Multi-dimensional data fusion for enterprise debt maturity risk assessment: a stacked autoencoder-based deep learning approach

This study proposes a novel data fusion-based feature selection and reconstruction (DFFSR) method for assessing debt maturity risk based on multi-dimensional data from publicly traded Chinese companies between 2000 and 2023. The DFFSR approach maps the fused data into a lower-dimensional embedding space using a stacked autoencoder (SAE), thereby enabling feature reconstruction while preserving data heterogeneity. It employs a CancelOut layer to identify a salient subset of features and reduce indicator redundancy. The DFFSR model also enhances managers’ understanding of the decision-making process related to debt maturity. The comprehensive ranking of the relative importance of risk factors enables enterprises to manage risk more effectively by focusing on key indicators, optimizing their liability structure, and improving overall performance.