FRM VaR explained: interpret thresholds, tails and risk reports
Value at risk (VaR) is a loss threshold for a stated confidence level and horizon. It does not cap the loss. Expected shortfall describes the average loss in the tail; stress testing asks what happens under specified adverse scenarios.
| Measure | Question answered | What it does not say |
|---|---|---|
| VaR | Where is the chosen loss quantile? | The largest possible loss or average loss beyond it. |
| Expected shortfall | How large is loss in the specified tail on average? | A guaranteed worst-case loss. |
| Stress loss | What loss follows from these shocks and assumptions? | How likely the scenario is unless separately estimated. |
- VaR
- A thresholdAlways state confidence and horizon.
- Expected shortfall
- A tail averageIncludes how severe the tail losses are.
- Stress test
- A scenarioIts probability is not implied by the loss.
Same VaR, different tail severity
Imagine 100 equally likely one-day loss outcomes, sorted from smallest to largest. Define the empirical 95% VaR as the 95th ordered loss. In both portfolios that loss is $10. Portfolio A’s five worst losses are $12, $14, $16, $18 and $20; portfolio B’s are $20, $40, $60, $80 and $100.
Scale: 0 to 60 dollars
- A: 95% VaR10
- B: 95% VaR10
- A: worst-5% average16
- B: worst-5% average60
Original worked example
Which portfolio has the more severe worst 5% of outcomes?
B: its worst-five average is $60, compared with $16 for A.
A averages (12 + 14 + 16 + 18 + 20) / 5 = 16; B averages 300 / 5 = 60. Both keep the same 95th outcome, so the VaR alone does not expose the difference.
Choose a model and name its assumption
Model choice trades tractability against how exposures and distributions are represented. More simulation does not rescue a misspecified model. Check the position, data window, dependence assumptions and valuation method.
| Method | Basic construction | Question to challenge it |
|---|---|---|
| Parametric | Use distributional and sensitivity assumptions. | Do tails and nonlinear positions fit those assumptions? |
| Historical simulation | Revalue using observed past shocks. | Does the history include the risk regime you need? |
| Monte Carlo | Generate scenarios from a specified model. | Are the dynamics and dependencies credible? |
Read the labels before comparing two VaR numbers
A reported VaR needs a portfolio, valuation date, horizon, confidence level and model basis. A one-day number and a ten-day number are not directly interchangeable. Neither are a dollar amount and a percentage of portfolio value. Establish those labels before deciding that one exposure is safer.
For a calculation from assumed returns, use the dedicated construction lesson. This overview focuses on what the finished risk number means and what a reader should ask before acting on it.
Ask how the reported model was checked
A risk report should explain its validation context. An observed loss beyond VaR is possible by construction; it is not automatically proof of a broken model. Repeated or clustered exceptions require investigation using aligned forecasts and outcomes. The dedicated backtesting lesson develops those diagnostics.
Keep a model’s reported risk separate from confidence in the model. A smaller number can reflect a smaller exposure, different assumptions or an omitted risk. It does not, by itself, demonstrate a better method.
Turn a VaR statement into a precise sentence
Suppose a report states that one-day 99% VaR is $2 million. Under the specified model and loss convention, that is a threshold associated with the 99th percentile of one-day loss. It does not mean the portfolio can lose no more than $2 million, and it does not describe the average size of the losses beyond that threshold.
The remaining tail probability is 1% under the model. That does not promise exactly one exceedance in the next 100 days or tell you on which day one will occur. Actual outcomes depend on the realized conditions, portfolio changes and model adequacy. A frequency interpretation should retain those qualifications.
Now compare the statement with a specified stress loss of $12 million. The stress loss answers what happens under the chosen scenario; it need not carry an estimated probability. It can therefore reveal a vulnerability without contradicting the VaR statement. Before deciding how to respond, identify which question each measure was designed to answer.
| Statement | What to correct |
|---|---|
| “VaR is the worst possible loss.” | It is a quantile, so more severe losses remain possible. |
| “One percent means exactly one loss every hundred days.” | It is a model probability, not a timetable. |
| “The stress loss is larger, so the VaR calculation must be wrong.” | The measures answer different questions under different assumptions. |
What to ask when the reported number changes
Check whether the portfolio changed before attributing a larger VaR to a model problem. A new position, a larger notional or a concentration can raise the estimated risk even when the methodology is unchanged. Conversely, an unchanged portfolio can produce a different estimate when volatility, dependence or the data window changes.
Ask for the comparison on a consistent basis: same currency, horizon, confidence and portfolio scope. Then separate exposure changes from assumption changes. A report that only shows the final number leaves the reader unable to identify the cause. The direction of the number alone is insufficient evidence about model quality.
Use this overview for interpretation. The construction lesson develops how to obtain a VaR estimate, the expected-shortfall lesson develops tail measures, and the backtesting lesson examines forecast performance. Following the appropriate method lesson is more useful than treating one overview as a full treatment of all three subjects.
Common questions
- Is VaR the maximum amount I can lose?
- No. Losses can exceed the VaR threshold.
- Is expected shortfall always just an average of losses above VaR?
- That intuition works cleanly for continuous distributions. In a discrete sample, state the quantile and tail-weight convention, especially when observations tie at the threshold.
- Does a good backtest remove model risk?
- No. It assesses behavior on observed data; different future conditions and omitted risks can still matter.
Sources and editorial notes
By MiloPrep editorial team. Source check: . Official requirements and product information can change. Examples and planning allocations are independent teaching material, not official questions or passing standards.
- Basel Committee: market risk framework ↗
Market risk, model validation and expected shortfall context; not an FRM exam blueprint.
MiloPrep is an independent study resource and is not affiliated with or endorsed by GARP or the named providers. FRM and provider names are trademarks of their respective owners. Editorial methodology.
Notes
- * The threshold misses the change in tail severity: Original discrete illustration. Quantile convention: 95th ordered outcome; each of the 100 outcomes has probability 1%. ↩ Return to content