3 Requirements for Fast-Tracking Machine Learning at Insurance Organizations
Machine learning is one of those buzzwords not everyone understands. It involves artificial intelligence (AI) … but what does that really mean in the context of an insurance company built on old IT networks and an even older set of ingrained processes?
“Machine learning is an approach to AI that gives a computer system the ability to obtain their own knowledge by extracting patterns from data instead of relying on hard-coded knowledge. In business, machine learning can help you predict what operational decisions will lead to an optimal outcome such as maximizing profitability based on the decisions you have made in the past. As more decisions are made and more data is collected, the algorithm continues to learn and make better predictions over time,” said Sean Naismith, Head of Analytics Services at Enova Decisions.
There are a multitude of operational decisions that feed the insurance process. What is an applicant’s exposure? What premium is competitive yet mitigates risk? Is a claim legitimate and compensable?
“At most insurance companies, decision-making is manual . Machine learning enables humans to offload some of the more repetitive yet complex decisions or helps them make more informed decisions. ” Naismith said.
Applications of machine learning include more precise risk classification and pricing, earlier fraud detection, and more efficient claims management.
But machine learning adoption among insurers remains slow. A lack of understanding is one challenge. Senior leaders can’t buy in without a clear use case and quantifiable ROI.
Lack of infrastructure is another. Not only do machine learning models need high availability deployment, but models also need access to clean, multidimensional data in real-time.
Lack of transparency is a third challenge. Without a means for explaining how machine learning models are making predictions, insurers are putting themselves at risk of being in noncompliance (e.g. discriminating against protected classes)
All of these challenges can be overcome, however, with the right planning. Here’s how insurers can fast-track machine learning at their organizations:
1. Build a use case around a specific decision.
“Most businesses make the mistake of starting with data when they should be starting with decisions,” Naismith said. “Prioritize those that have the greatest impact on business objectives.”
For example, an insurer’s objective may be to increase net income by 20%. Decisions that drive net income include risk classification, coverage, and price.
“Once you’ve identified the decisions you want to optimize, you need to determine what questions need to be answered in order to make those decisions and what information you need to answer those questions,” Naismith said. “Then, you’ll know exactly which machine learning models you need to build and what data you will need to train them.”
Back to the earlier example, a decision flow could be built using machine learning models to predict the optimal price that would maximize probability of acceptance while balancing probability of exposure based on historical performance to achieve the goal of increasing net income by 20%.
Those that define a specific use case and success metrics around business objectives will have an easier time getting executive buy-in for investment.
2. Leverage cloud computing to overcome infrastructure challenges.
“Implementing machine learning is complicated. There are many great analytics platforms out there that enable data scientists to build machine learning models. However, most IT teams aren’t equipped with the right tools and skills to get a machine learning model into production for real-time use,” Naismith said.
Even businesses that are more advanced in their digital transformation may only have the ability to automate simple business rules. Gartner’s 2019 survey (paywall) of insurance industry CIOs determined that many “still are building out the foundation for analytics and face legacy system challenges that will need to be overcome in the next couple of years.”
A commissioned study conducted by Forrester Consulting on behalf of Enova Decisions likewise found that although nearly 80 percent of business leaders believe automation of operational decisions is important, only 22 percent are very satisfied with their decisioning software today.
The time, resources, and cost required to build machine learning deployment capabilities is not feasible for many businesses. Not to mention what’s required for ongoing maintenance and security management. This is why Naismith believes partnering with a reputable cloud computing vendor is the sensible approach.
“If you leverage a cloud-based decision management software or platform as a service provider, you mitigate the need for additional infrastructure. The vendor will manage all those data connections, deployment of machine learning models with high availability, and ongoing IT maintenance,” Naismith said. “Plus, ‘as-a-service’ enables you to pay for what you need with minimal upfront cost.”
3. Create transparency around decision-making.
“Stakeholders will be hesitant to use machine learning if they don’t understand how decisions are being made, no matter how accurate the model’s predictions are,” Naismith. “There’s too much risk involved, especially in highly-regulated industries. For example, a model may be inadvertently discriminating against age, gender, or any other protected class.”
Here lies the paradox: How can businesses explain what’s happening in a black box? Fortunately, there are new techniques like SHapley Additive exPlanation (SHAP) being developed to help data scientists interpret the predictions being made by machine learning models. Explainable machine learning is a highly active area of research. As progress is made, insurers will need to be less concerned with balancing complexity and presumably accuracy with transparency. Regardless whether the machine learning model is built in-house or purchased, insurance companies must ensure that explanations are being captured alongside every decision. In addition, models deteriorate over time. Therefore ongoing monitoring and management is necessary so no bias creeps in. One way to do this is to routinely compare model output of protected classes to the rest of the population., Naismith believes. “Significant variations in the prediction distribution of your customers of a particular subclass from that of the general population should trigger an investigation of your model.”
Fast-Track Machine Learning with Enova Decisions
Enova Decisions is a leading analytics and decision management technology company that’s been “eating its own dog food” for over 14 years. “We were created to make the analytics and real-time decisioning expertise that has powered our parent company accessible to businesses like insurance companies,” said Naismith. Enova Decisions is part of Enova International, a publicly traded financial services company (NYSE:ENVA) that has extended over 22 billion dollars in credit online to over 5 million customers worldwide.
“We take a decision-first approach to analytics,” Naismith said. “We use decision model and notation (DMN) to help our clients conceptually understand how they’re currently making decisions and where opportunities for improvement are. We have a team of decision scientists that can tailor machine learning models in a regulatory compliant manner for the defined specific use cases. Those models can then be deployed directly from our cloud-based platform. In a nutshell, our decision management software and services are designed to manage the complexity of decision automation and optimization so that our clients can focus on servicing their customers well.”
To learn more, visit https://www.enovadecisions.com/insurance/.
This article was produced by the R&I Brand Studio, a unit of the advertising department of Risk & Insurance, in collaboration with Enova Decisions. The editorial staff of Risk & Insurance had no role in its preparation.