[R-SIG-Finance] Winsorization

[Patrick Burns]

Neil,

I don't recall having the opportunity to think of MVE
in finance.  But I have tried the moral equivalent in
regression (when building risk models).  The results were
that high-breakdown regression did worst.  Best was
Huber M-estimation with quite mild robustness.
Least squares was almost as good as the best, and
better than almost all of the robust regressions.

Pat

ngottlieb at marinercapital.com wrote:
> Patrick:
>
> Have you looked at Rousseau's, minimum volume ellipsoids (MVE) for
> handling outliers?
>
> Curious if so how you found this for handling outliers? 
>
> Neil
>
> -----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 Brian G.
> Peterson
> Sent: Thursday, September 18, 2008 6:47 AM
> To: Patrick Burns
> Cc: r-sig-finance at stat.math.ethz.ch; ??????
> Subject: Re: [R-SIG-Finance] Winsorization
>
> On Thu, 2008-09-18 at 11:00 +0100, Patrick Burns wrote:
>   
>> I disagree with Ajay about the value of Winsorization.
>> Yes, it is ad hoc but it is simple to understand and often results in 
>> reasonable answers.
>>
>> It certainly depends on the context but if we are talking about 
>> financial returns, then I haven't had positive experience with 
>> traditional statistical robustness.
>> (Given that my thesis was on robustness, I don't say this lightly.)  
>> Robustness often gives inferior answers in finance (in my experience) 
>> even when it is obvious that it "should" be the proper thing to do.  
>> This is a phenomenon that I don't understand.
>>     
>
> I have to agree with Patrick.  We proposed an extension above and beyond
> classic Winsorization that would only reduce the outliers that occured
> beyond a certain confidence level (e.g. 95% or 99%).  Traditional robust
> methods have a tendency to ignore outliers rather than simply reduce
> their influence.  In measuring risk, this is clearly quite dangerous.
> We found that by only cleaning outliers beyond a certain confidence, we
> got much more stable and accurate out of sample predictions on a variety
> of risk measures (as well as predictions that compared well to kernel
> estimation and Monte Carlo methods with lower computational burden).
>
> Like I said in my previous email, code and documentation available upon
> request or in the next version of PerformanceAnalytics.
>
> Regards,
>
>     - Brian
>
> --
> http://braverock.com/brian/
> Ph: 773-459-4973
> IM: bgpbraverock
>
> _______________________________________________
> 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.
> --------------------------------------------------------
>
>  
>  
> This information is being sent at the recipient's requ...{{dropped:18}}