Forecasting Tax Revenue in Cambodia: A Comparative Study of Traditional Time-Series Models and Machine-Learning Methods

The study examines the predictive ability of traditional econometrics and machine-learning models for Cambodian tax revenue estimates using monthly LNTAX from January 2009 to March 2026. Repeated out-of-sample forecasts were then produced through a rolling-origin forecasting design, across two test horizons: 12-months and 24-months. We evaluated different techniques: moving average, exponential smoothing, damped trend, ARIMA, CART, GRNN and KNN. The performance against the moving average benchmark was compared using regression analysis with clustered standard errors, and forecast accuracy measured in terms of mean squared error and symmetric mean absolute percentage error. The results provide evidence that LNTAX is non-stationary in levels, and then stationary at first differencing. The best overall predictive vision is achieved within ARIMA with the smallest errors in most of the evaluations and significantly largest decreases in MSE and sMAPE compared to moving average. Damped-trend exponential smoothing also does quite well, especially for percentage accuracy over the longer horizon. In general, machine-learning models outcomes are mixed: CART shows average performance, whereas KNN in some instances is able to improve sMAPE versus naive-based models but GRNN ranks the weakest. The results indicate that time-series models have a superior forecasting performance than the state-space framework-based univariate tax revenue forecasting model in this context. Hence, it is suggested that ARIMA is used as the main benchmark model and future research should be conducted in hybrid methods, more macroeconomic variables, structural breaks and a policy-oriented forecasting evaluation.