Netflix Customer Churn Analysis
MSIS 510: Data Analytics · University of Washington · December 2025
Team: Alysia Tran, Trevor Rubino, Yousef AlSadhan, Shraddha Mahangare
Overview
Netflix faces rising competition and needs a proactive, data-driven approach to reduce subscriber churn. This project analyzed 5,000 Netflix user records across 14 variables to identify high-risk subscribers, uncover churn patterns, and build predictive models that enable targeted retention strategies.
Objectives
- Demographics: Identify core customer demographics and their relationship to churn
- Subscriptions: Analyze subscription tiers and their impact on churn rates
- Prediction: Build a model to predict churn-risk subscribers
- Retention: Recommend targeted strategies to reduce churn and protect revenue
Dataset
| Attribute | Details |
|---|---|
| Source | Netflix User Data from Kaggle |
| Records | 5,000 rows |
| Variables | 14 columns |
| Key Features | age, gender, subscription_type, watch_hours, last_login_days, region, device, monthly_fee, payment_method, number_of_profiles, avg_watch_time_per_day, favorite_genre |
| Target Variable | churned (1/0) |
| Data Quality | No missing values, no duplicates |
Exploratory Analysis
Subscription Type vs. Churn Rate
| Subscription Tier | Churn Rate |
|---|---|
| Basic | 61.83% |
| Standard | 45.44% |
| Premium | 43.71% |
Basic-tier subscribers churn at significantly higher rates — nearly 20 percentage points above Premium.
Churn vs. Payment Method
Gift card users had the highest churn rate, while debit and credit card users had the lowest. This suggests that recurring payment methods correlate with higher retention.
Churn vs. Last Login
Churned customers averaged 40 days since last login, compared to 20 days for non-churned users — a clear signal that login recency is a strong churn predictor.
Key Finding
Age and gender were not significant predictors of churn. Behavioral features (watch hours, login recency, number of profiles) were far more predictive than demographic variables.
Feature Selection
Used the Logit function (statsmodels) to identify statistically significant features:
- Watch hours
- Last login days
- Monthly fee
- Number of profiles
- Average watch time per day
Age was dropped as it was not statistically significant (p = 0.692).
Predictive Modeling
Split data 75/25 into training and test sets. Compared two models:
Model Comparison
| Metric | Logistic Regression | Random Forest |
|---|---|---|
| Overall Accuracy | 89.12% | 94.32% |
| Recall (TP/TP+FN) | 0.9268 | 0.9379 |
| Precision (TP/TP+FP) | 0.8818 | 0.9484 |
| F1 Score | 0.9417 | 0.9432 |
| ROC AUC | 0.9607 | 0.9920 |
Baseline model accuracy was 0.5024. Both models significantly outperform the baseline, with Random Forest achieving the highest predictive accuracy at 94.32%.
Random Forest Confusion Matrix
| Predicted: No Churn | Predicted: Churn | |
|---|---|---|
| Actual: No Churn | 590 (TN) | 32 (FP) |
| Actual: Churn | 39 (FN) | 589 (TP) |
K-Means Clustering
Applied K-Means (k=3) to segment subscribers by monthly fee and watch hours:
| Cluster | Profile | Watch Hours | Churn Rate |
|---|---|---|---|
| Cluster 0 | Casual users at risk | ~25 hrs | 42% |
| Cluster 1 | Inactive customers | ~12 hrs | 90% |
| Cluster 2 | Highly engaged | ~59 hrs | 9% |
Cluster 1 (inactive) represents the highest-risk segment with 90% churn — the primary target for retention interventions.
Business Recommendations
- Utilize the churn model to segment subscribers into risk groups for proactive outreach
- Deploy targeted marketing strategies and personalized incentives to high-risk segments
- Create retention offers to upgrade Basic customers to Standard/Premium plans, reducing the 61.83% Basic churn rate
- Implement re-engagement campaigns triggered when last login exceeds 30 days
Challenges & Limitations
- Small Dataset: 5,000 records limits generalizability to Netflix's actual 250M+ subscriber base
- Synthetic Data: Kaggle dataset does not capture real human behavior patterns
- Missing Variables: No data on date-joined, cancellation reasons, session duration, price change impact, or marketing email response rates
Reflection
This project demonstrated the power of combining predictive modeling with customer segmentation to drive actionable retention strategies. Random Forest's 94.32% accuracy and the K-Means clustering revealed that behavioral signals (watch time, login recency) are far more predictive of churn than demographics — a finding with direct implications for how streaming platforms should prioritize their retention investments.