Evenin' all!I'm building a nice simple parametric VaR model, and need some advice.1) I'm trying to find a good lamda value, and think I might as well at least start off at 0.94 a la risk metrics etc. However, I'm not sure if it's as simple as that. My time horizon is one month, and I have daily data. Presumably the 0.94 is usually used in the context of daily data and one-day time horizon? In that case, do I want to use a lamda such that the most recent month in my scheme has the same weighting as the first day would for a one-day time horizon scheme? That is, should I be doing x^21=0.94 where x is my lamda weighting for daily data? Thus if a one-day time horizon scheme requires around 74 days for the weighting to decay to 0.01, I would need 74 months to decay to the same weighting?2) In the same context, I'm wondering whether it would be best to assume zero average returns when generating my covariance matrix. That is, if I assume zero, it's MeanSquare(R) and if I calculate it directly it's MeanSquare(R-avR). I get the feeling the more "correct" method would be to calculate the average, but this leads on to another question - should that calculation be done in the same exponential weighting way as the variance calculations? In other words, should recent returns count more to the average than long gone ones?I guess many would say that answers to these questions are either a matter of opinion, or simply whatever backtests better, but I'm interested to know if there is strong feeling that there is a right way of doing things...Thanks for any help!PD

if you have daily data, then why not do daily VaR. its always better to parametrize using more data than less data. you can useV(k)=0.94*V(k-1)+0.06*R(k-1)^2.. but still MLE will easily provide you your own number for lambda for monthly kinda horizon VaR.MLE procedure remains unchanged wrt frequency. over a very short horizon like one day, i think even riskmetrics takes zero returns.ewma takes care of weightage problem(more recent more important thing) although not in the most general sense

in line with amit7ul's previous comment...All this depends on your choice of time horizon (which is arbitrary).We find that weighted models perform pretty well with daily VaR. And the mean of the daily returns are quite small in this perspective.The original poster is correct to say that the exponential weight and decay factor depends on which has better backtest results.It is acceptable to use the risk metrics lambda (0.94) as a starting point.As a rule of thumb, if your asset class exhibits frequent and drastic changes in volatility, the larger the weight.If your asset class exhibits stable volatility, weighting is not necessary (all equal weights).

Thanks for the replies.You're right, the best solution would be to do daily VaR and scale up (or not even scale up - just run with daily VaR).Unfortunately, I wasn't quite clear in my first question. In fact, for some of the assets (freight FFAs and underlyings) I have daily historical data and for some, I only have weekly. Monthly VaR is chosen for practical purposes. I could interpolate my weekly to daily, but I'd be rather cautious about that - this market can do some silly stuff in a week!That leads me on to another question - this market is certainly not normal (in any sense!). I believe (although I have not proved it yet) that the high kurtosis is causing my gaussian parametric VaR to always underestimate the number of expected outliers. That is, my VaR uses a 5% tail, but backtesting so far shows consitently higher (6 to 7%) exceptions for all variations of lamda etc. _IF_ this is just due to kurtosis, or some other distribution quirk, then is it ssen as acceptable to whack a premium onto the VAR to make it right, or does this just smack of slap-dashery?! Of course, the idea solution is to analyse the real distribution and find the 5% tail position, but there is a certain amount of haste required just now, and oh so many ifs and buts regarding the distributions...It is interesting that jomni says that for markets with frequent drastic changes in volatility, a fast decay is more appropriate. I'd say that certainly describes this market. Is there a sensible lower limit to lamda? Would it be totally silly to set it such that the decay is finished within a historical timeperiod equal to the forward period for which you're forecasting the vol? Eg in my case, I am trying to predict a 1 month vol, so might it be OK to use only the previous month's returns to make this prediction? For me, I guess my weekly data will cause too much trouble for this, but in principle...?Ta for any help!

Oh, so you're talking about Freight Forwards... then a daily VaR horizon will definitely not make sense. Weekly-monthly might do.Have you tried Historical Simullation VaR instead of Parametric? This way, you will not assume normality.If you find it necessary to use weighting for your Historical Sim VaR, it is also possible to do so.Sorry I don't have actual experience with this asset class...

Hmm... I currently do a historic VaR, but don't weight it. Because of that, the enormous vol of the last few years crops up in current VaR, so I don't use it in a serious manner.What I'm planning is a lovely big MC engine which will solve all my problems - and generate a big pile of new ones...Right now, the parametric approach works well for a diversified portfolio, but not not so well for naked positions (shows a lower than real risk) or spreads (shows a higher than real risk), for obvious reasons.Would you say this is relatively acceptable for an initial model, used with some pragmatism, or that it should be shown to be accurate under all circumstances?

QuoteOriginally posted by: jomniin line with amit7ul's previous comment...All this depends on your choice of time horizon (which is arbitrary).I disagree completely with this statement. There is a correct time horizon (or at least a great deal of clearly incorrect time horizons) depending on strategy, trade structuring, liquidity, etc. If you are trading near-dated options (or at least have a lot of options with <1 month to expiry), using a 1 month MC VaR horizon would drastically understate your true VaR. Options would expire and your model (unless it is very sexy) would assume the positions are not rolled or replaced (even though it is likely they would be), therefore dramatically under-estimating risk. On the otherhand, if you are trading what are essentially private equity positions (like so many self-styled "hedge funds" do these days) even a 6mo horizon won't do because you couldn't get out of stuff that fast if things hit the fan.