Abstract
Financial engineering has a mature toolkit for modeling default risk under conditions of incomplete information—reduced-form intensity models originally designed for corporate credit, where bond markets cannot observe a firm's assets directly and must infer default risk from noisy accounting signals (Duffie and Lando, 2001). This paper asks whether the same class of model can be repurposed for a different information problem: retail borrowers in data-scarce markets who have no financial-statement or credit-bureau history at all, but who leave behind a different kind of trace—mobile-money transaction records, utility payment logs, airtime top-up patterns—that has been shown to carry predictive signal about repayment behavior (Björkegren and Grissen, 2020).
We formalize this as a Cox (1972) proportional-hazards default model in which the covariate vector is partitioned into a traditional block (loan and borrower characteristics ordinarily available even in thin-file settings) and an alternative-data block (behavioral covariates derived from digital footprints), and we study the properties of this model under a fully specified, parametrically known data-generating process. Because no de-identified loan-level dataset combining both blocks with verified default outcomes was available to us for this study, we deliberately do not present the results as an empirical finding about any real market; instead we treat the exercise as a Monte Carlo validation of the estimator and of the discrimination gain the framework can plausibly deliver, under assumptions we state explicitly and calibrate loosely to the order of magnitude reported in the alternative-credit-scoring literature.
Across 8,000 simulated loans, adding the alternative-data block raises the concordance index from 0.638 to 0.693 and is jointly significant against the traditional-only model (likelihood-ratio χ²₃ = 334.9, p < 10⁻¹⁶); the discrimination gain scales monotonically with the assumed strength of the alternative-data signal, degrades gracefully rather than catastrophically as alternative-data coverage becomes incomplete, and the maximum-likelihood estimator is essentially unbiased with correct confidence-interval coverage down to samples of 500 borrowers. We interpret these as necessary, not sufficient, conditions for using hazard-based alternative credit scoring in production, and we set out explicitly what empirical validation on real repayment data would still need to show.