insight is that missing Combine data is itself highly predictive: players who skip drills usually were not invited and rarely get drafted.

Impact

  • Reached 0.829 OOF ROC AUC with Optuna tuning (50 trials), 5-fold CV, 3-seed averaging, and an 85/15 LGBM-CatBoost blend.
  • Used leak-safe fold-internal target encoding for 236 schools and treated missing Combine drills as predictive signal.
  • Deployed as a serverless Modal /predict API with a Gradio UI, SHAP per-prediction factors, and Weights & Biases tracking.

Modeling craft

  • Leak-safe target encoding computed inside each CV fold so a player's own outcome never leaks into their features.
  • Missingness-as-signal design: per-drill missing flags, drills-missing counts, and full-combine indicators.
  • Two complementary models on purpose - LightGBM on encoded schools, CatBoost on raw categoricals - blended 85/15 by OOF AUC.
  • Variance control on a 696-row test set via 3-seed x 5-fold averaging, trusting OOF AUC over the noisy public leaderboard.

Engineering & deployment

  • Refactored a competition script into an installable src/ package with pinned dependencies and pytest coverage.
  • One shared feature pipeline used by both training and serving, so the model cannot drift between them.
  • Serverless Modal /predict endpoint returns probability + top SHAP factors; a Gradio thin client hosts the form.
  • Weights & Biases tracks Optuna trials and ensemble progression.

Stack