EViews Limits?

[elias41]

Hello,

I have the following problem:

When I import more than 100 time series with 257 observations each to estimate the BEKK Modell with

"sys01.arch(coef = beta,deriv = fa, tdist, m = 10) @diagbekk c(indef) arch(1) TARCH(1) garch(1)"

EViews 8 Enterprise Version takes a lot of time with my computer (RAM 8 GB, Intel Core i7 3.4 GHz, 64 Bit) for this operation and seems to get stucked at the step "Estimation setup". After roughly 1,5 h the programm breaks down. What causes EViews to take so much time and what causes the break down? My company thinks about buying a more powerful computer. So for us it is important to understand if it could be an EViews limit which prohibits the calculation.


Regards,
Elias

[EViews Glenn]

What equations are in SYS01?

[elias41]

Just the usual mean equations: series01=c(1), series02=c(2) ... , seriesn=c(n)

[elias41]

With 31 time series CPU Usage is on average at ~25%
With 100 time series CPU Usage is on average at ~62%, EViewsRAM Usage ~0,4GB
With 140 time series CPU Usage is on average at ~62%, EViews RAM Usage ~3,5 GB
With 180 time series CPU Usage is on average at ~62%, EViews RAM Usage ~5,5 GB

1. What is needed to avoid a break down?
2. What is needed to speed up the calculation?

[EViews Gareth]

Just so we're clear - you're estimating a single BEKK model with 180 variables in it?!

[elias41]

No! I estimate a BEKK model for 180 time series using TARCH like in the above mentioned line. That should amount to 17,010 parameters.
One estimation just got finished with 166 time series so I think the crash issue should be solved. But it took about 3 h and 50 minutes, so my second question remains. (Does a Xeon brings much more performance than my above described i7 pc?) Any help to boost performance is greatly appreciated.

[EViews Gareth]

I doubt switching to a Xeon will make much of a difference. A 180 variable BEKK is extremely unusual...

With 17,000 coefficients, the coefficient covariance matrix is huge.

[elias41]

Does the EViews team have experience with such large scale operations?
Would a server speed up this calculation?
Who is the right person I should talk to about this issue? Does EViews provide any technical support, e.g. a hotline, concerning our performance issue?

[EViews Gareth]

Throwing hardware at the problem won't speed it up any, unfortunately. At least not with current computers.

Multivariate ARCH was designed with, say, five variables in mind. Not 180.

[elias41]

I don't agree with your statement about multivariate ARCH models.
Are you sure about hardware bringing no change?
We observe EViews taking as much CPU as the system allows EViews to do with distributing utilization almost equally between the 4+4 cores of the machine.
Statcon told us the EViews provider would have to prepare a virtual installation in case we wish to change machines. So the EViews team might have expirience with more powerful machines. Could you therefore provide an answer to the question if EViews provides a technical support somehow? Many thanks.
We think of buying a workstation with more cores as we observed the above phenomena. Can you assure more power won't speed it up???

[EViews Gareth]

Your above posts seem to indicate that CPU usage was not at maximum, so obtaining a faster, or more CPUs will not necessarily improve the speed. Equally you're not using all of your RAM either, so getting more RAM probably won't help.

You are speaking to the EViews technical support team.

[startz]

Gareth,

Might messing with the multi-processor/multi-core option do any good for this kind of operation?

[EViews Gareth]

The default setting of Auto is probably the best setting.

With 17,000 coefficients, the covariance matrix has nearly 3m elements.

Multivariate ARCH algorithms simply weren't created for handling such large estimation projects. I'm somewhat surprised it estimates at all.

I don't want to recommend purchasing a new computer to improve speeds, when my hunch is that it will only improve speeds by a tiny amount.

[elias41]

Concerning RAM: The usage increases with the number of variables. So in the future we will even have more variables so we definitely need more RAM so we'll just plug more RAM in my pc, e.g. +8.
Concerning CPU: We assume the operating system (Windows 7) reserves about 40% CPU usage unless the user sets the process priority level to high. Then EViews takes up to 75-80% CPU usage and the calculation speeds up with about 25%. Additionally, in contrast to Excel, EViews utilizes all 4+4 cores. This is why we assume more power might help. Nevertheless, your doubts made us cautious again.

[startz]

You already have a very fast machine. There are some workstations that are 6+6 rather than 4+4, and which may be a little faster. So at the outside, you're probably talking about cutting time in half.

But as Gareth points out, if EViews isn't maxing out existing cpu or ram getting a better machine will make very little difference. Perhaps someone else in your organization has a faster machine that you could do a time trial on? Or if your budget is such that buying a state-of-the-art machine isn't a problem , we'd all enjoy hearing how much improvement you can get.