Abstract
Sequence of events that occur over time may have recurrent patterns. In this paper we develop a scalable methodology to uncover such time patterns and demonstrate their value in business applications. We build upon prior work by Magnusson (2000) and identify four limitations of the existing T-pattern algorithm that preclude it from being useful for typical business problems. The four categories of limitations are: (i) scalability, (ii) supervised learning, (iii) heterogeneous individuals, (iv) distributional assumptions, and propose a solution for each limitation. We use simulations to exhibit the properties of our proposed algorithm and its ability to uncover true T-patterns. The simulations demonstrate the gains accrued from our proposed algorithm when compared to the original T-pattern algorithm. We use insurance claims data from a well-known insurance company to test the algorithm. We show that the algorithm successfully detects T-patterns that routinely occur in the context of insurance claims. Using each T-pattern as a binary feature in machine learning models we classify the claims into the two groups of satisfied and dissatisfied customers. This reveals T-patterns that separate dissatisfied customers from satisfied ones and identifies ouch-points that could minimize customer dissatisfaction with the claims process.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have