Well, it's not MY VaR methodology, just the one I use for simple projects...From what I've gathered, the variance-covariance approach could be summarized in the following steps (nuts-and-bolts work here really):1)obtain historical prices2)calculate the log returns and test for normality. If normal, proceed....3)calculate the mean log returns vector (R)4)determine your portfolio & the vector of weigts of each instrument in the portfolio (W)5)calculate the expected log return of the portfolio by multiplying the transposed weights vector by the mean log returns vector. W'R (or R'W), whichever, which yields RETport (portfolio return).6)calculate the variance/covariance matrix from the log returns(S)7) obtain the portfolio variance by: W'SW8) Taking the square root of the portfolio variance, calculate the portfolio standard deviation. (SEport)9)determine a 95% confidence interval (or whichever alpha percentage you want) confidence interval by RETport+/-(1.65*SEport).10)Take the lower bound (LB) of that confidence interval = (RETport-(1.65*SEport)and stick it in the formula V1=V0*e^^(LB), where V1 is the portfolio value at the time 1, and V0 is the value at time 0 (which is given).11) Calculate V0-V1 to obtain the estimated VaR.Is this correct? Did I forget anything? And another question... if you have a portfolio with short entries, should you treat those as having a negative weight? Thanks for the help guys!

Nobody? Just a 'yes, that's right' or a 'no, step x is wrong' would be helpful. I'm in doubt here....

Ok, I need to clarify some things first before I dive into this...R is a column vector, the mean log return is the average of LN(S(i+1)/S(i)), where S(i) is the price of stock on day i.W is also a column vector. If some stocks are short, then yes, they have negative weights.VERY IMPORTANT POINT - You need to specify the time scale for your VaR. Is it the 10 day Var, one day, one month??? If you are using daily stock prices in your historical data then your mean returns and variances are all one day mean returns and variances. Therefore, your final VaR is a 0ne day VaR. You didnt mention this anywhere so I assume you have accounted for it!The main point that confuses me is why are you working with log returns??? Why not just stick to the actual price and actual returns?? Sam

This is basically correct, but I mention a few points.Steps (1) through (4) and (6) require that you have first broken your positions down into sensitivities to market factors. It is the market factors that have a mean and covariance matrix, not necessarily the log returns of your securities. In equities, people sometimes assume that the log return each equity is a market factor. But in other markets, including equity derivatives or equities with important FX exposure, you cannot make the securities and market factors the same thing.For example, if you had a USD corporate bond portfolio, your market factors might be treasury interest rates at various tenors, spreads for various credits and tenors and an idiosyncratic spread for each issuer. You would simulate these, then price each bond based on its yield. If you tried instead to use bond prices directly, you would get an unstable and inaccurate covariance matrix.For (2), the important issue is usually multivariate normality rather than univariate. In other words, you don't care much if individual asset return distributions are skewed or kurtotic (within reason), you care that all the codependance is captured by linear correlation.(9) and (10) assumes a linear exposure to the market factors. If your market factors are your securities, as you seem to assume above, then everything will be linear. But if you hold say, a stock and a put option on the stock, you cannot compute the confidence interval the way you indicate. Instead you have to generate the full distribution of P&L (almost certainly through simulation).(5) is usually not done. Expected P&L should be much less than VaR for the short time periods (1 day, 10 day) and high volatility relative to capital portfolios people analyze with VaR. Since you are using a simplistic model for simulating security returns, you don't really trust the expected profit anyway. However for long time periods or low volatility/high net investment portfolios, you could do (5). Similarly, (11) is usually neglected because you define the portfolio as zero value. In other words, you carry all securities at market, and you're interested in the change from that value.

@sam: all the things you mentioned are correct. As for the time horizon, I have taken that into account. I didn't mention it because it seemed so logical. But it IS important, so you're absolutely right. As for using the log returns; I did that becaues according to the RiskMetrics technical doc "there is strong evidence that log price changes are normally distributed". That would mean that prices would be lognormally returned, and using a variance based on such a ddensity function doesn't sound like a good idea to me... Which brings us to a pint Aaron made...@Aaron. Your first point is one that I had never thought or heard about before. Put like that, it makes a whole lot of sense to model the market facors over the securities. I had always assumed that mapping your instruments onto market factor was done to facilitate calculations and to limit the number of parameters to be estimated.As for the distributon of the return, is it not enough for the variables to be univariably distributed? Or does this assumption only imply that the ]variancesare relevant, and that it says nothing about the relevance of the covariances? (Sounds and is a newb question, I guess.)Non-linearity in my case is not an issue, as the portfolio is 100% equity.Point taken, but I felt like having an idea, even if somewhat inaccurate, of the range the portfolio return and VaR were going to be in.Anyway, thanks for your help guys!

I have another problem. I'm having trouble proving that the VaR methodology mentioned above yields the same result as: VaR= SQRT(VaRx' R VaRx) , where VaR is the portfolio Value-at-Risk, Varx is a column vector of the individual VaR's, measured as described above (only with a weight of 1, obviously) and R is the correlation matrix. Does anyone know if this proposition is correct, and if so, how to prove it?

Your VaR formula is correct only if VaR is proportional to standard deviation, i.e. if all your exposures are linear (you indicated earlier that was the case), the shape the distribution of all linear combinations of market factors is the same symmetric distribution (as for a Normal distribution) and your VaR's are directional.I suspect the last is your problem. Suppose I own $1,000 of stock A and $1,000 of stock B, both have VaRs of $50 and a correlation coefficent of 0.5. My combined VaR is $87, as your formula indicates. But what if I am short $1,000 of B? Your formula gives the same $87, although the correct VaR is $50.The solution is to use directional VaR's. A positive VaR, by convention, means a loss if the underlying market factor goes down (a long position). A negative VaR means a loss if the underlying market factor goes up (a short position). Then your overall directional VaR obeys your formula, but it is no longer directional so you cannot use it as part of another calculation to aggregate VaR. That is, you cannot compute VaR by trading desk this way, then use the formula on the trading desk VaR's to compute a firmwide VaR.