Diebold and Yilmaz 2012

[TomDoan]

Your series are mostly NA's and all in different patterns. If you do a 4 lag VAR, just to get a single usable entry, you need every single series to have five consecutive entries of actual data ending with that entry. USELEC barely has five consecutive entries itself. Why do the series have so many holes, and what were you planning to do about that?

[TomDoan]

You don't have seven day a week data, do you? If all Saturdays and Sundays are missing, you have five day a week data. As you've written that, you are only going to get one usable entry per week. Change your CALENDAR to reflect the data that you do have.

You're also not adapting that to the larger number of series. The first graph, for instance, has

spgraph(vfields=2,hfields=2,$

which allows for only four fields, when you're trying to graph eight series. They end up sitting on top of each other. Also a 100 point rolling window is rather small given the size of your model.

[TomDoan]

Anywhere between the names on those long lines, such as:

Code: Select all

data(format=xlsx,org=columns, sheet="us_sec",top=2,left=2) 2008:01:01 2017:06:30 USAIR USBANKING USBEV USBLDG USCABLE $
 USCHEM USCGM USCON USCSM USELEC USELTN USOF USFOOD USHLT USSUP USHOME USIND USLEI USINF USREITS USRT USRET $
 USPROP USOIL USGAS USMTLS USMACH USLOD USPUB $ USPHAR USRAIL USSTRS USSVS USTELECOM USTEXT USTOB USTRSP USTL USVEH

A 200 width window with four lags and 39 variables leaves almost no degrees of freedom. The rolling window analysis is suspect to start, but with windows too short for the data, it will be statistical nonsense.

[TomDoan]

You need to put <<enter>> after the $. The $ should be the last thing on a line.

[TomDoan]

Before you ask about it, your USLOD series is missing a huge chunk of data near the end which will make the moving windows fail. You should figure out what's up with that and either drop it, or cut the overall range of the analysis to stop where it (effectively) ends.

[TomDoan]

First of all, the "pairwise net spillovers" are really just junk numbers. (It's basically subtracting apples from oranges). In the original paper, there were four series, so there are six pairs. You have nine series, so there are 36 pairs, but there's no obvious way to organize those. If, notwithstanding the fact that they really don't make much sense, you want to graph them, you're probably better off doing a separate page for each series which would have graphs for the nets vs the other eight series.

[TomDoan]

Dear Sir,
I am doing some sensitivity analysis on my model to check its robustness. When it comes to direct and iterative forecasting, can they be applied to the DY methodology?

I'm not really sure what you mean. DY doesn't do any forecasting at all.

Also, when it comes to the window width given >2480 observations should I go for a width greater than 400, given the sample size?

You don't want the windows to be too small relative to the size of the model. The size relative to the whole data set isn't that important in comparison.

[TomDoan]

It looks like you have an error in your data set. CS_EQ is exactly equal to cs_o_eq{2}. (Look at the results for the ESTIMATE for CS_EQ---perfect fit with a 1.0 coefficient on CS_O_EQ{2}).