[R-SIG-Finance] change-point detection in highly dependent time series

[Johannes Moser]

Dear all,

in the context of a scenario analysis framework I manually selected some 
quiet and some stressful periods in the last 25 years of DAX returns. 
Now I wish to confirm this choices by means of a change-point detection 
test in R. The test ideally should pick changes in the conditional mean, 
more importantly in the conditional variance and covariance structure.

The test should be able to handle highly-dependent time series. So tests 
that assume e.g. structural breaks in iid gaussian noise or assume 
independence between the breaks in another way are not the right thing 
(chow test etc.). I've already tried to apply tests that assume 
independence to the standardized returns (after fitting a GARCH model 
and extracting the volatility estimates) as the standardized returns can 
be considered close to iid. But this approach didn't seem to work well 
(package 'changepoint', test 'segneigh.var.css').  The results seem 
intuitive when this test is applied to the rough returns series. But 
this is clearly a violation of the independence assumption.

I further tried the wavelet test for covariance stationarity (test 
'hwtos2' in the 'locits' package) but it is much too sensitive (e.g. it 
rejects covariance stationarity in a simulated Gaussian GARCH(1,1) ). 
Also, the test (and R syntax) that comes with the paper of Fryzlewicz 
(2014) "Multiple-change-point detection for auto-regressive conditional 
heteroscedastic processe" has been applied to the returns series, but 
I'm not quite satisfied with the results.

There are many more change-point detection packages and tests in R that 
I have considered. But so far none seems to be appropriate for financial 
returns and gives results that are interpretable in the context of 
knowledge of crisis periods and visual examination of the series.

Is there any test for change-points in highly-dependent data (such as 
financial returns) that someone could recommend from his own experience? 
I'd be more that thankful for any advice!

Kind regards, Johannes