Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting Customer Lifetime Value Using Recurrent Neural Net

Huigang Chen, Edwin Ng, Slawek Smyl, Gavin Steininger

TL;DR

The paper tackles predicting customer lifetime value in SaaS by explicitly modeling three interconnected time dimensions: cohort, age-in-system, and contemporaneous date. It introduces a dilated, multi-layer RNN with a novel dRNN cell that outputs both the forward signal and a controlling state to enhance gating, combined with data preprocessing to handle sparse transaction histories. The method is validated on two real-world tasks—Uber acquisition LTV and Meta Quest rolling LTV—showing superior accuracy over BTYD and light gradient-boosting baselines, with metrics such as RMSE and aSMAPE across multiple horizons. The work demonstrates that this approach yields more accurate, flexible LTV forecasts, enabling more timely and effective ROAS optimization in SaaS contexts.

Abstract

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

Predicting Customer Lifetime Value Using Recurrent Neural Net

TL;DR

The paper tackles predicting customer lifetime value in SaaS by explicitly modeling three interconnected time dimensions: cohort, age-in-system, and contemporaneous date. It introduces a dilated, multi-layer RNN with a novel dRNN cell that outputs both the forward signal and a controlling state to enhance gating, combined with data preprocessing to handle sparse transaction histories. The method is validated on two real-world tasks—Uber acquisition LTV and Meta Quest rolling LTV—showing superior accuracy over BTYD and light gradient-boosting baselines, with metrics such as RMSE and aSMAPE across multiple horizons. The work demonstrates that this approach yields more accurate, flexible LTV forecasts, enabling more timely and effective ROAS optimization in SaaS contexts.

Abstract

This paper introduces a recurrent neural network approach for predicting user lifetime value in Software as a Service (SaaS) applications. The approach accounts for three connected time dimensions. These dimensions are the user cohort (the date the user joined), user age-in-system (the time since the user joined the service) and the calendar date the user is an age-in-system (i.e., contemporaneous information).The recurrent neural networks use a multi-cell architecture, where each cell resembles a long short-term memory neural network. The approach is applied to predicting both acquisition (new users) and rolling (existing user) lifetime values for a variety of time horizons. It is found to significantly improve median absolute percent error versus light gradient boost models and Buy Until You Die models.
Paper Structure (14 sections, 5 equations, 4 figures, 3 tables)

This paper contains 14 sections, 5 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Problem Statement
  3. Acquisition LTV
  4. Rolling LTV
  5. Related Work
  6. Method
  7. The Recurrent Neural Network Approach
  8. Architecture
  9. dRNN Cell
  10. Data PreProcessing
  11. Results
  12. Acquisition LTV (Uber)
  13. Rolling LTV (Meta)
  14. Conclusion

Figures (4)

  • Figure 1: The daily user value as a function of age-in-system (left), and as a function of observation date (right). The colors indicate the system status (blue is an outage, green is an event).
  • Figure 2: An example RNN composed with 6 layers(cells), in two blocks, with the Res-net style shortcut over block number 2. Additionally, some layers are dilated, and there is a final linear “adaptor” layer that converts the top-most cell output into the needed size. Each dot represents a cell or input or output in a particular time step.
  • Figure 3:
  • Figure 4: Subsampling of periods with no transactions