SaaS Customer Churn Analysis - Hi, I'm Sebastian Marrero

SaaS Customer Churn Analysis

View on GitHub

End-to-end analysis of customer churn using a fictional SaaS dataset. This project combines exploratory data analysis, model development, and interpretability techniques to identify key churn drivers and propose data-backed retention strategies.

Project Overview

Purpose

  • Identify behavioral and usage-based predictors of churn.
  • Compare interpretable and ensemble models for prediction performance.
  • Extract actionable insights to inform customer success and engagement.

Dataset & Tools

  • Data Source: Kaggle’s SaaS Churn Dataset
  • Tech Stack: Python (pandas, matplotlib, seaborn, scikit-learn), Jupyter Notebook
  • Deployment: GitHub Pages + Jekyll

Key Business Questions

  1. What features best predict whether a SaaS customer will churn?
  2. How do behavior metrics like support calls, payment delays, and inactivity relate to churn?
  3. Which model provides the best trade-off between interpretability and accuracy?

Data Visualizations

Churn by Subscription Type

Churn by Subscription Type

  • Basic plan users showed slightly higher churn than Premium/Standard users.
  • Upselling strategies to retain lower-tier customers may help reduce churn.

Churn by Contract Length

Churn by Contract Length

  • Quarterly contracts had highest churn rates.
  • Annual plans promote loyalty—an opportunity for longer-term commitment strategies.

Tenure by Churn Status

Tenure by Churn Status

  • Shorter-tenure users were more likely to churn.
  • Highlights importance of early-stage engagement and onboarding efforts.

Support Calls by Churn Status

Support Calls by Churn Status

  • Churners placed more support calls, indicating frustration or dissatisfaction.

Last Interaction by Churn Status

Last Interaction by Churn Status

  • Churners tended to show longer periods of inactivity before leaving.

Modeling Results

Model Performance Summary

ModelAUCAccuracyChurn RecallStay Recall
Logistic Regression0.7971%77%67%
Random Forest (Default)0.6250%95%22%
Random Forest (Tuned)0.6856%93%25%

ROC Curve Comparison

ROC Curve - LR vs RF ROC Curve - LR vs Tuned RF

  • Logistic Regression outperforms both default and tuned Random Forest models.
  • Tuning improved RF performance, but Logistic Regression remained most reliable.

Feature Importance (Random Forest)

Feature Importance RF Original Feature Importance RF Tuned

  • Top drivers: Support Calls, Total Spend, Payment Delay.
  • Tuned model emphasized different behavioral features (e.g., Age, Gender).

Logistic Regression Coefficients

Logistic Regression Coefficients

  • Support Calls (+2.20), Payment Delay (+0.89), Inactivity (+0.52) increase churn risk.
  • Tenure (–0.13), Total Spend (–1.40), Usage Frequency reduce churn risk.

Conclusion

  • Logistic Regression provides strong AUC and interpretability, making it ideal for churn scoring.
  • Behavioral features outperform demographics—support volume, payment issues, and inactivity are key churn flags.
  • High-value users (long tenure, high spend) are more loyal—focus retention there.

Reproducibility

  1. Clone the repo: git clone https://github.com/SebastianMarrero/SaaS_Churn_Analysis.git
  2. Run CustomerChurnEDA.ipynb in Jupyter Lab or VS Code
  3. View outputs and visuals in /assets/images

Created by Sebastian Marrero — sebastianmarrero64@gmail.com — LinkedIn