Previsão de Dificuldade Financeira com XGBoost: Evidências no Setor de Utilidade Pública da América Latina

This study investigates the effectiveness of the XGBoost machine learning algorithm for forecasting financial distress (FD) in Latin American utility firms using financial data from 2000 to 2019. The dataset comprises firms from six Latin American countries, with FD defined through accounting indicators, such as negative Earnings Before Interest, Taxes, Depreciation, and Amortization (EBITDA), and market indicators based on declines in market value. The model was trained using undersampled data to address class imbalance and evaluated using metrics including Area Under the Curve (AUC), accuracy, Type I and Type II errors, and Brier Score. The results show that, within the sector-specific sample analyzed, the XGBoost model outperformed logistic regression and achieved performance comparable to or superior to that of Random Forest, whose reference results were obtained from a multi-sector sample. Overall, XGBoost achieved accuracy ranging from 88% to 95.5% and Type II error rates between 0% and 12%, surpassing the reference model. The most relevant variables differed according to the type of financial distress. For accounting-based FD, profitability indicators, especially earnings per share, were predominant. For market-based FD, variables associated with investor perceptions became more prominent, including year-over-year changes in the price-to-book ratio and location-related factors (country-level variables). The study confirms the potential of XGBoost both for forecasting financial difficulties with a one-year lead time and for interpreting the underlying factors, which evidences a potentially useful tool for analysts, regulators, and investors. More broadly, the results contribute to the ongoing discussion of forecasting performance in finance.