Forecasting Credit Card Defaults Using Light Gradient Boosting Machine with Dart Algorithm

Abstract

With the rapid development of financial services and technologies, credit cards have been increasingly used for personal daily consumption and small loans. However, bad debts caused by credit card defaults remarkably affect the healthy development of financial markets. Therefore, forecasting potential credit card defaults is of great significance with respect to financial stability and economic order. For this purpose, we propose a machine learning method based on Light Gradient Boosting Machine to detect credit card defaults in this paper. DART algorithm is utilized in our model instead of the traditional gradient boosting tree. The model is trained and evaluated using the dataset provided by American Express in the Kaggle competition American Express - Default Prediction. Based on feature analysis and engineering, raw data with 190 descriptors are transformed into data with 2358 descriptors, and are used to train 3 LightGBM models with different hyper-parameters. By applying the model ensemble and pseudo-label technique, the competition metric of our method reaches 0.80029/0.80767 on the public/private test set. This score ranks 106/4874 (top 2.2%), and can get a silver medal in the Kaggle competition.