An Attention-Based Stochastic Simulator for Multisite Extremes to Evaluate Nonstationary, Cascading Flood Risk

Climate change creates insurance blind spots: The paper identifies a critical gap in flood risk assessment—existing tools either provide short-term seasonal forecasts (months) or long-term climate projections (decades), but insurance contracts operate on 1-5 year timescales where climate-driven risks are most consequential yet least understood.

Floods don't happen in isolation: The research demonstrates that flood risk is "correlated in space and time," meaning when one location floods, nearby areas are more likely to flood simultaneously. This spatial clustering undermines traditional insurance diversification strategies that assume risks are independent across different locations.

AI can make climate risk more interpretable: Rather than treating the model as a "black box," the researchers used explainable AI techniques to link simulated flood patterns to specific large-scale climate drivers. This transparency helps stakeholders understand why floods cluster together, not just that they do.

Better tools enable better financial planning: By generating realistic flood scenarios across multiple sites over 1-5 year periods, the framework allows insurance companies and financial institutions to stress-test their portfolios against plausible climate-driven risks that current tools cannot adequately represent.

Physics matters for AI credibility: The model grounds its predictions in physical climate processes rather than pure statistical patterns, making the simulated flood scenarios more trustworthy for high-stakes financial and policy decisions affecting insurance coverage and disaster preparedness.