Using machine learning to assess risks for insurance policies
Case 13/16 of applied Artificial Intelligence
The idea that you can quickly find out exactly what the risk associated with a new policyholder is will be music to the ears of many an insurance company. At present, this type of risk assessment is still largely carried out with the aid of labour-intensive models, and it often costs a great deal of time to deliver a new risk model. Deloitte is working with new technologies that are helping insurers to make assessments with greater speed and accuracy.
Risk assessments are predominantly made on the basis of personal and objective characteristics. “If someone drives a large car, it is more likely that they will cause greater damage. If someone has a thatched roof, fire damage will be more severe, on average,” explains Jurjen Boog, Manager in Financial Risk Management at Deloitte.
Since the 1990s, insurers have been using statistically based Generalised Linear Models (GLMs) for these types of assessments. The models are developed by actuaries with many years of expertise and experience. Now that machine learning technologies are on the rise, it raises the question of whether intelligent algorithms are able to make even more accurate assessments.
“With machine learning, an algorithm makes a risk assessment based on pre-determined criteria, rather than estimating parameters for statistical models,” says Boog. “A conventional GLM can take account of the interaction between two, but no more than three variables, such as the relationship between a policyholder’s age and sex. Machine learning, on the other hand, can ‘understand’ thousands of variables and much deeper interactions.”
AI email alert
Get 16 AI cases in your emailAI email alert
Deloitte’s Financial Risk Management team compared the risk assessment of machine learning with that of a GLM. “We looked at a car insurance policy, and then specifically at third-party liability insurance,” explains Boog. “This is a component with a relatively small data set, which enabled us to explain exactly how the two predictions would differ.”
Surprisingly enough, machine learning and the conventional model generated predictions of comparable quality. “That’s not so strange; there aren’t very many deeper interactions that machine learning can take into account in such a narrow data set”, reasons Boog. What machine learning was able to do, however, is establish valuable new relationships. “This enabled us to map out clusters of policyholders with a higher risk of losses,” says Boog. “This knowledge can be used to manage the portfolio, such as by adjusting pricing or acceptance.”
More effective assessment
The team will soon be starting work using a broader data set. “Then we will be looking at combined policies,” explains Boog. “If someone combines car insurance with fire insurance, can we demonstrate that this is less risky than two separate insurance policies with exactly the same risk factors? And if so, how much less? By adding more interactions, we expect to achieve more accurate predictions for things like this using machine learning.”
Will machine learning radically alter the insurance system? “It will mainly result in more robust substantiation of decisions,” asserts Boog. “Insurance is about weighing things up. You want to know the risk posed by every potential new policyholder, and a more accurate risk model enables you to measure what the impact of particular choices will be.”