Improving CBO's Methodology for Projecting Individual Income Tax Revenues

Posted on
February 10, 2011

In preparing its annual report on the budget outlook and updates to that report during the course of the year, CBO projects revenues from the federal individual income tax. A background paper released today discusses two ways to use information about tax collections to improve those projections of tax receipts: Explicitly using the information provided by recent tax collections to adjust the projections, and basing the projections on multiple years of tax return information. The results of CBO’s analysis suggest that combining the two approaches modestly improves the near-term projection of individual income tax revenues. Based on those findings, CBO has used this combination of approaches in estimating revenues in recent years.

Each January in The Budget and Economic Outlook (you can find our most recent report here), CBO publishes its projections of federal revenues spanning 11 years, assuming that the laws affecting revenues remain unchanged.  Since 1997, receipts of individual income taxes have differed from CBO’s January projections for the current year—the fiscal year already underway at that point—by an average of about 6 percent. On average, receipts of individual income taxes for the budget year—the first full year after the year in which the projections are produced—have differed from the projections made in January of the previous year by about 11 percent. Those discrepancies have been calculated so that they exclude the effects of legislation enacted after the projections were prepared. They have also been calculated without regard to whether the discrepancies were positive or negative—that is, whether CBO overestimated or underestimated actual receipts.

An analysis of the errors in CBO’s past projections of individual income tax revenues suggests that, over time, the agency’s estimates are about as likely to be too high as too low. For the years since 1997, positive and negative errors have largely offset each other over that time. For example, the average net error—a calculation that incorporates the offsetting effects of positive and negative differences—over that time in CBO’s projections for the current fiscal year is 0.6 percent, and the average net error for the budget year is 1.3 percent. Nevertheless, CBO’s analysis also suggests that errors in the projections of revenues for the current and budget years tend to be correlated over time; that is, if the projections made in one January for the current year and budget year were too low, the projections made the next January would probably also be lower than actual revenues in those years.

This paper analyzes two approaches for improving the results from the model that CBO uses to project revenues from individual income taxes. The results of this analysis have been incorporated in CBO’s revenue forecasting. In the first approach, CBO sought to determine the best way to use the most recent information about tax collections to minimize errors in its forecasts. When actual collections differ from projections, forecasters must judge whether the discrepancy they observe is a temporary deviation or the beginning of a longer-term trend and thus whether and by how much to adjust the model’s projections of future revenues to reflect that discrepancy. Such determinations are complicated, though, because detailed tax return data that could identify the sources of the difference are not available for some time. In the absence of such data, CBO examined whether the accuracy of its projections would improve if it simply assumed that actual revenues would differ from the model’s projection in future years by the same percentage that they differed in the most recent year for which actual revenues were known.

In the second approach, CBO analyzed the improvement to be gained from using additional years of tax return data as the basis for its projections. CBO’s individual income tax model starts with the most recent year of detailed tax return data available. But those so-called base-year data may not be representative of longer-term trends. Therefore, CBO investigated whether the model might achieve more accurate results by producing multiple projections using tax return data from different years (rather than data from only a single year) as the starting point and averaging the projections.

Using a simplified version of its individual income tax model and examining the projections for the 1984–2005 period, CBO evaluated various versions of those two approaches as well as combinations of the two. CBO’s findings include the following:

  • Projections of tax liabilities (what taxpayers owe) were more accurate—but only in the very short term—when CBO adjusted the model’s results by the percentage error in those results in the most recent year for which actual revenues were known.
  • The use of multiple years of detailed tax return data reduced the size of the error in CBO’s projections compared with the use of only the most recent year of such data. Using four years of tax return data to produce four separate projections, and then averaging them, yielded the best overall results.
  • A combination of the two approaches was better than either one by itself.

Those improvements have limitations, however. The reduction in the size of the errors is small relative to the errors seen before making those adjustments. The reason is that the use of the most recent information can do little to reduce the uncertainty that surrounds the future course of the economy and the rate of growth of income in the tax base. Also, the degree of improvement in the projections’ accuracy declines rapidly over the 10-year forecast horizon.

For historical information on revenues, see my December testimony on Trends in Federal Tax Revenues and Rates.

Kurt Seibert, Mark Booth, and David Weiner of CBO’s Tax Analysis Division wrote the paper with assistance from Regis Barnichon, formerly of CBO.