tobbi3176/african-credit-scoring

# African Credit Scoring Challenge — USIU-Africa Hackathon Predicting the likelihood of loan default for customers in Kenya and Ghana. **Platform:** Zindi — african-credit-scoring-challenge1 --- ## Folder Structure ``` ├── model.py # Full training and prediction pipeline ├── predictions.csv # Final submission file (ID + target) ├── README.md # This file ├── environment.txt # Python dependencies │ │ (place these in the same folder before running) ├── Train.csv # Training data (from Zindi) ├── Test.csv # Test data (from Zindi) └── economic_indicators.csv # FRED economic indicators (from Zindi) ``` --- ## How to Run ### 1. Set up environment ```bash pip install -r environment.txt ``` ### 2. Place data files in the same folder as model.py - `Train.csv` - `Test.csv` - `economic_indicators.csv` All three are available on the Zindi competition data page. ### 3. Run the script ```bash python model.py ``` This will print progress to the terminal and save two output files: - `predictions.csv` — Zindi submission file (ID + target) - `predictions with probabilities.csv` — same but with raw default probabilities --- ## Features Used 28 features across 5 groups: **Loan financials (log-transformed)** - `log_Total_Amount`, `log_Total_Amount_to_Repay`, `log_Amount_Funded_By_Lender`, `log_Lender_portion_to_be_repaid` **Derived loan cost features** - `repay_ratio` — Total_Amount_to_Repay / Total_Amount (how expensive is the loan?) - `interest_amount` — Total_Amount_to_Repay minus Total_Amount - `funded_ratio` — lender's share of the loan - `Lender_portion_Funded`, `duration`, `loan_term_days` **Multi-lender and customer signals** - `num_lenders` — loans with 2 lenders default at 8.3% vs 1.4% for single-lender loans - `is_new_customer` — new customers default at 21% vs 1.7% for repeat customers - `customer_loan_count` — total loans per …