Backtesting VaR values from MC-simulationHelloI am trying to backtest Value at risk values.I have the portfolio for each time in (Value format) xxx $ ... for each day.Now when I backtest it versus 2000 tradingdays of data.What shall I look for?Shall I take my portfoliovalues and deduct the initial value. Pt-P0, there by know whatI can loose during a day. Then take these and sort them like sort(change in price)so they will go from example -10........5Then for alpha(confidential value) take skip the lowest % numbers of losses. So we will have a95% VaR at ex. -7.5This value is then compared/backtested to the change over the 2000 tradingdays who have losses for ex.-20.... to....5. The number of excess observations over the -7.5 should be as close to (5%*2000) 100 as possible in this case.Is this correct??Or should one compare the real prices vs. the backtested ones?Thank you in advance

I'm having trouble following your proposed method. In particular, I missed where you computed VaR.The basic idea of a backtest is to compute VaR at close of business day 1. Since you use a 1-day VaR, you compare it to the change in price from close of business day 1 to close of business on day 2. If Day 1 Value - Day 2 Value > Day 1 VaR, you have a VaR break.Since you use a 95% VaR, you expect 1 break every 20 days. If you backtest over a year (260 trading days) you expect 13 breaks. If the breaks are independent in time, the standard deviation of the number of breaks is about the square root of 13 or 3.6. If you had 20 or more breaks, there's likely something wrong with your VaR model. If you have fewer than 6, the model is probably too conservative. That won't worry a regulator, but from a methodology standpoint it's just as serious.If your VaR method assumes symmetrical price changes, you can increase the power of the test by looking at positive and negative breaks (just use an absolute value on the left side of the inequality). Then you expect 26 breaks with a standard deviation of 5.1. You can also increase the power by testing over shorter intervals and lower VaR percentiles, although you can't get much lower than 1-day/95%.

Hello!I have evaluated my VaR estimates based on the binomial distribution(Kupiec), but I would like to apply more backtesting techniques. Going through some papers I have found advantages and disadvantages for different methods. Which one would you recommend?Thank you in advanceE

Can you be more specific? How do you calculate VaR and what do you use it for? What do you hope to gain from the additional backtesting techniques?

I am calculating VaR through different models/methods and I want to see which one performs better...

Alternatively to backtesting (which is often problematic due to intraday trading, valuation errors due to bad price feeds, etc) you can try doing an attribution. If your risk model, whatever it may be, can explain a lot of the real daily or weekly performance given real market inputs (e.g. yield move, stock market move, vol move, ....), then it is a good model. What you will probably find is that every VaR model leaves an error term of unexplained variation/performance. From my experience, VaR is hard to bring in line with the real world (model and real world seem to be in different "units"), for many practical and theoretical reasons, but it can be useful to just use the VaR model output itself as a decision making variable without trying to reconcile it 100% with reality. (i.e. if the number is high vs history we are taking a lot of risk, and vice versa...develop an intuition for the model numbers).