Abstract
Commercial banks face a structural choice: maintain dense physical networks or thin them as digital intermediation grows. We ask how the rise of fintech firms reshapes the three pillars of bank outreach, namely branches per 100,000 adults, accounts per 1000 adults, and ATMs per 100,000 adults. Using a 103-country panel from 2005 to 2021, we estimate Pseudo Poisson Maximum Likelihood (PPML) models with country and year fixed effects and country-clustered standard errors. Because dependent variables are continuous rates per adult population, PPML is consistent under any non-negative conditional mean and avoids the integer count assumption that limits classical negative binomial estimators. Endogeneity between fintech entry and bank infrastructure is addressed through a Bartik-style shift-share instrument that combines pre-sample subsector composition with leave-one-out global subsector growth, paired with a Rivers-Vuong style control function. Across specifications, fintech entry is associated with fewer branches and ATMs, and the instrumented estimates point to substitution at both margins, while its effect on accounts is moderated by digital readiness and bank capitalization. Because the exclusion restriction cannot be verified directly, we state the assumptions under which the instrumented estimates admit a causal reading and treat that reading as conditional on them throughout. Regional patterns are heterogeneous in economically interpretable ways. Europe shows the clearest substitution, Africa shows account inclusion alongside ATM substitution and funded fintechs that co-develop physical infrastructure, and Asia and the GCC show muted effects consistent with bank-fintech coexistence. The regional estimates rest on materially smaller subsamples than the global model and are reported as indicative of heterogeneity. We relate these patterns to specific banking mechanics, namely electronic know-your-customer onboarding, agent banking, mobile money rails, and digital wealth interfaces, which the cross-country design does not measure directly and which we therefore advance as interpretation rather than as identified channels.