Company Bankruptcy Prediction

Course: CSE343 / ECE363: Machine Learning
Instructor: Dr. Jainendra Shukla
Institution: IIIT-Delhi
Semester: Monsoon 2024
Dataset: Company Bankruptcy Prediction Dataset – Taiwan Economic Journal

---

Project Overview

This project explores the application of machine learning models to predict the financial sustainability of enterprises. Given a high-dimensional dataset with 96 financial indicators from Taiwanese firms (1999–2009), we built a classification framework to predict whether a company is at risk of bankruptcy. Our motivation stems from the need for early financial distress warning systems that aid policymakers, investors, and stakeholders in mitigating risks.

---

Objectives

• Analyze financial health indicators to predict bankruptcy risks.
• Handle real-world challenges such as class imbalance and multicollinearity.
• Evaluate and compare multiple ML models including ensemble and deep learning methods.
• Identify key financial features through statistical feature selection techniques.

---

Dataset Description

• Source: Taiwan Economic Journal (1999–2009)
• Size: 6819 records, 96 features
• Target Variable: Bankrupt? (binary classification: 0 or 1)
• Feature Types: Financial ratios (e.g., ROA, Asset Growth, Revenue/Share)
• Kaggle Dataset

---

Methodology

Exploratory Data Analysis (EDA)

• Detected severe class imbalance (only ~100 bankrupt firms)
• Identified multicollinearity via correlation matrix
• Outlier detection in critical financial features

Data Preprocessing

• Class imbalance handled using:
  RandomOverSampler, RandomUnderSampler, SMOTE, ADASYN, CondensedNearestNeighbor, and Tomek Links
• Multicollinearity resolved using correlation-based feature pruning
• Feature selection via ANOVA (numerical) and Chi-square (categorical)

---

Models Implemented

Model	Highlights
Logistic Regression	Baseline linear model
Random Forest	Robust ensemble method with good generalization
Naive Bayes	Probabilistic model assuming independence
SVM (Linear, RBF, Poly)	Effective in high-dimensional space
MLP Classifier	Feed-forward neural network for non-linearity
Custom ANN	Custom deep learning implementation
XGBoost	Gradient boosting decision trees for accuracy & speed

---

Results (Test Set Highlights)

Model	Accuracy	Precision	Recall	F1 Score	ROC AUC
Logistic Regression	80.5%	0.49	0.88	0.63	0.88
Random Forest	89.4%	0.74	0.67	0.70	0.91
XGBoost	88.5%	0.69	0.69	0.69	0.91
MLP Classifier	85.4%	0.64	0.50	0.56	0.89

Note: Due to the critical nature of predicting bankruptcies, recall was prioritized over precision.

---

Key Takeaways

• High recall models (SVMs, Logistic Regression) are crucial for real-world bankruptcy risk minimization.
• Feature selection significantly improved performance and interpretability.
• Sampling strategies (especially CondensedNearestNeighbor) helped handle severe class imbalance.
• XGBoost and Random Forest provided the best trade-off between performance and generalization.

---

References

• Barboza et al., Machine learning models and bankruptcy prediction, ESWA, 2017
• Nanxi Wang, Bankruptcy Prediction Using ML, JMF, 2017
• Danilov, Corporate Bankruptcy: Assessment & Prediction, MIT Sloan
• Kaggle Dataset