Abstract
This study explores an advanced methodology for predicting loan defaults by integrating fuzzy community detection (FCD) with machine learning models, specifically targeting the complexities and uncertainties within borrower data in developing economies, exemplified by data from L&T Finance in India. Traditional credit risk assessments often rely on deterministic models that overlook the nuanced relationships among borrower attributes, leading to potential misclassifications and systemic financial risks. By using FCD, this research captures ambiguous borrower profiles, uncovering clusters characterized by shared risk factors, such as credit history, income level, and repayment behaviour, thus allowing a more adaptable approach to credit risk assessment. The integration of FCD with ensemble learning techniques, such as Neural Networks, XGBoost, and LightGBM, significantly enhances predictive accuracy and minimizes both false positives and negatives in borrower classifications. The study reveals the value of a differentiated lending strategy based on unique borrower segments, ranging from those with stable credit histories to those exhibiting higher-risk behaviours. This model's insights not only promote financial stability but also inform more inclusive lending practices by accounting for overlapping uncertainties in borrower data. Future directions suggest expanding FCD applications with dynamic data sources, like social media or alternative credit scores, to capture real-time economic shifts and improve the interpretability of complex model outputs for practical lending decisions.