TLDR: InsurAgent is a new AI system using large language models (LLMs) to simulate individual flood insurance purchase behavior. It combines empirical survey data with LLM common sense to accurately predict purchase probabilities, considering various factors like demographics, location, and personal experiences. Unlike general LLMs, InsurAgent can make quantitative predictions and model how decisions evolve over time, offering a powerful tool for policy analysis.
A new study introduces InsurAgent, an innovative system powered by large language models (LLMs) designed to simulate how individuals decide whether to purchase flood insurance. This research addresses the critical issue of low flood insurance participation rates among at-risk populations in the United States, despite flooding being one of the most devastating natural hazards globally.
The core idea behind InsurAgent is to bridge the gap between qualitative understanding and quantitative prediction in human behavior modeling. While previous studies have identified various economic, geographic, psychological, and socio-demographic factors influencing insurance decisions, traditional models often oversimplify complex human behavior. Large language models, known for their human-like intelligence, offer a promising avenue for more nuanced simulations.
The researchers first evaluated existing LLMs, such as the Llama-3.3 70B model, for their ability to understand and predict flood insurance purchase behavior. They found that while LLMs could qualitatively grasp the factors at play – for example, recognizing that older, more educated, or higher-income individuals are generally more likely to buy insurance – they struggled to provide accurate quantitative probabilities. These models often produced consistently high, generic probabilities (e.g., 80%) regardless of specific individual profiles, indicating a “knowledge-to-action gap.”
To overcome this limitation, InsurAgent was developed with a sophisticated five-module architecture: perception, retrieval, reasoning, action, and memory. This design mimics human cognitive processes, allowing for a structured, multi-stage decision-making approach.
How InsurAgent Works
The perception module first analyzes a user’s profile, extracting both standard factors (like age, income, education) and unique personal details. This information then goes to the retrieval module, which uses a technique called retrieval-augmented generation (RAG). This module queries a database of region-specific survey data, specifically from the U.S. Gulf Coast, to find population-level flood insurance purchase probabilities associated with the extracted factors. These empirical statistics serve as crucial reference points.
The reasoning module is the brain of InsurAgent. It takes the retrieved data and the user’s profile, then performs a first-person role-playing task. It organizes the survey data as quantitative benchmarks, weighs the influence of each factor, and then uses the LLM’s common sense to adjust the baseline probability based on unique contextual information not found in the survey data. For instance, it can consider a person’s specific city of residence, occupation, social environment, past flood experiences, or even their insurance claim history.
Finally, the action module generates the predicted flood insurance purchase probability as a percentage, along with a clear explanation of the reasoning. The memory module records all reasoning steps and final decisions with timestamps, creating an episodic history. This allows InsurAgent to simulate how an individual’s decisions might evolve over time in response to new life events or changing circumstances, providing a dynamic and realistic simulation.
Also Read:
- A Dual-Agent Framework for Aligning LLMs with Human Travel Behavior
- Urban-MAS: A New Approach to Understanding Cities with AI
Key Findings and Capabilities
InsurAgent demonstrated remarkable accuracy in aligning with empirical data. For individual factors, it precisely replicated the flood insurance purchase probabilities from the survey. When estimating probabilities based on two combined factors, it showed strong alignment with benchmark data, achieving a high correlation and low error rate. This indicates its ability to integrate multiple pieces of evidence effectively.
Beyond simply reproducing statistical trends, InsurAgent showcased its unique ability to extrapolate using the LLM’s common sense. It successfully incorporated contextual information that traditional regression models cannot handle. For example, it could predict varying purchase probabilities based on different residential cities (e.g., higher in Miami, lower in Phoenix), specific occupations (an insurance analyst vs. a chemistry researcher), social influences (proportion of insured peers), past flood experiences (catastrophic vs. minor), and even previous insurance claim histories (smooth vs. difficult). These extrapolations, while currently unvalidated by specific benchmarks, align with common sense and qualitative research findings.
The memory module also proved effective in modeling dynamic decision trajectories. The researchers simulated a “roller coaster” life story, where an individual’s flood insurance purchase probability changed over time in response to events like attending a coastal resilience workshop, reviewing historical flood records, witnessing a neighbor’s flood damage, learning about federal relief programs, reading climate reports, and experiencing negative claim processes. This demonstrates InsurAgent’s capacity for temporally consistent reasoning, reflecting how human beliefs and intentions evolve.
This groundbreaking work, detailed in the research paper “INSURAGENT: A Large Language Model-Empowered Agent for Simulating Individual Behavior in Purchasing Flood Insurance”, offers a valuable tool for behavioral modeling and policy analysis, potentially leading to better strategies for increasing flood insurance adoption and mitigating disaster-related losses.
