Value at Risk: the loss it promises you will not exceed — and what it leaves unsaid

Value at Risk answers one question for a trader: what loss the portfolio will not exceed over a chosen period, provided nothing from the category of rare events happens. It is a threshold figure, not a forecast and not an average outcome. You set the horizon (a session, a week, a month) and the tolerance — the share of cases you agree to leave out of the picture; the model returns a loss, in money or as a share of capital, that you will not breach while the market behaves normally. Everything that lies beyond this threshold VaR does not describe at all — and that is its defining property, not a quibble about the method. A short definition with the parameters is given in the glossary entry on Value at Risk.

What exactly VaR claims — and what it does not

The result is always stated in three parts: the loss, the horizon, the tolerance. Remove any part and you are left with a meaningless number: VaR without a horizon cannot be compared, VaR without a tolerance cannot be tested. A correct reading sounds like this: "over the horizon of a trading session, at the chosen confidence level, the portfolio's loss will not exceed such-and-such an amount".

What this statement does not contain. There is no promise that the threshold will not be breached — on the contrary, by construction it will be breached with a frequency equal to the tolerance. There is no indication of how deeply it will be breached: beyond the threshold there may be a slightly larger loss or a devastating one. There is no insurance against the market ceasing tomorrow to resemble the history you used for the calculation. VaR is a measurement of an ordinary day, and confusing it with a stress test is dangerous precisely because the figure looks convincing.

Input parameters: you choose them, not the model

The horizon is tied to how quickly you are able to exit. For intraday trading in liquid securities such as {{instrument:SBER}} a short horizon is appropriate; for a portfolio of illiquid issues it should be noticeably longer, because exiting them takes time, and measuring risk over a period shorter than the liquidation period is self-deception.

The confidence level determines how far into the tail of the distribution you look. The higher the confidence, the larger the number and the fewer the observations it rests on: a high confidence level relies on isolated historical episodes, and the reliability of the estimate falls exactly where the estimate itself rises.

The depth of history is the third choice. A short sample reacts quickly to a change of regime but remembers few crises; a long sample remembers many but blurs current volatility. There is no universally right answer, only a deliberate trade-off.

How it is calculated

The historical method takes the actual price changes over a past period, applies them to today's portfolio and cuts off the required percentile. It assumes nothing about the shape of the distribution, but it knows only what has already happened.

The parametric method (delta-normal) estimates volatilities and correlations and derives the threshold from an assumed distribution. It is fast and analytical, but fat tails and non-linear instruments — options, structured products — break it most severely.

The simulation method (Monte Carlo) generates scenarios from a specified model and revalues the portfolio in each of them. It is flexible and handles non-linearity, but it inherits in full the quality of the model that produces the scenarios.

Turning VaR into position size

Step 1 — define the loss of capital you can accept over a day or a week; this is a decision about yourself, not about the market. Step 2 — calculate VaR per unit of position for the instrument in question. Step 3 — the position size is obtained by dividing the acceptable loss by the VaR of a unit. Step 4 — check the portfolio as a whole rather than the sum of separate limits: when correlations are high, the diversification you were counting on disappears, and portfolio VaR ends up close to the sum of the individual figures. It is easier to compare risk across asset classes by looking at equities and OFZ at the same time: the yield curve and share prices produce different tail profiles.

Backtesting: the only test of honesty

A model is tested by counting breaches: how many times over the observed period the actual loss exceeded the threshold. If there are noticeably fewer breaches than expected, you are overestimating risk and running a smaller position than you could; if there are noticeably more, you are underestimating it, and the limit is fictitious. Clustering is examined separately: breaches that line up in a row mean the model is not keeping pace with a shift in volatility, even if the total count matched.

What to add on top

Conditional VaR (expected shortfall) answers the question "how bad is it beyond the threshold" — it averages the losses in the tail and is therefore sensitive to how heavy that tail is. A stress test abandons probabilities and runs the portfolio directly through specific historical shocks and hypothetical scenarios. A drawdown limit works on the result already realised, not on an estimate.

A separate source of risk that price-based VaR does not see: events in the security itself. Earnings reports, record dates and reorganisations move the price by a logic of their own — check the events calendar and the news flow before trusting the threshold on the eve of a publication.

{{callout:warning}}The platform's data does not include the methodology for calculating the exchange's risk parameters or the broker's margin requirements — these are their internal models. Your VaR and the broker's requirements are calculated differently and diverge; plan your position against both, not just one.{{/callout}}

Definitions of related concepts are in the glossary.