User’s Guide : Multiple Equation Analysis : Mixed Frequency VAR
Mixed Frequency VAR
Standard Vector Autoregression (VAR) models require all variables in the VAR to be of the same frequency—all variables in the VAR are monthly, or all are quarterly, etc. The single frequency requirement is generally not met in practice as macroeconomic researchers typically possess variables of different frequencies.
The traditional solution to mixed frequency estimation uses frequency conversion methods to convert the high frequency variables to the lowest frequency in the VAR. This aggregation leads to a loss of information as multiple observations are combined into a single data point.
In univariate models this loss of information has been overcome through the use of MIDAS estimation (see “Midas Regression” in User’s Guide II). MIDAS creates multiple low frequency variables for each high frequency variable, allowing retention of the high frequency fidelity. The downside to this approach can be a dramatic increase in the number of variables in the model and the corresponding loss of degrees of freedom.
Although MIDAS models can alleviate this by using a polynomial weighting scheme to reduce the number of variables to a more manageable level, in cases where the number of added variables is small, as with quarterly and monthly data, the unweighted, or U-MIDAS approach (Ghysels 2016), is generally employed.
In VAR models, different approaches to alleviate the data aggregation problem have recently become popular. There are three broad categories of methods for estimating these mixed frequency VARs:
State-space approaches such as Schorfheide and Song (2013).
U-MIDAS approaches as outlined in Ghysels (2016).
MIDAS with polynomial weights, Ghysels (2016).
EViews offers support for the unweighted U-MIDAS approach which has been expanded to include both Bayesian and Classical VAR techniques.
We begin with a discussion of how to estimate a Mixed Frequency VAR in EViews. Technical details are provided in “Technical Background”.