Public Expenditure Forecasting in Cambodia: A Hybrid Econometric and Machine Learning Approach

This study compared traditional time-series and machine-learning methods for the monthly forecasting of government expenditure of Cambodia with dataset. The study employed a natural logarithm of the monthly flows of government expenditure from January 2009 to March 2016. The dataset has 207 observations and was divided into train set, where the general model training took place (80%) and test set out of sample (20%). Mean Absolute Error (MAE), Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE) are the metrics that are used to evaluate forecasting performance. The study compares the benchmark against the Autoregressive Integrated Moving Average (ARIMA) and four machine-learning approaches: Hybrid ARIMA-Random Forest, Random Forest (RF), Extreme Gradient Boosting (XGBoost) and Support Vector Regression (SVR). The ARIMA model detects a statistical significant seasonal autoregressive behaviour and a drift, indicating that government expenditure is characterized by long-run growth and seasonal dynamics. Nevertheless, the forecasting accuracy using the traditional method is outmatched by results of machine-learning models in general. In particular, XGBoost outperforms all the other models with the lowest MAE, RMSE and MAPE metrics indicating that it is the best forecasting technique. The Random Forest is in the second position while Hybrid ARIMA-Random Forest model significantly outperforms standalone ARIMA model, though it remains below the tree-based ensemble methods. SVR also beats ARIMA, but fails to beat XGBoost and Random Forest. Overall, the results illustrate that machine-learning approaches, particularly XGBoost and Random forest, consistently outperform the traditional ARIMA model in terms of fitting government expenditure´s nonlinear, seasonal and trend pattern, which would enhance forecasts accuracy for fiscal planning and policy decisions.