Not sure I've learnt much more than I already knew, but I did glean something by reading between the lines. Thanks.

One thing I should have noted though is that there is no need to model individual underlyings in order to get delta hedges; once can simply take the spread delta and multiply by 1 or -1. But one thing is clear - I dont really think anyone really knows how to risk manage CSOs very well; no wonder they contributed to the blow up of Amaranth!

QuoteOriginally posted by: yetanotherquantOne thing I should have noted though is that there is no need to model individual underlyings in order to get delta hedges; once can simply take the spread delta and multiply by 1 or -1.In almost all sane models, you'd expect to see some very strong vega correlations with your effective deltas, and the vega reactions are almost always going to be more complicated than just the spread because not only the ATM vol-spread moves, but the skew can also react quite drastically.QuoteBut one thing is clear - I dont really think anyone really knows how to risk manage CSOs very well; no wonder they contributed to the blow up of Amaranth!I think Amaranth's problem was not that they didn't know how to risk manage - though they probably didn't - but more that they didn't think they needed to! We used to wonder about that a lot as we bought their portfolio. No matter - it kept us busy for many weeks.

Thanks - I guess you are referrring to smile-adjusted delta (partial_V/partial_F + vega * partial_vol/partial_fwd) versus unadjusted delta? That is very model dependent, and I suspect that will definitely change hedging ability.

Absolutely so. But I would argue that the key to a decent CSO model is how you handle joint skew/smile dynamics. If you aren't going to model this, you might as well just use Kirk. You can sometimes do it implictly and just model the spread rather than having to do a full model the separate underlyings, but you must at least think about it.

Just so that we are on the same page, I am not arguing in favor of a simplistic model - remember, I am just asking a question and trying to understand the answers! (btw, Kirk is useless).For European options I'd imagine "dynamics" to mattter only insofar as it affects the skew-adjusted deltas. But the larger question still remains: froma hedge analysis is it easy to say that a 2-factor model that captures this joint dynamics is indeed better than a "well-crafted" spread model?

Sorry - I certainly didn't mean to imply that you thought Kirk was a "good" model - I was just noting that, to do significantly better, you have to work quite hard.It's not so much the number of factors but how you spend them that matters. I could have two GBM factors in a joint-price model (like Kirk) or I could have two factors in some stochastic volatility spread model (one for the spread, one for the stochastic variance). I could cook up a billion possible alternatives. I have written 1-factor spread models that do a better job than the 2-factor Kirk model - as you probably have, too.Just adding factors doesn't help unless you make them count. I used to work with 3-factor IR models - they were still terrible at pricing and risk-managing yield-curve spread options. The same is true of a lot of the credit correlation models which performed so badly in the last few years. Similarly, Kirk isn't a bad model because it has too few factors, but because it forces your distribution to be bivariate normal in a case where the only reason people are really interested in trading the product is _precisely_ because no one believes that it is!If your higher dimension model captures empirically relevant effects without introducing a whole feast of un-measurable and/or unhedgeable things then, yes, I would say it is better than your lower dimensional model. If not, I would say the lower dimension model is better by Occam's Razor. But, all too often, people assume that just because they are adding factors and making things more complex, that they are meaningfully and usefully capturing richer dynamics - that is just not true. Usually all they are doing is making their model mis-specification less obvious so that it will bite them in a more painful place later on when things really blow up.The thought process should be "what do I need to model to make this product safe for my children/traders?" and not "how many factors should I use?".

I understand (and totally agree) with your statements regarding the number of factors. My question was more like "how does a "good" spread model compare with say a bivariate copula model (with appropriately calibrated correlation) in a hedge analysis"... maybe I have to do my own analysis, but I have a feeling the answer will not be straighforward.

I'm afraid so. But good luck!

HI , I am trying to price calendar spread options in commodities. My starting point is simplest form of Monte Carlo on individual prices, using correlation. However its not giving very good results. Can anyone please help with any kind of pointers on this.