Human vs. Machine – Which One Is Changing Insurance?
November 15, 2016| Von Guizhou Hu | L/H General Industry | English
According to Ingenin’s 2015 Insurance Innovation Survey, 65% of insurance executives believe new technology will disrupt the insurance industry over the next five years. Among the many disruptors making the scene, it appears predictive analytics has the greatest potential to reshape our business.
Multiple industries already use predictive analytics to lead various technology innovations; one example is Google’s self-driving car. Within insurance, predictive analytics promises to improve accuracy and efficiency across most aspects of the business – from product design to pricing, and from risk management to marketing, distribution, and customer service.
To better understand its influence on insurance, let’s define predictive analytics. “Traditional statistical modeling” is the classic definition of predictive analytics. There is also machine learning, which is an emerging predictive analytics technique that has grown in popularity due to the availability of big data, cheaper computers and the advances in mathematic algorithms. Both techniques have their purposes.
First let’s talk about the differences in these two techniques. The classic “human vs. machine” analogy comes to mind. Traditional statistical modeling, typically performed by a statistician trained on statistical theorem, explores data relationships not intuitively obvious to the human brain. Machine learning is typically performed by a data scientist trained on computer engineering and data management, and uses data to train a model with minimal human intervention, many times with a goal of achieving pieces of human brain function such as image/voice recognition and natural language processing, or artificial intelligence. There are applications for both techniques depending on the purpose.
Traditional Statistical Modeling
A statistical theorem is based on assumptions about distribution and probability. It defines association as anything that disproves a null hypothesis of random sampling error with 95% confidence. A traditional statistical model is usually derived from a carefully designed population study with well-defined variables or predictors.
An example of a traditional statistical model used in life insurance underwriting, especially in preferred risk classification, is the Framingham coronary heart disease model. This model predicts the likelihood of having a heart attack in 10 years by evaluating blood pressure, blood lipid, smoking status, etc. Another example is the generalized linear model (GLM), the foundation of traditional statistical models. Actuaries use GLM when conducting experience studies to improve the efficiency of data usage and to smooth the actuarial curves and tables.
One of the most publicized developments in machine learning is the advancement in artificial intelligence. IBM Watson technology is one familiar example. Originally well-known as a winner of the American television show Jeopardy!, Watson is a process that began with natural language processing, which includes voice recognition and interpretation of text based on grammar and context. It evolved into cognitive computing, which empowers machine capability to learn. Insurance applications of Watson technology are currently being explored and could include automatic questions and answers during the complex insurance sale process, as well as analysis of unstructured data during underwriting.
Another common feature of machine learning techniques is “models ensemble,” which combines multiple models and turns a weak learner into a strong learner. For example, random forest, a popular machine learning technique, is comprised of multiple individual decision trees by which the majority of votes produces the final prediction. By applying a similar principle at Gen Re, we found an ensemble of multiple underwriting models, developed in the past through traditional statistical modeling or machine learning, can generate a new model that significantly outperforms any individual models previously developed. As a result, we’ve offered this new ensemble underwriting model to our clients.
Embracing the Power of Predictive Analytics
Clearly, predictive analytics has significant potential to bring powerful change to insurance. To more deeply embrace it, insurers need a master data management strategy, advanced analytic platforms, and a center of excellence to process the analytics and then test and validate the performance of various models. (See my earlier blog on this important topic.)
We’ve been forming an advanced analytical team with expertise in data, analytics and applications for insurance. This team’s dual goal is to better evaluate risk, and to provide a resource for our clients’ own analytics endeavors.