Forecasting Tax Revenue in Indonesia using Single and Hybrid Methods

Tax is a vital component as a primary revenue for a nation. However, in Indonesia, the monthly tax target allocation for each vertical unit within the Directorate General of Taxes (DGT) is still conducted conventionally, without utilizing a specific methodology. This study aims to forecast tax revenue in Indonesia, comparing four single models- Trigonometric Seasonality Box-Cox Transformation ARIMA error Trends Seasonal (TBATS), Autoregressive Integrated Moving Average (ARIMA), SVR (Support Vector Regression), and NNAR (Neural Network Auto Regression)-and five hybrid models. Six types of monthly tax revenue data were used in this research, encompassing four types of income tax, value-added tax (VAT), and stamp duty. We employed Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) as evaluation metrics. Hybrid models, such as ARIMA-NNAR, ARIMA-TBATS, and TBATS-SVR, demonstrated superior performance across four types of tax data, while ARIMA and NNAR performed best for the remaining two. Futhermore, ARIMA-TBATS and TBATS-SVR are better because it can capture trends and patterns of the testing data well. Future research could explore more extensive datasets in different timeframes, such as daily or weekly forecasts.