Reserving with Machine Learning: Applications for Loyalty Programs and Individual Insurance Claims

Motivation Reserving is typically performed on aggregate claim data using familiar reserving techniques such as the chain ladder method. Rich data about individual claims is often available but is not systematically used to estimate ultimate losses. Machine learning techniques are readily available to unlock the benefits of this information, potentially resulting in more accurate reserve estimates.

Method In this paper we introduce a reserving framework that leverages machine learning to incorporate rich granular information that is not captured when data is analyzed at the aggregate level. The framework relies on the snapshot date triangle as the format for organizing the data, which enables us to incorporate all available information in the prediction of ultimate values. A mix of machine learning algorithms is applied to the snapshot date triangle to create segments of claims with homogeneous development patterns. Standard triangular methods can then be applied on each segment to estimate the ultimate values.

Results This method was developed in the context of reserving for loyalty programs. Within the loyalty context, reserving refers to estimating future redemption patterns for points issued to-date, producing an estimate of the loyalty program liability. We show how this framework can be used to create segments of members with homogeneous redemption behaviors, which facilitates the reserving exercise.

Conclusions We see a clear analogy between a loyalty program member’s redemption pattern and a claim payment pattern. Consequently, the applicability of this framework for loyalty program reserving suggests there may be an opportunity to apply this framework for insurance claim reserving.

Keywords Data Mining, Predictive Modeling, Reserving Methods, Individual Claims Reserving, Claims Triage, Loyalty Program Liability, Data Organization, Ultimate Redemption Rate, Breakage Estimation, Snapshot Triangle