Bank Customer Churn: Predictive Modeling & Analysis

Predicting customer defection using Random Forest and Decision Tree classifiers.

---

📌 Project Overview

This project addresses the challenge of predicting client churn to help banks assess risk and improve customer retention. Using a dataset of 10,127 clients, I developed a machine learning pipeline to identify key factors influencing bank turnover.

---

🛠 Data Synthesis & Preprocessing

To ensure model integrity and computational efficiency, I performed the following steps:

• Null Value Handling: Used .isnull().sum() and .dropna() methods to verify and ensure zero missing values within the dataset.
• Feature Pruning: Identified and removed redundant columns, including the unique CLIENTNUM and two Naïve Bayes classifier columns provided by the original dataset author.

---

🔍 Data Inspection & Distribution

Before engineering features, I inspected the underlying structure and class balance of the dataset:

• Statistical Summary: Utilized .describe() and .info() to identify variable types and check for anomalies in the numeric features.
• Class Balance Check: Performed .value_counts() on categorical columns to understand the distribution of the target label (Attrition_Flag) and other key demographics like Income_Category and Education_Level.

---

🧬 Feature Engineering & Categorical Encoding

To prepare the data for the Random Forest and Decision Tree classifiers, I transformed text-based categories into numerical formats:

• Ordinal Encoding: Manually mapped ranked variables such as Education_Level, Income_Category, and Card_Category to a numeric scale (e.g., Graduate = 4, Doctorate = 6).
• One-Hot Encoding: Used pd.get_dummies for nominal variables like Gender and Marital_Status to create binary identifiers.
• Data Integration: Concatenated the newly encoded features back into the primary dataframe and removed the original text columns to finalize the feature set.

---

📊 Exploratory Data Analysis (EDA) & Visualization

To understand the underlying drivers of churn, I utilized Seaborn and Matplotlib to visualize correlations and distributions across the 10,000+ customer records.

• Correlation Heatmap: Generated a comprehensive heatmap to identify strong associations, such as the relationship between Credit_Limit and Avg_Open_To_Buy.
• Feature Distributions: Plotted histograms for all variables to identify skewness. While Age showed a unimodal distribution, features like Total_Trans_Amt exhibited left-skewed multimodal patterns.
• Churn Drivers (Box Plots): Analysis revealed that customers with low average utilization ratios and low transaction counts/amounts were at the highest risk of leaving the bank.
• Demographic Impact: Evaluation indicated that customers earning less than $40K annually and those with three or fewer dependents were key attrition segments.

EDA Gallery

Correlation Heatmap	Label Distribution
Identifying strong associations between features	Visualizing the 5.25:1 data imbalance

Transaction Amounts	Transaction Counts	Demographic Impact
Total Transaction Amounts	Total Transaction Counts	Impact of Gender/Income

---

📏 Data Slicing & Feature Scaling

To ensure the machine learning models could generalize to new data, I implemented a rigorous preparation pipeline:

• Feature/Label Isolation: Separated the target variable (Attrited Customer) from the feature set.
• Feature Scaling: Implemented StandardScaler to normalize the feature set, ensuring features with larger numerical ranges (like Credit_Limit) do not disproportionately influence the model.
• Train-Test Split: Sliced the data into training and testing sets using an 80/20 split with a random_state of 0 for reproducibility.

---

🤖 Machine Learning Implementation & Evaluation

I implemented and evaluated a Random Forest Classifier for its robustness against data imbalance and ability to rank feature importance.

Model Training & Benchmarking

• Algorithm Selection: Built the primary model using RandomForestClassifier.
• Baseline Scoring: Performed Tenfold Cross-Validation to assess variance and reliability.
• Performance Metrics: Automated evaluation using accuracy_score, classification_report, and confusion_matrix.

Random Forest Results

• Accuracy: The model achieved an initial accuracy of 95.31% on the test set.
• Predictive Success: Correctly identified 1,931 outcomes (1,695 true positives and 236 true negatives) out of 2,026 test cases.

---

💡 Feature Importance & Business Insights

A primary objective was to determine variables most heavily influencing customer defection using the Random Forest algorithm's native ranking capability.

• Top 3 Predictors:
  1. Total Transaction Amount: The most significant predictor of churn.
  2. Total Transaction Count: Frequency of account use was highly correlated with retention.
  3. Total Revolving Balance: Lower balances showed a higher propensity for attrition.

---

🌳 Comparative Analysis: Decision Tree Classifier

I implemented a Decision Tree Classifier as a baseline for performance comparison.

• Implementation: Initialized with default parameters.
• Accuracy: Achieved 93.83%, slightly lower than the Random Forest's 95.31%.
• Validation: Tenfold Cross-Validation showed higher variance in scores compared to the ensemble method.

Decision Tree Metrics

• Confusion Matrix: Correctively predicted 1,658 non-churning and 243 churning customers.
• Weighted F1-Score: 0.94, confirming strong predictive power but validating Random Forest as the superior model.

---

🏁 Conclusion & Recommendations

The comparative results confirm that while the Decision Tree is effective, the Random Forest provides superior accuracy and stability for this dataset. Both models consistently identified transaction count and revolving balance as the primary churn drivers.

Recommendation: The bank should focus CRM efforts on high-value customers who frequently perform sizable transactions. These regular users represent the highest risk for attrition if their engagement levels drop.

Full project available at Bank Churn.docx