Key Findings from Hidden Structure in Term Coverage End Dates: An Unsupervised Exploration of In-Force Term Life Data

Proformex Insights

At Proformex, we spend a lot of time asking questions about life insurance and annuity data. We are able to combine our domain knowledge and data expertise to offer guidance to our customers on how to interpret their own data and work together to find solutions to longstanding pain points. A key component of this process is discovery and experimentation to further develop our understanding of the data and how it can be leveraged by organizations across the industry. In this and a series of future papers, we’d like to share some of our research to bring better visibility to these insights and hopefully inspire curiosity and conversations across the industry. This article summarizes our first research paper, which you can find on our website.

Our first topic covers an experiment we conducted to learn more about the term data in our data platform. While there is a seemingly limitless series of pointed questions we’d like to ask, we find that there’s tremendous value in sometimes taking a more open-ended approach to learn things about the data we never would have thought to ask. By designing experiments with less rigid objectives, we provide ourselves the opportunity to learn unrestrained.

Our experiment was conducted on 730k active term policies across 11 carriers and aimed at finding unique characteristics amongst these policies that we’d not previously considered. We applied this data to two clustering algorithms which allowed us to take an unsupervised approach to finding less obvious patterns in the data. More details on the models and parameters for the experiment can be found in the full version.

We used a scoring technique to select our best-performing model and began to investigate how the model labeled the data. We found that the model separated the data into two clusters (Cluster 0 and Cluster 1) and after reviewing how the labels were applied to different features, we found that there was notably clean separation based on a policy’s term duration.

Image Caption: Segmentation of data by cluster and term duration for the best-performing model in our experiment. Cluster 1 had term durations beyond what we would expect for standard term policies.

The policies in Cluster 0 had term durations that aligned with our expectations for a typical term policy, while those in Cluster 1 extended many years beyond expectation. Upon further examination, we found this to be a characteristic of several carriers’ term data. Given prior experience with other data elements, we recognized this may be due to different interpretations for how the Term Coverage End Date is provided through carrier data feeds.

In order to test our suspicions, we determined a method to derive the term coverage end date for policies in Cluster 1 to infer a Consolidated Term Duration. What we found was that this derived approach resulted in term durations that aligned Cluster 0 and now Cluster 1 with our expectations for typical life policies.

Image Caption: Data segmentation by cluster and consolidated term duration after substituting our derived term coverage end date for the carrier provided date for policies in Cluster 1. We found that the derived term coverage end date / consolidated term duration aligned with our expectations and those found for policies in Cluster 0.

We see this as a meaningful outcome from this open-ended experiment. The term coverage end date is one of the core data elements for any term policy and is critical for any industry professional to understand when additional coverage conversations may be needed with the policyholder. Since the conclusion of this experiment, we’ve developed internal methods to address this and other data quality issues we’ve identified in the data. More importantly, we’ve been able to proactively help our customers understand these nuances to better manage their books and collaborate with carriers to address these.