22-pinkee/Home-Credit-Risk-Analysis

🏦 Home Credit Risk Analysis — EDA, Feature Engineering & ML Models

This project explores the Home Credit Default Risk dataset through EDA, statistical insights, feature engineering, clustering, regression, and classification modeling.

The goals of the project are to:

• —Understand borrower behavior and key risk drivers
• —Predict missing creditworthiness scores (EXTSOURCE2)
• —Develop a classification model to predict loan default
• —Deploy both models via HuggingFace for reproducibility

📊 Dataset Overview

After extensive cleaning, the dataset contains:

• —307,511 rows, with all features engineered and cleaned
• —Strong target imbalance: 92% non-default vs 8% default

✔ Data Cleaning Summary

To prepare the data for analysis and modeling:

• —Removed irrelevant or unusable columns
• —Dropped features with excessive missing values
• —Imputed missing values (numeric & categorical)
• —Treated outliers via percentile capping
• —Added missing-value indicators
• —Ensured consistent feature formats for ML

This provided a stable, reliable dataset for EDA and modeling.

📌 Project Structure

1. 1.EDA & Research Questions
2. 2.Data Cleaning Summary
3. 3.Feature Engineering
4. 4.Clustering (K-Means + PCA)
5. 5.Regression Model — Predicting EXT_SOURCE_2
6. 6.Classification Model — Predicting Default (TARGET)
7. 7.Model Deployment (HuggingFace)
8. 8.Key Insights, Limitations & Future Work

🔍 1. Exploratory Data Analysis (EDA)

We explored the dataset through seven core research questions.

1️⃣ Does EXT_SOURCE_2 differentiate between borrowers who default?

[image]

Interpretation: Borrowers who did not default show consistently higher EXT_SOURCE_2 scores. This variable is a strong risk separator.

2️⃣ Which borrower characteristics most strongly influence EXT_SOURCE_2?

[image]

Interpretation: Income, employment stability, demographic factors, and document counts correlate significantly with EXTSOURCE2.

3️⃣ Does borrower age affect default risk?

[image]

Interpretation: Younger borrowers default more often. Default probability declines steadily with age until ~60.

4️⃣ Does education level impact loan default risk?

[image]

Interpretation: Lower education levels correlate with a higher chance of default.

5️⃣ Is family status associated with different default probabilities?

[image]

Interpretation: Single and separated borrowers show elevated risk compared to married borrowers.

6️⃣ Does occupation type influence default likelihood?

[image]

Interpretation: Laborers and low-skill job categories have substantially higher default rates.

7️⃣ Does income level predict default probability?

[image]

Interpretation: Higher income strongly reduces default risk, making it one of the strongest predictors.

🧱 2. Feature Engineering

A structured pipeline generated several predictive feature groups:

✔ Financial Ratios

• —Credit / Income
• —Annuity / Income
• —Debt burden indicators

✔ Stability Features

• —Employment duration
• —Stability measures
• —Dependents / family load

✔ Missingness Indicators

• —MISSING_COUNT
• —Boolean missing flags

✔ Document Indicators

• —DOCUMENT_COUNT

✔ K-Means Cluster Features

• —CLUSTER_ID
• —CLUSTER_DIST

These clusters captured non-linear borrower patterns and improved model performance.

🧭 3. Clustering (K-Means + PCA)

Borrowers were clustered to reveal hidden behavioral groups.

PCA Cluster Spread

[image]

PCA Cluster Projection

[image]

Insight: Clusters displayed distinct financial & demographic traits. Including CLUSTER_ID and CLUSTER_DIST improved downstream model performance.

📈 4. Regression Model — Predicting EXTSOURCE2

Model: Random Forest Regressor

Trained on a 50,000-row sampled dataset for robustness and efficiency.

Performance (Test Set)

• —MAE: ~0.138
• —RMSE: ~0.170
• —R²: ~0.21

Interpretation: The regression model captures meaningful signals in a noisy variable and significantly outperforms linear methods.

📌 Model Deployment

🔗 https://huggingface.co/guyshilo12/guyshilo-loan-default-regression-model

🛑 5. Classification Model — Predicting Default (TARGET)

Model: Random Forest Classifier

Trained with strong emphasis on recall, the most important metric in credit-risk decisioning.

Performance (Test Set)

• —Accuracy: ~0.71
• —Precision: ~0.165
• —Recall: ~0.638
• —ROC-AUC: ~0.739

Interpretation: Recall is prioritized because missing a risky borrower (false negative) is far more costly than misclassifying a safe borrower.

📌 Model Deployment

🔗 https://huggingface.co/guyshilo12/home-credit-default-classifier

🌟 Key Insights & Conclusions

• —EXTSOURCE2 and income are among the most influential predictors.
• —Age, education, and occupation show strong default risk patterns.
• —Missing values themselves contain valuable predictive information.
• —Non-linear algorithms (Random Forest) outperform linear models significantly.
• —Cluster features improved both regression and classification models.

⚠️ Limitations

• —EXTSOURCE2 contains intrinsic noise → limits regression accuracy.
• —K-Means assumes spherical clusters; real borrower structure is more complex.
• —Random Forests lack full model interpretability.
• —Class imbalance reduces precision despite optimized recall.

🧩 Project Files

If you wish to explore the analysis further, the full notebook is included in the repository. [image] [image]

video walkthrough

you can watch the full video walkthrough here: https://drive.google.com/file/d/1NoT0LF52VE7FY8g7UtvIwXnPUXLgJM_i/view?usp=sharing

📄 License

MIT License