- slacknoise
**Posts:**49**Joined:**

Hi all,I would like to measure the risk of my equity portfolio via a VaR model incorporating a Student-T Copula and Extreme Value theory. What is an appropriate historical time span to consider for the equities in order to so? I appreciate any help. kind regards,

Basel recommends a historical time period of minimum 250 trading day's.your VaR metric computed with the following parameters-for general MR management : 1-day @ 95% quantile-for Trading Book Capital : 10-day @ 99% quantile.industry practice to apply square root of time rule to scale VaR in time and space.howeverit is Model dependent.

- slacknoise
**Posts:**49**Joined:**

thanks for the insight and the fast reply. very much appreciated

"It depends"What's in your portfolio, and what's your forecast horizon?If all your equities are US equities, then all the close-of business prices happen at the same time. You can look at calculating VaR based on daily return timeseries. You want a good number of observations to make the volatility estimate of any one stock reasonable. So maybe 252 days is good.If your equities are global, then close-of-business prices are staggered across time zones. You're probably better basing the calculation on weekly returns. So to keep a good estimate of individual stock volatilities, you're going to need a longer historical period (252 weeks?).If you're predicting a 1-day VaR, then you want to try & base it on the recent past to try and 'capture the current volatility regime'. Maybe 1 year of daily returns is good.If you're predicting a 1-month VaR, then maybe you want to use 5 years of data to allow for changing volatility regimes.If you use exponetially weighted observations in your model, you also need to use a time period that is consistent with your half-life. Riskmetrics gives a rule of thumb that your history shoul dbe 3 to 6 times your half life. This means that the earliest observation gets 0.13 to 0.02 times as much weight as the latest observation.

- slacknoise
**Posts:**49**Joined:**

QuoteOriginally posted by: MHill"It depends"If your equities are global, then close-of-business prices are staggered across time zones. You're probably better basing the calculation on weekly returns. So to keep a good estimate of individual stock volatilities, you're going to need a longer historical period (252 weeks?).thanks for the advise. I am dealing with chileanen equities and ADR's and i am interested in 1day VaR and 1 month VaR so thats really helpful

- DominicConnor
**Posts:**11684**Joined:**

It is worth trawling through to find the worst case that your data supports, not just what is required to keep risk management happy

Depends also on what you are measuring. I.e. I have seen 500 day historic, 250 day Stress with 250 day antithetic scenarios applied.

how do you define antithetic scenarios ?

"how do you define antithetic scenarios ?"Say you have N observations of historical price changes.Extend the vector by multiplying each one of those changes by -1.Now you have 2*N observations.The same tactic is often used in Monte Carlo simulation - you draw your random number and then generate another one by multiplying it by -1.I am not sure "scenarios" is the right word to use here, what we are really talking about are historical observations, I suppose one could generously call them scenarios.

- slacknoise
**Posts:**49**Joined:**

Has anyone here experience with real time VaR models concerning this topic?

QuoteOriginally posted by: DavidJN"how do you define antithetic scenarios ?"Say you have N observations of historical price changes.Extend the vector by multiplying each one of those changes by -1.Now you have 2*N observations.The same tactic is often used in Monte Carlo simulation - you draw your random number and then generate another one by multiplying it by -1.I am not sure "scenarios" is the right word to use here, what we are really talking about are historical observations, I suppose one could generously call them scenarios.so in terms of P&L ,we are dealing with antithetic P/L ,and that is L/P , it makes VaR more intuitive as it becomes a positive number for a positive quantile under antithetictransformation. is this the correct logos ?

QuoteOriginally posted by: slacknoiseHas anyone here experience with real time VaR models concerning this topic?Not directly. I know a group at a large German Firm were looking into it, but I don't know how far they got. There were looking at using FPGA to get the necessary increase in performance. I don't know if it went anywhere.

QuoteOriginally posted by: slacknoiseHas anyone here experience with real time VaR models concerning this topic?Incrementally, what are you trying to pick up with the "real time" version? Intraday trading? I would think this would be relatively straightforward if all you trade is cash equities? Going back to your original question on window length etc. I have two suggestions: First, use more than one (a lot of folks use short, medium and long horizon based VaR numbers) and second, backtest backtest backtest! That way you will have a bit of an early warning if the regime is changing. I hope this helps.

- slacknoise
**Posts:**49**Joined:**

Going back to your original question on window length etc. I have two suggestions: First, use more than one (a lot of folks use short, medium and long horizon based VaR numbers) and second, backtest backtest backtest! That way you will have a bit of an early warning if the regime is changing. I hope this helps.thanks i will consider it in my model

QuoteOriginally posted by: DavidJN"how do you define antithetic scenarios ?"Say you have N observations of historical price changes.Extend the vector by multiplying each one of those changes by -1.While this is useful to reduce noise in a Monte Carlo context, I dont see why this would make sense with historical data. Wouldn't it just distort your dataset?

GZIP: On