Confusion matrix.

This study analyzes 284 publicly listed companies first designated as ST or *ST between 2015 and 2023. It utilizes two types of textual indicators: Management’s Discussion and Analysis (MD&A) and stock forum comments. PCA and MLP are employed for dimensionality reduction. The study compares the recognition performance of single-class models with ensemble learning models while also examining the impact of various base learners and meta-learners on the performance of the ensemble learning model. The findings show that using the two types of textual indicators significantly enhanced the model’s accuracy in recognition. The single-class and ensemble learning models demonstrated average improvements of 1.24% and 1.75%, respectively. Notably, stock forum comments outperformed MD&A text. Additionally, the MLP proved more effective in feature processing than PCA. The D-M-BSA-FT model achieved an accuracy of 88.89%. Ensemble learning models outperform single classification models. After introducing textual features, the ensemble learning model achieved an average recognition accuracy of 85.31%, compared to 82.09% for the single classification model. Therefore, the financial warning model developed in this study provides valuable insights for enhancing the accuracy of financial warning identification.

本研究分析了2015年至2023年间首批被列为ST（Special Treatment）或*ST（Delisting Risk Warning）的284家上市公司。本研究采用两类文本指标：管理层讨论与分析（Management’s Discussion and Analysis, MD&A）以及股票论坛评论。研究采用主成分分析（Principal Component Analysis, PCA）与多层感知机（Multi-Layer Perceptron, MLP）进行降维处理。本研究对比了单分类模型（single-class models）与集成学习模型（ensemble learning models）的识别性能，同时考察了各类基学习器（base learners）与元学习器（meta-learners）对集成学习模型性能的影响。研究结果表明，引入这两类文本指标可显著提升模型的识别准确率。单分类模型与集成学习模型的平均识别准确率分别提升1.24%与1.75%。值得注意的是，股票论坛评论的表现优于管理层讨论与分析文本。此外，MLP在特征处理方面的表现优于PCA。D-M-BSA-FT模型的识别准确率达88.89%。集成学习模型的表现优于单分类模型。在引入文本特征后，集成学习模型的平均识别准确率达85.31%，而单分类模型的平均识别准确率为82.09%。因此，本研究构建的财务预警模型（financial warning model）可为提升财务预警识别准确率提供极具价值的参考思路。