Why regulation holds auto insurers back from using advanced modelling

Auto insurers could potentially use artificial intelligence to improve risk segmentation but some of the newer modelling techniques may not be a good fit with the way rates are currently regulated, a data science expert for one insurer suggests.

Ontario’s Financial Services Regulatory Authority announced July 21 it is creating a Technical Advisory Committee for Auto Insurance Data and Analytics Strategy.

This committee is a “great initiative,” said Baiju Devani, chief data officer and senior vice president of data science for Aviva Canada, in an interview with Canadian Underwriter.

“There are a number of ways the regulators can help insurers take greater advantage of data and analytics in improving risk segmentation and pricing accuracy in auto insurance. One is having flexibility in accepting different modelling techniques when it comes to risk segmentation,” Devani said.

Most Canadian provinces have a rate filing process. In Ontario, insurers must submit proposed rate changes to FSRA for approval. Among the rate filing processes is a requirement for the insurer to submit its current and proposed rating algorithm and highlight the changes. With its new standard rate riling process introduced in 2019, FSRA has 25 days to respond if it rejects a rate filing.

One reason FSRA is setting up a technical advisory committee is to get advice on regulatory implications of using AI and big data analytics in the auto insurance system to protect consumer interest while promoting market innovation.

“There are new predictive modelling techniques under the umbrella of artificial intelligence and machine learning,” Devani said. “These new techniques can allow more sophistication in the risk segmentation process and improve accuracy in terms of charging the right price to the right customers. However, these newer modelling techniques do not necessarily fit in well with the traditional auto rate filing process.”

This, Devani says, is because if insurers were to try to use some of these new modelling techniques with the existing rate filing process, this would make the process more onerous.

Take decision trees for example. A simple decision tree might ask if the motorist is older than 40. If the answer is yes, then you go to the left of the tree. If the motorist is under 40 then you go to the right of the tree, explained Devani. The next steps in moving down the tree involve applying different criteria.

“A lot of the new predictive modelling methods would generate hundreds of decision trees. But the existing [auto rate filing] models tend to generate coefficients that you can put in a filing. The regulator staff can look at those coefficients and say whether or not they agree. If we were to use decision trees instead of models that produce coefficients, we would have to figure out how we use those decision trees to file proposed rate changes.”

A lot of people assume these new modelling techniques are black box models, but Devani says this is not the case.

“I think these models can be explained, but it just requires a different way for us to communicate and express those kinds of models. That is where the regulators can play a part in terms of considering changes to the rate filing process so that they can accommodate more advanced methodologies,” Devani said.

“Some regulators are open to new modelling techniques but this requires an investment on the part of the regulators. For example, some regulators may need to bring in new technologies as well as new skills and talent into their organization, so they can work with the insurers and get comfortable with these new modelling techniques.”

Feature image via iStock.com/hakule