Predicting corporate defaults using machine learning with geometric-lag variables

ABSTRACT

This study examines whether corporate default prediction techniques based on machine learning can achieve better performance by using geometrically declining weighted average values of the time series variables, that is, geometric-lag variables. We test four machine learning algorithms: logistic regression, random forest, support vector machine, and feedforward neural network. The geometric-lag financial variables capture each company’s historical financial information. Using such variables reduces the computation time and improves the prediction performance. The actual default rates increase with the predicted default probabilities, suggesting that our model predictions can help investors make better investment decisions.

KEYWORDS:

• classification
• corporate default prediction
• geometric lag
• machine learning
• risk measure

JEL CLASSIFICATION:

• G12
• G17
• G33

Acknowledgments

This work was supported by the 2019 Yeungnam University Research Grant.

Disclosure statement

No potential conflict of interest was reported by the authors.

Notes

1 Inactive firms in the Compustat database include the following: i) trading suspended or halted; ii) called, expired, or matured; iii) merger; iv) acquisition; v) untraded or unquoted; and vi) delisted. DRLSN (“Reason for Deletion”) codes include i) acquisition or merger, ii) bankruptcy, iii) liquidation and other miscellaneous items.

2 A memory overflow occurs when the training process exhausts all memory available to the training machine, generating an out-of-memory error. To avoid this error, we recommend using fewer observations and feature variables.