Loan Default Prediction using Machine Learning

Retail lenders lose money when borrowers default, so scoring each application by its probability of default is one of the highest-value uses of machine learning in banking. Most student projects on this topic train on tiny synthetic datasets and report accuracy on data that never resembles a real loan book — leaving the student unable to defend their approach in a viva. This project works with the real LendingClub loan dataset of 2.26 million funded loans: 145 features per record, real class imbalance, real missing data. The build covers the full pipeline — data cleaning and feature engineering, handling imbalance with class weights, training a gradient-boosted-tree plus logistic-regression ensemble, calibrating probabilities, and choosing an operating threshold from business cost trade-offs. The demo dashboard makes the model tangible: enter an applicant profile in Indian context (₹ amounts, Indian employment categories) and see the default probability, risk band, recommended decision and which features drove the score.

1. The LendingClub dataset is loaded and filtered to closed loans with a clear outcome: fully paid vs charged off/defaulted.
2. Features are engineered: loan-to-income ratio, debt-to-income, employment length, delinquency counts, revolving utilization and grade/sub-grade encodings; leakage-prone columns are removed.
3. The data is split by time (train on earlier loans, validate and test on later loans) so evaluation reflects how the model faces new originations.
4. An ensemble is trained — gradient-boosted trees for non-linear interactions plus a logistic-regression head — with class weights to handle the ~4:1 imbalance; probabilities are calibrated with isotonic regression.
5. The operating threshold is chosen from the cost trade-off between missed defaults and declined good applicants, not from accuracy alone; AUC-ROC, precision/recall and the KS statistic are measured on the held-out test split.
6. In the dashboard, an applicant profile is transformed with the same feature pipeline and passed through the calibrated model; the probability, risk band and per-feature contributions are displayed with the recommended decision.

Tech stack:

• Python 3, scikit-learn
• XGBoost (gradient-boosted trees)
• Pandas, NumPy, Matplotlib
• Jupyter Notebook (training & evaluation)
• HTML/CSS/JavaScript (scoring dashboard demo)
• LendingClub loan dataset (public, via Kaggle)

Dataset & model details

• Dataset: LendingClub loan data 2007–2018 — 2,260,701 funded-loan records, 145 features per loan, public release distributed via Kaggle. Classes: approximately 80% fully paid vs 20% charged off/defaulted (binary task).
• Task: Binary classification — input = applicant/loan feature vector (loan amount, term, grade, employment length, annual income, DTI, delinquencies, revolving utilization, home ownership), output = calibrated 12-month probability of default.
• Model: Ensemble of gradient-boosted trees (XGBoost) with a logistic-regression layer; probability calibration via isotonic regression; class weights and SMOTE evaluated for imbalance handling.
• Metrics: AUC-ROC design target 0.86, precision/recall at the chosen operating threshold, KS statistic and calibration curve. No metric is presented as measured — the shipped report documents the numbers the actual built-to-order training run produces.

• [Real-dataset training] The model is trained on the public LendingClub loan dataset (2007–2018, 2.26M records) — real feature distributions, real imbalance (~80% fully paid / ~20% charged off), real missing values.
• [Live risk-scoring dashboard] Enter loan amount, income, credit score, employment type, tenure, existing loans and past delinquencies; get a default-probability gauge, Low/Moderate/High risk band and an Approve/Review/Decline recommendation.
• [Feature-contribution explanations] Every prediction breaks down which features pushed the probability up or down — loan-to-income ratio, credit score and past defaults are typically the dominant drivers.
• [Full training notebook] Data cleaning, feature engineering, imbalance handling, model training, calibration and threshold selection in one reproducible Jupyter notebook.
• [Honest evaluation report] AUC-ROC, precision/recall at the operating threshold, KS statistic and calibration curve — all from the actual training run, with design targets stated up front and never presented as pre-measured.
• [Imbalance handling documented] Class weights and SMOTE are compared properly on a validation split, so the viva answer on "why not just accuracy" is backed by the build itself.
• [Batch-scoring view] A sample applicant table runs multiple realistic profiles through the same pipeline, showing how scores differ across employment types and credit-score buckets.
• [Viva-ready documentation] Report explains probability calibration, threshold choice and the cost of false approvals vs false declines.

• Complete training & evaluation Jupyter notebook on the LendingClub dataset
• Trained, calibrated scoring model with the feature pipeline
• Interactive risk-scoring web dashboard (single-file HTML/JS)
• Evaluation report: AUC-ROC, precision/recall, KS, calibration curve from the actual run
• Project report PDF (background, dataset analysis, methodology, results, threshold rationale)
• PPT presentation for final review
• Viva Q&A preparation document (calibration, imbalance, threshold trade-offs, credit-risk basics)