Johanna Schmidle
Has to go to the bank
a couple of times a year

Goal: build a machine learning model to predict if a bank's customer will churn (leave the bank) or not.
For each instance in the test set, you must predict a 0 or 1 value for the target variable (Classifier).

Steps

1. EDA
   • data.describe()
   • Pie chart
   • Correlation heatmap
   • Histograms
   • KDE plots
   • Univariate Exploration
2. Feature Engineering (note I will train models on original features and engineered features separately)
   • Age Group: grouping customers into age categories to capture non-linear relationships and improve the model’s ability to identify patterns.
   • Balance to Estimated Salary Ratio: provides insight into a customer's financial stability. A higher ratio may indicate financial strain, which could increase the likelihood of the customer leaving the bank.
   • Tenure Group: grouping tenure into categories to capture non-linear effects and reduce noise.
   • Loyalty Index: a composite feature reflecting customer loyalty, combining tenure, the number of products held, and whether they are an active member.
   • Balance Above 0: a binary feature indicating whether the customer's account balance is at or above zero.
3. Preprocessing
   • Split data
   • Create pipeline
     • OneHot encoding
     • StandardScaler
4. Create Models
   • Model 1: Random Forest Classifier
   • Model 2: Logistic Regression
   • Model 3: Support Vector Machine
   • Model 4: K-Nearest Neighbours
   • Model 5: XGBoost
   • Visualize models:
     • Confusion matrix
     • Display feature importance (when applicable)
5. Create Models with Engineered Features
   • Repeat above (10 models total)

Result of Model Evaluations

• Best Performing Model: Random Forest Classifier
• Worst Performing Model: SVC Model (Features)
• Interestingly, models trained on the original features outperformed those utilizing feature-engineered data. Investigating the reasons behind this unexpected result could provide valuable insights into the feature engineering process and its impact on model performance.
• Age appears to be a significant predictor of customer churn. This finding suggests that the bank should explore the underlying reasons for why age is such a strong indicator. Implementing targeted promotions or incentives for customers in high-risk age groups could effectively reduce churn and retain these valuable customers.

Data

The dataset used in this project is available publicly on Kaggle: https://www.kaggle.com/datasets/shrutimechlearn/churn-modelling

Technologies

• pandas, numpy, matplotlib, seaborn
• sklearn (OneHotEncoder, StandardScaler, make_column_transformer, ColumnTransformer, Pipeline, RandomForestClassifier, LogisticRegression, SVC, KNeighborsClassifier, XGBClassifier, cross_val_score, GridSearchCV, ConfusionMatrixDisplay)
• XGBoost