mahami7/Loan-Default-Prediction-Model-Using-Deep-Learning

A deep learning model to predict loan default probabilities for the African Credit Scoring Challenge. # Predicting Loan Defaults with Deep Learning A deep learning model to predict loan default probabilities for the African Credit Scoring Challenge. ## Motivation This project was developed as part of a competitive challenge to address real-world financial risk assessment in African markets. It showcases my skills in deep learning, data preprocessing, and feature engineering while tackling a socially impactful problem. ## Description This project develops a robust neural network to predict the likelihood of loan defaults using financial data from the African Credit Scoring Challenge. Built with TensorFlow, the model processes diverse features (e.g., loan amounts, durations, and borrower demographics) to provide accurate risk assessments for financial institutions in Africa's dynamic markets. Key techniques include feature engineering, SMOTE for class imbalance, and L2 regularization for robust generalization. ## Dataset - Source: African Credit Scoring Challenge - Features: Loan amounts, durations, country IDs, loan types, and more. - Target: Binary (0: No default, 1: Default). ## Model Architecture - Layers: 4 dense layers (256, 128, 64, 32 neurons) with ReLU activation. - Regularization: L2 regularization and dropout (0.5, 0.4, 0.3). - Optimization: Adam optimizer (learning rate: 0.0003). - Callbacks: Early stopping and learning rate reduction on plateau. ## Features - Predicts loan default probabilities using a deep neural network. - Handles imbalanced data with SMOTE oversampling. - Incorporates feature engineering (e.g., date-based features, log transformations). - Visualizes data distributions, correlations, and model performance (e.g., ROC AUC, F1 Score). ## How It Works ## Exploratory Data Analysis (EDA) - Inspected missing values, duplicates, and class distribution. - Visualized data using histograms, boxplots, and heatmaps. - Analyzed correlations to understand feature relationships with the target. …