An ELM-based Classification Algorithm with Optimal Cutoff Selection for Credit Risk Assessment

Lean Yu, Xinxie Li, Ling Tang, Li Gao

In this paper, an extreme learning machine (ELM) classification algorithm with optimal cutoff selection is proposed for credit risk assessment. Different from existing models using a fixed cutoff value (0.0 or 0.5), the proposed classification model especially considers the optimal cutoff value as one important evaluation parameter in credit risk modeling, to enhance the assessment accuracy. In particular, using the powerful artificial intelligence (AI) tool of ELM as the basic classification, the simple but efficient optimization algorithm of grid search is employed to select the optimal cutoff value. Accordingly, three main steps are included: (1) ELM training using the training dataset, (2) cutoff optimization via the grid search method using the training and validation datasets, and (3) classification generalization based on the trained ELM and optimal cutoff using the testing dataset. For illustration and verification, the experimental study with two publicly available credit datasets as the study samples confirms the superiority of the proposed ELM-based classification algorithm with optimal cutoff selection over other some popular classification techniques without cutoff selection.