Currently I am using historical vol(std) as a measure of risk;I am looking for practical approach to take fat-tail into our account of measurement of risk.It should be a good tradeoff between theoretical complexity and real-world usage...Any pointers?Thanks a lot!

Far out CVaRs and CF-VaRs?

QuoteOriginally posted by: mizhaelCurrently I am using historical vol(std) as a measure of riskThere has never been a fat tail in history?Just make something up. In case you didn't notice, that's the way it has always been done. If you can do better than "Hmmm, the risk is probably about here," then you deserve a 10% salary increase.

Maybe this is the problem with risk management. Only an idiot would go on a web forum an ask where should I buy the euro if I want to make money? Not unless, at the very least, he plans to do the opposite of what everyone else is doing!So how can you harness the devious ingenuity of a trader in a risk-manager role? Maybe select the top trader from the previous year, and give the guy deep-in-the-money calls plus out-of-the money covered calls on $20 million in company stock?But this is not the first time I have advocated on behalf of diversity in risk management.

Last edited by farmer on June 24th, 2010, 10:00 pm, edited 1 time in total.

You might look at p.14 of this journal:http://www.soa.org/library/newsletters/ ... ss17.pdfIs shows how to apply extreme value methods based on historical data. The idea is that past extremes can offer a (scientific) hint at the magnitude and probability of potential 'black swans'. The example uses monthly total returns for a bond index.I think it's very well written but I am certainly biased!

- spacemonkey
**Posts:**443**Joined:**

The most practical thing you can do is to stop thinking about them. 'Fat-tails' are the financial equivalent of phlogiston, and arise from a fundamental misunderstanding of probability theory. Worry about non-stationarity and liquidity, those are a problem.

can you elaborate on what this misunderstanding is?

- spacemonkey
**Posts:**443**Joined:**

QuoteOriginally posted by: pcasperscan you elaborate on what this misunderstanding is?The belief that probability is ontological rather than epistemological. Asset returns cannot 'have' a fat-tailed probability distribution, because a probability distribution is a summary of your information. It isn't a physical property.So, for example, you can't measure the tails of a probability distribution.

A classical fat tail model: Pareto distribution. It was originally proposed for modeling of distribution of household with very high wealth or incomes. Generally, household (or family) income distribution is close to lognormal and is modeled accordingly - except the lognormal model grossly underestimates the number of rich. This is not in finances, of course, but still in economics, where data quality, object complexity, and modeling mindset are generally similar. Pareto distribution is used also to model the probability of severe losses in actuarial calculations. This should be rather close to risk management in finances.

Last edited by yurakm on June 26th, 2010, 10:00 pm, edited 1 time in total.

easiest way is to use historical simulation, isn't it? It is like estimating the distributions and dependency structure (copula) of your risk factors parametric free (not very sophisticated, but nevertheless...) and is in this regard superior to the classical normal distributed / gaussian copula - linked procedures.

Hi folks,Thanks for your replies...I guess for practical usage, I am just looking for a method to generate a variable and the variable can replace my current standard deviation. So hopefully, I can have a new variable named "Modified Standard Deviation" or "Modified Risk Measure" and plug it into my existing application and replace the currently used "standard deviation" ...Any thoughts?Thanks

quite heterogeneous discussion here... mizhael, perhaps you can describe how you estimate your standard deviation and what you do with that afterwards (calculate a VaR in a gaussian setting?).

QuoteOriginally posted by: pcasperseasiest way is to use historical simulation, isn't it? It is like estimating the distributions and dependency structure (copula) of your risk factors parametric free (not very sophisticated, but nevertheless...) and is in this regard superior to the classical normal distributed / gaussian copula - linked procedures.No, the measure is used somewhere else as an input parameter...Cannot see how I can get a "modified/improved standard deviation" number from simulation...

QuoteOriginally posted by: pcaspersquite heterogeneous discussion here... mizhael, perhaps you can describe how you estimate your standard deviation and what you do with that afterwards (calculate a VaR in a gaussian setting?).How I estimate my standard deviation?Just the simplest standard deviation, past 250 days...that's all.Then this std measure is used somewhere else as an input parameter...

- frenchyWill
**Posts:**59**Joined:**

I'm doing an internship on the computation of risk measures using heavy tails distribution. If you want, I can send you a couple of very good references.

GZIP: On