A novel method for financial distress prediction based on sparse neural networks with regularization

Area of Science:

• Corporate Finance
• Machine Learning
• Accounting Theory

Background:

• Corporate financial distress poses significant risks to enterprises and stakeholders.
• Existing prediction models are often limited by variable selection or reliance solely on financial data.
• Integrating diverse predictors is essential for robust financial distress prediction.

Purpose of the Study:

• To develop a novel financial distress prediction (FDP) model using sparse neural networks (FDP-SNN).
• To investigate the predictive power of both financial and non-financial variables in corporate financial distress.
• To enhance prediction accuracy and model interpretability.

Main Methods:

• Screening financial and non-financial variables based on accounting and finance theory.
• Developing a sparse neural network (FDP-SNN) with L1 regularization for variable selection.
• Comparing FDP-SNN performance against classic prediction models.

Main Results:

• Non-financial variables (e.g., investor protection, governance) are key predictors, especially for longer forecast periods.
• The FDP-SNN model demonstrates superior accuracy, precision, and AUC compared to traditional methods.
• The model's sparsity enhances interpretability by identifying important predictors.

Conclusions:

• Sparse neural networks offer a powerful approach for financial distress prediction.
• Non-financial factors significantly contribute to predicting corporate financial distress.
• FDP-SNN provides a more accurate, interpretable, and robust method for financial risk assessment.