Bankruptcy risk prediction: A new approach based on compositional analysis of financial statements
Year: 2025
Type: article
Abstract: The development of models for bankruptcy risk prediction has gained much attention in recent years due to the great availability of financial statement data. Most existing predictive models rely on financial ratios, which are performance-based measures expressing the relative magnitude of two accounting items. Despite the popularity of financial ratios, their use is notoriously accompanied by serious practical drawbacks, like the occurrence of outliers and redundancy, making data preprocessing necessary to avoid computational problems and obtain a good predictive accuracy. Isometric log ratios can potentially overcome these problems because they are designed to represent compositional data efficiently and have a logarithmic form that limits the occurrence of outliers. However, although they are not novel in the analysis of financial statements, no study has ever employed them to predict bankruptcy . In this article, we show the effectiveness of isometric log ratios to detect bankruptcy events in a sample of 138,720 Italian firms (127,420 active and 11,300 bankrupted) belonging to different industries and with different size and age. For this purpose, we use logistic regression with adaptive LASSO regularization and random forests to construct several predictive models featuring either financial ratios or isometric log ratios, and combining different horizons and lag structures. The results show that a set of 8 isometric log ratios provides, without preprocessing, almost the same predictive accuracy as a selection of 16 financial ratios that requires dropping 3.6% of the data. Also, the adaptive LASSO regularization reveals that redundancy for isometric log ratios is always below 20%, and in some cases near 0%, while it ranges from 12.5% to 46.9% for financial ratios. The predictive accuracy of models based on logistic regression is in line with and even higher than the one reported by recent studies, and random forests achieve a gain in the area under the Receiver Operating Characteristic (ROC) curve ranging between two and three percentage points.
Source: Big Data Research
Author: Alessandro Magrini
Institution: University of Florence
Language: English
Subfield: Accounting
Domain: Social Sciences
Open Access status: hybrid
Estimated APC paid: $2,840
OpenAlex ID: W4410629587