📊 Business Bankruptcy Risk Prediction (ML)

A complete machine learning pipeline that predicts company bankruptcy risk using financial ratios. Built with Python + R, featuring 5 ML models, statistical analysis, and an interactive prediction web app.

Overview

This project builds and compares multiple machine learning models for predicting business bankruptcy, following the methodology from:

• Olson & Wu (2017) — Enterprise Risk Management Models (Decision Tree, Logistic Regression, Neural Network)
• Sundararajan (2025) — Multivariate Analysis and Machine Learning Techniques (Ensemble methods, Feature Selection)

Dataset

Taiwan Economic Journal Company Bankruptcy Prediction (Kaggle)

Property	Value
Records	6,819 companies
Features	95 financial ratios
Target	Binary (Bankrupt: 3.2%, Survived: 96.8%)
Period	1999–2009

Results

Model	Accuracy	Recall	F1-Score	AUC-ROC
Logistic Regression	0.872	0.773	0.281	0.917
Decision Tree	0.890	0.818	0.324	0.906
Random Forest ★	0.932	0.773	0.422	0.945
XGBoost	0.930	0.682	0.387	0.918
Neural Network	0.960	0.432	0.409	0.881

★ Best model: Random Forest (highest F1-Score + AUC-ROC + cross-validation stability)

Project Structure

business-risk-ml/
├── data/
│   ├── data.csv                    # Raw Kaggle dataset (download separately)
│   ├── README.md                   # Dataset download instructions
│   └── processed/                  # Preprocessed train/test splits
│       ├── X_train.csv
│       ├── X_test.csv
│       ├── y_train.csv
│       ├── y_test.csv
│       └── metadata.json
├── notebooks/
│   ├── 01_eda.py                   # Step 2: Exploratory Data Analysis
│   ├── 02_preprocessing.py         # Step 3: Data Preprocessing Pipeline
│   ├── 03_model_training.py        # Step 4: Train 5 ML Models
│   └── 04_evaluation.py            # Step 5: Model Evaluation & Comparison
├── src/
│   ├── __init__.py
│   ├── data_loader.py              # Data loading & validation utilities
│   └── preprocessor.py             # Reusable preprocessing functions
├── R_analysis/
│   ├── install_packages.R          # R package installation
│   ├── load_data.R                 # Load Python-processed data into R
│   └── model_comparison.R          # R statistical analysis (Step 5)
├── models/
│   ├── best_model.pkl              # Best performing model (Random Forest)
│   ├── logistic_regression.pkl
│   ├── decision_tree.pkl
│   ├── random_forest.pkl
│   ├── xgboost_model.pkl
│   ├── neural_network.pkl
│   └── scaler.pkl                  # StandardScaler for new predictions
├── app/
│   └── streamlit_app.py            # Step 6: Interactive prediction web app
├── figures/                        # All generated charts (01-15)
├── config.py                       # Project configuration
├── requirements.txt                # Python dependencies
├── .gitignore
└── README.md

Quick Start

1. Clone & Install

git clone https://github.com/OVuyo/business-risk-ml.git
cd business-risk-ml

# Python
pip install -r requirements.txt

# R (optional, for statistical analysis)
Rscript R_analysis/install_packages.R

2. Download Data

Download data.csv from Kaggle and place it in data/.

3. Run the Pipeline

# Step 2: Exploratory Data Analysis
python notebooks/01_eda.py

# Step 3: Preprocessing (outliers, feature selection, SMOTE, scaling)
python notebooks/02_preprocessing.py

# Step 4: Train all 5 models
python notebooks/03_model_training.py

# Step 5: Evaluate and compare models
python notebooks/04_evaluation.py

# Step 5b: R statistical analysis (optional)
Rscript R_analysis/model_comparison.R

# Step 6: Launch the prediction web app
streamlit run app/streamlit_app.py

Pipeline Details

Preprocessing (Step 3)

Raw Data (6,819 × 96)
  → Winsorization (cap outliers at 1st/99th percentile)
  → Multicollinearity removal (drop features with |r| > 0.95)
  → Feature selection (top 30 by Mutual Information)
  → Stratified train/test split (80/20)
  → SMOTE oversampling (176 → 5,279 bankrupt examples)
  → StandardScaler normalization
Clean Data (10,558 × 30 training / 1,364 × 30 testing)

Models (Step 4)

1. Logistic Regression — Linear baseline with L2 regularization
2. Decision Tree — Interpretable rules (max_depth=8, like Rattle/R)
3. Random Forest — 200 trees with √features per split (bagging)
4. XGBoost — Gradient boosted trees (sequential boosting)
5. Neural Network — MLP with 3 hidden layers (64→32→16, ReLU)

Key Findings

• Top predictors: Persistent EPS, Net Income ratios, Borrowing dependency
• Class imbalance (30:1) addressed with SMOTE — critical for recall
• Random Forest achieves best balance of recall (77.3%) and precision (29.1%)
• Decision Tree has highest recall (81.8%) but lowest precision
• Results align with Olson & Wu's finding that ensemble methods outperform single models

Technologies

Python: pandas, scikit-learn, XGBoost, Keras/TensorFlow, matplotlib, seaborn, Streamlit, imbalanced-learn

R: caret, randomForest, rpart, ggplot2, corrplot, pROC

References

1. Olson, D.L. & Wu, D.D. (2017). Enterprise Risk Management Models. Springer.
2. Sundararajan, S. (2025). Multivariate Analysis and Machine Learning Techniques. Springer.
3. Olson, D.L., Delen, D. & Meng, Y. (2012). Comparative analysis of data mining methods for bankruptcy prediction. Decision Support Systems, 52(2), 464–473.
4. Dataset: Taiwan Economic Journal via Kaggle.

License

MIT License — see LICENSE for details.