Hello,
I have generated 511 correlation estimates using a DCC GARCH model with a VECM mean specification. I am wondering what test I can use, if any, to see if the correlations are statistically different from one another (between wheat/corn and corn/ethanol). The generated correlations are not normally distributed. Would something like a Mann-Whitney test be appropriate?
DCC GARCH Correlations Test
Re: DCC GARCH Correlations Test
No. The DCC correlations are most emphatically NOT data---there is no test that would be appropriate for data that could be used.
If the correlations are the same, CC would work. Now the CC and DCC models don't properly nest but if CC (which has more free parameters than DCC once N>2) gives you a lower log likelihood, then you could safely conclude that the correlations aren't constant.
If the correlations are the same, CC would work. Now the CC and DCC models don't properly nest but if CC (which has more free parameters than DCC once N>2) gives you a lower log likelihood, then you could safely conclude that the correlations aren't constant.
Re: DCC GARCH Correlations Test
Hello Tom,
Thanks for the reply. I have already compared the CCC model with my DCC model and my conclusion is that the DCC model is more appropriate. What I am trying to do is compare the relative magnitudes of the mean correlations (for example, the mean correlation for wheat/corn is greater than corn/ethanol from 2008 to 2018) and rank them in order of strongest to weakest. But to do that, I need to ensure that the mean correlations are statistically different from one another.
So if I can not use some parametric or non-parametric test, could I do something like construct confidence bands for each correlation series and see if they overlap?
Thanks for the reply. I have already compared the CCC model with my DCC model and my conclusion is that the DCC model is more appropriate. What I am trying to do is compare the relative magnitudes of the mean correlations (for example, the mean correlation for wheat/corn is greater than corn/ethanol from 2008 to 2018) and rank them in order of strongest to weakest. But to do that, I need to ensure that the mean correlations are statistically different from one another.
So if I can not use some parametric or non-parametric test, could I do something like construct confidence bands for each correlation series and see if they overlap?
Re: DCC GARCH Correlations Test
GARCHMVDCCGIBBS.RPF does Gibbs sampling for a DCC model, which will produce a series of correlations for each pair for each draw. You can do a statistical analysis of those across draws.