Kaggle Competition: American Express Default Prediction

Overview

Led a predictive analytics project achieving Top 0.4% ranking (20th out of 4,874 teams) and Silver Medal in the American Express Default Prediction Kaggle competition.

Key Technologies

• Machine Learning: LightGBM (DART), XGBoost (GPU-accelerated), CatBoost
• Ensemble Methods: Linear weighted averages, correlation-based weighting
• Feature Engineering: Time-series features, lag features, rolling statistics

Achievement

🏆 Ranked 20th out of 4,874 teams (Top 0.4%)
🥈 Silver Medal

Technical Approach

• Developed weighted ensemble of LightGBM (DART) and GPU-accelerated XGBoost models
• Processed 16 GB tabular time-series data covering transactions, balances, delinquencies, and repayments
• Optimized hyperparameters via grid search and stratified 5-fold cross-validation
• Designed diverse feature sets including:
  • Lag features
  • Rolling statistics
  • Trend indicators
  • Transaction patterns
  • Temporal aggregations

Model Strategy

• Ensemble Architecture: Combined multiple model types with correlation-based weights
• Stability Enhancement: Trained multiple seeds to boost prediction stability
• Feature Diversity: Extracted temporal patterns from time-series credit data
• Efficient Design: Delivered compact, high-performing solution

Impact

• Demonstrated ability to work with large-scale financial data
• Showcased expertise in ensemble learning and feature engineering
• Achieved top-tier performance in highly competitive global competition
• Led model tuning and ensemble strategy for the team

Competition Details

American Express – Default Prediction
Platform: Kaggle
Timeframe: May 2022 – August 2022
Dataset: 16 GB tabular time-series credit card data