Sensitivity Analysis: Improving Inclusive Credit Scoring Algorithm Through Feature Weight and Penalty-Based Approach

Recent advancements in artificial intelligence (AI) and machine learning (ML) have enhanced credit scoring systems by improving prediction accuracy, fairness, and transparency. However, limited attention has been given to financial inclusion, particularly how algorithmic adjustments can improve access for underserved borrowers. This study addresses this gap by evaluating how parameter tuning in credit scoring models impacts borrower classifications and promotes inclusivity in online lending systems. Using a micro-enterprise loan dataset from Indonesia, we simulate three approaches: Feature Weight Adjustment, Penalty-Based Models, and a novel Hybrid Feature Penalty Tuning (HFPT) that combines both. Each method is evaluated using custom metrics that measure changes in borrower distribution, inclusion gains, and reclassification risks. Results show that HFPT consistently improves inclusive outcomes without increasing high-risk concentrations. These findings underscore the importance of regulatory oversight and transparent algorithm design, as small changes can easily influence the outcome of models' parameters. For governments, adopting clear guidelines for implementing and evaluating such models is crucial to ensure responsible lending practices. The flexibility of algorithm design, as manifested in this work, offers a pathway for policymakers to balance inclusivity with risk management in credit scoring systems.