Essays in Industrial Organization and Finance

Abstract

The first two chapters study the long-run implications of bank branch closures stemming from the digitization of banking services and increased competition from fintech mortgage lenders. In the last decade, nonbank lenders have doubled their market share and now originate more than half of US mortgages. Meanwhile, banks have closed more than 13,000 (13%) branches. Reduced-form evidence shows this rise in nonbank mortgage origination is responsible for a large proportion of the observed decline in bank branching. Furthermore, this nonbank growth led to spillovers in banks' deposit-taking and small business loan origination, reducing both substantially. To evaluate the significance of the branching response for policy analysis, I build a structural model of dynamic bank branching decisions. Banks earn flow profits from taking deposits and providing lending services while dynamically adjusting their branch networks in response to market conditions. As each bank faces substantial heterogeneity in its competitive environment over time, the branching equilibrium is characterized by a high-dimensional dynamic game. Banks must properly account for the evolution of state variables, including demographics, fintech competition, and their rivals' branch networks. I develop a new dimensionality-reduction technique for estimating and solving the model and evaluate the long-run impact of policies such as capital requirements and conforming loan limits. My work suggests that endogenous branching responses can have significant implications for policy analysis in banking markets.

The third chapter, coauthored with Andreas Schaab, discusses the aforementioned dimensionality-reduction technique in detail. We show how adaptive sparse grids can be used to solve high-dimensional dynamic programming problems in economics such as the one from my second chapter. We then apply our method to high-dimensional continuous-time variants of the models from Krusell and Smith (1998) and Ericson and Pakes (1995) and find that we achieve significant computational gains. Beyond these example cases, our method is highly scalable and portable, potentially allowing for broad applicability.