Forecasting State Tax Revenue: A Bayesian Vector Autoregression Approach By

This paper compares alternative time-series models to forecast state tax revenues. Forecast accuracy is compared to a benchmark random walk forecast. Quarterly data for California is used to forecast total tax revenue along with its three largest components, sales, income, and corporate tax revenue. For oneand four-quarter-ahead forecasts from 2004 to 2009, Bayesian vector autoregressions generally forecast best based on root mean squared errors compared to standard vector autoregressions or a random walk model. Similar to the macroeconomic forecasting experience, Bayesian vector autoregressions should be consider as a useful and cost effective revenue forecasting model for state governments. *This paper was written while on sabbatical leave during the Fall 2009 semester. The author thanks the College of Business and Economics for financial support. I thank Shirley Svorny for helpful comments.