The Laboratory for Comparative Social Research announces the next regular seminar, which will be held as a zoom session on February, 6th, at 02:30 p.m. CET (04:30 p.m. Moscow time, GMT+3). Dmitriy Skougarevskiy & Sergey Bondarkov (Institute for the Rule of Law, European University at Saint Petersburg) will deliver a report "Non-accounting Data Improves Bankruptcy Predictions: Evidence from Benford’s Law and Litigation". To participate, please, register via the link. 

Abstract

Bankruptcy prediction models typically use as features financial ratios derived from accounting statements. We extend this feature set by novel variables measuring good standing of a firm: conformity of the first significant digits in its financial statements with Benford's law and incidence of bankruptcy-unrelated litigation. The former measure relies on the long known observation that deviations from Benford's law may signal data manipulation. The latter builds on the intuition that would-be bankrupts face litigation related to their contractual non-performance before filing for bankruptcy. In a novel data set covering the near-universe of 2~million Russian firms in 2012--2018 we find feature importance and predictive capability of litigation-based variables on par to most of the financial ratios traditionally used in bankruptcy prediction, both for larger firms and small and medium-sized enterprises (SMEs). In contrast, conformity of statements with Benford's law is revealed as a modest bankruptcy predictor and an unimportant feature. We estimate that lenders of larger firms employing gradient-boosted decision trees-based bankruptcy prediction models with litigation variables could have increased their Expected Maximum Profit (EMP) by USD815 mln worth of total liabilities in 2018 in Russia in comparison with traditional financial ratio-based models. This is not the case for SME lenders: their EMP gain is estimated at only USD14 mln. This suggests that the search for improved bankruptcy models for SMEs may have little to no practical significance in certain real-life contexts.

Everyone interested is invited!
Working language is English.