VAR Model Estimation

Preparing VAR Models for Fitting

To create a model of multiple time series data, decide on a VAR model form, and fit parameters to the data. When you have a fitted model, check if the model fits the data adequately.

To fit a model to data, you must have:

• Time series data, as described in Multivariate Time Series Data

• At least one varm model object configured for estimation, as described in Vector Autoregression (VAR) Model Creation

There are several Econometrics Toolbox™ functions that aid these tasks, including:

• estimate, which fits VARX models.

• summarize, which displays and returns parameter estimates and other summary statistics from fitting the model.

• lratiotest and aicbic, which can help determine the number of lags to include in a model.

• infer, which infers model residuals for diagnostic checking.

• forecast, which creates forecasts that can be used to check the adequacy of the fit, as described in VAR Model Forecasting, Simulation, and Analysis

Fitting Models to Data

estimate performs parameter estimation for VAR and VARX models only. For definitions of these terms and other model definitions, see Types of Stationary Multivariate Time Series Models. For an example of fitting a VAR model to data, see Fit VAR Model of CPI and Unemployment Rate.

How estimate Works

Before fitting the model to data, estimate requires at least Mdl.P presample observations to initialize the model, where Mdl is a varm model object and P is the property storing the model degree. You can specify your own presample observations using the Y0 or Presample name-value argument. Or, by default, estimate removes the first Mdl.P observations from the estimation sample. Therefore, if you let estimate take requisite presample observations from the input estimation sample data, then the effective sample size decreases.

estimate finds maximum likelihood estimates of the parameters present in the model. Specifically, estimate estimates the parameters corresponding to these varm model properties: Constant, AR, Trend, Beta, and Covariance. For VAR models, estimate uses a direct solution algorithm that requires no iterations. For VARX models, estimate optimizes the likelihood using the expectation-conditional-maximization (ECM) algorithm. The iterations usually converge quickly, unless two or more exogenous data streams are proportional to each other. In that case, there is no unique maximum likelihood estimator, and the iterations might not converge. You can set the maximum number of iterations with the MaxIterations name-value argument of estimate, which has a default value of 1000.

For numeric array data inputs, estimate removes entire observations from the data containing at least one missing value (NaN). For table or timetable data inputs, estimate issues an error when any observation is missing. For more details, see estimate.

estimate calculates the loglikelihood of the data, giving it as an output of the fitted model. Use this output in testing the quality of the model. For example, see Select Appropriate Lag Order and Examining the Stability of a Fitted Model.

Examining the Stability of a Fitted Model

When you enter the name of a fitted model at the command line, you obtain a object summary. In the Description row of the summary, varm indicates whether the VAR model is stable or stationary.

Another way to determine stationarity of the VAR model is to create a lag operator polynomial object using the estimated autoregression coefficients (see LagOP), and then passing the lag operator to isStable. For example, suppose EstMdl is an estimated VAR model. The following shows how to determine the model stability using lag operator polynomial objects. Observe that LagOp requires the coefficient of lag 0.

ar = [{eye(3)} ar]; % Include the lag 0 coefficient.
Mdl = LagOp(ar);
Mdl = reflect(Mdl); % Negate all lags > 0
isStable(Mdl)

If the VAR model is stable, then isStable returns a Boolean value of 1, and 0 otherwise. Regression components can destabilize an otherwise stable VAR model. However, you can use the process to determine the stability of the VAR polynomial in the model.

Stable models yield reliable results, while unstable ones might not.

Stability and invertibility are equivalent to all eigenvalues of the associated lag operators having modulus less than 1. In fact, isStable evaluates these quantities by calculating eigenvalues. For more information, see isStable or Hamilton [102].