Tax Revenue Forecasts withPossible Structural Changes

During the past three consecutive years, over-estimations in forecasts of tax revenue have led to revenue shortfalls. These developments have motivated this study, which examines the current practices in tax revenue forecasting and the existing literature, and conducts tests for the presence of possible structural changes in tax revenue and other relevant time series. We then comparatively evaluate the performance of each forecasting model to select the best- performing models by tax item, using these to make 5-year forecasts for the revenues of the national tax and other key tax items. While differences in political characteristics and budget processes across countries does not allow for a direct comparison, We find that the forecast errors for Korea’s tax revenue over the past decade are not particularly serious in terms of bias or instability when compared to that of other countries. Nevertheless, it is evident that the average bias has been widening recently, while the tax revenue has been overestimated by nearly 5% during the past three years. To avoid the systematic over-estimation of tax revenue, it is important to consider the slowdown of tax revenue growth. We check the validity of the log linear system and strongly reject the log linear system. There is a clear sign of structural change in the linear trend or nonlinear trend in national tax revenue series and related macro series. However it is uncertain what kind of nonlinear trend is appropriate for the purpose of forecasting because several nonlinear specifications give similar in-sampling fit but make quite different out-sample forecasts. We incorporate the element of structural breaks or nonlinear trend in making our forecasts through employing dummy variables, cubic spline models, exponential smoothing models, piecewise regression models combined with structural change tests, as well as a combination of the forecasting models. Because of the characteristics of each tax item’s data and the differences in the associated tax systems, different models were found to perform better for each tax item. For example, piecewise regression and exponential smoothing are effective for national tax revenue forecasting but vector autoregressive model and error correction model with dummy variables are best-performing for corporate tax revenue forecasting. Due to the high possibility of recent structural changes in the national tax and corporate tax, the effects of such changes must be examined more closely. As there are still uncertainties regarding the existence of recent structural changes and their effects, using exponential smoothing models which take account of such uncertainties or a combination of models may contribute to gains in predictive power. Along with the aforementioned technical treatment of tax revenue forecasting models, we would like to point out the necessity for a review of when the forecasts are made during the budget process. In addition to the fact that the circumstances in both the domestic and international economy, within which taxation must take place, have become increasingly complex, there have been calls for more accurate revenue forecasts to secure fiscal soundness. These developments suggest that we must reconsider whether it is adequate to present the 1-year horizon forecast (for individual tax items) when the budget is proposed to the National Assembly, and the 5-year horizon forecast (for the national tax) for the National Fiscal Management Plans. Rather than keeping revenue forecasts a one-off affair prior to the drawing up of the budget bill and its deliberation, making it an ongoing effort throughout the budget process and linking it to the next year’s budget process will allow for better amendment of errors and overall improvements in structural predictive power.