Feature Selection Methods as Attribute Weighting Schemes for Clustering Corporate Financial Statements

Abstract

Financial clustering plays an important role in uncovering hidden patterns and supporting decision-making in high-dimensional financial statement analysis. However, clustering performance is often constrained by noisy, redundant, and heterogeneous attributes that reduce cluster quality, interpretability, and robustness. Although previous studies in financial analytics have extensively explored clustering algorithms, limited attention has been given to systematically evaluating feature-selection-based attribute weighting strategies for improving clustering effectiveness in complex financial datasets. This study investigates the effectiveness of 6 feature-selection-based weighting strategies for improving financial statement clustering using annual financial data from 604 publicly listed companies on the Indonesia Stock Exchange during 2020-2023. Rather than relying solely on feature filtering, the evaluated methods were utilized to prioritize informative financial attributes and improve clustering structure. Clustering performance was assessed using internal validation metrics, including the Silhouette Coefficient, Dunn Index, Calinski Harabasz Score, and Davies Bouldin Index to evaluate cluster compactness, separation, and overall quality. The results demonstrate that feature-selection-based weighting substantially improves clustering quality compared with the baseline. Fisher Score achieved the strongest overall performance with a Silhouette score of 0.9899, Dunn Index of 1.7238, and Davies Bouldin Index of 0.0034, outperforming the baseline values of 0.8678, 0.3509, and 0.8716, respectively. Autoencoder-based ranking also produced highly competitive results, achieving a Silhouette score of 0.9757 and a Calinski Harabasz Index of 37,105.45. These findings indicate that prioritizing informative financial attributes significantly enhances cluster compactness, separation, and interpretability. The study contributes a comprehensive comparative evaluation of feature-selection-based weighting schemes and provides practical insights for financial segmentation, risk profiling, decision support, and data-driven financial analytics. This study further offers a foundation for developing more robust clustering frameworks for complex financial datasets.