Estimating Financial Constraints with Machine Learning

Abstract

We use random decision forests, a statistical machine learning technique, to classify firms’ equity and debt constraints using only firm-level financial information. We train our model on the Hoberg and Maksimovic (2015) text-based measures, which are informative, but lack coverage. By mapping to financial variables we are able to extend the coverage of the text-based measures both in the cross-section and the time series, increasing the number of classified firm-years by 245%. Our method captures important non-linearities and interactions between financial variables and constraints and exhibits significant out-of-sample performance. We assess the informativeness of our constraint classifications using multiple tests – tests that commonly-used indices have previously been shown to fail. Our classifications perform well. They even outperform the Hoberg and Maksimovic (2015) measures over a similar sample period likely due to the increased coverage.

Keywords: financial constraints, machine learning, random forests, equity recycling

JEL Classification: G00, G30, G35, G1, G3