Machine Learning for Zombie Hunting: Predicting Distress from Firms' Accounts and Missing Values

Abstract

In this contribution, we propose machine learning techniques to predict zombie firms . First, we derive risk of failure by training and testing our algorithm on disclosed financial information and non-random missing values by 304,906 firms active in Italy in the period 2008-2017. Then, we spot highest distress conditional on the predicted risk being in the last decile of the risk of failure distribution as this is the threshold after which the observed chances of firms transiting to a lower risk of failure are negligible. We identify zombies as firm that persist in a status of high risk of failure based on their permanence in an above-the-last-decile predicted risk status. For our predictive purpose, we implement a rework of the Bayesian Additive Regression Tree with Missingness Incorporated in Attributes (BART-MIA) which is specifically useful in our setting as we provide evidence that patterns of undisclosed accounts correlate with firms’ failures. We show that BART-MIA outperforms (i) proxy models like the Z-scores and the Distance-to-Default, (ii) traditional econometric methods, and (iii) other widely used machine learning techniques. Eventually, we document that zombies are on average less productive, smaller, and they tend to increase in times of crisis. In general, we argue that our application can be of help to financial institutions and public authorities in the design of evidence-based policies – e.g., optimal bankruptcy laws.

Keywords: zombie firms, Machine Learning, Bayesian statistical learning, financial constraints, bankruptcy, missing data