In this article series, we will explore how new behavioural data is generated and early signs of how this data could be the next wave of predictive power for issues like fraud, and maybe even claims costs. This month we’ll be hosting a live webinar where we will discuss these trends and review a software solution to capture this kind of data. To register, click here.
In our article last week, we showed how digital behavioural data can enhance the fraud models of insurers and lenders by boosting the ability of those models to identify inaccurate data. We also learned that such data has predictive power as an associative data element.
The answer an applicant submitted in an online questionnaire may have been perfectly accurate, but we saw that the user exhibited certain behaviours (like typing fast or correcting the field content multiple times) and it has been shown that these behaviours are themselves quite predictive.
How robust is this new data and how stable is its predictive power?
This is a new type of data being collected, so it is in very early stages for many of these business models. Discoveries are being made every day and naturally, there are challenges in teasing out the signals and deciding if and how to use the data. Consider one company’s experience using the behaviours seen when entering an applicant’s mobile phone number. The company found a positive correlation for fraud when applicants entered the number many times or used the backspace key a lot. They also found a positive correlation for fraud if a user spent a minuscule amount of time on the field or typed excessively slowly. And time hovering over the field had no correlation at all. What is a company to make of these signals? Should they act on them? Should they try to interpret them?
With newness comes risk
Early modelling with the data indicates that the predictive models of businesses are enhanced and improved, and in some cases, the increase in predictive power is significant. This raises operational questions: Should an insurer or lender really drive new applicants into a post-sale investigation based on four features related to how that person keyed in their mobile phone number? What might the first-mover advantage be for those who try? There are also issues of feature engineering and variable interaction. This just might be a goldmine for predictive modellers, who have longed for truly new data to use in their predictive models.
Next week: Why mis-spelling my address led to higher rates for collision coverage.
Read part one of this series here and read part two here.
Learn about ForMotiv and the new partnership created with FTI Consulting here.