VARMAX Procedure

Engle and Kroner (1995) propose a general multivariate GARCH model and call it a BEKK representation. Let be the sigma field generated by the past values of , and let be the conditional covariance matrix of the k-dimensional random vector . Let be measurable with respect to ; then the multivariate GARCH model can be written as

where C, , and are parameter matrices.

Consider the bivariate GARCH(1,1) model

or, representing the univariate model,

For the BEKK representation of the bivariate GARCH(1,1) model, the SAS statements are

model y1 y2;
garch q=1 p=1 form=bekk;

The multistep forecast of the conditional covariance matrix, is obtained recursively through the formula

where for .

Bollerslev (1990) proposes a multivariate GARCH model with time-varying conditional variances and covariances but constant conditional correlations.

The conditional covariance matrix consists of

where is a stochastic diagonal matrix with element and S is a time-invariant correlation matrix with the typical element .

The element of is

Note that .

If you specify CORRCONSTANT=EXPECT, the element of the time-invariant correlation matrix S is

where T is the sample size.

By default, or when you specify SUBFORM=GARCH, follows a univariate GARCH process,

As shown in many empirical studies, positive and negative innovations have different impacts on future volatility. There is a long list of variations of univariate GARCH models that consider the asymmetricity. Four typical variations follow:

For more information about the asymmetric GARCH models, see Engle and Ng (1993). You can choose the type of GARCH model of interest by specifying the SUBFORM= option.

The EGARCH model was proposed by Nelson (1991). Nelson and Cao (1992) argue that the nonnegativity constraints in the GARCH model are too restrictive. The GARCH model, implicitly or explicitly, imposes the nonnegative constraints on the parameters, whereas these parameters have no restrictions in the EGARCH model. In the EGARCH model, the conditional variance is an asymmetric function of lagged disturbances,

In the QGARCH model, the lagged errors’ centers are shifted from zero to some constant values,

In the TGARCH model, each lagged squared error has an extra slope coefficient,

where the indicator function is one if and zero otherwise.

The PGARCH model not only considers the asymmetric effect but also provides a way to model the long memory property in the volatility,

where and .

Note that the implemented TGARCH model is also well known as GJR-GARCH (Glosten, Jaganathan, and Runkle 1993), which is similar to the threshold GARCH model proposed by Zakoian (1994) but not exactly the same. In Zakoian’s model, the conditional standard deviation is a linear function of the past values of the white noise. Zakoian’s model can be regarded as a special case of the PGARCH model when .

The following formulas are recursively implemented to obtain the multistep forecast of conditional error variance and :

In the preceding equations, for . Then, the multistep forecast of conditional covariance matrix , is calculated by

where is the diagonal matrix with element .

Engle (2002) proposes a parsimonious parametric multivariate GARCH model that has time-varying conditional covariances and correlations.

The conditional covariance matrix consists of

where is a stochastic diagonal matrix with the element and is a time-varying matrix with the typical element .

The element of is

Note that .

As in the CCC GARCH model, you can choose the type of GARCH model of interest by specifying the SUBFORM= option.

In the GARCH model,

In the EGARCH model, the conditional variance is an asymmetric function of lagged disturbances,

In the QGARCH model, the lagged errors’ centers are shifted from zero to some constant values,

In the TGARCH model, each lagged squared error has an extra slope coefficient,

where the indicator function is one if and zero otherwise.

The PGARCH model not only considers the asymmetric effect but also provides another way to model the long memory property in the volatility,

where and .

The conditional correlation estimator is

where is the element of S, the unconditional correlation matrix.

If you specify CORRCONSTANT=EXPECT, the element of the unconditional correlation matrix S is

where T is the sample size.

As shown in the CCC GARCH models, the following formulas are recursively implemented to obtain the multistep forecast of conditional error variance and :

In the preceding equations, for . Then, the multistep forecast of conditional covariance matrix , is calculated by

where is the diagonal matrix with element , and is the matrix with element ,

The log-likelihood function of the multivariate GARCH model is written without a constant term as

where is calculated from the first-moment model (that is, the VARMAX model or VEC-ARMA model). The log-likelihood function is maximized by an iterative numerical method such as quasi-Newton optimization. The starting values for the regression parameters are obtained from the least squares estimates. The covariance of is used as the starting value for the GARCH constant parameters, and the starting values for the other GARCH parameters are either or , depending on the GARCH model’s representation.

In multivariate GARCH models, the optimal (minimum MSE) l-step-ahead forecast of endogenous variables uses the same formula as shown in the section Forecasting. However, the exogenous (independent) variables, if present, are always assumed to be nonstochastic (deterministic); that is, to predict the endogenous variables, you must specify the future values of the exogenous variables. The prediction error of the optimal l-step-ahead forecast is , with zero mean and covariance matrix,

where is the h-step-ahead forecast of the conditional covariance matrix. As emphasized by the subscript t, is time-dependent. In the OUT= data set, the forecast standard errors and prediction intervals are constructed according to . If you specify the COVPE option, the prediction error covariances that are output in the CovPredictError and CovPredictErrorbyVar ODS tables are based on the time-independent formula

where is the unconditional covariance matrix of innovations. The decomposition of the prediction error covariances is also based on .

Define the multivariate GARCH process as

where , , and . This representation is equivalent to a GARCH() model by the following algebra:

Defining and gives a BEKK representation.

The necessary and sufficient conditions for covariance stationarity of the multivariate GARCH process are that all the eigenvalues of are less than 1 in modulus.

The following DATA step simulates a bivariate vector time series to provide test data for the multivariate GARCH model:

data garch;
   retain seed 16587;
   esq1 = 0; esq2 = 0;
   ly1 = 0;  ly2 = 0;
   do i = 1 to 1000;
      ht = 6.25 + 0.5*esq1;
      call rannor(seed,ehat);
      e1 = sqrt(ht)*ehat;
      ht = 1.25 + 0.7*esq2;
      call rannor(seed,ehat);
      e2 = sqrt(ht)*ehat;
      y1 = 2 + 1.2*ly1 - 0.5*ly2 + e1;
      y2 = 4 + 0.6*ly1 + 0.3*ly2 + e2;
      if i>500 then output;
      esq1 = e1*e1; esq2 = e2*e2;
      ly1 = y1;  ly2 = y2;
   end;
   keep y1 y2;
run;

The following statements fit a VAR(1)–ARCH(1) model to the data. For a VAR-ARCH model, you specify the order of the autoregressive model with the P=1 option in the MODEL statement and the Q=1 option in the GARCH statement. In order to produce the initial and final values of parameters, the TECH=QN option is specified in the NLOPTIONS statement.

proc varmax data=garch;
   model y1 y2 / p=1
         print=(roots estimates diagnose);
   garch q=1;
   nloptions tech=qn;
run;

Figure 81 through Figure 85 show the details of this example. Figure 81 shows the initial values of parameters.

Figure 82 shows the final parameter estimates.

Figure 83 shows the conditional variance by using the BEKK representation of the ARCH(1) model. The ARCH parameters are estimated as follows by the vectorized parameter matrices:

Figure 84 shows the AR parameter estimates and their significance.

The fitted VAR(1) model with the previous conditional covariance ARCH model is written as follows:

Figure 85 shows the roots of the AR and ARCH characteristic polynomials. The eigenvalues have a modulus less than one.