Hi,Any of you have any experience or tips on calculating long term VaR, say for a horizon of 1 year? A simpler rule is to calculate 1day VaR and extrapolate it by multiplying it by the square root of time. However, I've read a few papers that say this is not valid, and this rule might lead to a over or under-estimation of the actual VaR. A general guideline is to use returns of the same horizon as the VaR that we need. However, for horizons of 1 year, we typically do not have enough pass returns to do the calculation. Any ideas? My current way is to calculate monthly VaR and scale up to a year.

Besides the fact that VaR generally makes little sense, when you look at long horizons you run into the problem of having to estimate the center of the distribution. This requires you to essentially forecast how the portfolio is going to perform. For 1 day VaR, you generally assume the distribution is centered at zero, which is ok. But for one year, you can no longer do that. I would say, 1 day VaR has little meaning already (given all the arbitrary assumptions that go into it), but 1 year VaR is absolutely useless. I am sorry if this sounds too harsh, but it is an honest opinion of someone who actually invested meaningful lifetime into VaR models (and really really regrets it....).

Two thoughts:1. You might consider doing an n-Day VaR, scaled to 1 year and then look at how the scaled value changes as n approaches 1 year. That might give you a feel for both the stability of estimating the 1-year VaR from shorter-term VaR and the general trend of 1-year value as the value of n approaches 1 year. You might be able to find the approximate optimum n that is small enough to provide a stable estimate and large enough that it doesn't require too much scaling.2. Take a look at Hurst's Rescaled-Range or R/S analysis, which considers the issue of the change in the span of extrema as a function of the scale of the data interval. The Hurst exponent is an empirical estimate of the scaling properties -- it will be 0.5 if the square-root rule is valid. If 0.5>H>1.0, then it suggests a time-series with trending properties. 0<H<0.5 suggests a time series with reversion.

Thanks for the replies. I do know that 1 year VaR does not make much sense, but this is what is required of me.