10th U6+ Consortium of African Universities International Conference

Digital credit markets in East Africa have expanded rapidly, driving financial inclusion for thin-file borrowers via mobile money ecosystems. Conventional credit scoring relies heavily on correlation-based machine learning, which propagates structural biases, conflates spurious associations with actual default drivers, and fails systematically during macroeconomic volatility. This paper evaluates CreditPulse, an open-source Causal AI framework designed to decouple statistical correlations from genuine causal default mechanisms to facilitate equitable risk assessment. Validated on 307,511 loan applications, the framework integrates a multi-layered pipeline combining causal inference (Difference-in-Differences [DiD] and Regression Discontinuity [RD] designs), Bayesian uncertainty quantification via bootstrap ensembles, regulatory Natural Language Processing, and production-grade machine learning classifiers (XGBoost, Random Survival Forest, and Isolation Forest). The predictive pipeline achieved an XGBoost AUC-ROC of 0.7415 and a survival C-index of 0.8130. Econometric evaluations isolated a statistically significant income-shock causal coefficient of -0.1125 (p = 0.002) via DiD, and a Local Average Treatment Effect of 0.0058 (p = 0.0028) via RD. Furthermore, segment-level fairness audits verified that demographic approval disparities reflect genuine risk differentials rather than systemic algorithmic bias, ensuring alignment with IFRS 9 and Central Bank of Kenya regulations. Fusing rigorous causal econometrics with predictive machine learning proves that high diagnostic power, structural transparency, and regulatory compliance can successfully coexist within a unified open-source architecture. Digital lenders and financial regulators in emerging markets should transition from purely associative models to causal AI frameworks to mitigate systemic credit risk, improve portfolio resilience, and guarantee algorithmic fairness.