A multi-level sparse attentive fusion network integrating hard and soft information for firm-level loan default prediction

Abstract

Firm-level loan default prediction (FLDP) deserves much attention from both academic and industry. Even a small improvement in the accuracy of FLDP could lead to significant savings by reducing credit risk. While previous studies have utilized deep learning models for FLDP task, they failed to well handle the intra-type ambiguity and inter-type interaction simultaneously facing with combined hard and soft information, thus remaining an area of ongoing development. By this perspective, we seek to design a novel Multi-level Sparse Attention (MLSA) based deep learning fusion framework for FLDP, aiming to fully capture default signals conveyed from both hard and soft information. First, multiple types of information are extracted grounded in 5P theory and LAPP theory, ensuring the sufficiency and rationality of the features. Second, Sparse Attentive MLP (SA-MLP) and Sparse Attentive GRU (SA-GRU) module are proposed to handle the intra-type ambiguity embedded in hard and soft information separately. Further, the Attentive Fusion (AF) module including Differential Enhancive module and Common Selective module is proposed to explore inter-type interaction among hard and soft information. Last, we adopt the focal loss function to mitigate the adverse effects of imbalanced data. The proposed MLSA informs future FLDP research about how to fully exploit the value of hard and soft information by considering their intra-type ambiguity and inter-type interaction. Empirical evaluation of the MLSA on a real-world dataset demonstrates its outperformance of state-of-the-art benchmarks in the FLDP task. Our results also contribute to the growing literature on this topic by highlighting the roles of hard and soft information and improving interpretability.