Page 3 of 3

Practical solution for fat-tail risk management?

Posted: August 9th, 2013, 5:22 pm
by And2
QuoteOriginally posted by: frenchyWillJe vois qu'il y a quelques français par ici (ou du moins des francophones) : nous finirons par conquérir toutes les salles des marchés ...C'était vraiment fabuleux!(...Sorry, could not resist )

Practical solution for fat-tail risk management?

Posted: August 11th, 2013, 10:14 pm
by purbani
Hi MizhaelYou may be interested in Why Distributions Matter and;Four Moment Risk DecompositionIn the latter we originally tried and partly succeeded to come up with a 'modified vol' number but the Cornish Fisher expansion cannot model sufficiently high levels of excess skewness and kurtosis to represent extreme / fat tail events. These do seem to be better represented by a Type I Extreme Value distribution viz Gumbel, Pareto or Weibull type distribution with even the skew T not being sufficient. Obviously as all the anti-frequentest / Taleb types point out there is no point in doing any of this unless your sample contains at least a few extreme events. Banks basing their 'stress tests' on the past 12 months of historic data thus not very 'stressful'.Kind regards,Peter Urbani

Practical solution for fat-tail risk management?

Posted: September 8th, 2014, 3:55 pm
by slacknoise
There is a recent paper about tail risk which i found quite interesting: "Robust and Practical Estimation for Measure of Tail Risk" http://papers.ssrn.com/sol3/papers.cfm? ... id=2444381

Practical solution for fat-tail risk management?

Posted: September 10th, 2014, 4:08 am
by Samsaveel
to answer your question about fat tail dynamics ,you have to assume non-Gaussian governing dynamics for your asset return distribution.in a Gaussian setting you Normal-VaR is proportional to your SD,i.e Pr[Loss > x] =0.01 @ 99%.here you are assuming Gaussian dynamics.basel adds a multiplier of [3,4 ] depending on the robustness of your risk management infrastructure.the question is where does the multiplier come from ?if you only assume that the underlying P&L distribution is symmetric then an upper bound on VaR would be applying Jensen Pr [ L > x] <= sigma^2/( 2*x^2).rearranging gives x < = sqrt(1/0.02) * sigma so that the upper bound loss <= 7.07 sigma ,i.e should be less than 7.07 SD ,this is under the assumption of symmetric P&Lso clearly at the 99% under Gaussian assumption we know that the 99% is 2.33 SD ,if you take the ratio of 7.07/2.33 = 3.03 approx ,an that's how you adjust your VaR.