Credit Scoring Model in e-commerce for Buy-now-pay-later services

KAIM Week 6 Challenges

Project Overview

This repository contains all the necessary notebooks, scripts, and resources required to build and deploy a Credit Scoring Model for Bati Bank. The model assesses credit risk, predicts fraudulent transactions, and classifies users based on their likelihood of default. The entire workflow adheres to Basel II regulatory standards and uses advanced machine learning techniques and feature engineering.

The project is divided into several tasks, from exploratory data analysis to model deployment on Render.

Check it out yourself: Deployed API

Folder Structure

├── data/                       
├── notebooks/                  
│   ├── kaim-week6-task-2.ipynb         
│   ├── kaim-week6-task-3.ipynb         
│   ├── kaim-week6-task-4.ipynb         
├── src/  
│   ├── load_data.py                 
│   ├── eda.py
│   ├── 
├── scripts/ 
│   ├── main.py
├── tests/ 
│   ├── test_load_data.py                 
├── app/                        
│   ├── app.py                 
│   ├── model.pkl
│   ├── requirements.txt
├── .gitignore                   
├── README.md                   
├── requirements.txt

Installation and Setup

To get started, clone this repository to your local machine:

git clone https://github.com/tedoaba/KAIM-W6.git
cd KAIM-W6

Install Dependencies

Ensure you have Python 3.8+ installed. Install the necessary libraries using:

If you want to set up a virtual environment:

# Create a virtual environment
python3 -m venv .week6
# Activate it
source .week6/Scripts/activate   # On Linux: venv\bin\activate
# Install dependencies
pip install -r requirements.txt

Dependencies include:

• pandas, numpy: Data manipulation and numerical operations.
• scikit-learn: Machine learning models and evaluation metrics.
• seaborn, matplotlib: Visualization tools.
• flask: For model deployment via a web API.
• xverse, woe: Advanced feature engineering and Weight of Evidence transformations.

Tasks Breakdown

Task 2: Exploratory Data Analysis (EDA)

• Dataset structure overview.
• Summary statistics for numerical and categorical variables.
• Visualization of distributions and correlations.
• Identifying missing values and outliers.

Notebook: task2_EDA.ipynb

Task 3: Feature Engineering

• Creation of aggregate features like TotalTransactionAmount, TransactionCount, and SpendScore.
• Extraction of date-time features from TransactionStartTime.
• Categorical encoding and handling missing values.
• Normalization and scaling of numerical features.

Notebook: task3_feature_engineering.ipynb

Task 3 (Extended): Default Estimator and WoE Binning

• Creation of RFMS features (Recency, Frequency, Monetary, SpendScore).
• Classifying customers into Good and Bad risk categories using SpendScore.
• Applying Weight of Evidence (WoE) binning for better interpretability of key features.

Notebook: task3_extended_woe_binning.ipynb

Task 4: Model Selection and Training

• Training and evaluation of models: Logistic Regression, Decision Trees, Random Forest, Gradient Boosting.
• Hyperparameter tuning using Grid Search for Random Forest and Gradient Boosting.
• Evaluation metrics include Accuracy, Precision, Recall, F1 Score, ROC-AUC, and confusion matrix.

Notebook: task4_modelling.ipynb

Model Deployment

The model is deployed using Flask on Render, enabling real-time credit scoring predictions via an API. The deployed model takes transaction data and outputs a classification (Good/Bad risk) or a default probability score.

Deployment Folder Structure

The app/ directory contains files required for deployment:

• main.py: The Flask web application that serves the trained model.
• xgb_model.pkl: The trained model saved as a pickle file.
• requirements.txt: Lists the necessary libraries for deployment.

Running Locally

To run the Flask app locally:

1. Navigate to the app/ directory:

   cd app

2. Install app-specific dependencies:

   pip install -r requirements.txt

3. Run the Flask app:

   python app.py

4. The API will be hosted locally at http://127.0.0.1:5000/.

Deployed on Render

The model is deployed live on Render for real-time predictions. You can send transaction data to the endpoint and get predictions on credit risk. The URL will look something like this:

https://kaim-w6.onrender.com/

API Endpoints

Results and Outputs

Results from model evaluation, such as confusion matrices, performance metrics, and hyperparameter tuning results, are saved in the results/ directory. This includes:

• Accuracy, Precision, Recall, F1 Scores for each model.
• Feature importance plots that show key features influencing predictions.
• Confusion matrices to visualize classification errors.

How to Run Notebooks

You can run the notebooks locally or on platforms like Kaggle or Google Colab. Follow these steps:

1. Download the dataset and place it in the data/ directory.
2. Open any notebook from the notebooks/ folder.
3. Run all cells to execute the EDA, feature engineering, and model training steps.

Contributions

Contributions to improve the project are always welcome! Feel free to open issues or pull requests if you have suggestions or enhancements to propose.

Future Work

Future improvements and extensions for this project include:

• Model Monitoring: Implement model drift detection to ensure performance over time.
• Explainability: Add tools like SHAP or LIME for feature-level explainability.
• Additional Data: Incorporate more external data sources (e.g., credit history, behavioral data) to improve predictions.
• Deployment Scaling: Scale the API for production using tools like Docker or Kubernetes.