Lag selection in VAR

1. What should be optimim lag size for VAR?
Tringa Ymeraga‎ posted the following figure.


















Saad Riaz commented> why you use max lag of 3? use 4 to 8 lags and than
apply this lag selection criteria....

Zia Eco Marwat commented> Use different lags ..... when u find the lowest AIC
value then that would be Ur proper lags.

Muhammad Anees commented> Commonly VAR with 4/5 will give you a
refernece. Changing the lag length in estimation of VAR will not affect the
model due to efficiency and parsimoniousness reasons. The model with * on
the AIC or BIC for the above two reasons can be selected. With more than 4
variables, select the model with BIC and less than 4 variables, use the model
with AIC.

Zia Eco Marwat comented> Sir here 0 lag has the smallest value... so now?

Olasehinde Timilehin commented> If you are to strictly follow this result.... It
means there was no sign of dynamic in the model. The best option is to use
OLS or any other GLS methods. Moreover... No lags implies long run. Please
read more on long run estimation.

Tringa Ymeraga commented> Thank you guys. In fact, I tried to use more lags
and the same thing happened till in the 6th lag. While, when I choose 7 lags,
the lag order is 7. Is this okay?

Sayed Hossain commented> 7 lags are not recommended. According to the
result, zero lag (as * is there) is the best option. In that case it is not dynamic in
nature. I would suggest to use simple regression model.

Moulana Naykrasyvishyy Cholovik There is no need to over-parameterize a
model with such small sample (29) by including 7 or more lags for yearly data.
Inclusion of too many lags than true lag length is same crime as exclusion of
lags from true lags. You can perform ‘VAR lag exclusion WALD test’ for
additional support to your results. See the IMF working paper on External
Shocks to Inflation in Sri Lanka by Duma, Nombulelo

Muhammad Anees commented> Including more than 3 lags with a model for
more than 4 variables and 40 observations creates efficiency issues. How do
people believe not adding more than 4 variables to a time series of 40
observations while we get a model of 7 or 10 variables. Please check the
comment of Moulana Naykrasyvishyy Cholovik in addition to one of my reply to
your questions.


2. How to choose lag length in VAR?
Professor Shahnawaz Karim has commented as such >> VAR model is built by
expressing each endogenous variable in the model in terms of past values of
itself and other variables. Hence, VAR models essentially include lagged
values of all model variables. But choosing the appropriate 'lag length' is the
real problem. How can we choose the appropriate 'lag length'? In this case,
'data frequency' comes before even 'lag selection' criteria, as we can't even
utilize the 'lag selection' criteria to choose the most appropriate 'lag length'
before estimating a VAR model. Depending on whether the data frequency are
annual, quarterly, or monthly helps us to estimate the initial VAR model to start
with. In case of quarterly and monthly data, we need to use 4 and 12 lags
respectively, to estimate the initial VAR models. Then, how about the annual
data? In this case, the researcher should use his or her intuition. As in most
cases, we're not blessed with long enough annual observations, we should
use a lag length, which is neither too small nor too big, since lags are
important, as they eliminate residual autocorrelation in VAR models. The
question is, if the lag length is optimal or not. Lag length criteria indicate a
definite way of selecting the optimal lag after estimating the initial VAR model.
Under the guideline of 'lag length criteria', we need to use the 'lag length',
which is selected by most of the 'lag selection criteria' named after the
econometricians who developed them, like HQ, SIC, AIC and LR, etc.
Generally, we choose the lag length for which the values of most of these lag
length criteria are minimized, indicated by asterisks in the EViews output. But
the selection of optimal lag length does not stop there. We need to check the
residual autocorrelation of the estimated VAR model, at least for the chosen
lag length. Furthermore, due to multiple lags, VAR models are
overparameterized and lags erode the degrees of freedom and weakens the
strength of 'diagnostic tests'. That's why, we need to perform the 'lag exclusion'
tests to select the minimum lags that eliminates VAR residual autocorrelation.

Lag selection for VECM model
3. Professor Olasehinde Timilehin commented as such > Lag selection in
VECM is rigorous than that of VAR. My valuable advice is to select a lag that
will make the model stable.... and the other assumptions will not be violated...
POPOLAR BLOGS
Dave
Meo School of Research
Shishir Shakya
Noman Arshed
Lag selection in VAR
Hossain Academy Note
Univariate Models
Multivariate Models