In the highly competitive lending industry, incorrectly identifying defaulters leads to significant financial losses, while rejecting reliable borrowers results in missed revenue opportunities. Financial institutions must strike a careful balance between risk mitigation and growth.
This project develops a robust machine learning–based credit risk assessment system designed to:
- Accurately identify high-risk borrowers
- Prioritize recall to minimize costly false negatives
- Maintain acceptable precision to avoid excessive rejection of safe applicants
- Align model optimization directly with business risk objectives
The final solution leverages XGBoost with Bayesian hyperparameter optimization, achieving improved F-beta performance while handling severe class imbalance.
- Business Context
- Project Structure
- Data Overview
- Exploratory Data Analysis
- Data Cleaning & Feature Engineering
- Modeling & Evaluation
- Handling Imbalanced Data
- Hyperparameter Optimization
- ROC & Performance Analysis
- Installation
- Usage
- Business Impact
- Future Improvements
Traditional credit risk assessment relies on:
- Manual underwriting
- Fixed credit score cutoffs
- Income thresholds
- Rule-based systems
These approaches fail to capture:
- Non-linear borrower behavior
- Complex feature interactions
- Evolving risk patterns
As a result, institutions either:
- Approve risky borrowers (financial loss), or
- Reject reliable customers (lost revenue)
Predict loan status:
0→ Non-default1→ Default
Business Priority:
Missing a defaulter (False Negative) is more costly than incorrectly flagging a safe borrower.
Therefore, the model was optimized using:
- F-beta score (β = 1.5)
- Emphasis on Recall
- Balanced Precision
This reflects a real-world 70/30 tradeoff favoring recall.
The project includes the following files and directories:
Credit_Risk_Assesment.ipynb: Jupyter notebook containing the complete codebase and analysis for assessing credit risk.Credit_Risk_AI.ipynb: Jupyter notebook integrating a language model AI to address your questions and concerns about the credit risk process.Credit_Risk_KPI.ipynb:Jupyter notebook focused on data analysis, showcasing critical Key Performance Indicators (KPIs) related to credit card data.Credit_Risk_Test.ipynb: A notebook where you can interactively test the developed model to predict whether a customer (or yourself) is likely to default.README.md: Detailed description of the project.datasets: Directory designated for storing datasets and associated pickle files.models/: Directory for saving trained machine learning models.images/: Directory to save plots and visualizations.
Dataset Source: Kaggle (anonymized financial lending dataset)
Features include:
- Age
- Income
- Employment length
- Loan amount
- Loan interest rate
- Loan intent
- Loan grade
- Home ownership type
- Credit history length
- Historical default status
All personally identifiable information was removed.
The correlation matrix revealed:
- Moderate correlations between loan amount and income percentage
- Positive relationship between interest rate and default risk
- No severe multicollinearity issues
Key Observation:
The dataset is heavily imbalanced, with significantly more non-defaulters than defaulters — a common real-world credit dataset challenge.
- Employment length → Mean imputation
- Loan-type missing entries → Removed
- Duplicate records → Dropped
Unrealistic values identified via box plots:
- Age > 120 years
- Employment length > 120 years
These were treated as data-entry errors and replaced using mean values.
- Label Encoding used for categorical variables:
- Loan intent
- Loan grade
- Home ownership
One-hot encoding was avoided to prevent dimensionality explosion.
All features were standardized to improve performance of:
- Logistic Regression
- SVM
- KNN
- 75% Training
- 25% Testing
- K-Nearest Neighbors
- Support Vector Machine
- Logistic Regression (L1 & L2)
- Decision Tree
- Random Forest
- Extra Trees
- LightGBM
- XGBoost
Because defaulters are the minority class but most important for risk mitigation:
Applied:
- SMOTE (Synthetic Minority Over-sampling Technique)
- Class weighting
- F-beta scoring
SMOTE significantly improved minority class learning.
Used:
GridSearchCVto optimize SMOTE ratio (Optimal: 0.66)BayesSearchCV(Bayesian Optimization) to tune XGBoost
Bayesian Optimization was selected because it:
- Learns from previous trials
- Models performance probabilistically
- Finds optimal hyperparameters faster than brute-force search
Optimized Parameters:
- Max depth: 14
- Learning rate: 0.26
- Estimators: 81
Result:
- ~2% improvement in F-beta score
- Final F-beta ≈ 0.79
Final Model Performance:
- Accuracy: ~93%
- F-beta (β = 1.5): ~0.79
- Strong Recall (business priority)
- High ROC-AUC
The model successfully improves identification of high-risk borrowers while maintaining acceptable precision.
git clone https://github.com/your-username/credit-risk-analysis.git
cd credit-risk-analysis
pip install -r requirements.txt