[R-SIG-Finance] Riskmetrics volatility and correlation estimation

[Patrick Burns]

The operation is essentially the same whether univariate
or multivariate -- just do an outer product of the returns
at each time.  For the univariate case this reduces to the
squared return.


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")

Murali Menon wrote:

>Brian,
>
>Thanks for the suggestions and tips. Indeed, after nabbling away in the archives, I found a Zivot statement to the effect that EWMA (function available in fMultivar) on squared returns is equivalent to RiskMetrics vol. No equivalent methodology as far as I can see for correlations, so I thought I'd post a query before attempting to implement it myself.
>
>As for RiskMetrics, I've seen the following paper that discusses why RiskMetrics seems to work reasonably well in practise even if it is mis-specified for the data (algorithms given herein):
>
>http://www.colbud.hu/pdf/Kondor/riskm.pdf
>
>The idea is that for the short-horizon, 95%-confidence level sort of VaR computation, even non-normal distributions have quantiles close to the normal one, so RiskMetrics works okay; but will fail to do so for longer horizons and/or 99%-ile queries.
>
>Engle's DCC paper: http://pages.stern.nyu.edu/~rengle/dccfinal.pdf shows that RiskMetrics methodology outperforms the rolling technique in estimating and forecasting correlation, even if not always so well as dynamic conditional correlations (which I don't think is available in R either).
>
>Thanks,
>
>Murali
>
>  
>
>>Date: Wed, 12 Dec 2007 06:05:07 -0600
>>From: brian at braverock.com
>>To: feanor0 at hotmail.com; r-sig-finance at stat.math.ethz.ch
>>Subject: Re: [R-SIG-Finance] Riskmetrics volatility and correlation estimation
>>
>>Murali Menon wrote:
>>    
>>
>>>Are there functions available to compute the Riskmetrics (1996)
>>>volatilities and correlations for financial time-series?
>>>I refer to the exponentially weighted moving average vols
>>>and exponentially smoothed correlations (with lambda = 0.94).
>>>I looked in the VarModelling part of fPortfolio, but this
>>>stuff doesn't seem to be there?
>>>      
>>>
>>Not that I know of, but they shouldn't be too hard to construct.
>>
>>So, if you want help constructing them from this group:
>>
>>1> post the link to the RiskMetrics algorithms
>>
>>2> do a little research on Google and the list archives
>>
>>There have been several examples posted of exponentially weighted moving
>>averages in R which should go a long way toward solving the volatility
>>question above, for example.
>>
>>R contains many different smoothing algorithms, as you can see with
>>help.search("smooth") or help.search("smoothing")
>>
>>3> suggest an approach, try an approach, post your failures
>>
>>and I'm sure someone here can probably help you out.
>>
>>Another interesting (to me anyway) question is "Why do you care?" What
>>in the literature leads you to want to try these techniques? Are you
>>just trying to replicate a set of RiskMetrics algorithms in R? Have you
>>looked at other research on smoothing and rolling windows? I ask this
>>trailing set of questions because often when I go looking in the
>>literature, I find that there are several newer techniques which have
>>been shown to work better than the older methods. Sometimes these newer
>>techniques are already implemented in R, other times I have to do it
>>(but at least I'm then implementing a more modern approach).
>>
>>Regards,
>>
>>- Brian
>>    
>>
>
>_________________________________________________________________
>[[replacing trailing spam]]
>
>_______________________________________________
>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.
>
>
>  
>