Anonymized Loan Default Prediction

Project Description

This project focuses on predicting loan defaults using a dataset with anonymized features. The workflow encompasses data preprocessing, exploratory data analysis (EDA), handling missing values with MCMC imputation (via mice), feature selection with Boruta, and building a series of classification models to predict the binary target variable repay_fail. The target indicates whether a loan was repaid ("No") or defaulted ("Si").

Dataset

The dataset Anonymize_Loan_Default_data.csv includes various financial and applicant-specific features:

• Selected Features for Modeling: int_rate, term, emp_length, home_ownership, annual_inc, verification_status, dti, purpose, total_acc, revol_bal, funded_amnt, total_pymnt.
• Target Variable: repay_fail (mapped to "Si" for default and "No" for paid).

---

Project Phases

1. Data Preprocessing & EDA

• Converted the target variable repay_fail into a binary factor ("Si" / "No").
• Visualized overall class distribution and missingness using the VIM package.
• Missing Data Imputation: Utilized the mice package with the CART (Classification and Regression Trees) method to impute missing values for the selected covariates.
• Collinearity & Near-Zero Variance: Checked for highly correlated predictors and near-zero variance features using caret. Handled a specific issue with installment and funded_amnt.

2. Feature Selection

• Boruta: Applied the Boruta algorithm to identify the most important features and confirmed the selection of the core variables.
• A baseline selection of variables was established for model training.

3. Model Training and Validation

Multiple classification models were trained and tuned using 10-fold Cross-Validation in caret. The evaluation metric optimized was Sensitivity ("Sens").

• Logistic Regression (GLM)
• Naive Bayes
• Lasso Regression (GLMNET)
• K-Nearest Neighbors (KNN)
• Partial Least Squares (PLS)
• Decision Tree (CART/Rpart)
• Random Forest (RF)
• Gradient Boosting Machine (GBM)
• Neural Network (NNET)

4. Model Evaluation & ROC Analysis

• Extracted predictions on the validation set.
• Generated ROC curves for all models using the pROC and ROCR packages.
• Overlaid ROC curves and calculated the Area Under the Curve (AUC) to compare models visually.
• Analyzed Lift and Gain charts using funModeling for the best performing models (e.g., GBM, GLM, NNET).

5. Threshold Optimization & Scoring

• Iterated over probability thresholds (0 to 1) to calculate Precision, Recall (Sensitivity), Specificity, and the F1-score for the Neural Network model predictions.
• Visualized the performance metrics against thresholds to find the optimal cut-off for classification.
• Scoring: Applied the chosen model (GBM) on the hold-out score_set, defining a custom threshold (0.05) to generate the final predictions and evaluated them using a Confusion Matrix.