Document Type

Thesis

Comments

Editorial Reviewer: Alicia Lamere, Ph.D

First Faculty Advisor

Son Nguyen, Ph.D

Keywords

Bankruptcy; bankruptcy prediction; classification; macroeconomic variables

Publisher

Bryant University

Rights Management

This work is licensed under a CC BY license.

Abstract

At its core, bankruptcy prediction is a binary classification problem where a researcher attempts to model a company’s financial status, defined as either bankrupt or non-bankrupt, based upon a slew of financial ratios, market indicators and even macroeconomic variables. Several studies (Altman and McGrough, 1974; Moyer, 1977 and Mensah, 1984) have noted that such models tend to suffer reduced accuracy when predicting bankruptcy for time periods other than the one in which they were trained. A possible solution to this problem is to include macroeconomic variables in the model, since such variables fluctuate over time and are suspected of impacting the stresses felt by firms (Altman and McGrough, 1974; Lev, 1974 and Mensah, 1984). This study tests this theory by examining the performance over time of several gradient boosting machine (GBM) models, with and without macroeconomic variables. Models in this study were trained on data collected between 1998 and 2000 and then tested on data extending from the training period up until 2018. Ultimately it was found that macroeconomic variables do little to improve the predictive power of bankruptcy prediction models. Furthermore, little evidence was found that model performance fluctuates over time, which suggests that the factors which determine bankruptcy do not change across time periods.

Share

COinS