CEU Electronic Theses and Dissertations, 2025
| Author | Molnar, Marton |
|---|---|
| Title | A Comparative Analysis of Distress Prediction in Public Firms: The Impact of Macroeconomic Conditions, Industry Dynamics, and Nonlinear Modeling |
| Summary | Reliable prediction of corporate financial distress is critical for effective risk management and maintaining economic stability, especially given the systemic vulnerabilities revealed by the 2008 financial crisis. Traditional models, that rely on accounting ratios, are limited by linear assumptions and poor adaptability across varying economic conditions. This thesis compares classical econometric models (logistic regression) with advanced machine learning algorithms (extreme gradient boosting and shallow neural networks) under a standardized framework to address three research questions: Do macroeconomic indicators meaningfully improve predictive accuracy beyond accounting metrics alone? Do sector-specific models outperform general cross-industry models? Do advanced machine learning models offer sufficiently better performance to justify their complexity? To answer these questions, a novel binary distress indicator is defined that reflects significant financial underperformance rather than legal insolvency and is scaled to align with historical large-cap bankruptcy rates. This label is then applied to a panel dataset of 490 listed U.S. firms (2015-2024) in sectors with historically higher bankruptcy rates. Results show that XGBoost significantly outperforms other models, achieving accuracy over 97% and precision-recall (PR-AUC) between 65-69% during expansionary periods, with slightly reduced but robust performance in volatile conditions (94% accuracy; 55% PR-AUC). While macroeconomic and sector-specific variables occasionally improved predictive accuracy, they generally introduced unnecessary complexity and reduced interpretability without significant practical advantages. Practitioners are recommended to prioritize flexible machine learning models, like XGBoost, relying on core financial ratios, and periodically retrain these models to ensure consistent accuracy in changing economic conditions. Key limitations, including moderate dataset size, class imbalance, and indirect validation of distress cases, suggest ways for further research, such as incorporating external distress validation or exploring recurrent neural architectures. |
| Supervisor | Békés, Gábor |
| Department | Undergraduate Studies BA |
| Full text | https://www.etd.ceu.edu/2025/molnar_marton.pdf |
Visit the CEU Library.
© 2007-2025, Central European University