➤ Insurers suffered more than $8 billion in losses from the 2020 wildfire season in the western U.S. With 2021 shaping up to be another year for hefty fire losses, insurance carriers, especially those doing business in California, are searching for ways to keep those losses under control.

➤ Zesty.ai's Z-FIRE wildfire model, which combines property details and actual loss data with machine learning to produce a predictive risk score, has caught on with some insurers in the Golden State. The model has been accepted by California's insurance regulator and is being used by multiple carriers.

Attila Toth, co-founder and CEO of Zesty.ai, talked with S&P Global Market Intelligence about the development of Z-FIRE, how it differs from more traditional simulation models and what it could mean for the future of the business.

The following conversation has been edited for length and clarity.

S&P Global Market Intelligence: Why are companies turning to Z-FIRE for their risk modeling?

Zesty.ai CEO Attila Toth: Those old-school models are very often 30 years old. They are not looking at the risks for individual properties, and they are not leveraging current data. That is why we decided to build Z-FIRE, a new model leveraging a lot of loss history and property-specific risk modifiers. It includes everything from vegetation density in multiple defensible zones, roof material, building material, how densely neighborhoods are populated and what slope houses are built on. There are a lot of factors that our research and our research partners, including the California Department of Forestry and Fire Protection and the Institute for Business and Home Safety, have determined as individual risk factors.

What makes Z-FIRE unique?

We are not just running simulations on future losses and then altering them. We have actually harvested loss data from 1,400 historical events. That is a very unique factor that, with the advent of artificial intelligence and with satellite imagery, we can do today. With AI-enabled models, you can harvest a lot of loss history and you can establish links between losses and property-specific features. You can prove that there is a very strong correlation between losses and those features.

What requirements did California's regulator want you to meet before it approved Z-FIRE for wildfire modeling?

The department hired an independent actuarial consulting firm to do a stress test. They wanted to determine that the model does not use any data that would unfairly discriminate. They also needed to see if the model performs the task that it is set out to perform. There was also a transparency requirement from the insurance commissioner that the model was explainable; an insurance agent needs to be able to tell an insured why he or she is determined to be at higher risk and what could be done to mitigate that risk.

What do you think was the deciding factor for the regulator?

The fact that this is not a simulation model and because there is a lot of loss history that's being observed in this model. The California FAIR Plan commissioned Milliman to do an actuarial study where it looked at 250,000 properties in the FAIR Plan over the last five years with its incumbent method of risk assessment, then used the [Z-FIRE] AI-enabled model and found that it outperformed the incumbent model.

Has your model changed how carriers write insurance in California, particularly in areas deemed high risk?

Based on the "old school" way of looking at wildfire, close to 20% of California's homes are in danger zones and therefore should lose insurance coverage. Our view is that it's not as bad as that because we can create models with surgical precision. When you can understand what the house is built from, what the individual condition of the house is, we believe that maybe a mid-single-digit percentage of homes are at high risk. Do they require some type of treatment? Yes. Do they require mitigation? Yes. They are in high danger zones, but we're not talking about 20% of California homes.

The FAIR Plan said it is going with us because we pinpoint the risks at the individual property level.