RavenStack: Behavioral SaaS Churn Analysis

Author: Harini Mukesh | Psychology Graduate & Data Analyst
Tools: Python · SQL (MySQL) · Power BI
Dataset: RavenStack Synthetic SaaS CRM — Credit: River @ Rivalytics (MIT License)

---

What Makes This Project Different

Most churn analyses compare lifetime metrics between churned and active customers.
This project applies a behavioral psychology lens — examining how customer behavior changes in the period leading up to churn, and connecting those patterns to established psychological frameworks.

"Customers who are about to churn don't dramatically change their usage — they quietly stop seeking support. The absence of tickets in the pre-churn window, not the presence, is the warning signal."

---

Project Overview

RavenStack is a stealth-mode SaaS startup delivering AI-driven team tools. This end-to-end analysis investigates what drove customer churn before their public launch — combining SQL exploration, behavioral pattern analysis, machine learning, and business intelligence dashboards into a single connected project.

Analysis Pipeline

Raw Data (5 tables)
      ↓
SQL Exploratory Analysis
      ↓
Python: Behavioral Pattern Analysis
      ↓
Python: Churn Prediction Model
      ↓
Power BI: Business Intelligence Dashboards
      ↓
Customer Success Action Export

---

Behavioral Frameworks Applied

Framework	Where Applied
Engagement Decay Theory	Pre-churn usage pattern analysis — does engagement drop before churn?
Expectation-Disconfirmation Theory	Explains why highly engaged customers still churn
Learned Helplessness	Explains the support ticket drop in the 90 days before churn
Cognitive Dissonance	Explains why customers who upgrade are more likely to churn

---

Key Findings

1. Engagement alone does not predict churn
Churned customers show similar feature usage to active customers — they were not disengaged. This points to unmet expectations rather than loss of interest.

2. Support ticket drop is the real warning signal
In the 90 days before churn, support tickets drop significantly (1.26 vs 2.97 in the earlier period). Customers stop seeking help before they leave — a pattern consistent with learned helplessness.

3. Early lifecycle is the highest risk window
85% of churn occurs within the first year. Onboarding experience and early value realization are the most critical retention levers.

4. Upgrades precede churn
A significant portion of customers churn shortly after upgrading their plan — suggesting the upgraded experience failed to meet raised expectations.

5. Feature breadth increases churn risk
Customers using more unique features are more likely to churn. Broader adoption raises expectations that the product ultimately fails to meet.

---

Dashboard Preview

Executive Dashboard

Customer Behavior & Churn Drivers

Revenue Intelligence

Churn Risk Radar

Customer Success Action Center

---

Tech Stack

Tool	Purpose
MySQL	Database creation, data cleaning, exploratory analysis
Python (Pandas, Scikit-learn, Matplotlib, Seaborn)	Behavioral analysis, feature engineering, churn prediction model
Power BI	Interactive dashboards and business intelligence
Git / GitHub	Version control and project documentation

---

Churn Prediction Model

Algorithm: Logistic Regression with class balancing
Why Logistic Regression: Interpretable coefficients allow us to explain why a customer is flagged as high risk — essential for translating model output into business action.

Key decisions:

• Used class_weight='balanced' to handle class imbalance (78% active, 22% churned)
• Used one-hot encoding for plan_tier to avoid false ordinal relationships
• Prioritized recall over accuracy — in a business context, missing a churner is more costly than a false alarm

Risk Scoring Output:

Risk Level	Threshold	Accounts
High Risk	≥ 60% churn probability	39
Medium Risk	45% – 59%	327
Low Risk	< 45%	134

---

Project Outputs

churn_risk_predictions.csv — Full risk scores for all 500 accounts (used in Power BI)

customer_success_actions.csv — Prioritized action list for the customer success team including:

• Churn probability and risk level per account
• Top behavioral signal driving their risk
• Recommended action based on that signal

---

Repository Structure

Behavioral_SaaS_Churn_analysis/
│
├── data/
│   ├── raw/
│   │   ├── ravenstack_accounts.csv
│   │   ├── ravenstack_subscriptions.csv
│   │   ├── ravenstack_feature_usage.csv
│   │   ├── ravenstack_support_tickets.csv
│   │   └── ravenstack_churn_events.csv
│   └── processed/
│       ├── churn_risk_predictions.csv
│       └── customer_success_actions.csv
│
├── images/
│   ├── 01_Executive_Dashboard.png
│   ├── 02_Customer_Behavior_Churn_Drivers.png
│   ├── 03_Revenue_Intelligence.png
│   ├── 04_Churn_Risk_Radar.png
│   ├── 05_Customer_Success_Action_Center.png
│   ├── pre_churn_behavioral_signals.png
│   └── feature_importance.png
│
├── notebooks/
│   └── churn_prediction_model.ipynb
│
├── powerbi/
│   └── Analysis_Dashboards.pbix
│
├── sql/
│   ├── Creating_database.sql
│   └── Exploratory_analysis.sql
│
├── requirements.txt
├── .gitignore
└── README.md

---

How to Run

SQL:

1. Import raw CSVs into MySQL using Table Data Import Wizard
2. Run sql/Creating_database.sql to create and populate tables
3. Run sql/Exploratory_analysis.sql for exploratory analysis

Python:

1. Install dependencies: pip install -r requirements.txt
2. Open notebooks/churn_prediction_model.ipynb
3. Run all cells top to bottom
4. Outputs saved to data/processed/

Power BI:

1. Open powerbi/Analysis_Dashboards.pbix
2. Update data source paths to your local data/processed/ folder
3. Click Refresh

---

Dataset

Synthetic SaaS CRM data representing 500 customer accounts across 5 tables: accounts · subscriptions · feature_usage · support_tickets · churn_events

Credit: River @ Rivalytics — Rivalytics Medium Blog
License: MIT-like — free to use for learning and portfolio purposes with attribution.

---

Author

Harini Mukesh
Psychology Graduate (M.Sc., University of Madras) | Data Analyst
Combining behavioral science with data analytics to build human-centered insights.