In March, as California was still in the middle of its rainy season, a startup called Kettle was using machine learning to predict where wildfires might happen later in the year, mapped out in half-square-mile resolution across the state. Out of thousands of larger 7-by-7 mile grids, the areas that the model predicted were most at risk of burning—the top 10%—11 out of 14 major fires later happened. (All 14 occurred in the areas that the model said were in the top 20% for risk.)
The company, which announced a $4.71 million round of funding on October 20, aims to help the insurance industry grapple with the increasing disasters caused by climate change. Cofounder Nathaniel Manning, a software engineer who worked in the Obama-era White House on open data for humanitarian response and later ran Ushahidi, a software platform for crisis response, had seen, firsthand, how critical insurance was to recovery. “I just had this realization that insurance was really this very beautiful and elegant solution to protecting people from crises,” he says. “I got obsessed, but I hadn’t worked in insurance, I worked in tech.” He partnered with Andrew Engler, who had worked in the reinsurance industry, to begin to work on the challenge: How can insurance companies survive and continue to offer homeowners coverage as disasters like wildfires come with larger and larger costs?
Reinsurance companies, which insure insurance companies, have been fleeing California over the last few years, though the state has prevented insurance companies themselves from dropping policyholders. But the costs of insurance have been rising. New homeowners also have more trouble getting policies, which may make it more difficult to get a mortgage. The value of some houses may fall. By better predicting where wildfires may get out of control, the startup believes it can set the right prices for insurance companies and stabilize the system.
The current models for prediction are outdated, Engler says. “The industry has been using the same tool set, stochastic modeling for over 120 years,” he says. “If you’re using stochastic modeling, you’re basically saying, I’m going to take the past 500 years, and I’m going to say how many times did a wildfire hit Los Angeles? If it happened twice, now I’m going to price all of my next year’s policies and everything for the next 10 years off of the probability that there’s a one in 250 chance of a wildfire hitting Los Angeles. And that is just no longer true. We’ve seen a complete nonlinear increase in the severity and the frequency of these events.”
Two changes make better tech possible now. Massive amounts of data are available, and extreme computing power is cheap enough for a startup like Kettle to use. The company’s algorithms use more than 7 billion lines of data from satellites, weather, and real-time data from fires, running three tredecillion (3 x 1042) calculations per run, and use an approach called particle swarm optimization that is more commonly used in robotics.
Since the insurance industry writes policies that last one year before they’re updated or renewed, the company makes predictions a year into the future. Kettle runs simulations biweekly and wants to open up that data to others—like fire agencies and utility companies—that can use it to take action. “The thing about insurance is that we’re actually all on the same side,” he says. “Everyone’s incentive is aligned. So it is very much in our incentive to give away all of our learnings and our data . . . Instead of just saying Angeles National Forest is [at high risk of fire] today, what you need to know is there’s a 250-square-foot patch underneath this 50-year-old power line, and if you see winds getting above 20 miles an hour, this is the right place to shut it down early.” The results could also be used to decide where an area is unsafe to build, or the government should encourage residents to move.
Similar models could also be used for hurricanes, flooding, and other disasters. While the current model of insurance isn’t sustainable as disasters grow—with a sharp increase over the last 20 years, with 7,000 extreme weather events globally—Engler believes that the industry can adapt. “When you look at California, about 12% of the homes are actually built in the middle of forests or dangerous areas,” he says. “You’re talking about 12% driving 90% of that risk rate. That’s really where you can see a manageable effect. It’s definitely not sustainable in the current model, and the paradigm that has been set up. But it is completely manageable in the future.”