Abstract: To address the limitation of traditional binary financial distress prediction models in providing fine-grained, tiered early warnings, this paper constructs an interpretable multi-class financial crisis early warning model based on the fusion of financial and non-financial information. First, management discussion and analysis (MD&A) tone information is incorporated to enrich data sources for small and medium-sized enterprises (SMEs). Subsequently, random forest (RF), light gradient boosting machine (LightGBM), and support vector machine (SVM) models are utilized to predict the financial performance of SMEs, which are then further integrated using an improved Dempster-Shafer (D-S) evidence theory. Finally, the SHapley Additive exPlanations (SHAP) is introduced to facilitate interpretable analysis. The results show that the information-fusion model exhibits a 1.3% improvement in the F1 score compared with the best-performing base classifier, effectively avoiding prediction “disaster points.” The model also identifies key early warning indicators such as the debt-to-asset ratio, undistributed earnings per share, and return on equity. Overall, the proposed model achieves more accurate financial crisis classification and more stable predictive performance, thereby providing a novel perspective for financial crisis early warning research on SMEs.

Key words: multiple-class classification, financial crisis early warning, information fusion, SHapley Additive exPlanations (SHAP), decision support

摘要： 针对传统二分类财务困境预测模型难以提供细粒度分级预警问题，本文构建了一个基于财务与非财务信息融合的可解释多分类财务危机预警模型。首先，通过引入管理层讨论与分析（MD&A）语调信息，丰富中小企业数据源；其次，采用RF、LightGBM和SVM对中小企业财务状况进行初步预测，并运用改进的D-S证据理论对结果进行二次融合；最后，借助SHAP框架对模型进行可解释性分析。研究发现：基于信息融合模型的F1值相比最优基分类器提升了1.3%，能够有效避免预测“灾难点”的出现，同时揭示了资产负债率、每股未分配利润和净资产收益率等指标在财务预警中的重要作用。本文模型具备更精准的财务危机定位能力和更稳定的预测效果，为中小企业财务危机预警研究提供了新视角。

关键词: 多分类, 财务危机预警, 信息融合, SHAP, 决策支持