Association for Public Policy Analysis & Management

The primary goal of this study is to apply the logic of machine learning algorithms to the prediction of fiscal stress in local governments. Generated from a combination of computer science and statistics, machine learning algorithms are receiving much attention as the next generation of mechanisms for unlocking the potential of big data (Anastasopoulos & Whitford, 2018). In the new era of big data, these algorithms can provide methods for handling the variety and volume of data constantly being generated by public organizations (Qiu et al., 2016; Helsby et al., 2018).

Today, machine learning algorithms as predictive analytics are being used for finding solutions to pressing public policy problems in collaboration with government and business (Glaeser et al., 2016; Kleinberg et al., 2015). For instance, Glaeser et al. (2016) collaborated with the City of Boston, Yelp, and DrivenData to run an open tournament for predicting restaurant hygiene and sanitation violations based on Yelp reviews. Using voluminous data from the criminal justice system, Kleinberg et al. (2015) provided evidence that predictive analytics can help judges decide whether to detain or release arrestees and substantially reduce the probabilities of recidivism.

Our study extends this line of inquiry by using the case of fiscal stress in local governments. Ensuring fiscal sustainability has been a continuing policy concern for different layers of governments. One of the emerging practices used to prevent fiscal stress is the early-warning fiscal system, which measures fiscal health, assigns fiscal distress labels, and provides immediate fiscal management assistance for local governments. These fiscal stress-monitoring systems have been adopted by several state governments in the US, such as New York, North Carolina, Ohio, and Michigan, as well as central governments abroad, such as South Korea and Spain. One of the first tasks in this system is to measure the level of fiscal stress that localities face.

Long-standing literature on fiscal stress indicators has suggested various measures, ranging from a pioneering work by the Advisory Commission on Intergovernmental Relations (ACIR) to the recent development of metrics such as Wang’s cash solvency index and the Brown 10-point fiscal index. There is, however, little consensus on how to construct the best metrics of fiscal stress indices for public entities. The main research question is whether machine-learning algorithms from the data-driven approach can improve the prediction of fiscal stress, which can offer helpful guidance to fiscally stressed localities. Specifically, we collect the data on 150 of the largest municipalities in the U.S. from 1977 to 2015 from across more than 120 categories of revenues, expenditures, and debts. Then, we compare the two different approaches such as the classical asymptotic inference techniques and machine learning approaches (e.g., random forest, neural networks, and anomaly detection) and suggest some pitfalls and perils of machine learning approaches. Our study would be a step toward understanding the adapting mechanisms of machine learning in predicting policy problems.