guyshilo-loan-default-regression-model

Home Credit – EXT_SOURCE_2 Regression Model (Random Forest)

This repository contains the winning regression model trained to predict the external credit score (EXT_SOURCE_2) in the Home Credit Default Risk dataset.
The model uses extensive feature engineering and captures non-linear financial and demographic patterns using a Random Forest Regressor.

---

📁 Files Included

• regression_winner.pkl — trained Random Forest regression model
• feature_columns.json — ordered list of all input features the model expects

Both files must be used together for inference.

---

🧠 Model Description

Task: Regression
Target: EXT_SOURCE_2 (external credit score proxy)
Model: RandomForestRegressor

Why Predict EXT_SOURCE_2?

EXT_SOURCE_2 is a powerful creditworthiness indicator in the Home Credit dataset.
However, it contains many missing values, and being able to predict it from other borrower features helps improve downstream models.

---

🏗️ Feature Engineering Summary

The model was trained on a heavily engineered dataset including:

✔ Financial Features

• Total income
• Credit amount
• Annuity
• Credit/income ratio
• Annuity/income ratio

✔ Demographic Features

• Age
• Employment duration
• Family size
• Dependents

✔ Stability Indicators

• Employment stability
• Income consistency
• Relative wealth vs. region

✔ Missing Value Indicators

• MISSING_COUNT
• Boolean missing indicators for key features

✔ Document Indicators

• DOCUMENT_COUNT (how many documents applicant supplied)

✔ Cluster Features (K-Means)

• CLUSTER_ID
• CLUSTER_DIST

These help the model capture non-linear borrower profiles that ordinary linear regression cannot.

---

⚙️ Training Configuration

RandomForestRegressor(
    n_estimators=80,
    max_depth=12,
    min_samples_leaf=50,
    n_jobs=-1,
    random_state=42
)

---

Why these hyperparameters?

• max_depth=12 → prevents overfitting

• min_samples_leaf=50 → forces smoothing and stability

• n_estimators=80 → balances accuracy and training speed

• n_jobs=-1 → uses all CPU cores

Model trained on 50,000 sampled rows for fast and robust performance.

---

📊 Model Performance (Held-Out Test Set)

---

Interpretation

An R² ≈ 0.21 means the model captures some meaningful signal but the dataset contains a lot of noise.

Performance is much stronger than a linear model, which scored ~0.18 R².

The model benefits significantly from:

Ratio features

Missing-value patterns

K-Means cluster structure

---

Why this model won

✔ Outperformed Linear Regression on all metrics ✔ Handles non-linear interactions between income, credit, stability, and documents ✔ Much less sensitive to noise and missingness ✔ Best trade-off between accuracy, generalization, and speed

---

⬇️ How to Download the Model

from huggingface_hub import hf_hub_download

model_path = hf_hub_download(
    repo_id="guyshilo12/guyshilo-loan-default-regression-model",
    filename="regression_winner.pkl"
)

cols_path = hf_hub_download(
    repo_id="guyshilo12/guyshilo-loan-default-regression-model",
    filename="feature_columns.json"
)

---

📦 How to Load the Model

import pickle, json

# Load model
with open("regression_winner.pkl", "rb") as f:
    model = pickle.load(f)

# Load feature list
with open("feature_columns.json", "r") as f:
    feature_cols = json.load(f)

---

🔮 How to Run a Prediction

import pandas as pd

# Example input (replace zeros with real values)
X = pd.DataFrame([{col: 0 for col in feature_cols}])

pred = model.predict(X)[0]

print("Predicted EXT_SOURCE_2:", pred)

---

🏆 Intended Use

This model is ideal for:

Academic ML projects

Feature engineering demonstrations

Credit-risk modeling research

Improving missing-value imputation strategies

Not intended for production lending decisions without regulatory review.