This Capstone Project develops a solution for banks to monitor their exposure to climate risk impacts on the U.S. property and casualty (P&C) insurance market, combining national-scale predictive modeling with a geographically granular Bayesian forecasting framework for California. Project Goal and Context The core problem is that increasing frequency and severity of natural disasters due to climate change lead to higher insurance premiums and, in some high-risk areas, insurers withdrawing coverage. This exposes banks to three main risks: increased mortgage default rates, loss of insurance protection on collateral, and decline in home values — all of which raise credit risk. The project objectives are to develop a predictive model for home insurance premium rate increases by region and to identify regions with potential home price declines associated with climate factors. The final solution aims to predict the future effects of catastrophic weather and climate risk on Homeowners Insurance Premiums (HOI) and the Home Price Index (HPI). Methodology and Findings National HPI Modeling. The team built a unified panel dataset linking various climate hazard indicators with the Home Price Index (HPI) and macroeconomic control variables (e.g., unemployment, GDP per county). Modeling progressed through several stages: • Initial Modeling: Simple linear models (OLS, Ridge Regression) proved ineffective, yielding very low R² scores (around 0.015–0.15). Time-series models (AR(1) and mixed-effects models) performed better by capturing broad housing cycles, including the 2008 crash. • Machine Learning Performance: The XGBoost gradient-boosted trees model provided the strongest forecasting signal, achieving an R² of approximately 0.98 on the test set for year-over-year HPI changes. California Premium Forecasting. A Bayesian hierarchical forecasting framework was developed for homeowner insurance premiums at the ZIP code and county level. Using observed insurance, climate, and financial data from 2018 to 2020, the model estimates how baseline premium levels and time trends vary across geographies and quantifies the associations between ten climate hazard and financial predictors and earned premium per housing unit. Validated on held-out data from 2021 to 2023, the model achieved an overall R² of 0.46, a mean absolute error of $242, a MAPE of 14.6%, and a 95% posterior predictive interval coverage of 95.0% — matching the nominal calibration target. Accuracy is highest in the nearest out-of-sample year (2021: R² = 0.77, MAPE = 6.8%) and degrades at longer horizons, consistent with the inherent difficulty of extrapolation beyond the observed period. For long-term planning, the framework generates scenario-based premium projections for 2025 through 2030 under three climate-risk pathways calibrated to Network for Greening the Financial System (NGFS) guidance on climate hazard growth rates: optimistic (1.5 °C warming, Paris Agreement goals achieved), moderate (2.0 °C, current policies continued), and pessimistic (3.0 °C, limited climate action). The framework is implemented in Python using open-source tools and delivers three ZIP-level scenario forecast CSV files covering 2025–2030. Key Challenges The central methodological challenge is a structural break in housing market behavior caused by the COVID-19 pandemic. Models trained on pre-2020 data (the period with complete climate and financial variables) fail to accurately capture the sharp HPI increase observed between 2021 and 2024, creating a trade-off between model stability and real-world realism.
Project poster (PDF)
Mentor: Alex Fisher
