Latest posts by Justin Lyon (see all)
The goal of this project was to produce a high-resolution agent based model (ABM) suitable for the assessment of life and health insurance portfolio risk under various possible flu pandemic scenarios.
The model used company data provided to us by an ‘Insurance Company Client’ (ICC) and linked with U.S. census data. ICC was an industry leader in group and individual income protection and related coverages and services, insuring more than 22 million people. At the time of the project, ICC reported revenue of more than $10.5 billion which included $7.95 billion in premium income.
The disease was modelled using agent based modelling techniques. Agent mortality and infection susceptibility characteristics in the agent based model (ABM) were determined by a set of configurable risk factors initialised from US census data and ICC data.
As the model ran, it generated time-series data describing the number of infections in various regions (at a zip code and individual level). That is, you could say that Person A at Home Y was infected and died, but Person B at Home X recovered.
This data could then be overlaid with insurance portfolio distribution data to determine how their insurance portfolio compared with the general population. This comparison was performed both at the level of static risk assessment as well as a risk assessment based on the likely progression of the disease through the population over time.
As shown in the image below, in the black square, we plotted each individual as a dot, coloured according to their disease state. In the Disease Evolution graph, we showed the population numbers for those people who were Susceptible, Exposed, Infected and Recovered (SEIR). In the Population State chart, we provided bar chart comparisons of the number of people in each disease state. It is also possible to see the age distribution of this particular region, as we knew that the age is an influential factor in the disease progression.
The main simulation interface was called the Insurance Risk Assessment panel and is shown in the picture below.
On the top left of the panel we showed data related to ICC's insured population: the total number of people insured (1,994,942), the number of zipcodes they live in (16, 264), the percentage of life vs standard insured and the age distribution for ICC's customers.
On the top right, we included controls to manage some parameters of the disease model like infection probability for the exposed and infectious individual, the incubation and infection duration times, social network parameter like number of contacts per person and also a value representing the likelihood that a person is connected to one of its neighbours. The bigger this value the less mixing there is in the population. This enabled them to quickly and easily adjust parameters when conducting 'what-if' experiments.
Clicking the “Run the model and switch to Main View” button started the simulation by initializing the population. When the simulation run ends (meaning the disease has run its course), the risk to ICC was assessed.