How Predictive Analytics is Transforming Insurance Industry
By Jon Holtan, CEO, SpareBank 1 Skadeforsikring
Different machine learning algorithms excel at different tasks. Some algorithms can be successful with little data, others require more. Some algorithms need pre-processed and cleaned data sets; others can work directly on raw data and still extract valuable insights. While machine learning models may not be new to the field of insurance, the developments in recent years in available data, computational power, and algorithms may soon change where models are applied, who is applying them, and pave the way for innovations such as fully automated real-time pricing and underwriting, and parametric private market insurance. The main drivers for the increased importance of machine learning in insurance can be summed up in three parts:
I. Increased adoption of digital channels in every aspect of our lives, from work and banking to shopping and entertainment, where data is created and captured in every interaction. Processes in these channels, being “digital-native”, are naturally suited for machine learning based automation, and regulations such as the EU’s GDPR and PSD2 ensures that consumers may make their data available to third parties if they so wish.
Traditionally most of the data used for predictive modeling in insurance comes from the day-to-day handling of the insurance portfolio
II. Data gathering sensors are appearing in more and more of our material goods; Apps on our cell phones and smart watches measure our physical activity and movement patterns, cars log driving behavior, and appliance producers are racing to add Internet-of-Things functionality to everything from refrigerators to juice presses. Third parties can get valuable insights from such data through partnerships with app makers and producers.
III. Cheaper computational resources and improvements in the machine learning algorithms themselves are opening up new, more unstructured data formats for analysis, such as images, voice recordings, chat logs, emails, and other freeform text. Established companies may find treasure troves of “new” data in old archives of data that was once thought too messy to be of use for analysis.
This threatens to change the long-standing “data asymmetry” of insurance, where incumbents have their large portfolio data sets available for risk analysis and modeling, while new challengers have to make do with less: A startup insurance challenger may soon supplement their limited portfolio data with data from potential customers’ transaction histories or even social media behavior (given, of course, the customers’ informed consents). Car manufacturers may search for patterns in their growing sensor data sets and use these to price motor insurance, including insurance in their list of offered new car add-ons. For many such challengers it would make sense to partner with incumbent insurers for the backend processes such as capital planning and optimization, regulatory reporting and in some cases claims handling. Indeed, it is possible to envision a future where incumbent insurance companies, while not going out of business, are reduced to producing a backend “insurance handling” commodity as efficiently as possible, leaving most of the insurance pricing and sales to manufacturers and retailers of the products and services to be insured.
On the other hand, incumbent insurers may be better placed than many others to embrace the opportunities offered by new data sources and new machine learning algorithms: Many of the most interesting new data sources depend on the consent of the customer. Customer trust in incumbent insurance brands may make it more acceptable for consumers to share their activity-, transaction-, and other personal data for analysis. Incumbents may also have an advantage in the fight for machine learning talent: Within the existing teams of actuaries, analysts, software developers and architects there are often advanced machine learning enthusiasts to be found. These key resources can be utilized when building the competencies, infrastructure, development, and operations methodologies required to effectively develop and deploy more advanced machine learning models and services.
Finally, let us not forget that the incumbents’ large insurance portfolio data sets still contain very valuable information: New data and new algorithms may enable challengers to build deep insights in their customers’ general behavior and object use, but to apply this to insurance pricing they have to translate this insight into estimates of insurance risk. And for now nothing can beat the incumbent insurers’ large historical claim data sets for the task of translating customer and object data into insurance risk estimates. New machine learning algorithms and data sources may disrupt the insurance business in the years to come, but don’t write off the incumbents just yet.