Association for Public Policy Analysis & Management

Forecasting future tax revenue is crucial for setting a budget. Under-predicting has adjustment costs as the budget is revised, while over-predicting can force agencies to make disruptive mid-year budget cuts. In this research project, we began by examining state tax revenue forecasting methods, the causes of error, and possible improvements. We plan to extend this analysis to the federal level.

First, we examine the effects of business cycles on forecast error rates and the relationship between revenue volatility and error, as well as methods to predict recessions in time for revenue forecasting. We have shown that much of the increase in errors during recessions can be attributed to increasing revenue volatility, but that nonwithholding (i.e., capital gains, proprietor, rental, and interest) income tax revenue does not fit this this pattern. We are examining possible reasons for this—one likely candidate is the behavioral effect of taxpayers responding to real or expected changes in policy.

Second, we examine different ways of modeling post-forecast policy tools such as the Nonwithholding Collar, which restricts how much of forecasted nonwithholding income tax revenue the state can budget for. We show for Virginia that nonwithholding revenue is inherently hard to predict, but that its ratio to total General Fund revenue has closely followed a sinusoidal curve which may be useful in improving nonwithholding forecasts.

Third, we test the current forecasting models being used against different model specifications and forecasting methods: different specifications, sample sizes, data transformations, and use of cross-validation; models from time series and Bayesian econometrics; machine learning methods; and methods of forecast averaging. So far we have been able to improve significantly on the current Virginia Tax Department models using the above methods.

Moving forward, we will be testing hybrid linear/non-linear forecasting models, incorporating signal decomposition methods, expanding our dataset, examining whether the components of nonwithholding income might be better forecast on their own, attempting to automate model specification, and investigating methods for early detection of turning points in time series.