This project investigates how climate-related disasters translate into household credit stress by estimating whether borrowers in disaster-affected counties become more likely to fall into serious delinquency or default, and how quickly that risk rises after an event. The central idea is to combine two levels of information: (1) county-level disaster exposure capturing the type of event (e.g., flood, hurricane, wildfire, severe storm), its timing, and recent frequency; and (2) loan-level mortgage performance outcomes that allow us to identify when a loan first becomes severely delinquent and/or transitions into a default-like state.

We will build the analysis in layers. First, we will construct “exposure windows” (e.g., 1–3–6–12 months after a disaster) and engineer features such as recency, cumulative disaster count, and event-type indicators for each county-month. Next, we will estimate an interpretable baseline using logistic regression to quantify the marginal effect of each disaster type on default probability while controlling for loan and borrower characteristics. We will then train an XGBoost model to capture nonlinear effects and interactions (for example, whether disasters disproportionately increase risk for high-LTV loans or certain loan ages). Model explainability outputs will be used to communicate which disaster features and borrower/loan attributes drive risk changes and to produce clear, regulator-friendly narratives.

The project is compelling because it connects physical climate shocks to micro-level credit outcomes in a way that supports real-world risk monitoring and climate stress-testing. Planned datasets include FEMA disaster declaration records for county-level disaster events and a loan-level mortgage performance dataset that provides monthly performance histories from which default timing can be derived.