Abstract
This paper introduces a recurrent neural network approach for predicting userlifetime value in Software as a Service (SaaS) applications. The approachaccounts for three connected time dimensions. These dimensions are the usercohort (the date the user joined), user age-in-system (the time since the userjoined the service) and the calendar date the user is an age-in-system (i.e.,contemporaneous information).The recurrent neural networks use a multi-cellarchitecture, where each cell resembles a long short-term memory neuralnetwork. The approach is applied to predicting both acquisition (new users) androlling (existing user) lifetime values for a variety of time horizons. It isfound to significantly improve median absolute percent error versus lightgradient boost models and Buy Until You Die models.