Copy of BUAN357 - Final Presentation.pptx
- 2. Table of contents 01 04 03 Data Overview Predictive Models Analysis Findings
- 3. Data Overview Telecom Customer Churn 01
- 4. Telecom Customer Churn ● Multiple Lines ● Internet Service ● Online Security ● Online Backup ● Device Protection 19 Predictor Variables, 7032 rows 16 categorical ● Gender ● Senior Citizen ● Partner ● Dependents ● Phone Service 3 Numerical ● Tenure ● Monthly Charges ● Total Charges → Target Variable: Churn (binary) ● Tech Support ● Streaming TV ● Streaming Movies ● Contract ● Paperless Billing ● Payment Method
- 5. Data Exploration - 1,869 out of 7,032 customers churned - 26.6% Churn rate
- 6. Data Exploration - The longer the customers stay the more loyalty they build - Less churn, greater the tenure
- 7. Data Exploration (Service related) - Fiber optic has the highest churn rate - Month-month contract seems to be the worst type - No service and no online security impact the churn
- 8. Data Exploration - People with monthly rates between 75-110 tend to churn the most
- 9. Data Exploration (gender) - Gender doesn’t seem to be affecting churn as much
- 10. Data Exploration (gender) - Electronic check seems to have the worst churn rate - Over 80%, way higher than anything else
- 12. Later we realised we are just chill guys who took chill kofi’s class
- 14. 3 Analysis Questions What factors are contributing to churn? What prediction models can we make? How accurate are they? What customer segments are most at risk for churning? Cause Analysis Prediction Preventing Churn
- 15. 1: Cause Analysis ● Monthly contracts are the most important dictator of churn ● Total charges has an impact on churn, but neither a high or low amount is particularly linked to churn ● A higher monthly charge increases churn likelihood
- 16. 2: Prediction Models ● Logistic Regression and Gradient Boosting have the highest accuracy and balanced accuracy ● Model accuracy rarely reaches above 82% ● Models have higher accuracy in predicting negative cases → Models are better at predicting when a customer will not churn than customers who will churn
- 17. 3: Preventing Churn ● Used elbow method to determine best K value ● Using K Means, we found 4 main clusters Cluster sizes and churn rate: ● Cluster 0: 1973, 18.1% ● Cluster 1: 1473, 13.9% ● Cluster 2: 680, 25% ● Cluster 3: 2906, 39.1% → Cluster 3 has the highest churn rate, followed by Cluster 2
- 18. 3: Preventing Churn Cluster 3: ● Not senior citizens ● Mean Tenure: 15.5 ● Mean Monthly Charge: 61.59 ● Either Gender ● Less likely to have partners ● Less likely to have dependents ● Has phone service ● More likely to have paperless billing Cluster 2: ● Not senior citizens ● Mean Tenure: 31.8 ● Mean Monthly Charge: 42 ● Either Gender ● Equally likely to have partners ● Less likely to have dependents ● No phone service ● Equally likely to have paperless billing Cluster Profiles can be used to create target customer segments - Cluster 3 and 4 are most at risk for churn
- 20. Findings 1) Churn Overview 7032 Total Customers 1 in 4 Customers Leave 26.6% (1,869 out of 7,032 customers churned)
- 21. Findings 2) Key Predictors Contract Type Month-to-month contracts have the highest churn rate Service Related Factors Fiber optic customers have a notably higher churn rate Financial Factors Higher monthly charges increase the likelihood of customers leaving
- 22. Findings - Senior citizens and customers without partners show slightly higher churn tendencies - Customers with dependents tend to have lower churn rates 3) Demographic Insights
- 23. Findings - Electronic check payment method is associated with higher churn rates 4) Payment and Billing - Paperless billing is correlated with lower churn, suggesting more engaged customers
- 24. Recommended Actions - Incentives for long term contracts - Develop targeted retention programs for high-risk clusters - Provide value-added services for customers with fiber optic internet 1) Retention Strategies
- 25. Recommended Actions - Review pricing structure, especially for high-risk customer segments - Bundle additional services like online security and tech support - Create personalized service packages 2) Pricing and Services
- 26. Recommended Actions Monitor key indicators like: ● Tenure (especially in first 15-20 months) ● Changes in monthly charges ● Service usage patterns ● Payment method shifts 3) Early Warning System
- 27. Recommended Actions Before Telco took BUAN 357 After Telco took BUAN 357
- 28. Limitations 05
- 29. Limitations ● Asymmetry: The dataset has more non-churners than churners, making it difficult for models to accurately predict customer churn. ● Missing Values: Some columns, like TotalCharges, had missing values, which were dropped, reducing the dataset size and potentially introducing bias. ● Limited Feature Use: Key behavioral or satisfaction-related metrics (e.g., customer feedback) are not included ● Non-Timed Data: The dataset is a snapshot rather than time-series data, limiting the ability to analyze how churn risks evolve over time or respond to customer lifecycle events ● Lack of External Context: The analysis does not incorporate external factors such as market competition, or economic conditions
Editor's Notes
- #1: 5 - 7 minute presentations that follow the same structure as report, worth 5 points out of 30 for the final project Prioritiz answering our business questions, exclude kaggle explanations if we need to ( does kaggle explanations mean our dataset overview or explanation of other solutions?)