The Minneapolis Mixed Frequency Vector Autoregression (MF-VAR) model is the most recent in a line of statistical forecasting models that began with research conducted at the Federal Reserve Bank of Minneapolis and the University of Minnesota in the 1970s and early 1980s. Rising computer power led to resurgent interest in these models in the 1990s, which in turn brought about improvements along several dimensions. The current model reflects many of these advances. The model is used by the Minneapolis Policy Group in forecasting and other policy projects. Please note that the forecasts produced by the MF-VAR model do not represent official forecasts of the Minneapolis Fed, its president, the Federal Reserve System, or the FOMC.
A noteworthy feature of the model is that it combines data measured at both monthly and quarterly frequencies. The primary advantage of the mixed-frequency approach is that it can use more timely monthly data to help forecast quarterly variables—primarily GDP and associated national income and product account concepts—that are available on a less timely basis. The algorithm used to solve the model uses all available monthly information to construct forecasts of the quarterly variables.
The “vector” is the collection of economic concepts, or variables, to be forecast. The Minneapolis model contains 14 of these variables; 5 are of immediate interest to the Federal Open Market Committee: real GDP, the unemployment rate, PCE prices, PCE prices excluding food and energy, and the federal funds rate. The remaining 9 variables include personal consumption expenditures (PCE), fixed investment, government purchases, aggregate hours worked, average hourly earnings, industrial production, the 10-year yield on U.S. Treasury bonds, the yield on Moody’s Baa-rated corporate bonds, and the S&P 500 index of equity prices. The “autoregression” part of VAR means that the forecasts of all model variables depend on their own past values as well as the past values of every other variable in the model.
Like its predecessors from the 1970s, the model is couched in the Bayesian framework of statistical analysis, which allows researchers to introduce a priori extra-sample information into the forecast—typically, commonly held views on the long-term properties of the variables. This information reduces the tendency for the model to find false patterns in the data, thereby aiding forecast accuracy.
The primary output of the model is a set of “predictive densities,” which provide a comprehensive probability assessment of future values of the model variables given current and past data observations. They can be used to compute probability bands around the median forecast of a variable and then displayed as “fan charts.” Here we show fan charts corresponding to four key model variables. Each chart displays the mean forecast and the 70 and 90 percent probability bands surrounding the median.
The Minneapolis MF-VAR is a revised version of the original model by Frank Schorfheide and Dongho Song of the University of Pennsylvania. For more information on the details of that model, see the original working paper. A paper describing the revised and current version is also available.