💸 Loan Default Risk Prediction - End-to-End ML Project

📌 Objective

Build a machine learning model to predict loan default risk using LendingClub data (2007–2018). The goal is to help financial institutions make informed lending decisions by identifying high-risk borrowers before disbursing loans.

🧰 Tools & Technologies Used

• Python (Pandas, NumPy, Matplotlib, Seaborn)
• Machine Learning: Scikit-learn, XGBoost, LightGBM, Random Forest
• Hyperparameter Optimization: Optuna
• Model Interpretability: SHAP
• Environment: Google Colab

🔍 Dataset Overview

• Dataset Source: LendingClub via Kaggle
• Raw dataset: accepted_2007_to_2018Q4.csv (~2 million records)
• Target variable: loan_status → binary (Fully Paid → 0, Charged Off → 1)

🔬 Data Preprocessing & Feature Engineering

• Selected relevant features (loan amount, term, interest rate, income, credit history, etc.)
• Encoded categorical variables using Label Encoding
• Converted % fields (like int_rate, revol_util) to float
• Handled missing values using median imputation
• Added derived features like credit_history_length (from issue date and earliest credit line)

📊 Exploratory Data Analysis (EDA)

• Analyzed class imbalance: ~20% of loans are defaults
• Visualized default rates by:
  • Grade & Sub-grade
  • Employment Length
  • Home Ownership
  • Loan Purpose
• Observed key default patterns (e.g., higher interest & lower grades → higher default)

Modeling Approaches

Three tree-based models were trained and compared:

Model	AUC-ROC	Accuracy	Recall (Default)	F1-Score (Default)
Random Forest	0.738	0.68	0.66	0.45
XGBoost	0.741	0.68	0.67	0.46
LightGBM	0.743	0.68	0.68	0.46

✅ LightGBM was selected as the best performing model (highest AUC, best recall for default class).

🧪 Hyperparameter Tuning (Optuna)

• Used Optuna to maximize AUC by optimizing:
  • learning_rate, num_leaves, max_depth
  • min_child_samples, subsample, colsample_bytree
• Applied early stopping during training
• Best AUC achieved via Optuna: 0.7379

📈 Threshold Tuning (Precision vs Recall)

• Evaluated model across thresholds (0.0 to 1.0)
• Identified best threshold: 0.23
• Final classification metrics at optimal threshold:

Precision (default): 0.37
Recall (default):    0.60
F1-score (default):  0.46
Accuracy:            0.72
AUC-ROC:             0.7419