The RATS Software Forum

1. That isn't much data for a model that complicated
2. Monthly data tends to have fairly weak GARCH effects, and a "spillover" model doesn't make sense unless those are fairly strong.

power23 wrote:thanks Tom!

I tried I/O split which is very helpful!
I applied different variance specifications rather than spillover, but I have still no convergence..Is there any example about how to implement preliminary simplex iterations?

Note that VARIANCES=VARMA is rather difficult model to fit which is why we use some rather high-end adjustments to the non-linear estimation process using NLPAR. It has numerical problems because the coefficients can be either positive or negative as you can see below. (If you try to impose non-negativity, the model fits only barely better than the model without the cross variable terms). If high residual or high volatility entries don't align very well among variables, a variance for some variable can approach zero at some observation. For instance, here A(2,3) is fairly large negative, so a very big residual in the Swiss data (variable 3) might push the variance of France (variable) 2 negative in the next entry. Any zero (or negative) value for H at any data point results in an uncomputable likelihood resulting in quite a few "dead-end" parameter paths, where you approach a local maximum near a computability boundary. The more dissimilar the data series are, the harder it will be to get estimates with VARMA variances.

power23 wrote: thanks also for the volatiltiy graphs...is it a similar code for correlations?
for "set corr_ad %regstart() %regend() = %cvtocorr(hh(t))(1,2)
GRAPH(STYLE=LINE,HEADER="DCC:Canada-France") 1
# corr_ad 1999:02:01 2016:10:05"
because it has to be created 21 graphs for the pairs

do i=1,7
  do j=1,i-1
      set corrtemp %regstart() %regend() = %cvtocorr(hh(t))(i,j)
      graph(style=line,header="DCC:"+glabels(i)+glabels(j)) 1
      # corrtemp
   end do j
end do i

debbysoraya wrote:Hi anyonem
Right now I try to implement bivariate dcc garch developed by engle (2002).
So what I want to do is analyzing the conditional correlation between two countries using bivariate dcc. (using market indicies in the X and Y, terms of return already)

1. In the literature that I read, sometimes they mention that I should pre white first before runninf my dcc estimation, does it mean that I need to do:
system(model=var1)
variables x y
lags 1
det constant
end(system)

debbysoraya wrote: 2. I got a little confuse so when I want to run the code, should I use
a. GARCH(P=1,Q=1,MV=DCC,VARIANCES=SPILLOVER,rvectors =rr, hmatrices=hh,ROBUST,noprint) / X Y
my question for this one is that in the estima's instruction it says that for engle two step the feature can only be variances=simple or variance= exponential
Does it mean code that I use is garch(p=1,q=1,mv=dcc) / X Y will be enough or if the pre whitean mean including the var, and it will be garch(p=1,q=1,mv=dcc, model=var1) / X Y

OR

b. GARCH(P=1,Q=1,MV=DCC,REGRESSORS) / Lx Ly
# Constant Lx Ly (i got this from few early post from you in this thread

I dont understand in here why there has to be L, in front of the variable. Does it mean if I am doing my stock indices return in two countries my code supposed to be garch(p=1,q=1,mv=dcc, model=var1) / LX L Y ?

debbysoraya wrote: 3. Actually the way you put order for terms variables X Y , does it matter?

TomDoan wrote: With just two variables, there is no good reason not to do a joint estimate. What you're describing sounds like "3-step" estimator (pre-whiten, then univariate GARCH, then DCC). If there's a concern with serial correlation, you just use a VAR for the mean model.

system(model=ar1)
variables rus rfr
lags 1
det constant
end(system)
*
garch(p=1, q=1, model=ar1, mv=dcc, pmethod=simplex, hmatrices=hd, rvectors=rd, piters=10)

debbysoraya wrote:Thanks lot for your help Tom! I have few more question

TomDoan wrote: With just two variables, there is no good reason not to do a joint estimate. What you're describing sounds like "3-step" estimator (pre-whiten, then univariate GARCH, then DCC). If there's a concern with serial correlation, you just use a VAR for the mean model.

So you suggest I just do garch(p=1,q=1,mv=dcc) / X Y ?

debbysoraya wrote: There is one other thing,
so the paper that I follow (the method)
r_t=γ_0 + γ_1 r_(t-1)+ γ_2 r_(t-1)^us + ε_t
the author us AR (1) process in the mean equation
for instance if I am doing DCC Garch between france and US,
the return equation for france at time t = constant + france return ar(1) + US return ar(1)

I looked the user guide and the code for multivariate including ar(1) process will give me only the ar(1) process for the return.
In my case they they will give me mean equation : france return at time t = constant + france return ar(1) without the US return ar(1)

So what I put in the code is

Code: Select all

system(model=ar1)
variables rus rfr
lags 1
det constant
end(system)
*
garch(p=1, q=1, model=ar1, mv=dcc, pmethod=simplex, hmatrices=hd, rvectors=rd, piters=10)

is this correct?

TomDoan wrote: No. That would only be appropriate if there were no serial correlation in the series. It sounds like you need some dynamics in the mean model. It's just that it makes more sense to include the dynamics of the mean directly into the estimates instead of "prewhitening" and using the residuals only.

system(model=var1)
variables rus rfr
lags 1
det constant
end(system)
*
garch(p=1, q=1, model=var1, mv=dcc, pmethod=simplex, hmatrices=hd, rvectors=rd, piters=10)