Using Machine Learning Algorithms to Improve Fiscal Distress Prediction Models: The Case of U.S. Local Governments

ABSTRACT This study aims to improve the prediction of fiscal distress in local governments (LGs) using machine learning (ML) methods. We apply six ML algorithms, including Logistic Regression, XGBClassifier, RUSBoost, C4.5 Decision Tree, Exactly Balanced Bagging, and Roughly Balanced Bagging, which are combined with and without two undersampling methods—Majority Sampling and Optimal Ratio Sampling—to predict fiscal distress of LGs. Analyzing financial and socioeconomic indicators from LGs across 49 U.S. states from 2015 to 2017, we find that regardless of undersampling strategy, the Exactly Balanced Bagging Classifier generally outperforms other ML algorithms including Logistic Regression: its F1-scores range from 53.11 percent to 55.33 percent, significantly higher than 27 percent to 58 percent of Logistic Regression. Furthermore, Optimal Ratio Sampling offers the most significant benefit to Logistic Regression. These findings offer valuable insights into the analysis and prediction of municipal fiscal distress, particularly within the context of imbalanced datasets. Data Availability: Data are available from the public sources cited in the text. JEL Classifications: G31; G32; G33; M21.