Robustness of Machine Learning-Based Forecasting Models for Local Tax Revenues in Volatile Economic Conditions: Evidence from the COVID-19 Pandemic

Accurate revenue forecasting aid local governments in ensuring fiscal stability and providing timely public services. We examine the robustness of machine learning (ML) models, particularly the gated recurrent unit (GRU), in forecasting local tax revenues during the COVID-19 pandemic. Using annual revenue data in 1994–2021, we compare GRU-based forecasts against initial budget estimates for 2017–2021, focusing mainly on pandemic years. Conventional forecasts substantially underestimated tax revenues during the pandemic, with errors doubling in 2019–2021. By contrast, the GRU model generated lower forecast errors and greater resilience to economic disruption. This study highlights the accuracy and robustness of GRU and the importance of resilient forecasting tools for fiscal planning under uncertainty. It also underscores institutional and behavioral influences that persist regardless of modeling sophistication. ML models are worth integrating into public finance systems as decision-support tools to promote data-driven governance and enhance fiscal transparency and long-term policy credibility.