Skip to content

EMM: how to make forecast using EMM methods?

11 messages · elton wang, Mark Leeds, Patrick Burns +4 more

#
Hi all,

We followed some books and sample codes and did some EMM estimation,
only to find it won't be able to generate forecast.

This is because in the stochastic volatility models we are estimating,
the volatilities are latent variables, and we want to forecast 1-step
ahead or h-step ahead volatilities.

So it is nice to have the system estimated, but we couldn't get it to
forecast at all.

There is a "Reprojection" Method described in the original EMM paper,
but let's say we reproject to a GARCH(1,1) model, then only the
GARCH(1, 1) parameters are significant, which basically means we
degrade the SV model into a GARCH model. There is no way to do the
forecast...

Could anybody give some pointers?

Thanks!
#
I've heard opinions that GARCH/SV volatility models
are not better on forecasting than simple exponential
moving average volatilities or even rolling window
historical vol.
Any practitioners mind comment?
--- Michael <comtech.usa at gmail.com> wrote:

            
____________________________________________________________________________________
Looking for last minute shopping deals?
#
I can't say much about Garch/SV being better or worse but I know that's
there an approximate functional equivalence between exponential smoothing
and a regular moving average ( i.e: rolling window ). It's something like
lambda = 1/(2n +1) or something like that but I don't remember. It's in any
decent technical analysis book and it's true empirically because I've played
around with it in the past.



-----Original Message-----
From: r-sig-finance-bounces at stat.math.ethz.ch
[mailto:r-sig-finance-bounces at stat.math.ethz.ch] On Behalf Of elton wang
Sent: Thursday, February 28, 2008 4:36 PM
To: r-sig-finance at stat.math.ethz.ch; r-help
Subject: Re: [R-SIG-Finance] EMM: how to make forecast using EMM methods?

I've heard opinions that GARCH/SV volatility models
are not better on forecasting than simple exponential
moving average volatilities or even rolling window
historical vol.
Any practitioners mind comment?
--- Michael <comtech.usa at gmail.com> wrote:

            
____________________________________________________________________________
________
Looking for last minute shopping deals?

_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only. 
-- If you want to post, subscribe first.
#
Michael,

If I understand correctly, you've used some EMM algorithms to estimate
the parameters of a stochastic volatility model.

If this is the case you should now be able to use Monte Carlo methods to
generate forecasts from your model.

That is, you will generate random variables (according to the
specifications of your model), feed them into your model and hence
simulate your stochastic volatility process.

Note sure what references you have been using but perhaps these would be
helpful:

Gallant, Hsieh and Tauchen (1997). "Estimation of stochastic volatility
models with diagnostics", Journal of Econometrics, 81, 159-192. 

Andersen, T.G. H.-J. Chung, and B.E. Sorensen (1999). "Efficient Method
of Moments Estimation of a Stochastic Volatility Model: A Monte Carlo
Study," Journal of Econometrics, 91, 61-87.

Best,

-- G



-----Original Message-----
From: r-sig-finance-bounces at stat.math.ethz.ch
[mailto:r-sig-finance-bounces at stat.math.ethz.ch] On Behalf Of Michael
Sent: Thursday, February 28, 2008 12:56 PM
To: r-sig-finance at stat.math.ethz.ch; r-help
Subject: [R-SIG-Finance] EMM: how to make forecast using EMM methods?

Hi all,

We followed some books and sample codes and did some EMM estimation,
only to find it won't be able to generate forecast.

This is because in the stochastic volatility models we are estimating,
the volatilities are latent variables, and we want to forecast 1-step
ahead or h-step ahead volatilities.

So it is nice to have the system estimated, but we couldn't get it to
forecast at all.

There is a "Reprojection" Method described in the original EMM paper,
but let's say we reproject to a GARCH(1,1) model, then only the
GARCH(1, 1) parameters are significant, which basically means we
degrade the SV model into a GARCH model. There is no way to do the
forecast...

Could anybody give some pointers?

Thanks!

_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only. 
-- If you want to post, subscribe first.
#
Hi Guy,

Thanks for your help! Yes, we have the coefficient estimated using
EMM. And we followed those papers.

Just want to check my understanding about your suggestion:

Do you mean that after we obtain the estimated coefficients,

we run one simulation to obtain the whole sequence of latent variable
(the volatility time series, from time 0 to time t+1),

where time t is today, and t+1 is tomorrow(one step forecast);

And that's one simulation.

And we run such simulation for N times, let's say N=10000,

and obtain 10000 such volatility time series, each ending at time t+1,

and then we take average of the 10000 data points at t+1,

the average will be the mean-forecast of the volatility tomorrow(i.e.
that's the one step forecast that we want)...

Am I right in doing these procedures?

Thanks



On Thu, Feb 28, 2008 at 4:30 PM, Guy Yollin
<guy.yollin at rotellacapital.com> wrote:
#
(This is not being sent to R-help.  It is considered
impolite to cross-post messages, especially on topics
that are purely financial.)

As always, the answer is, "That depends."

The key question is the time frame of the
prediction.  If the prediction is for a month or
more, then there's unlikely to be much advantage
in a fancy model.  If the time frame is a few
days, then something like a garch model will
vastly outperform a rolling window.  How much
a garch model would outperform an exponential
smooth depends on the smoothing parameter (an
exponential smooth is a degenerate form of a
garch model).  As far as I know, there is not a
clear winner between garch models and stochastic
volatility models, but with some evidence that garch
might be better.  Corrections to this impression are
certainly welcome.


Patrick Burns
patrick at burns-stat.com
+44 (0)20 8525 0696
http://www.burns-stat.com
(home of S Poetry and "A Guide for the Unwilling S User")
elton wang wrote:

            
#
Hi Michael,

Yes, this is what I'm suggesting.  Bear in mind, your model estimation
process should have also resulted in volatility estimates for t-1, t-2,
etc.

Your simulation will require one or more of these terms as input (in
addition to the random innovations) since your stochastic volatility
model will have lagged volatility terms.

Good luck.

-- G


-----Original Message-----
From: Michael [mailto:comtech.usa at gmail.com] 
Sent: Thursday, February 28, 2008 5:46 PM
To: Guy Yollin; r-help; r-sig-finance at stat.math.ethz.ch
Subject: Re: [R-SIG-Finance] EMM: how to make forecast using EMM
methods?

Hi Guy,

Thanks for your help! Yes, we have the coefficient estimated using
EMM. And we followed those papers.

Just want to check my understanding about your suggestion:

Do you mean that after we obtain the estimated coefficients,

we run one simulation to obtain the whole sequence of latent variable
(the volatility time series, from time 0 to time t+1),

where time t is today, and t+1 is tomorrow(one step forecast);

And that's one simulation.

And we run such simulation for N times, let's say N=10000,

and obtain 10000 such volatility time series, each ending at time t+1,

and then we take average of the 10000 data points at t+1,

the average will be the mean-forecast of the volatility tomorrow(i.e.
that's the one step forecast that we want)...

Am I right in doing these procedures?

Thanks



On Thu, Feb 28, 2008 at 4:30 PM, Guy Yollin
<guy.yollin at rotellacapital.com> wrote:
estimate
to
be
volatility
Method
estimating,
#
But I doubt this is not a one-step forecast.
For one-step cast, you only need start from today's
value and simulate one step ahead. you need to use the
orignal innovations as of today instead of simulating
from day 1.
--- Guy Yollin <guy.yollin at rotellacapital.com> wrote:

            
____________________________________________________________________________________
Be a better friend, newshound, and
#
Hi Guy and Elton,

Thanks for the replies.

However this is exactly the weird thing about EMM. It simulates the
latent variables when doing the estimating itself. So there is no
clear estimates of the latent variables itself. That's to say, I don't
have a "today's value" to work out the 1-step ahead forecast...

That's kind of strange...

Any thoughts?
On Fri, Feb 29, 2008 at 7:37 AM, elton wang <ahala2000 at yahoo.com> wrote:
#
For simple SV models (e.g. log normal ar(1)), the model can be written in
state space form and the the Kalman filter may be used to forecast the
latent volatility. See Harvey, Ruiz and Shephard's paper in ReStud for
details. However, the Kalman filter is only the best linear forecast. In
general, the SV models are non-linear and non-gaussian state space models
and the optimal forecasting algorithms are given by the particle filter. I
have a short paper that describes how to do this on my webpage

 http://faculty.washington.edu/ezivot/research/Creal_Gu_Zivot_2007.pdf

-----Original Message-----
From: r-sig-finance-bounces at stat.math.ethz.ch
[mailto:r-sig-finance-bounces at stat.math.ethz.ch] On Behalf Of Guy Yollin
Sent: Thursday, February 28, 2008 4:30 PM
To: Michael; r-sig-finance at stat.math.ethz.ch
Subject: Re: [R-SIG-Finance] EMM: how to make forecast using EMM methods?

Michael,

If I understand correctly, you've used some EMM algorithms to estimate the
parameters of a stochastic volatility model.

If this is the case you should now be able to use Monte Carlo methods to
generate forecasts from your model.

That is, you will generate random variables (according to the specifications
of your model), feed them into your model and hence simulate your stochastic
volatility process.

Note sure what references you have been using but perhaps these would be
helpful:

Gallant, Hsieh and Tauchen (1997). "Estimation of stochastic volatility
models with diagnostics", Journal of Econometrics, 81, 159-192. 

Andersen, T.G. H.-J. Chung, and B.E. Sorensen (1999). "Efficient Method of
Moments Estimation of a Stochastic Volatility Model: A Monte Carlo Study,"
Journal of Econometrics, 91, 61-87.

Best,

-- G



-----Original Message-----
From: r-sig-finance-bounces at stat.math.ethz.ch
[mailto:r-sig-finance-bounces at stat.math.ethz.ch] On Behalf Of Michael
Sent: Thursday, February 28, 2008 12:56 PM
To: r-sig-finance at stat.math.ethz.ch; r-help
Subject: [R-SIG-Finance] EMM: how to make forecast using EMM methods?

Hi all,

We followed some books and sample codes and did some EMM estimation, only to
find it won't be able to generate forecast.

This is because in the stochastic volatility models we are estimating, the
volatilities are latent variables, and we want to forecast 1-step ahead or
h-step ahead volatilities.

So it is nice to have the system estimated, but we couldn't get it to
forecast at all.

There is a "Reprojection" Method described in the original EMM paper, but
let's say we reproject to a GARCH(1,1) model, then only the GARCH(1, 1)
parameters are significant, which basically means we degrade the SV model
into a GARCH model. There is no way to do the forecast...

Could anybody give some pointers?

Thanks!

_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only. 
-- If you want to post, subscribe first.

_______________________________________________
R-SIG-Finance at stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-finance
-- Subscriber-posting only. 
-- If you want to post, subscribe first.