Microcredit Churn Classifier

Welcome to my Churn Prediction project, where I explored the challenge of identifying customers at risk of leaving a business. By applying machine learning techniques, I built a pipeline capable of delivering accurate predictions and valuable insights to support decision-making.

1. Exploratory Data Analysis (EDA):

   • Identify cohorts by first transaction and retention rate:

2. Feature Engineering:

   • Set Churn as > 90 days without activity
   • Create a Dataframe from transactional data to analytical data for churn prediction with the following correlation between features:

3. Machine Learning Models:

   • Developed and evaluated Logistic Regression and Random Forest Classifier models,
   • Achieved over 90% accuracy in both of them measured by F1-Score, ROC-AUC-Curve and Confusion Matrix

   RandomForestClassifier Confusion Matrix:

   ROC Curve:

Results

1. Model Comparison:

   • The RandomForestRegressor performs slightly better, with both models achieving over 90% accuracy.

2. Data Limitations:

   • Despite the high accuracy, the models do not appear to be overfitting, suggesting they are well-generalized.
   • However, the dataset covers only 1 year, which may limit capturing long-term churn trends.

3. Feature Importance:

   • Recency strongly correlates with churn, aligning with the 90-day threshold.

4. Cohort Impact:

   • Customer cohorts from first transactions significantly affect churn predictions, highlighting the importance of segmentation.

Files

• Churn Prediction - churn_prediction.ipynb