Author(s)

Cong Zhang

Corresponding Author

Cong Zhang

ABSTRACT

In recent years, small loan businesses are growing rapidly in emerging markets. And due to lack of traditional customer risk related information and development of big data related techniques, more and more financial loan companies apply machine leaning models to manage loan risks. However, traditional machine learning algorithms do not financially optimize or evaluate models and thus the optimal models we get may not be the best ones in term of financial perspective. In this paper, the author adds CVA components into model objective functions on three popular GBDT-based machine learning models: Xgboost, LightGBM, Catboost; and applies those models on Kaggle’s European credit card fraud detection dataset and Lending Club data loan dataset to verify whether this technique can lead to a financially preferred result. As a result, it is found that using CVA components to adjust model objective functions during training process will enable models to predict more accurately for the loans that are more likely to make large gains/losses, and thus give us financially optimized models.

KEYWORDS

Default risk, machine learning model, GBDT, CVA