Distributional Deep Learning for Flood Risk Under Compound Climate Events: An Application to Index Insurance, (with Peng Shi).
Abstract: Catastrophic flood events are increasingly driven by compound climate processes, where interactions across atmospheric, hydrological, and geographic systems generate complex and spatially structured loss patterns. These features pose challenges for property-level loss modeling and limit the effectiveness of traditional risk management approaches. We develop a hierarchical distributional deep learning framework that captures temporal evolution of weather conditions, spatial structure induced by geographic proximity and hydrological connectivity, and heterogeneous exposure characteristics. The model integrates sequence-based representations, graph-based spatial learning, and permutation-invariant aggregation of property-level features, and parameterizes the full conditional loss distribution using a generalized gamma family.
Using data from named flood events under the National Flood Insurance Program, we show that the proposed approach improves predictive performance at both the property and portfolio levels relative to existing methods, including under forward-looking evaluation. Beyond prediction, we demonstrate the value of distributional modeling through an application to index insurance design for catastrophic risk management. These findings illustrate how improved modeling of losses arising from compound climate events can support more effective and responsive risk mitigation strategies.
Bridging the Disaster Protection Gap with Index Insurance, (with Peng Shi). Under Review. SSRN.
Abstract: Natural disasters have become increasingly frequent and severe, leading to rising financial costs. Yet, the insurance protection gap remains substantial, with over 60% of global economic losses uninsured between 2012 and 2021. Index insurance, a relatively new approach, offers a potential solution by providing payouts based on pre-specified indices, significantly reducing both costs and settlement times compared to traditional indemnity insurance. This raises the question: can index insurance complement indemnity insurance in bridging the disaster protection gap? While promising, index insurance faces the challenge of basis risk—the discrepancy between the index and the actual losses. This paper develops a conceptual framework comparing an indemnity-only market with a joint market offering both indemnity and index insurance, showing how index insurance complements indemnity coverage in bridging the protection gap. Our empirical analysis focuses on flood insurance. By leveraging rich, yet complex, weather data and advanced deep learning techniques, we develop a modeled index designed to forecast ultimate flood losses. Specifically, to capture the intricate effects of compound weather events, we propose a neural-network-based predictive model. This model features a recurrent neural network with an attention mechanism to capture the temporal weather dynamics, complemented by a feedforward network to handle nonlinear dependencies and complex interactions between weather variables and static information. The proposed index outperforms the benchmark indices and improves average consumer welfare in the joint market relative to the indemnity-only market. These findings offer valuable insights for policymakers, insurers, and policyholders on how risk management innovations can enhance disaster resilience.
A Copula Model for Marked Point Process with A Terminal Event: An Application in Dynamic Prediction of Insurance Claims, (with Lu Yang and Peng Shi), Annals of Applied Statistics, 18(4), 2679-2704, (2024). Link.
Storm CAT Bond: Modeling and Valuation, (with Ken Seng Tan, Jinggong Zhang, and Wenjun Zhu), North American Actuarial Journal, 28(4), 718–743, (2024). Link.
Epidemic Financing Facilities: Pandemic Bonds and Endemic Swaps, (with Jinggong Zhang and Wenjun Zhu), North American Actuarial Journal, 28(3), 626–657, (2024). Link.