Predicting Loan Payback (Kaggle Playground Series S5E11)

📌 Project Overview

This project focuses on predicting whether a loan will be paid back using structured tabular data. The solution was developed entirely with Python scripts (.py), without notebooks, following a clean and reproducible end‑to‑end machine learning pipeline.

The project is based on the Kaggle competition Playground Series – Season 5, Episode 11 and is designed as a portfolio‑ready Data Science / Machine Learning project.

---

🧠 Problem Statement

Given customer financial and demographic data, predict the probability that a loan will be paid back (loan_paid_back).

This is a binary classification problem, evaluated using ROC AUC.

---

🗂️ Project Structure

playground-series-s5e11/
│
├── data/                 # Raw competition data (train.csv, test.csv)
├── outputs/              # EDA reports, metrics, trained models, submission
│   ├── eda_report.txt
│   ├── metrics.txt
│   ├── catboost_model.cbm
│   └── submission.csv
│
├── src/                  # Source code (pure Python, no notebooks)
│   ├── config.py
│   ├── load_data.py
│   ├── eda.py
│   ├── train_catboost.py
│   └── inference.py
│
├── requirements.txt
└── README.md

---

⚙️ Tech Stack

• Python 3.11+
• Pandas, NumPy
• Scikit‑learn
• CatBoost (native categorical feature handling)

---

🔍 Exploratory Data Analysis (EDA)

EDA is performed via a standalone Python script and saved as a text report:

• dataset shapes
• column overview
• target distribution
• missing values check
• data types

Output:

outputs/eda_report.txt

---

🤖 Model

CatBoostClassifier was selected due to:

• native handling of categorical features
• strong performance on tabular data
• minimal preprocessing requirements
• robustness and stability

Categorical Features

['gender', 'marital_status', 'education_level',
 'employment_status', 'loan_purpose', 'grade_subgrade']

---

📊 Results

Metric	Score
Public Leaderboard AUC	0.92293
Private Leaderboard AUC	0.92385
OOF AUC	0.92338 ± 0.00069

These results indicate a stable and well‑generalizing model.

---

▶️ How to Run Locally

1️⃣ Create virtual environment

python -m venv venv
venv\Scripts\activate

2️⃣ Install dependencies

pip install -r requirements.txt

3️⃣ Run EDA

python -m src.eda

4️⃣ Train model

python -m src.train_catboost

5️⃣ Generate submission

python -m src.inference

---

🧪 Key Design Decisions

• ❌ No Jupyter notebooks
• ✅ Script‑based, reproducible pipeline
• ✅ Clear separation of concerns (EDA / training / inference)
• ✅ Local development (VSC‑friendly)
• ✅ Ready for extension (SHAP, feature engineering, hyperparameter tuning)

---

🚀 Future Improvements

• Feature engineering (ratio & interaction features)
• SHAP‑based model interpretability
• Hyperparameter optimization
• Model ensembling

---

📎 Kaggle

Competition: Playground Series S5E11

Submission performed via Late Submission (learning & portfolio purposes).

---

👤 Author

Grzegorz

Focused on Data Science and Machine Learning with emphasis on clean pipelines, reproducibility, and production‑ready code.