This thesis deals with bankruptcy prediction, and in particular compares the most recent machine learning models with the classical models that finance has produced in this field. The study focuses on Italian SMEs active between 2015 and 2017, extracting from the AIDA database all firms, both bankrupt and active, that meet the SME parameters, thus providing a real dataset on which the performance of the models can be compared in a realistic way. Financial ratios were calculated year by year from the financial statements, with the addition of dynamic variables such as the standard deviation of the ratios, with the aim of predicting which firms went bankrupt the following year. The ratio of healthy to bankrupt firms in the dataset is 67:1, and the total number of firms analysed after applying several control filters is 141,377, with 139,288 healthy and 2,089 bankrupt firms. The dataset was then divided into a training set and a test set with a 70/30 ratio. The imbalance was not corrected through resampling; instead, the classification threshold of each model was set at the optimal point of the ROC curve on the training set. The machine learning models tested in the thesis are GLM, LASSO, KNN, decision tree, random forest, gradient boosting, support vector machine, neural network and stacking, while the classical models are Ohlson, Altman Z′, Altman Z″ and Zmijewski. The machine learning models were trained on the training set, then tested on the test set (without the target variable available) and compared with the results of the classical models. The results show that the machine learning models outperform the classical ones on average, with gradient boosting and KNN reaching a ROC AUC of 0.90 and 0.87 respectively, while stacking stands out from all the others with a PR AUC of 0.427. The classical models still produced good results, with a ROC AUC always above 0.75 and, in the case of Zmijewski, a PR AUC of 0.210, higher than that of all the individual machine learning models. These results show us that today's machine learning models can improve the prediction of corporate bankruptcy, while still using classic models such as Zmijewski's as a benchmark.
Questa tesi tratta la previsione della bancarotta e, in particolare, mette a confronto i più recenti modelli di machine learning con i modelli classici che la finanza ha prodotto in questo ambito. Lo studio si concentra sulle PMI italiane attive tra il 2015 e il 2017, estraendo dal database AIDA tutte le aziende, sia fallite sia attive, che rispettano i parametri di PMI, fornendo così un dataset reale su cui confrontare in modo realistico le performance dei modelli. Dai bilanci sono stati calcolati gli indici anno per anno, con l'aggiunta di variabili dinamiche come la deviazione standard degli indici, con l'obiettivo di prevedere quali aziende sarebbero fallite l'anno successivo. Il rapporto tra aziende sane e fallite nel dataset è di 67:1, e il totale delle aziende analizzate dopo l'applicazione di vari filtri di controllo è di 141.377, di cui 139.288 sane e 2.089 fallite. Il dataset è stato poi diviso in training set e test set con un rapporto 70/30. Lo sbilanciamento non è stato corretto tramite ricampionamento; la soglia di classificazione di ciascun modello di machine learning è stata invece fissata nel punto ottimale della curva ROC sul training set. I modelli di machine learning testati nella tesi sono GLM, LASSO, KNN, decision tree, random forest, gradient boosting, support vector machine, rete neurale e stacking, mentre i modelli classici sono Ohlson, Altman Z′, Altman Z″ e Zmijewski. I modelli di machine learning sono stati allenati sul training set, poi testati sul test set (senza la variabile target disponibile) e confrontati con i risultati dei modelli classici. I risultati mostrano che i modelli di machine learning superano in media quelli classici, con il gradient boosting e il KNN che raggiungono rispettivamente una ROC AUC di 0,90 e 0,87, mentre lo stacking si distingue da tutti gli altri con una PR AUC di 0,427. I modelli classici hanno comunque prodotto buoni risultati, con una ROC AUC sempre superiore a 0,75 e, nel caso di Zmijewski, una PR AUC di 0,210, superiore a quella di tutti i singoli modelli di machine learning. Questi risultati mostrano che i modelli di machine learning attuali possono migliorare la previsione del fallimento aziendale, mentre modelli classici come quello di Zmijewski restano un valido benchmark.
Bankruptcy Prediction on Italian SMEs: A Comparative Analysis from Z-Score to Machine Learning
FERRARI, RICCARDO
2025/2026
Abstract
This thesis deals with bankruptcy prediction, and in particular compares the most recent machine learning models with the classical models that finance has produced in this field. The study focuses on Italian SMEs active between 2015 and 2017, extracting from the AIDA database all firms, both bankrupt and active, that meet the SME parameters, thus providing a real dataset on which the performance of the models can be compared in a realistic way. Financial ratios were calculated year by year from the financial statements, with the addition of dynamic variables such as the standard deviation of the ratios, with the aim of predicting which firms went bankrupt the following year. The ratio of healthy to bankrupt firms in the dataset is 67:1, and the total number of firms analysed after applying several control filters is 141,377, with 139,288 healthy and 2,089 bankrupt firms. The dataset was then divided into a training set and a test set with a 70/30 ratio. The imbalance was not corrected through resampling; instead, the classification threshold of each model was set at the optimal point of the ROC curve on the training set. The machine learning models tested in the thesis are GLM, LASSO, KNN, decision tree, random forest, gradient boosting, support vector machine, neural network and stacking, while the classical models are Ohlson, Altman Z′, Altman Z″ and Zmijewski. The machine learning models were trained on the training set, then tested on the test set (without the target variable available) and compared with the results of the classical models. The results show that the machine learning models outperform the classical ones on average, with gradient boosting and KNN reaching a ROC AUC of 0.90 and 0.87 respectively, while stacking stands out from all the others with a PR AUC of 0.427. The classical models still produced good results, with a ROC AUC always above 0.75 and, in the case of Zmijewski, a PR AUC of 0.210, higher than that of all the individual machine learning models. These results show us that today's machine learning models can improve the prediction of corporate bankruptcy, while still using classic models such as Zmijewski's as a benchmark.| File | Dimensione | Formato | |
|---|---|---|---|
|
Ferrari_Riccardo.pdf
accesso aperto
Dimensione
7.48 MB
Formato
Adobe PDF
|
7.48 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/118990