| 基于时空图卷积的上市企业财务困境预测 |
| Financial Distress Prediction for Listed Companies Based on Spatiotemporal Graph Convolution |
| 摘要点击 73 全文点击 0 投稿时间：2025-08-21 修订日期：2026-05-18 |
| 查看/发表评论 下载PDF阅读器 |
| 中文关键词 时空图卷积；财务困境预测；违约特征；深度学习解释 |
| 英文关键词 Spatiotemporal Graph Convolution, Financial Distress Prediction, Default Characteristics, Deep Learning Interpretation. |
| 基金项目 |
| 投稿方向 金融数据分析与图深度学习 |
| 作者 | 单位 | 邮编 | | 张晓黎* | 上海对外经贸大学 | 201102 |
|
| 中文摘要 |
| 以往研究聚焦财务指标和提升机器学习整体正确率，对多数类预测的准确性掩盖了对少数类别预测的精准性，对机器学习结论缺乏解释。本研究建立时空动态图卷积模型，对多维特征经XGBoost模型筛选加权和深度学习，提高对财务困境少数类预测的准确率。创新体现在：①基于中国上市企业真实数据集，用消融和逐步实验方法构建财务、金融交易、创新、数字化程度、信息披露情感和关联关系六个维度的特征体系。②用最小生成树构建业务和地理关联的动态时空图，使图卷积既耦合当期邻居特征，又耦合前期自我特征。比较XGBoost特征筛选加权、时空图动态相连、当期交易状态和全局网络邻居相似度特征等对模型性能贡献度。③以创新投入最多、区块链程度最高和交易异常企业为例，从关联关系角度探讨企业陷入财务困境的诱因和影响。 |
| 英文摘要 |
| Previous studies have often focused on financial indicators and improving the overall efficiency of machine learning models, neglecting the accuracy of predictions for minority categories, with insufficient explanatory in the conclusions. This study establishes a spatiotemporal dynamic graph convolutional model, which enhances the accuracy of predicting the minority class of financial distress by using XGBoost model for feature selection and weighting, followed by deep learning of multidimensional features. Innovations are reflected in: ①Based on a real dataset of Chinese listed companies, a feature system across six dimensions is constructed using ablation and step-by-step experimental methods: finance, financial transactions, innovation, digitalization level, emotional sentiment of information disclosure, and association relationships. ②The dynamic spatiotemporal graph is constructed to allows the graph convolution to couple not only the features of neighbors in the current period but also itself-features from the previous period. The contributions to model performance by the feature selection and weighting of XGBoost, the dynamic connection of the spatiotemporal graph, one's own current transaction status, and global network neighbor similarity features are compared.③Taking companies with the highest levels of innovation investment, the highest degree of blockchain adoption, and those with abnormal transactions as examples, this study explores the causes of enterprises falling into financial distress from the perspective of association relationships. |
| 关闭 |
|
|
|
|
|
