Forecasting Corporate Bankruptcy Through Class‐Rebalanced Self‐Training Semi‐Constrained Matrix Factorization

Forecasting corporate bankruptcy and financial distress is a crucial and intriguing research topic within the realm of business and finance. Recently, numerous models for forecasting bankruptcy and financial distress have been developed using artificial intelligence techniques under a supervised learning paradigm. However, in practical applications, generating large amounts of labeled samples for training supervised learning models is a highly inefficient and labor‐intensive process. In this paper, by taking practical application scenarios into consideration, we propose a novel bankruptcy and financial distress prediction method called Class‐Rebalanced Self‐Training Semi‐Constrained Matrix Factorization (CRST‐SemiCMF), which integrates the class‐rebalancing technique and the self‐training strategy within a constrained matrix factorization framework. The CRST‐SemiCMF method can alternatingly train a semi‐constrained matrix factorization model on both labeled and unlabeled datasets, progressively expanding the labeled dataset by sampling the unlabeled dataset with pseudo‐labels. Crucially, our method can adaptively determine the sampling rates for different classes of imbalanced datasets. Moreover, we also introduce a novel modification to the self‐training strategy employed in our proposed method to effectively address class‐imbalance issues commonly encountered in datasets for bankruptcy and financial distress prediction. To validate the proposed method, we conduct extensive experiments using multiple UCI benchmark datasets and a real‐world financial dataset of Chinese listed companies. The results demonstrate that (1) our method achieves the highest classification accuracy across nearly all proportions of available labeled samples; (2) it obtains a false negative rate on par with supervised‐CMF while surpassing SemiCMF; and (3) it gets better Recall and F1‐scores than SemiCMF, with performance comparable to or exceeding supervised CMF. These findings consistently confirm our method's effectiveness for bankruptcy prediction and financial distress forecasting across multiple evaluation metrics.