Answers are grounded in this paper and cite their sources.
The practical challenge posed by corporate bankruptcy prediction does not stem from the limitations of any particular classifier but rather from the very structure of the data. Financial datasets are typically plagued by numerous missing values, cluttered with partially redundant ratios, and structurally imbalanced, with failing firms representing only a small minority of observations. This paper proposes DAWN (Directional Asymmetry-Weighted dissimilarity for Nearest-case retrieval), a novel framework resting on four pillars: (i) Missing values are imputed using Case-Based Reasoning (CBR), where the nearest donor case is identified according to the proposed DAWN dissimilarity measure rather than conventional Euclidean, Manhattan, or cosine distances. (ii) Feature selection is performed using DAWN-FS, which ranks variables through a hybrid criterion combining mutual information with directional discriminability, retaining variables only when their class-specific medians exhibit the empirically identified direction of association with financial distress. (iii) Class imbalance is addressed using DAWN-SMOTE, an oversampling strategy that constructs minority-class neighborhoods according to the DAWN dissimilarity measure and introduces a directional interpolation mechanism that, on a ratio-specific basis, biases synthetic samples toward the identified risk trajectory. (iv) The SVM hyperparameters are optimized using a hybrid Open Competency Optimization–Particle Swarm Optimization (OCO–PSO) metaheuristic, guided by a G-mean-based fitness function. The proposed DAWN-SVM framework is evaluated as an integrated end-to-end methodology against baseline pipelines on one real-world UK dataset and three public benchmarks (Taiwan, Russian, and Slovak), using stratified 10-fold cross-validation across nine metrics, Wilcoxon signed-rank tests, and Holm–Bonferroni correction. All preprocessing parameters, including CBR-DAWN imputation, are estimated within each training fold to prevent data leakage. The four pillars are designed to operate jointly within one prediction pipeline, sharing the same robust scale, risk direction, and informational weight; the performance claims in this paper therefore concern the complete DAWN-SVM framework. DAWN-SVM ranks first on all nine metrics, with 27 of the 36 comparisons against the baselines remaining significant after Holm–Bonferroni correction and none favoring a baseline. The improvements are particularly pronounced on the dataset with the most missing values and on the severely imbalanced Russian dataset. These end-to-end results support the effectiveness of the integrated framework under challenging missing-data and class-imbalance conditions. A dedicated analysis of the dissimilarity itself shows that DAWN retrieves firms with the same outcome more accurately than the Euclidean, Manhattan, and cosine measures and that this advantage is robust to the number of ratios and to the strength of the directional penalty.
No takes yet. Share an insight, caveat, or question.
Tamouh et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: