This paper aims to predict corporate insolvency by comparing machine learning (ML) and analytical formulas applied on 8 financial ratios.
This study uses a dataset of 388,145 solvent and 842 insolvent Spanish companies in a total of 27 simulations, employing 3 analytical formulas (Springate, Grover, Amat), 3 ML models (K-nearest neighbours (KNN), random forest (RF), multi-layer perceptron (MLP)) and 3 sampling approaches (original dataset, downsampling and upsampling).
The dataset imbalance between solvent and insolvent companies matters in the prediction capabilities of the ML models; when balanced through downsampling of solvent companies or upsampling the insolvent ones, the ML models outperform the results of the analytical ones. The best performing ML models are built upon the ratios defined in Amat model, a particularization of Altman’s Z-score to Spanish companies.
This paper compares different sampling approaches, whereas literature is more focused on downsampling due to information availability. This paper also predicts insolvency from a small amount of financial ratios, rather than using many financial variables, more common in literature.
