It’s all about the data.
That’s the message I kept hearing at the CAS Ratemaking and Project Management conference in San Diego this spring.
The message was so strong, that I was asked to present my seminar, “Multi-Product Optimization: Challenges and Opportunities” not once, but twice along with LaCapitale’s Bruno Tremblay at the annual conference. As a Principal Consultant, Insurance North America at Earnix, a Tel Aviv-based provider of predictive analytics solutions for the financial services industry, I spend a lot of time focused on the data – specifically customer behavioral data.
Insurers are currently focused on pulling the pricing level, but there’s a whole other series of other ways to influence a customer’s behavior. The starting point is to understand your customer’s experience – from when a consumer first begins shopping for coverage to eventually terminating the policy. Each company may have a slightly different customer journey, but there is always a beginning and end based with other customer actions along the way. As I told the CAS RPM conference attendees, each of these decisions can be represented by a statistical model. When we string together all of these behaviors we can really start to see the bigger picture of how our customer engage with our company.
Unfortunately, while many industry insiders and experts like to talk about multi-product optimization, nobody’s really doing it. That’s because the biggest hurdle standing in the way is the data. However, a lack of data should not be an excuse for inaction; we need to stop making business decisions based on judgments and instead use science.
As the science of statistical modeling continues to evolve, new techniques are being made available to simplify the statistical modeling process. Machine learning focuses on a series of statistical modeling techniques, many of which are geared toward predictive modeling. This modeling relies on computer programing to develop, test and refine algorithms with the least amount of human intervention. While the algorithms themselves might not be very simple, the overall application of ML is relatively straight forward.
There is enormous potential for North American insurers when it comes to Machine Learning, according to a recent Earnix survey of nearly 200 global insurance professionals from companies that provide personal lines coverage. Our online survey found that only 4% of respondents from North America had adopted ML as a core strategy, compared with over 20% across the rest of the world.
The goal of the survey, “Machine Learning – Growing, Promising, Challenging,” is to help insurance executives understand how companies across the industry are currently leveraging Machine Learning and how they expect it to impact the industry in the near future. The survey found that Machine Learning is currently being employed by more than half (54%) of respondents.
When asked about the main benefits their organizations have realized from Machine Learning, 57% said “greater analytical accuracy” as one of the top three. Respondents also reported that the most significant promise and expected benefits of Machine Learning are “greater automation, productivity and cost savings.”
Most companies (70%) that use Machine Learning are using it to build risk models, according to the survey. The next biggest response was building demand models (45%) and fraud detection (36%).
Barriers to Break
Unfortunately, there are barriers preventing many insurance firms from achieving the full potential of Machine Learning. The biggest barrier is the knowledge gap. Despite growing adoption, 82% of respondents in the survey described themselves as still beginners (29%) or having intermediate knowledge (52%) when it comes to Machine Learning. Larger companies are further along in the learning curve, according to the survey, with 22% describing themselves as beginners versus 33% of smaller firms.
Another barrier to greater use of machine learning is the lack of an experienced hiring pool. It is extremely difficult to attract analytic talent to insurance companies. As you may have already realized, in today hiring market, you will be lucky to find one or two real data scientists after a prolonged search process. One of the reasons for this is due to the newness of these techniques very few universities in the United States are actually teaching courses relevant to Machine Learning.
Your best bet might actually be to hire and train talent right out of college. It will be a slow process, but a better investment for your firm. In addition, I have found that employee turnover is much lower for data scientists who are hired and trained straight from graduating from university.
Talent and techniques are only two pieces in this analytics puzzle. Strategically, companies need to move away from traditional methods of managing products and focus more attention on managing customers – specifically, understanding the drivers of customer behavior. The good news is that companies are realizing that they need to work on their data to complete this analysis. Unfortunately, no one ever said modeling human behavior was easy. Especially when we start considering decisions that happen when multiple products are considered.
With a better understanding of customer behavior, we will be better able to service our customers through more informed pricing decisions and product design. Some companies are well underway in their customer centric pricing journey. With the continually raising frequency and severity, it looks like more rate increases are ahead for many companies. Using science to adapt to these trends will always yield better results than going with your gut.
It all starts with the data.