Compound Risk & Infrastructure

Insurance, reinsurance, and finance sectors have seen mounting losses from convective storms due to inadequate risk models that fail to capture compounding hazards. Here, we will develop a compound risk assessment methodology to reduce uncertainties in risk estimates for housing infrastructure subjected to convective storms. To accomplish that, we will develop a compound risk framework to assess housing vulnerability under convective storms, implement computational models to analyze the effects of wind, hail, and rain interactions on structural performance, and derive compound fragility functions for wood house archetypes across progressive damage states. 

Team: Gonzalez-Dueñas (GMU)

This research will enhance and sharpen a data-visualization platform - the System for the Triage of Risks from Environmental and Socioeconomic Stressors (STRESS) - which provides an array of metrics designed such that they can be consistently combined according to user-specified priorities and interests. In this way, STRESS can inform researchers, stakeholders, and citizen scientists on the co-evolving challenges across landscapes of the natural, managed, and built environmental systems as well as social and demographic gradients. The work will extend and amplify STRESS’ to represent risks that may be directly compounding (drought and extreme heat) and/or impose “whiplash” extreme impacts (drought year followed by severe flooding).  Several case studies can be constructed according to the CCRA IAB interests and priorities. 

Team:  Schlosser (MIT), Morris (MIT), Gurgel (MIT), Fiore (MIT), and Paltsev (MIT)

Generative AI (GenAI) is becoming pervasive, yet full adoption faces numerous barriers in the finance and insurance industry. Businesses do not want to place their proprietary data into another company's GenAI platform (e.g., ChatGPT). Additionally, pre-trained Large Language Models (LLMs) are impressive but not specifically optimized for producing corporate risk analytics. This research project tests reward functions on synthetic datasets to subject pre-trained LLMs to additional reinforcement learning activities, specifically for developing enhanced risk analytics. The R&D output is a set of software development toolkits enabling companies to deploy these enhanced LLMs internally on their proprietary, commercially sensitive data.

Team: Oughton (GMU)