The answer is blowing in the wind: Harnessing real-time data to predict storm damage

By Glenn Fung, American Family Insurance data science research director

When catastrophic storms strike, American Family Insurance deploys claims teams as quickly as possible to assess damage and help impacted customers begin to rebuild their lives. We’re constantly monitoring the weather to help us anticipate areas that might be hit with severe storms and plan our claims response accordingly.

For example, we use radar detection and other technologies to forecast hail-related storm damage. This helps us pinpoint the geographic areas with the greatest needs and determine the proper number of claims adjusters, support personnel and equipment to dispatch.

Real-time situational awareness allows us to make sound estimates about the size and scope of a severe weather event. It also leads to better outcomes for our customers.

Technology has come a long way in helping us more accurately forecast hail damage before our claims teams are deployed — and because hail causes extensive — and expensive — damage every year, the sooner we can respond, the better it is for our customers and our company.

Wind is also highly destructive to our customers’ property, but our ability to estimate wind damage has historically been more limited.

We recognized there’s an opportunity to improve the measurement of surface winds in near real-time. That’s why we’re partnering with Dr. Michael C. Morgan, a professor in the Department of Atmospheric and Oceanic Sciences at the University of Wisconsin-Madison, along with researcher Brett Hoover, from the same department, to make it part of the overall weather patterns we monitor.

Specifically, we are exploring the use of updated numeric weather prediction output in conjunction with traditional measurement techniques — a first in the insurance industry — to generate a framework that will provide improved understanding of the surface wind field in near real-time, and more importantly, to be used as the foundation for a machine-learning methodology for forecasting insurance claims.

Training machines to predict the impact of severe weather

The common practice throughout the insurance industry is to glean information from disparate and sparse weather sensor instruments that measure and record wind velocity. We think, through a combination of tools, including machine learning, that we can more effectively measure wind as part of our overall weather monitoring.

Our goal was to use various types of weather data as the foundation for a machine-learning methodology that would forecast the number and severity of storm-related insurance claims.

American Family currently uses an app by Opterrix, an American Family spinoff, to quantify and visually show the potential customer impact of severe weather events.

Here’s how we are making our forecast model more predictive: Multiple sources of weather-related data such as radar-derived estimates of hail size, damaging wind speed swaths and crowd-sourced storm reports are fed into Opterrix. That information is then fused with our policy and claims data to provide a comprehensive view of potential impact immediately following a storm.

In addition, weather forecast models generated in Amazon Web Service’s cloud environment provide a 48-hour forecast four times a day. Each forecast model is initialized with updated weather observations from NOAA that provide predictions of multiple weather variables including precipitation, wind, temperature and pressure.

The output of these forecast models is then fed into Opterrix where several spatial algorithms will be used to simulate potential severe weather impact to our portfolio.

Having control of our own weather forecast model provides the opportunity to generate reanalysis data that can train machine learning algorithms to accurately predict potential severe weather outcomes. The process of reanalysis includes gathering archived weather forecasts for significant severe weather events and comparing it to our historical claims data from those events. Adding wind also gives us a fuller picture of the damaging storms.

By fusing historic claim and payment data with historic forecast models, we will be able to more accurately predict how future severe weather will impact our customers and our business and make better decisions based on this information.

Glenn Fung is a data science and artificial intelligence (AI) expert with American Family Insurance. His main interests are optimization approaches to machine learning and AI, with emphasis in kernel methods. For more than a decade he has worked in industry — including at Siemens, Amazon and American Family Insurance — developing and applying novel machine learning techniques to solve challenging industry-related problems.