# Predicting Customer Lifetime Values -- ecommerce use case

**Authors:** Ziv Pollak  
**arXiv:** [2102.05771](https://arxiv.org/abs/2102.05771)  
**Date:** 2021-02-10  
**Summary page:** https://www.summarizepaper.com/en/paper/2102.05771v1/

## Key points

- The study focuses on predicting customer future purchases and lifetime value in e-commerce.
- Two approaches are compared: a 'buy-till-you-die' statistical model and a neural network model.
- Historical purchase data is used for predictions due to the noncontractual nature of customer-firm relationships.
- Clearly's customer purchase historical data, consisting of approximately 2 million customers over 5 years, is utilized.
- Various features are extracted for each customer, including lifetime duration, number of purchases, average gaps between purchases, average revenue from purchases, days since first and last buy, and number of coupons used.
- The goal is to predict customer purchases within the next 6 months.
- The training dataset spans from June 2015 to June 2020 while the testing dataset covers June 2020 to December 2020 (a 6-month holdout period).
- Lifetimes Python library is used for implementing the statistical model with detailed purchase date and amount data.
- Good predictions can be achieved quickly with this approach.
- A fully connected neural network model is built using summarized data after testing multiple configurations.
- The study includes quantitative and qualitative analysis of both methods.
- Insights into their performance are provided along with recommendations for different scenarios.
- Opportunities for future research are identified.

## Summary

The study focuses on predicting customer future purchases and lifetime value in the context of e-commerce. The task is challenging due to the noncontractual nature of the customer-firm relationships, which requires relying mostly on historical purchase data for predictions. The research compares two approaches: a 'buy-till-you-die' statistical model and a neural network model. To conduct the comparison, the researchers utilize Clearly's customer purchase historical data, consisting of approximately 2 million customers over a period of 5 years. Various features are extracted for each customer, including lifetime duration, number of purchases, average gaps between purchases, average revenue from purchases, days since first and last buy, and number of coupons used. The goal is to predict customer purchases within the next 6 months. The training dataset spans from June 2015 to June 2020 while the testing dataset covers June 2020 to December 2020 (a 6-month holdout period). For the statistical model implementation Lifetimes Python library is employed using detailed purchase date and amount data. It is noted that reasonably good predictions can be achieved quickly with this approach. On the other hand a fully connected neural network model is built using summarized data. Multiple configurations are tested before settling on the final one. The study includes both quantitative and qualitative analysis of these two methods providing insights into their performance and offering recommendations for different scenarios as well as identifying opportunities for future research.