<p>Forecasting VAT Revenue Under Structural Change in Rwanda: Comparing Econometric and Machine Learning Models</p>

This paper evaluates how econometric and machine learning models forecast Value Added Tax (VAT) revenue under structural change in Rwanda. Using 156 monthly VAT observations from January 2012 to December 2024, with official macroeconomic indicators, the study compares ordinary least squares, vector autoregression, seasonal autoregressive integrated moving average with exogenous regressors, generalized autoregressive conditional heteroskedasticity, least absolute shrinkage and selection operator, extreme gradient boosting, Prophet, and long short-term memory networks. Forecast accuracy is assessed through rolling-origin evaluation using root mean squared error, mean absolute percentage error, and symmetric mean absolute percentage error. Results show that adaptive models generally outperform static regressions at the aggregate level. Prophet, long short-term memory, and seasonal autoregressive integrated moving average with exogenous regressors generate the strongest aggregate forecasts, while sectoral disaggregation does not consistently improve accuracy because several sectors are sparse, volatile, or heterogeneous. The findings reveal a trade-off between accuracy and interpretability: machine learning improves prediction in some settings, whereas econometric models remain useful for policy explanation and fiscal accountability. The paper recommends a hybrid framework combining transparent econometric models for budget communication with adaptive models for operational forecasting, contributing evidence on fiscal forecasting under structural change in Africa.