Study for the bankruptcy prediction using machine learning: modeling bankruptcy risk of companies based on financial ratios, using supervised learning to support credit and investment decisions

In an economic environment characterized by increasing uncertainty, competitive pressure, and financial complexity, the ability to monitor corporate performance and identify early signals of financial distress has become strategically relevant for managers, investors, regulators, and policymakers. The assessment of firms financial conditions has long played a central role in business decision-making, and its importance has further increased with the development of quantitative models for bankruptcy prediction. The objective of this thesis is to investigate the practical applicability and interpretability of predictive bankruptcy models by combining traditional financial analysis with modern data-driven techniques. The empirical analysis is based on a dataset of more than 9,000 firms operating in the Lombardy region of Northern Italy over the period 2016–2019. The dataset includes a broad set of accounting-based indicators, such as profitability measures (ROA, ROE, ROI, ROS), leverage and capital structure ratios, liquidity and working capital indicators, as well as coverage and solvency measures. Following an extensive process of data cleaning, normalization, and feature engineering, an exploratory data analysis is conducted to identify structural differences between bankrupt and non-bankrupt firms. The core of the thesis focuses on the development and comparison of several predictive models, including penalized logistic regression (LASSO), Random Forest, XGBoost, CatBoost, and Multilayer Perceptron neural networks. Model performance is evaluated using multiple metrics accuracy, precision, recall, specificity, and AUC with particular attention to class imbalance and early-warning capabilities. The results show that all models achieve overall satisfactory and consistent performance, with machine learning approaches generally outperforming traditional methods when supported by appropriate feature engineering. Nevertheless, the findings should not be interpreted as fully deterministic, as corporate bankruptcy is influenced by numerous external and qualitative factors not captured by financial statements. Consequently, the proposed framework should be regarded as a decision support tool rather than a definitive predictive system