Decoding Machine Learning Performance in Credit Risk Assessment: A Comparative Analysis Based on the Home Credit Default Risk Challenge

Abstract

Accurate credit risk assessment underpins modern financial risk management. Meanwhile, advanced machine learning algorithms have largely replaced traditional linear models in default prediction. However, real-world consumer credit data is scattered across complex relational tables and exhibits extreme class imbalance. As a result, conventional coarse-grained aggregations often fail to capture these nuances, leading to the loss of crucial micro-level default signals. Accordingly, this paper conducts a comparative analysis of the machine learning workflows that performed the best in the Home Credit default risk challenge. In particular, through qualitative deconstruction of pipeline architectures, three representative high-level parallel processing pipelines are analyzed and compared: one based on domain knowledge, one based on heterogeneous stacking framework, and one based on microscopic target aggregation scheme. Furthermore, the key differences among them in terms of feature engineering, handling of imbalanced data, and integration architecture are compared and evaluated. The results indicate that different architectures can effectively extract sparse default signals via their inherent mechanisms, yielding significant gains on imbalanced credit data. It further demonstrate the high effectiveness of feature dimension reconstruction and target dimension reduction in handling extremely imbalanced credit data, providing certain references for industrial credit risk modeling.