gauravku Wrote: ------------------------------------------------------- > For second one, I believe A is an advantage they > have over the variance-covariance method, because > they don’t have any such underlying assumptions, > MC is just random number generation and historical > is just a crude measure on the historical data, I > didn’t choose B because it was very general and > all the models (even variance covariance) suffer > that risk; I chose C as historical needs a lot of > data and MC also uses a lot of data for analysis. > False, historical VaR needs requires very little data. We spent a lot and I mean a lot of time on absolute VaR in my risk management course and this is not true. Even though “MC is just random number generation” it still requires covariance specifications in determining the stochastic framework. As Pylon said, regardless of whether there is an explicitely calculated covaraiance matrix in the historical method it is clearly assumed that the matrix continues into the future.
plyon Wrote: ------------------------------------------------------- > mcpass Wrote: > -------------------------------------------------- > ----- > > 1C. I like it because you can set the SD > relative > > to the benchmark and calculate VAR as relative > to > > the benchmark and of course the normal way as > an > > absolute measure. In fact, I know some > > institutions who do it. > > > > I don’t remember this from the curriculum at all > – do you? Can you explain this method a little > better? What do you mean by “set the SD relative > to the b’mark”??? Do you mean that if the > assets under question have an E® = benchmark and > a SD < b’mark, then there will be negative VAR? > Again… not in the curriculum was it? > > So you’ve made a weak case that “C” might be the > right answer, but done nothing to tell me why the > answer “B” would be wrong. I’m not persuaded. > Sorry, I must have covered relative benchmark VaR in my MSF then, although I’ll check the LII stuff when I get back to my house because I can almost picture it in the material. In the mean time here’s an excerpt from a page I googled. "As mentioned, Relative VaR fixes a target, such as an Index or Benchmark Portfolio. The return is then computed with regards to the Benchmark by computing the difference in returns between benchmark and portfolio rather than as a standalone return. If our Benchmark was the S&P500 we would come up with a result that would translate into “19 out 20 days our portfolio will fall short from the S&P index of no more than perhaps say 1% of the S&P500. The power of this approach lies in the information that can be extracted from the results. Each risk can then be broken down into risk factors (equity, currency, interest rate, spreads, volatility, commodity) but also according to user defined attributes such as industry sectors, regions, desks, customers, counterparties, etc. The final result produces a probability of falling short from the Benchmark’s returns according to the state-space defined” http://www.financial-risk-manager.com/risks/market/benchmarkvar_b.html
> False, historical VaR needs requires very little > data. We spent a lot and I mean a lot of time on > absolute VaR in my risk management course and this > is not true. Agree with you, I don’t have experience on working on a historical VaR model, I checked the curriculum and they dont discuss at all about the amount of data required for this. However, my quote was based on common sense that when you are doing a historical analysis, typically you need more data as it increases your statistical correctness. Even though “MC is just random > number generation” it still requires covariance > specifications in determining the stochastic > framework. As Pylon said, regardless of whether > there is an explicitely calculated covaraiance > matrix in the historical method it is clearly > assumed that the matrix continues into the future. The stochastic framework itself is random distribution and non deterministic, so I belive its independent of the covariance-variance matrix, they can change as its all random.
There are constraints programmed into the MC model for random number generation. Think of it practically, if you are testing VaR of a portfolio that is 50% alternative assets with 0 historical correlation and 50% S&P and someone else submits a portfolio for VaR calculation that is 50% Russell 2000 (higher correlation) and 50% S&P, the MC VaR output will be different for both or else the model would be useless.