Customer Lifetime Value Prediction

Predictive modeling to estimate the total value of customers over the entire relationship.

Predicting Future Value: CLTV Modeling

Executive Summary

Understanding the future value of a customer is crucial for setting Customer Acquisition Cost (CAC) targets. If you know a user will spend $500 over their lifetime, you can afford to spend $50 to acquire them. This project implemented a robust Customer Lifetime Value (CLTV) prediction model using the Pareto-NBD / BG-NBD probabilistic frameworks, allowing the finance and marketing teams to optimize spend with confidence.

Problem Statement

The company was using “Historic CLTV” (sum of past purchases) as a proxy for value. This is backward-looking and undervalues new users who might have high potential but haven’t spent much yet. We needed a predictive measure: “How much will this user value vary in the next 12 months?”

Methodology

1. “Buy ‘Til You Die” (BTYD) Models

We utilized the Lifetimes library in Python to implement probabilistic models:

  • BG/NBD (Beta-Geometric / Negative Binomial Distribution): Models the transaction process (frequency and recency). It assumes users buy at a constant rate until they “die” (churn).
  • Gamma-Gamma Model: Models the monetary value (average order value). It assumes the monetary value of a customer’s transactions varies randomly around their average transaction value.

2. Feature Data

  • frequency: Number of repeat purchases.
  • recency: Age of the customer when they made their last purchase.
  • T: Age of the customer (time since first purchase).
  • monetary_value: Average value of a customer’s purchases.

3. Validation

  • Calibration vs. Holdout: We split the data into a calibration period (e.g., 2023) and a holdout period (e.g., 2024). We trained on 2023 data and predicted 2024 transactions.
  • Metrics: RMSE, MAE between predicted and actual revenue in the holdout period.

Implementation Details

  • Batch Processing: The model is retrained monthly on the updated transaction history.
  • Scalability: While BTYD models are efficient, for very large datasets (10M+ rows), we utilized distributed implementations or sampled datasets.
  • Dashboarding: A Tableau dashboard visualizes the “Predicted CLTV” distribution, helping execs see the health of the customer base (e.g., “Are we acquiring higher quality users this year vs last?”).

Use Cases Enabled

  1. CAC Optimization: Marketing can bid higher for “High LTV” lookalike audiences on Facebook.
  2. Resource Allocation: Customer Success teams prioritize tickets from “High Future Value” clients.
  3. Financial Planning: Finance uses the aggregated predicted revenue for quarterly forecasting.

Challenges & Solutions

  • Challenge: The model struggled with one-time operational anomalies (e.g., a massive clearance sale that spiked frequency artificially).
  • Solution: Implemented cohort-based adjustments and outlier removal for anomalous periods to stabilize model parameters.

Results and Impact

  • Accuracy: The BG/NBD model predicted aggregate revenue within 5% of actuals for the 6-month holdout period.
  • Marketing ROI: By shifting budget to acquire “High Predicted LTV” users, the 6-month payback period improved by 20%.

Future Work

  • Deep Learning (DNN): Experimenting with deep neural networks (using features beyond just RFM, like clickstream data) to see if they outperform the probabilistic BTYD models for individual-level prediction accuracy.