Every December the Minneapolis Fed releases a Ninth District economic forecast for the upcoming year. This past December the forecast predicted moderate economic growth for 2012.
But as with any regular projection, it’s useful to know how accurate past forecasts have been. A review of Minneapolis Fed forecasts going back to 1998 shows that they’ve been pretty good, but have some soft spots that are inherent in current forecast models. That is, economic forecasting is not an exact science.
The Minneapolis Fed’s regional models forecast nonfarm employment growth, unemployment rate, personal income growth and housing units authorized growth for states in the Ninth District. Research department staff use a technique called Bayesian vector autoregression (BVAR). The forecasting models base projections on historical trends and recent movements in each of the data series combined with a BVAR forecast for the national economy.
This forecasting technique shares an unfortunate feature common to all forecasting models—they don’t predict turning points (i.e., recessions) very well. For example, in Chart 1 the forecast error, or difference between the forecast and actual data, in nonfarm employment for Minnesota was much larger during the recession periods when economic growth dipped. The forecasting model either didn’t anticipate the drop, as during the 2001 recession, or didn’t anticipate the size of the decrease, as during the 2007-09 recession.
A similar picture is found in Chart 2 for the unemployment rate. Here the forecasts for Minnesota also didn’t anticipate increases during recessions; forecast errors grew.
Table 1 shows that the 12-month unemployment rate forecast is within 0.7 percentage points on average from the actual unemployment rate. Meanwhile, the 12-month nonfarm employment growth forecast is within 1.3 percentage points of actual employment figures. As the time period between the forecast and the actual data narrows, forecast errors decrease since the models have more contemporary data on which to base the forecast; the exception is housing forecasts (see Table 1). (For those interested in technical details, the BVAR models calculate a confidence interval to estimate a range in which 70 percent of the expected outcomes would fall. You can see these ranges in the most recent forecast table.)
Forecasts for personal income growth and especially housing units authorized have larger errors. (Housing units authorized is the total number of units authorized in permits for single- and multi-unit housing projects.) The vigorous climb in housing units authorized prior to the recent recession and subsequent steep drop were difficult for forecasting models to account for as authorizations in Minnesota and Wisconsin dropped below levels observed over 30 years ago. Meanwhile, personal income growth is often affected by volatile levels of farm income, particularly in North Dakota. In Table 1, North Dakota is removed due to volatile changes in farm income; average errors would be about 1 percentage point higher with North Dakota in the mix.
Given the uncertainty, the Minneapolis Fed doesn’t rely only on forecasting models, but also conducts surveys of business and community leaders throughout the year to hear what they expect at their own companies and communities going forward. To learn more about surveys and other data collected by the Minneapolis Fed, go to our district data web page.