Corporate leverage in the US: economic explanation and machine learning exploration
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
This dissertation studies corporate leverage in the United States from the dual perspectives of economic explanation and machine learning prediction. Corporate leverage in our study is defined as the ratio of debt to assets. Our economic study proposes and investigates an explanation for the coexistence of two leverage patterns across US firms: low (or even zero) leverage for some firms and also some firms with very high leverage, even as the mean leverage ratio remains fairly stable over time. These patterns of the cross-sectional distribution of leverage have become even more pronounced in recent years. ☐ The explanation we propose centers on production flexibility, defined as the ability of firms’ to quickly adjust their output in response to demand conditions. Production flexibility is found to have increased in recent years, and is a companion pattern to the low-leverage pattern and the increasing high-leverage pattern. We posit that the combined influence of production flexibility along with the state of a firm’s financial constraints and its lenders’ concerns about being taken advantage of by managers or shareholders leads to a plausible explanation for the patterns in leverage. The core idea is that when financial constraints do not exist, high production flexibility indicates a low operation risk, which enables firms to take a high financial risk from debt as a substitute. Furthermore, firms where agency problem concerns are not so great can easily benefit from production flexibility to achieve growth, and the growth leads to lower leverage. However, for firms where agency problems are of greater concern to lenders, the production flexibility might be used to the detriment of lenders, leading to a positive relationship between leverage and production flexibility in the cross-section. ☐ Our model predicts that, as production flexibility has increased over the decades, we should observe a stronger low-leverage pattern for firms with less serious agency problems, and we should observe growing leverage for firms with more serious agency problems in extreme cases. We conduct an empirical investigation of these issues, and find results consistent with this reasoning, especially for firms listed in Nasdaq. We do not propose that growing production flexibility is the only reason for the leverage patterns we investigate, but we believe the reasoning and evidence suggests it is part of the explanation. ☐ Secondly, we conduct a study from a machine learning perspective. Our empirical finance study is conducted using standard panel-data regressions methods. Such methods typically involve controlling for a large set of unexplained ``fixed effects” in the cross-section, leaving only part of the variation to be explained by economic factors. The fixed effects are, mechanically, dummy variables whose coefficients need to be estimated. In the context of relatively short time series relative to the breadth of the cross-section, panel-data regressions are not usually considered to be useful for prediction. For prediction purposes, we reason that it may be useful to avoid using the information in the data to estimate pre-specified fixed effects, but instead to introduce clustering dummies only as needed in the model training process sequentially. The training algorithm we suggest is a variation on the traditional machine learning method “Gradient Boosting Machine” (GBM for short) and is called "PanelGBM" in our study. This approach relaxes assumptions of linearity which are inherent in a standard panel-data regression approach as well as other assumptions regarding clusters and data in other machine learning models that are designed for panel data analysis. Via a simulation study, we establish that our algorithm provides good accuracy and stability relative to other prediction methods for a variety of underlying true models and these clustering dummy variables that are learned in the model training process help improving model performance significantly. The algorithm could have applications in a variety of settings; in the context of this dissertation, we apply the algorithm to predict leverage for US firms and find that it performs much better than other models. Even a linear model with the clustering dummies learned from the PanelGBM performs much better than traditional panel regression model.
Description
Keywords
PanelGBM, Agency problem, Capital structure, Financial constraints, Machine learning, Production flexibility