Calistus-design/Africa-Credit-Scoring

# Africa Credit Challenge: Predicting Loan Default This project tackles a real-world loan default prediction challenge using a dataset of loans from Kenya and Ghana. The goal was to build a machine learning model that can accurately identify high-risk loan applicants. The project covers the entire data science lifecycle: from in-depth exploratory data analysis and advanced feature engineering to model comparison and the creation of a practical, business-focused credit scoring system. The final, champion model (XGBoost) was saved as a deployment-ready artifact. --- ## The Business Problem: Finding the Needle in the Haystack The core challenge of this project lies in the **severe class imbalance** of the dataset. A staggering **98.2%** of loans were successfully paid back, while only a tiny fraction (**1.8%**) resulted in a default. This creates two major problems: 1. **A "dumb" model can achieve 98% accuracy** by simply guessing "No Default" every time, making accuracy a useless metric. The true challenge is finding the rare defaulters. 2. The model's success must be measured by its ability to correctly identify this tiny minority of high-risk customers, making metrics like the **F1-Score** the primary indicator of performance. ## My Approach: A Step-by-Step guide I followed a structured, end-to-end process to move from raw data to a functional, business-ready solution. ### 1. Exploratory Data Analysis (EDA) I began with a deep dive into the data to uncover predictive patterns. The key findings were: - New customers were found to be **over 10 times more likely to default** (a >20% default rate) compared to "Repeat Loan" customers ( - **Loan Characteristics Matter:** Certain loan_type categories (like Type_15 and Type_23) had extremely high default rates, some exceeding 70%, identifying them as high-risk products. - **Financial Behavior Signals Risk:** On average, customers who defaulted tended to take out larger loans and for longer durations …