Bankruptcy prediction using ensemble of autoencoders optimized by genetic algorithm.

Area of Science:

• Financial modeling
• Machine learning for business analytics
• Computational finance

Background:

• Predicting corporate bankruptcy is crucial for financial institutions and stakeholders.
• Bankruptcy prediction is challenging due to complex influencing factors and imbalanced datasets.
• Existing methods often struggle with the skewed nature of bankruptcy data.

Purpose of the Study:

• To develop a robust bankruptcy prediction model.
• To address the challenge of imbalanced datasets in bankruptcy prediction.
• To improve the accuracy and reliability of identifying companies facing financial distress.

Main Methods:

• Utilized a shallow autoencoder ensemble for learning data distributions.
• Employed a genetic algorithm to optimize autoencoder thresholds for classification.
• Trained models on imbalanced datasets of small and medium-sized enterprises.
• Evaluated performance using geometric mean scores across multiple datasets.

Main Results:

• The autoencoder ensemble demonstrated strong performance in identifying bankrupt companies.
• Geometric mean scores ranged from 71% to 93.7%, indicating effective prediction capabilities.
• The genetic algorithm successfully optimized classification thresholds for improved accuracy.
• The method proved effective across different industries and evaluation periods.

Conclusions:

• The proposed shallow autoencoder ensemble optimized by a genetic algorithm offers a powerful solution for bankruptcy prediction.
• This approach effectively handles imbalanced data, a common challenge in financial risk assessment.
• The model's high performance suggests its utility for banks, government agencies, and business owners.