Local governments operate under tight budget constraints, balanced budget requirements, and increasing responsibilities, making it difficult to maintain consistent services during economic downturns. Fund balances are often used to provide stability, but the appropriate level of savings and the factors influencing deficits remain unclear. This research uses machine learning tools to identify the characteristics of county governments that predict deficits during economic recessions and help inform financial management and guide fund balance policies.
Drawing on over 900 variables and twenty years of data from North Carolina’s 100 counties, this study employs multiple machine learning methods – including LASSO, random forests, and decision trees – to build predictive models of local government deficits during recessions.
Despite almost 500 models being run, no model was able to predict which counties would have a deficit or the approximate magnitude of that deficit using pre-recession covariates. This finding is an important lesson that reveals that machine learning techniques are limited in this context and highlights the continued value of traditional analytic approaches.
This study challenges the “one-size-fits-all” approach to local government savings by examining predictors of deficits during the dot-com and Great Recessions and finding that socio-economic, fiscal, and financial health indicators are insufficient to predict the presence or magnitude of deficits during recessions. Additionally, it cautions practitioners, policymakers, and researchers against over-relying on machine learning. Ultimately, the research offers both a framework for applying machine learning in public finance and a reminder that these tools have limits when tackling complex, context-dependent problems.
