Sitonce
Country: US
Show exams for United States Hong Kong
Sign in

Value at Risk: FRM Part I

Updated 9 min read
Key takeaway

Value at Risk estimates a loss quantile for a stated horizon and confidence level under a chosen model.

  • It is not a maximum possible loss.
  • Compare historical, parametric and simulation methods by their assumptions, data needs and tail limitations.
  • State the horizon, confidence level, sign convention, method, and model limitations.
On this page10 sections
  1. Define the measure before calculating
  2. Worked normal approximation
  3. Horizon scaling and its limits
  4. Read the loss distribution and the quantile
  5. Historical simulation step by step
  6. Parametric VaR for a linear portfolio
  7. Why tail averages add information
  8. Check a backtest without overclaiming
  9. Choose the VaR method to match the portfolio and question
  10. Relate VaR to capital, limits, and decisions carefully

Define the measure before calculating

Value at Risk (VaR) is a quantile of a loss distribution over a stated horizon and confidence level. A one-day 99% VaR of $8,000 means the modeled loss is not expected to exceed $8,000 in 99% of one-day outcomes under the model assumptions. It does not say the maximum loss is $8,000.

State whether VaR is reported as a positive loss amount or a signed return quantile. If a portfolio value is $400,000 and a return quantile is -2%, the equivalent positive loss amount is $8,000. Mixing sign conventions can reverse the interpretation.

Always identify the horizon, confidence level, portfolio scope, valuation basis, and estimation method. VaR is not meaningful as a free-floating number. A one-day estimate cannot be compared directly with a ten-day estimate unless the scaling method and assumptions are understood.

VaR summarizes one point in the loss distribution. It does not describe how severe losses are after that point. Expected shortfall focuses on the average tail loss beyond the VaR threshold, while stress tests explore selected adverse scenarios.

Worked normal approximation

Suppose a portfolio is worth $400,000, daily return volatility is 1%, mean return is assumed zero, and the desired confidence level is 99%. For a standard normal distribution, the lower-tail z value is approximately -2.326. A simple loss VaR is 2.326 × 1% × $400,000, or about $9,304.

The computation uses a positive loss convention. The corresponding return quantile is about -2.326%. The example assumes normal returns, stable volatility, no mean adjustment, a one-day horizon, and a linear mapping from return to portfolio value.

If daily volatility were 1.2% with other assumptions unchanged, the estimate would rise to 2.326 × 1.2% × $400,000, or about $11,165. The model responds proportionally to volatility in this simple setup.

This arithmetic does not show that the portfolio cannot lose more than the estimate. A fat-tailed distribution or volatility jump can create more frequent exceedances than the normal model predicts.

Horizon scaling and its limits

A common square-root-of-time approximation scales one-day VaR to T days by multiplying by the square root of T. This relies on restrictive assumptions such as independent, identically distributed returns with stable variance and a suitable aggregation structure.

For ten days, the simple scaling factor is √10, approximately 3.162. Multiplying $9,304 by that factor gives about $29,420. This is an approximation, not a universal rule for every portfolio or regulatory purpose.

Autocorrelation, changing volatility, nonlinear exposures, compounding, and liquidity effects can make square-root scaling inaccurate. A model that is calibrated to one horizon should not be extrapolated blindly.

In an exam question, use the stated scaling rule and assumptions. If the question asks for limitations, explain which dependence or distribution assumptions may fail.

Read the loss distribution and the quantile

A loss distribution places probability on possible portfolio losses. To find a confidence-level VaR, order losses from the smallest to the largest and locate the selected percentile. At 95% confidence, the threshold is the 95th percentile of losses under this positive-loss convention. About 5% of modeled observations lie beyond it, subject to how the sample quantile is defined. In a return distribution, the corresponding cutoff appears in the lower tail because negative returns represent losses.

Check whether a question asks for a loss quantile, a return quantile, or a dollar amount. If the portfolio is $2 million and the lower return quantile is -1.5%, then the positive loss threshold is $30,000. If the question instead asks for the return quantile, preserve the negative sign. A correct magnitude with the wrong sign answers a different question.

Historical simulation step by step

Historical simulation begins with a portfolio today and a set of historical risk-factor moves. Apply each observed move to current positions, revalue the resulting portfolio, and calculate the loss relative to today's value. Sort the hypothetical losses and take the requested quantile. The method can preserve observed skew and co-movement in the selected sample, but it does not make the sample representative by itself.

For a two-factor portfolio, imagine that the factors are an equity index and an interest rate. A historical day with the index down and rates down produces one joint scenario. Keeping that joint move matters: independently sampling each factor could create combinations that never occurred and discard historically observed dependence. Conversely, a crisis outside the chosen lookback period will not appear unless a separate stress scenario is added.

Historical simulation also raises a weighting choice. An unweighted window treats an old observation like a recent one; a weighted approach can emphasize recent data but may understate a regime change or overreact to a short period. The exam-relevant conclusion is that historical VaR is assumption-light compared with a normal parametric model, not assumption-free. Window length, portfolio revaluation, data quality and quantile selection all affect the result.

Parametric VaR for a linear portfolio

For a linear position in a normally distributed return, a one-sided loss VaR can be approximated by portfolio value multiplied by volatility and the positive normal critical value. With nonzero expected return, the tail return also includes the mean term; many short-horizon examples assume that mean is negligible. State the assumption before using the shortcut. A portfolio's dollar exposure, return volatility, time horizon and confidence quantile must all use compatible units.

For example, a $750,000 linear portfolio with daily volatility of 0.8% and a 95% one-sided normal quantile near 1.645 has an approximate loss VaR of 1.645 × 0.008 × $750,000 = $9,870 when the mean is zero. If someone instead inserts a two-sided 95% critical value, the answer changes because that value allocates probability to both tails. A loss VaR normally uses the one-sided tail relevant to losses.

For several risk factors, portfolio variance depends on both individual volatilities and their covariances. Diversification reduces variance when correlations are below one, but correlation is not a fixed shield. In stress periods, dependence may change. An exam calculation may provide a covariance matrix; check that it is symmetric, that variances sit on the diagonal, and that weights and units match the stated positions before multiplying.

Why tail averages add information

VaR gives a threshold, not the expected size of losses beyond that threshold. Consider two portfolios with the same 99% VaR of $10,000. In one, the worst 1% of outcomes cluster just above $10,000; in the other, they reach $100,000. The threshold alone does not distinguish them. Expected shortfall summarizes the average loss in the tail beyond the selected quantile, subject to a clear convention for discrete distributions and probability mass at the threshold.

This distinction changes how a risk report should be read. VaR is useful for a concise threshold and for comparing positions on a common horizon and confidence basis. Expected shortfall provides more information about tail severity. Neither replaces stress testing, liquidity analysis or judgment about exposures that are poorly represented in the data. If an exam asks for a VaR limitation, a precise answer is that it gives no severity information beyond the chosen quantile.

Check a backtest without overclaiming

If a correctly calibrated 99% one-day model is tested on independent days under stable conditions, an exceedance is expected on roughly 1% of days. A sample with zero exceedances does not prove the model is perfect, and one exceedance does not prove it failed. Statistical evidence depends on the number and clustering of observations, the treatment of profit and loss, and whether the forecast and realized outcome use the same horizon and portfolio scope.

When exceedances cluster, the model may be missing changing volatility or dependence. But first inspect operational details: was the risk forecast frozen at the right time, did the portfolio change, and are market data aligned? A backtest compares a forecast with a defined outcome; mismatched definitions can look like model error. Exam responses score better when they name the possible cause rather than merely saying the model should be improved.

Choose the VaR method to match the portfolio and question

Historical simulation is transparent about observed scenarios but is only as informative as its history and portfolio revaluation. A parametric estimate can be efficient for approximately linear exposures, but a distributional assumption can miss skew or tail heaviness. Monte Carlo can revalue nonlinear positions across simulated factors, but it introduces model and calibration choices. The decision depends on the use case, not a universal ranking.

For an option-heavy book, a linear normal approximation may misstate risk because small factor changes do not map linearly to portfolio value. Full revaluation in a simulation may better capture option curvature, but only if the volatility dynamics and dependence model are credible. A more complex model can still produce a misleading number when its inputs are poorly chosen.

A risk report should disclose method, horizon, confidence level, portfolio scope, data window, and any material assumptions. If an estimate changes sharply after a model revision, compare those choices before concluding that market risk itself moved. Model changes and market moves can affect the metric in different ways.

Relate VaR to capital, limits, and decisions carefully

MethodStrengthMain limitation
Historical simulationUses observed joint scenariosDepends on sample window
ParametricEfficient under suitable assumptionsDistribution and covariance risk
Monte CarloCan revalue nonlinear positionsModel and calibration risk
Stress testingExplores selected severe casesScenario choice and coverage

VaR can help compare exposures or monitor a limit, but it should not be treated as a complete risk appetite statement. A threshold without stress scenarios can understate severity beyond the quantile. Liquidity, concentration, basis risk, and tail dependence may require additional measures and judgment. Stress tests complement VaR by exploring selected severe scenarios, though scenario choice and revaluation still require judgment.

When a VaR limit is breached, the result should trigger an investigation rather than an automatic conclusion. Check positions, data, market moves, and model inputs. A change in VaR may come from volatility, correlation, portfolio composition, or a rolling window dropping an old observation. Understanding the driver is necessary before choosing a response.

Common questions

Does VaR show the maximum possible loss?

No. VaR is a loss quantile at a stated horizon and confidence level. Outcomes beyond the threshold can be larger.

Which VaR method is best?

No method is universally best. Historical, parametric, and simulation approaches rely on different data and modeling assumptions.

What does a 99% one-day VaR mean?

Under the model, the one-day loss is not expected to exceed the threshold in 99% of outcomes. The remaining tail probability is not a loss cap.

How is VaR different from expected shortfall?

VaR gives a quantile threshold. Expected shortfall summarizes average loss beyond the selected tail threshold, subject to its definition and convention.