Causal Counterfactual Underwriting: A Neural Structural Discovery Engine for Equitable MSME Loan Access in African FinTech

We propose a neural structural causal discovery framework that recasts MSME loan underwriting within African FinTech ecosystems as a causally grounded counterfactual decision process. Conventional credit scoring models depend on opaque statistical correlations that frequently fail under distributional shifts caused by economic shocks or policy interventions, particularly in low-formalization markets where alternative data streams are plentiful yet noisy. To address this limitation, we introduce a three-module architecture that jointly learns, disentangles, and intervenes on causal mechanisms underlying repayment behavior. The first module employs a differentiable causal structure learner with a continuous acyclicity constraint to infer a directed acyclic graph from heterogeneous digital footprint features such as mobile money transaction frequencies and utility payment punctuality. This graph captures statistically robust, intervention-invariant relationships between features and default risk. The second module merges a variational autoencoder with an information bottleneck and an interventional consistency regularization term to generate a twenty-dimensional latent space in which each dimension corresponds to a distinct causal factor, such as cash flow stability or social collateral strength. This disentanglement is enforced by simulating do-interventions on latent dimensions and minimizing the divergence between predicted and observed counterfactual outcomes. The third module constitutes a counterfactual policy generator that identifies sparse modifications to this latent space, thereby yielding actionable and individualized loan policy recommendations for denied applicants, such as raising savings deposits by a specific percentage within a defined timeframe. The system connects with existing digital wallet and CRM infrastructure, substituting a scalar credit score with a causally meaningful vector that is subsequently processed by a monotonic classifier for approval decisions. We further embed a regulatory audit mechanism that calibrates the discovered causal graph against historical monetary policy shocks, holding a divergence metric below a predefined threshold to guarantee alignment with known economic interventions. This work contributes a structural, explainable, and intervention-grounded paradigm for equitable credit access in resource-constrained environments. Keywords: African FinTech, Alternative Data, Causal Discovery, Counterfactual Reasoning, MSME Lending.