In this study, we propose a cosine similarity under-sampling method for financial distress prediction. This method calculates the similarity between financially healthy company samples and iteratively adjusts a similarity threshold to filter out excess healthy samples, which not only balances the dataset but also retains the key information from original data. We conduct an empirical analysis using data from 129 ST (Special Treatment) listed companies and 1,425 non-ST companies in China, evaluating the model’s performance with three different classifiers: Support Vector Machine (SVM), Logistic Regression (LR), and Backpropagation Neural Networks (BPNN). The results show that the cosine similarity under-sampling method not only effectively alleviates the class imbalance problem but also improves the performance of various classifiers. This method provides a reliable tool for financial distress prediction, helping enterprises, financial institutions, and investors take more targeted measures to avoid potential financial losses.