Can Machines Learn Capital Structure?
51 Pages Posted: 22 Oct 2019 Last revised: 3 Dec 2019
Date Written: December 2, 2019
Yes, they can! Machine learning models that exploit big data can consistently predict corporate leverage better than classical methods over time and identify its determinants. Using a large sample of U.S. firms from 1972 to 2018, we apply random forests, gradient boosting machines, neural networks, and generalized additive models to predict corporate leverage and the contributing factors. Results show that machine learning models that allow for nonlinearities and complex interactions boost the out-of-sample R-squared from 36% to 56% over linear methods such as LASSO. The superior predictive performance occurs every year of the out-of-sample period at the aggregate level as well as subsamples such as firms undergoing a major capital restructuring. Additionally, machine learning methods consistently identify the determinants of corporate leverage over time. Our best performing model, a random forest, selects market-to-book, industry median leverage, cash & equivalents, Z-Score, profitability, stock returns, and firm size as robust and reliable predictors of market leverage. These findings suggest that despite leverage climbing to record highs, firm fundamentals drive corporate leverage because the determinants and predictability of corporate leverage have remained stable.
Keywords: Machine Learning, Capital Structure, Corporate Leverage, Random Forest, Gradient Boosting Machine, Neural Network, Generalized Additive Model, LASSO
JEL Classification: G0, G17, G30, G32, C10, C50
Suggested Citation: Suggested Citation