All Articles

Our office and the Department of Finance (DOF) each produce two formal forecasts of state General Fund revenues every year: one in the middle of the fiscal year and another in May, toward the end of the fiscal year. In the months between these formal forecasts, new revenue and economic data become available. These data can provide useful information about whether the outlook has shifted since the most recent formal forecast. Nonetheless, despite reporting on the new data, our offices generally have not quantified how the recent data changes the state’s revenue outlook.

To fill this void, our office recently developed a model to provide a monthly update of the forecast of current year collections from the state’s “big three” revenues (personal income, sales, and corporation taxes). The big three comprise the vast majority of state General Fund collections.

Model Uses Most Recent Data to Update Less Frequent Formal Forecasts. Our new monthly model draws on information from four key sources to predict where revenues will end up at the end of the current year: (1) the most recent formal forecasts from our office and DOF, (2) monthly cash collections from the big three, (3) monthly data on employment, retail sales, stock prices, home prices, building permits, and venture capital investment (quarterly), and (4) the most recent consensus forecast of the U.S. unemployment rate and U.S. gross domestic product from the Survey of Professional Forecasters. We examine the relationships between these components and big three revenues during the period 1999-00 through 2019-20. We then use these historical relationships to make predictions of current year revenues. Importantly, out of practical necessity, our approach takes the most recent data at face value and mechanically executes a statistical model. This is in contrast to our formal forecasts, which also incorporate professional judgements about data oddities, extraordinary circumstances, and policy changes.

Limitations of the Model. It seems reasonable to expect that incorporating the most recent data would improve forecast accuracy. That being said, other factors might cancel out the benefit of up-to-date data. Perhaps monthly data simply is too “noisy” (fluctuating because of random events) to offer useful prediction. Or maybe our statistical model fails to incorporate elements that affect the revenue outlook that can only be captured via application of deliberate professional judgment.  

More Frequently Updated Forecasts Offer Improved Accuracy, But Remain Far From Perfect. To gauge whether the benefits of up-to-date data balance out opposing factors, we looked at how our model would have performed in prior years. To do this, we removed a particular year from our dataset and then measured how well our model would have predicted that year without having any information about what actually happened. We then repeated this process for every year during the period 1999-00 through 2019-20. The results of this analysis are summarized in the graphic below.

Our results offer several important considerations for interpreting our monthly forecasts:

  • Up-To-Date Data Improves Forecast Accuracy. Updating the formal forecasts for recent data appears to offer increased forecast accuracy. Our estimates suggest that an updated forecast based on data available up through September was about 50 percent more accurate than the formal forecasts made in May of the same calendar year. Similarly, a March update was an estimated 15 percent more accurate than the formal forecasts made by our office in November and DOF in January.
  • More Deliberate Process of Formal Forecast Has Advantages. Despite the apparent benefits of more frequent forecast updating, our analysis also finds evidence that the more deliberate formal forecasts capture some elements relevant to the revenue outlook that are missed by our monthly model. When we do not allow our model to factor in the most recent formal forecasts, its accuracy declines notably. For example, more than half of the accuracy improvement of the September forecast update mentioned above is erased if the model has no information about the formal forecasts from May. 
  • Revenue Forecasting Is Hard. A final important takeaway from our analysis is that even the best forecast model will be wrong, often by sizeable amounts. For example, our model’s March predictions, which draw on data on actual economic and revenue outcomes covering two-thirds of the fiscal year, still had an average error of several billion dollars. In part, this challenge stems from the inherent unpredictability of the state’s largest revenue source: the personal income tax. The Legislature should always carefully consider the risks inherent in relying on a particular forecast.


  Article Tags