The relative movement of stars and planets is well understood today. Before Copernicus, scientists believed Earth was the center of the universe. Copernicus theorized heliocentrism, but it took another century before Galileo and his telescope (a better tool) would gather data to justify the theory. Copernicus was right, but it was the presence of better tools that allowed for more accurate forecasting.
Today we know that Halley’s Comet will be visible from Earth on July 28, 2061. Such accuracy was always there. Planets and stars have always moved in the same predictable ways. We understand this well today with tools much better than Galileo’s.
So it is with other aspects of the natural and even the manmade world. In life insurance, we’ve been forecasting outcomes for as long as we’ve been pricing and underwriting risk. Fortunately, with the improvement of tools for gathering and analyzing data continuing, we will be able to forecast outcomes more accurately.
Introduction to predictive modeling
We will generally define predictive modeling (PM) as using data to predict an outcome or calculating the likelihood of a future event. PM includes traditional inferential techniques which actuaries and statisticians have mastered.
New tools are available to solve the same sorts of problems, and more are created each day. We will refer to these new tools generally as machine learning (ML) but not limit the idea. This also invokes the terms artificial intelligence and data science.
With high volumes of data, ML is better at making predictions than earlier tools, much as a modern telescope is superior to Galileo’s refractive lens. Selecting the right tool for the job is a critical component to any successful ML solution. Artificial intelligence, for example, is just one type of tool, and it may or may not be the best choice.
The future of PM in life insurance
Predictive models using machine learning can and will be used to help optimize every aspect of the insurance value chain over the next five years:
- Targeted distribution and risk selection
- Underwriting (traditional, automated, hybrid, simplified…)
- Actuarial assumption setting and calibration
- Increasingly individualized pricing
- In-force management and retention
- Claims processing and fraud detection
- Overall health and wellbeing through wearables and continuous data
Machine learning is here to stay, and in order to use it effectively it is critical to combine the talents of specialists from outside the industry with those inside. Resistance and/or conservatism from both sides has slowed the industry’s response to the science.
The ability to accurately identify the risk of a given individual is the foundation of a successful life insurance company. To do this the current underwriting paradigm needs to change. The cornerstone of challenging traditional underwriting rests in leveraging mortality feedback. As the number of deaths increases, more and more features can be evaluated, and their predictive power can be compared to traditional evidence.
SCOR has formed a cross functional team of underwriters, actuaries and modelers with a variety of backgrounds including actuarial, statistical or data science. These experts know our industry and are embracing the changing landscape to create a new underwriting process.
Intersection of tools and data
Over the last decade the growth of open-source programming languages and platforms coupled with easy-to-obtain data from third-party data sources have created an environment in which modeling can thrive. Meanwhile cloud computing solutions have become cheaper, easier and more secure.
SCOR actuaries and data scientists have become skilled at using both the new tools and data sources. We will continue to develop these skills as we execute our strategic plan, which has a large focus on data and analytics. We use data from laboratories, the federal government, EHR aggregators, credit bureaus and continuously emerging non-traditional sources.
Barriers to adopting PM methods
- Proper change management is critical. Some individuals may perceive their skill set is less relevant and, therefore, resist change. In fact, the opposite is true as the methods used and lessons learned of older PM methods need to be fully integrated into a strategy going forward. Organizations need to work towards filling technical skills gaps of experienced employees while simultaneously providing business knowledge to resources coming from outside of the industry.
- Some methods will be considered “black boxes.” In fact, many tree-based methods lead to easily explained algorithms. Conversely, a coefficient-based model which is usually regarded as highly interpretable can be very misleading when covariates are correlated. Most importantly, new techniques can be used to provide very clear explanations of individual predictions. These techniques must be understood and embraced by regulators for PM to be successful.
Challenges to using ML and other non-traditional methods
ML carries some challenges. Many of these methods are completely free (open source) but still state of the science. While they are easy to use, the temptation of cargo cult science is always present, and we must be on guard against choosing easy over best.
- A forecast may not be accurate because it was not adequately represented in the training dataset. The analyst must know what the dataset represents and to whom it can be applied. These are not always cases of bad algorithms or bad techniques but can be due to non-representative data.
- Sometimes poorly performing models actually can advance the analysis. Things that are of low incidence and hard to forecast, e.g., the probability an individual will die during a certain term, may still be modeled. The resulting individual forecasts may be quite inaccurate but completely applicable to the analyses in larger swaths — a portfolio or subset or even the whole company’s data. Accepting an algorithm’s ability to provide analytical insights while struggling to make individual classifications requires progressive thinking.
- Different (sometimes very different) models can provide similar results. Because two very different resulting algorithms can make very similar predictions doesn’t discredit one or the other. Business considerations can be a differentiator in these situations, where one model may be less disruptive than the other.
- As the number of human decisions is reduced (not eliminated) the impact of a few individuals is amplified, with the benefit of offsetting errors going away. Actuaries must continue to define the guardrails for predictive models.
Predictive modeling already influences almost every aspect of the insurance process. The degree to which models influence decision making will continue to increase, and actuaries are well positioned to drive that trend in the direction that benefits both policyholders and business outcomes.
Through its partnerships, SCOR is continually evaluating a wide variety of data sources using advanced machine learning techniques executed by a cross-functional and business savvy innovation team. We are using this information to create a ML-driven underwriting framework which is flexible enough to incorporate new information at the same pace that it is made available.
While we know exactly where Halley’s comet will be 42 years from now, the future of underwriting in the life insurance industry is in some sense certain to be uncertain: the data used, the methods used and the degree to which ML is deployed will be entirely different from what it is today. Actuaries are extremely well positioned to adapt to and champion this change, but we need to come together with other experts to embrace a new way of thinking.