Behavioural Segmentation of Credit-Constrained Households Using Unsupervised Machine Learning

Résumé

Credit-constrained households are frequently treated as a homogeneous group in financial analysis, despite substantial heterogeneity in their financial behavior. This simplification limits the effectiveness of financial inclusion policies by obscuring important behavioral differences across households.

This study addresses this limitation by developing an interpretable unsupervised machine learning framework to uncover latent behavioral segments among credit-constrained households using FinScope microdata from the Rwanda National Institute of Statistics (NISR). A purposive filtering approach is applied to isolate credit-constrained households, followed by trimmed variance feature selection to retain the most informative financial variables while mitigating the influence of extreme outliers. The selected variables are standardized and analyzed using K-Means clustering.

Model evaluation results indicate that a three-cluster solution provides the optimal segmentation structure, as supported by the Elbow Method and a Silhouette Score of approximately 0.68, reflecting satisfactory cluster separation and internal cohesion.

The resulting clusters reveal distinct behavioral profiles, including low-income debt-burdened households, high-income high-leverage households, and moderate-income wealth-accumulating households. These findings demonstrate that credit constraint is not a uniform condition but a multidimensional phenomenon shaped by diverse financial structures, highlighting deviations from traditional assumptions of homogeneous household behavior.

In conclusion, the study provides a scalable and interpretable analytical framework that enhances behavioral segmentation in household finance and supports more targeted, evidence-based financial inclusion policies and institutional decision-making.