<![CDATA[Class imbalance is a classic yet critical challenge in credit risk modeling, where the number of creditworthy applicants significantly outweighs those at risk of default. Standard classification models often become biased toward the majority class, resulting in deceptively high accuracy while failing to effectively detect actual default risk. This study proposes a Cost-Sensitive Learning approach combined with a Threshold Moving strategy applied to the Random Forest algorithm to minimize False Approvals (representing the greatest financial risk). Using the CRISP-DM methodology, we compare the performance of a baseline model with an optimized model based on an asymmetric cost matrix (cost ratio of 10:1 between False Positive and False Negative). Experimental results demonstrate that adjusting the decision threshold from 0.5 to 0.6 successfully eliminates all False Approvals without significantly compromising overall accuracy (remaining at 98%). Further financial simulation indicates that the cost-sensitive model can improve estimated profitability by up to 25% compared to the standard model. These findings highlight that cost-based evaluation metrics are more relevant for strategic business decision-making than mere statistical accuracy.]]>
Copyrights © 2026
