📊 Customer Churn Prediction with Profit-Optimized Machine Learning

🧠 Project Overview

This project is an end-to-end machine learning system for customer churn prediction, designed not only to predict churn but to optimize business decisions under different operational strategies.

Instead of focusing solely on predictive performance, the system emphasizes:

• Probability calibration
• Profit-driven decision thresholds
• Business constraint trade-offs (coverage vs efficiency)
• Customer segmentation
• Model explainability (SHAP)

The final output is not just a model — it is a decision-making framework for customer retention.

---

🎯 Business Objective

The objective is to identify customers likely to churn and design a retention strategy that maximizes profit (ROI).

Business Assumptions

• 💰 Cost per retention action: €50
• 💰 Value of retained customer: €500

---

📊 Dataset

• Telco Customer Churn dataset
• Binary classification:
  • 1 → Customer churns
  • 0 → Customer stays

---

🤖 Model Performance

Classification Metrics

• ROC-AUC: 0.8376
• Accuracy: 0.78

Class	Precision	Recall	F1-score
0 (No churn)	0.82	0.91	0.86
1 (Churn)	0.63	0.45	0.52

Key Insight

The model performs better at identifying non-churners than churners, which is typical in imbalanced classification problems.

---

🧠 Explainability (SHAP)

• SHAP output shape: (1407, 35)
• Provides global feature importance
• Ensures interpretability of predictions

---

🎯 Decision Strategy (Threshold Optimization)

This project does NOT use a fixed 0.5 threshold.

Instead, it evaluates multiple operating points based on business value.

---

⚙️ Two Operating Strategies

🔵 Strategy 1: High-Precision Targeting

• Best threshold: 0.74
• Retained customers: 83
• Value saved: €41,500
• Campaign cost: €5,100
• ROI: 7.13

Interpretation

• Focuses only on highest-risk customers
• Maximizes efficiency per customer
• Low coverage, high precision strategy

---

🔴 Strategy 2: High-Coverage Targeting

• Best threshold: 0.10
• Retained customers: 345
• Value saved: €172,500
• Campaign cost: €42,050
• ROI: 3.10

Interpretation

• Captures more potential churners
• Higher total business impact
• Lower efficiency but higher coverage

---

📊 Key Insight: Trade-off Between Strategies

This project demonstrates a fundamental business trade-off:

Strategy	Precision	Recall	ROI per Customer	Total Impact
High Threshold (0.74)	High	Low	High	Medium
Low Threshold (0.10)	Low	High	Medium	High

👉 There is no single “best threshold” — only different business strategies

---

📈 ROI vs Threshold Analysis

The ROI curve shows how business performance changes depending on the decision threshold.

Key observations:

• Higher thresholds reduce campaign size but increase efficiency
• Lower thresholds increase coverage but reduce ROI efficiency
• Optimal operating point depends on business constraints

---

👥 Customer Segmentation

Customers are grouped into risk segments based on predicted churn probability:

Segment	Customers	Churn Rate	Avg Probability
Low	356	1.7%	0.02
Medium	354	11.8%	0.11
High	368	32.6%	0.33
Very High	329	62.6%	0.65

Key Insight

• Risk is strongly separated across segments
• High and Very High segments represent the core revenue opportunity

---

📉 Lift Curve

The lift curve demonstrates how effectively the model prioritizes churners compared to random targeting.

• The model significantly outperforms random selection
• Highest lift is observed in the top-ranked customers

---

🔍 Key Functions

• find_best_threshold_profit() → optimizes business profit
• business_impact() → evaluates campaign ROI
• plot_roi_curve() → threshold sensitivity analysis
• plot_lift_curve() → targeting effectiveness
• segment_analysis() → customer risk segmentation

---

🧠 Key Insights

• Accuracy is not sufficient for churn problems
• Threshold selection is a business decision, not a modeling decision
• Different thresholds correspond to different business strategies
• Profit optimization provides better decision-making than classification metrics
• Customer segmentation enables targeted retention strategies

---

🛠 Tech Stack

• Python
• pandas, numpy
• scikit-learn
• matplotlib, seaborn
• SHAP

---

📌 Conclusion

This project demonstrates a complete machine learning decision system, where predictions are directly translated into business actions.

Instead of building a static classifier, we build a dynamic decision framework that adapts to business goals such as:

• Maximizing ROI
• Maximizing coverage
• Balancing cost vs value

👉 The key takeaway:
In real-world data science, the best model is not the most accurate one — but the one that makes the best decisions.