Granular Support Vector Machine for Bankruptcy Prediction

The occurrence of corporate defaults has manifested as a significant issue for financial entities and stakeholders. Within an increasingly complex economic environment, there is a pressing need to develop innovative approaches for accurate financial risk assessment. One promising approach to preempt crises—particularly bankruptcies—is leveraging insights from other distressed or bankrupt firms. This necessitates the systematic collection and rigorous analysis of corporate financial data. However, such data presents a significant processing challenge: class imbalance. Imbalanced datasets are characterized by a disproportionate distribution of samples across classes. In practice, this means that the information available for one class (typically the bankrupt class) is insufficient—its observations are vastly outnumbered by those of the majority class (non-bankrupt firms). Crucially, the minority class is often the primary focus of interest. The scarcity of samples for this class would be less problematic if conventional classification techniques, such as Support Vector Machines (SVMs), did not inherently bias their predictions toward the majority class due to their generalization tendencies. It is imperative to acknowledge that the intricate arrangement of data significantly intensifies the issue at hand. Not only does data scarcity in the minority class impair machine learning performance, but the intrinsic complexity of the data (e.g., overlapping features, nonlinear separability) can also degrade model accuracy. These twin challenges—class imbalance and data complexity—are key obstacles in bankruptcy prediction. To address these issues, this paper employs Granular Computing techniques, including fuzzy sets, rough sets, shadowed sets, and Quotient Space Theory which offer a robust framework for modeling the nuanced membership functions relevant to the minority class (bankrupt firms). We further propose a hybridization of these techniques to optimize performance. Empirical validation conducted on practical datasets demonstrates that our Fuzzy Shadowed Support Vector Machine (Fuzzy Shadowed SVM) significantly outperforms traditional machine learning methodologies, achieving enhanced predictive accuracy.