A Global Conversation
Model Behavior

“I think after two very challenging years in 2017 and 2018, from a natural catastrophe perspective, we may be reaching an inflection point in 2019, when it comes to capacity, rates, and terms and conditions.

In the insurance industry, we're very used to severe natural catastrophes, but there were some elements of the events over the last couple of years that, to a degree, caught the industry by surprise. By that, we're talking about in late last year, the severity of the California wildfires, which is a peril that's not overly well modeled by the modeling companies.

We had a combination of less than robust modeling data for the insurance companies to rely on, and then just an unprecedented level of severity for those California wildfires last year. The year before, with Hurricane Maria hitting Puerto Rico so hard, I don't think the cat models or those of us that are following the insurance industry ever really expected that level of business interruption. The complexity of those losses, and the fact that those reserves are proving inadequate, even a year-plus post the event, we've seen adverse reserve development from many players in the market on the primary and the reinsurance side. That's also had implications for alternative capacity. There's been some headlines about collateralized programs, where the collateral is not being released because of this problem with adverse development on losses that happened over a year ago. That's led some ILS fund managers, ILS investors to start thinking about how much capacity do I really want to put into insurance risk or insurance as an asset class. All of those factors coming together could have implications for the amount of capacity that's available on the reinsurance and the retro side.”

Stefan Holzberger
Senior Managing Director and Chief Rating Officer
AM Best

“In most typhoon models, we actually capture the heavy precipitation because, very often, it's actually the inland flooding from the precipitation that causes the most damage, sometimes more than the wind or even the storm surge.

Most catastrophe models capture that currently. What's more uncertain are the landslides and mud slides, because they don't always happen, necessarily, at the time of the typhoon. That's a little bit more complex to capture. At the moment, we're still trying to fully understand that correlation of peril, looking at that closely. We certainly see that there's a correlation when typhoons happen, that you do have in several instances. In the Philippines with Typhoon Mangkhut, we saw there were very large landslides that killed 156 people. It does happen, but the difficulties in modeling that and representing it for the industry is a little bit harder. We haven't got there yet with landslides, but I think in the future it may well be something we model, as well.”

Iain Willis
Managing Director
JBA Risk Management Singapore

“There actually is a lot of data [concerning cyber losses]. I think that's a misconception that people have, that it's a data-poor type of problem. There's surprisingly a lot of data out there. In fact, we have data on about 77,000 historical incidents to use to build a model. That's a lot of incidents. This is just over the past decade or so. Things are happening all the time. Because of how quickly things are happening, there's a lot of data flowing in. Not only do we have that data on historical incidents just about the incidents, but we get real cyber claims data. This allows us to calibrate the model to what cyber insurers are suffering today, as opposed to just looking at a pure cyber security-focused model. We're doing that for sure, but we take it that step further to build the cyber insurance aspect into the model.”

Scott Stransky
Assistant Vice President and Director of Emerging Risk Modeling
AIR Worldwide

“Insurers are exploring ways to use machine learning techniques for pricing. They are using machine learning techniques to look for ways to improve their generalized linear models which can be used for pricing. For example, to find new features or combinations of variables that could improve those models.

Secondly, they're using them to develop models that support pricing decisions, such as inspection models. We still see insurers being hesitant to use machine learning techniques for rating because they simply don't have a complete understanding of how they'll impact prices and there's that reputational risk.”

Ben Williams
Director, Insurance Consulting and Technology
Willis Towers Watson

“The insurance sector is not only a potential victim to the cyber threat attacks but also underwrites insurance. We have the cyber insurance exposure, we have the cyber insurance underwriting and accumulation of risks. There are question marks whether sufficient data and sufficient models exist right now in the market to actually support these exposures.”

Dimitris Zafeiris
Head of Risks & Financial Stability Department
European Insurance and Occupational Pensions Authority