An end-to-end SaaS churn analytics project: it integrates five relational tables, engineers behavioral features, predicts churn risk, estimates customer lifetime value, and turns the results into a retention-targeting strategy and an interactive Power BI dashboard.
The focus is the full analytical cycle — from messy multi-table data to a business-ready retention framework, rather than a single high-accuracy model.
Subscription businesses lose recurring revenue through preventable churn. The project answers:
- Which behavioral and operational factors are associated with churn?
- Can we score each account's churn risk?
- Which high-value accounts deserve proactive retention?
- How should a limited retention budget be prioritized for the best return?
Five relational SaaS tables (synthetic, 500 accounts), merged to account level:
| Table | Description |
|---|---|
accounts.csv |
Customer profile, seats, plan tier, churn flag |
subscriptions.csv |
Subscription lifecycle and billing |
feature_usage.csv |
Product engagement and activity |
support_tickets.csv |
Support burden, satisfaction, escalations |
churn_events.csv |
Churn timing and cancellation reasons |
Note: synthetic dataset. The data is well-suited for demonstrating the full pipeline; as is common with synthetic data, the predictive signal is limited, discussed honestly in the results below.
| Notebook | Focus |
|---|---|
01_EDA_and_Data_Cleaning |
Multi-table merge, feature engineering, EDA |
02_Churn_Prediction_and_Strategy |
Logistic Regression baseline, risk segmentation |
03_Advanced_Models |
Random Forest, model comparison, churn probability, CLV, retention matrix |
- Merged five relational tables on
account_id, resolving granularity (subscription → account level) - Cleaned missing values; engineered behavioral features:
inactive,high_value,low_satisfaction,support_heavy - Explored churn patterns by inactivity, satisfaction, support load, and account value
- Built a churn KPI framework and a behavioral
risk_segment - Logistic Regression as an interpretable baseline
- Strategic segmentation for retention action
- Random Forest, compared head-to-head with Logistic Regression on the same split
- Full metric comparison (accuracy, precision, recall, F1, ROC-AUC)
- Churn probability score per account
- Customer Lifetime Value (CLV) estimation and revenue at risk = churn probability × CLV
- 2×2 retention targeting matrix: Immediate Action / Monitor / Nurture / Standard
Two models were compared on the same 80/20 split:
| Metric | Logistic Regression | Random Forest |
|---|---|---|
| Accuracy | 0.67 | 0.63 |
| Precision | 0.67 | 0.68 |
| Recall | 1.00 | 0.85 |
| F1 | 0.80 | 0.76 |
| ROC-AUC | 0.51 | 0.42 |
Honest interpretation: the ROC-AUC values are close to 0.5, which means the models do not separate churners from non-churners much better than chance on this data. The high recall of the Logistic Regression is misleading with a 70% churn base rate, predicting "churn" for almost everyone yields high recall but little real discrimination.
The most likely reason is the synthetic nature of the dataset: the engineered features don't carry a strong real signal for churn. This is itself a finding, it shows the importance of validating whether the data actually supports prediction, rather than trusting metrics at face value.
What remains valuable is the full pipeline: the five-table integration, the feature engineering, the CLV and revenue-at-risk framework, and the retention-targeting matrix, all of which would apply directly to a real dataset with genuine signal.
| Feature | Importance |
|---|---|
| total_usage | 37.6% |
| seats | 28.4% |
| ticket_count | 15.2% |
| avg_satisfaction | 14.4% |
| total_escalations | 4.4% |
Interactive dashboard across 3 pages: Churn Overview · Behavioral Risk Analysis · Retention Targeting
Interactive Dashboard: 
- Behavioral features (usage, support burden) carry more weight than account size in the model
- Inactivity and low satisfaction are associated with higher churn in the EDA
- On this synthetic data the models show limited predictive power (AUC ≈ 0.5) — an honest result
- The CLV + churn-probability matrix provides an actionable retention framework regardless of model strength
| Area | Tools |
|---|---|
| Data & Analysis | Python, Pandas, NumPy, Matplotlib, Seaborn |
| Machine Learning | scikit-learn — Logistic Regression, Random Forest, predict_proba, feature importance |
| Business Analytics | CLV estimation, revenue-at-risk, 2×2 retention matrix |
| BI & Reporting | Power BI |
| Workflow | Jupyter Notebook, GitHub |
- Validate the pipeline on a real dataset with genuine churn signal
- Add time-based cohort retention analysis
- Test gradient boosting once meaningful signal is present
- Integrate real billing data for accurate revenue-at-risk
Gabriele De Carlo · Python · SQL · Power BI — Data Analyst Portfolio Project, 2025