Skip to content
Free previewYou are reading 1 of 100 lessons, in the same interface members use. Everything else unlocks with full access.Get full access · $150

1.1.1 - Loss Distributions and Risk Measures

What you'll learn

  • Turn portfolio values, cash flows and returns into a consistent series of currency losses.
  • Read a loss quantile from equally or unequally weighted scenarios.
  • Interpret a value-at-risk report using its horizon, confidence level and reference value.
  • Calculate average tail loss when several scenarios share the cutoff loss.
  • Choose a measure that answers a desk's question about frequency, severity or concentration.

Estimated reading time: 16 minutes.

WHY

Why it matters

Two desks can report the same one-day value at risk even if a severe loss would cost one desk ten times as much. On the exam and in a risk report, the distinction turns on which part of the loss distribution enters the calculation: one ranked outcome, a probability-weighted group of outcomes, or the result of a particular shock.

KEY TERMS

Key terms

Profit and loss (P&L)
The economic gain or loss over a specified interval, expressed with gains positive.
Signed loss
A monetary outcome expressed with adverse changes positive. A profitable outcome remains in the distribution as a negative loss.
Probability mass
The probability assigned to an exact outcome. Several scenarios can contribute mass to the same loss amount.
Quantile
A cutoff located by accumulating probability from the smallest outcome upward.
Value at risk (VaR)
A loss quantile for a specified portfolio, holding period and confidence level.
Expected shortfall (ES)
The average signed loss in the worst specified fraction of the distribution.
Exception
A realized loss strictly above its corresponding VaR forecast under the strict-exceedance convention.
CORE

Core content

1.1.1.1 Construct the portfolio loss in each scenario

Begin with what happened to the owner's wealth. A bond that pays a coupon can finish the day at a lower quoted price while earning a positive total return. Conversely, a client deposit can increase account value on a losing day. Decide which cash flows the closing value already includes before assigning a sign.

Let \(V_0\) and \(V_h\) be opening and closing portfolio values over horizon \(h\), in the same currency. Let \(C\) be net investment cash paid out and excluded from closing value, and \(F\) be net external capital contributed. Positive \(F\) means the owner put money in. Economic P&L, \(\Pi\), and signed loss, \(L\), are

\[\Pi=V_h-V_0+C-F,\qquad L=-\Pi.\]

If income stays inside the portfolio as cash, it already contributes to \(V_h\). Counting it again would create fictitious profit. For a simple total return \(R\) on the opening capital with no external flows, \(L=-V_0R\). Enter a 2% decline as \(R=-0.02\). For a short position, first calculate P&L using the signed position size. An underlying price increase can then become a positive portfolio loss.

Table 1 reconciles three accounts that start with USD 100 million and finish with the same reported value.

Table 1. Cash-flow reconciliation in USD millions
Account eventClosing valueEconomic P&LSigned loss
No cash flow99−11
Coupon of 2 paid out991−1
Client contributes 299−33

For multiple positions, add the signed currency losses within the same scenario. If a rate shock costs the bond book USD 8 million and earns the hedge USD 5 million, the portfolio loss in that state is USD 3 million. Matching each position's worst day to every other position's worst day constructs a different joint scenario. A hedge can be effective precisely because its worst day occurs when the bond book gains.

A Convert the return before applying the position value

A desk reports a one-day log total return of \(-0.10\) on a USD 20 million portfolio with no external flows. A log return \(r\) gives a closing-to-opening wealth ratio of \(e^r\). The signed currency loss is therefore \(V_0(1-e^r)=20(1-e^{-0.10})\approx1.90\) million dollars. USD 2 million would be exact for a negative 10% simple return. The word log determines which conversion to use.

For an unhedged foreign holding, first express both valuations in the reporting currency. Suppose a euro-denominated asset gains 5% locally while the dollar price of one euro falls 10%. Each opening dollar becomes \(1.05(0.90)=0.945\) dollars, a 5.5% reporting-currency loss. A USD 20 million holding therefore loses USD 1.10 million. Adding the two percentage changes would miss their product. If the exchange rate is quoted in euros per dollar, invert the exchange-rate levels before computing the dollar value of the holding.

1.1.1.2 Read probability, location and dispersion

A loss distribution assigns probabilities to the portfolio outcomes. In a scenario distribution, distinct states can have identical losses. Their probabilities add. Table 2 gives the entire one-day distribution of a hypothetical trading book. Read the cumulative probability as the chance of losing the amount in that row or less, including all profitable outcomes.

Table 2. One-day trading-book loss distribution
StateSigned loss (USD millions)State probabilityCumulative probability
Rally−250%50%
Little change040%90%
Spread widening48%98%
Dislocation142%100%

Write the cumulative distribution function as \(F_L(x)=P(L\le x)\), where \(x\) is a loss threshold in currency units. The probability of losing more than USD 4 million in Table 2 is \(1-0.98=0.02\). The probability of losing at least USD 4 million includes the spread-widening state and is 10%. For a discrete distribution, changing one word can change the event substantially.

The expected signed loss is \(\mu_L=\sum_i p_iL_i\), where \(L_i\) is state \(i\)'s loss and \(p_i\) its probability. Here it is \(0.50(-2)+0.40(0)+0.08(4)+0.02(14)=-0.40\) million dollars. The book earns USD 400,000 on average under the model, even though it can lose USD 14 million. Averaging only losing days would answer a different question and would require dividing by the probability of a losing day.

Standard deviation is \(\sigma_L=\sqrt{\sum_i p_i(L_i-\mu_L)^2}\). It has the same units as the losses. The square gives equal weight to favorable and adverse deviations of the same magnitude. Reversing the sign of every observation reverses the mean and skewness but leaves standard deviation unchanged. Thus a long right tail in a loss chart appears as a long left tail in a P&L chart. A symmetric heavy-tailed distribution can also produce extreme losses: symmetry describes the balance between the two sides, while tail thickness describes how much probability lies far from the center.

Match each quantity to the operation that produces it.

Expected signed loss:
Weight every outcome by its probability and sum.
Standard deviation:
Take the square root of the mean squared distance from the center.
Cumulative probability:
Add state weights up to and including the specified amount.
Strict exceedance probability:
Add state weights above the specified amount.

A Read a loss band from a density or a cumulative distribution

On a continuous density plot, probability is area. A tall, narrow region can contain less probability than a low, wide region. A single exact loss has zero probability even where the density is high. On a cumulative distribution plot, the height already represents probability. For thresholds \(a<b\), the probability of the band \(a<L\le b\) is \(F_L(b)-F_L(a)\). This difference also works for discrete losses.

Suppose a desk pays a fee when its loss exceeds zero but stays at or below USD 4 million. Table 2 gives a trigger probability of \(0.98-0.90=0.08\). Subtracting the cumulative probability at zero removes every break-even outcome. If the fee also applies at zero, subtract only the probability strictly below zero, giving \(0.98-0.50=0.48\).

Fig 1 compares the two ways to read the probability of the same loss band.

Fig 1. One loss band on density and cumulative plots
Key point: Shaded area on the density plot equals the vertical probability difference on the cumulative plot.
{"type":"normal_vs_abnormal","describe":"The upper plot shows a smooth bell-shaped density with a band shaded between two loss thresholds. Directly below, the cumulative curve rises between those same thresholds, and a vertical bracket shows the probability in the band.","teaching_point":"A density area and a cumulative probability difference measure the same interval probability.","prompt":"Paired textbook comparison of two representations of one continuous normal loss distribution. Front view, vertically aligned plots with identical horizontal loss scales and increasing loss from left to right. Above, draw a bell density with the region from its center to one standard deviation above its center shaded orange. Below, draw the matching increasing S-shaped cumulative curve with its inflection aligned to the upper center. Carry the two threshold guides vertically through both plots; extend thin horizontal guides from their intersections with the lower curve to a vertical bracket spanning the two cumulative heights. MARKERS: 2 numbered circular markers, each a white digit on a solid dark circle, each joined to its target by a short thin leader line: 1 on the shaded density area; 2 on the vertical cumulative difference bracket. Use each digit exactly once, large and easy to read. NO words, captions or text labels anywhere in the image: digits 1 to 2 only. Use cool blue curves and orange for the area and bracket. Clean textbook illustration on white; omit arrows, gridlines and decoration.","caption":"Fig 4. One loss band on density and cumulative plots","size":"1024x1024"}
  1. Probability density. The shaded area gives the probability that the loss falls between the thresholds.
  2. Cumulative distribution function. The bracket measures the increase in cumulative probability over the same band.

B Interpret excess kurtosis

For a loss with positive standard deviation and a finite fourth moment, kurtosis is \(E[(L-\mu_L)^4]/\sigma_L^4\). A normal distribution has kurtosis three and excess kurtosis zero. A report of kurtosis five therefore implies excess kurtosis two. Large deviations on either side receive fourth-power weight, so reversing the P&L sign leaves kurtosis unchanged. Positive excess kurtosis alone cannot tell the committee which tail produced the extreme observations or determine a particular loss percentile.

1.1.1.3 Locate the VaR cutoff

Imagine moving a currency threshold upward from the best portfolio outcome. Stop the first time the accumulated probability reaches the requested confidence level. Formally, for confidence \(\alpha\) between zero and one, \(\mathrm{VaR}_{\alpha}(L)=\inf\{x:F_L(x)\ge\alpha\}\). Confidence is a probability. The resulting VaR is a loss amount. A model can assign probability to worse outcomes of any magnitude beyond that cutoff.

The following sequence locates a quantile without assuming a smooth distribution.

Fix the loss variable. Use one currency, one horizon and one gain-or-loss convention for every state.
Order the losses. Start with the most negative signed loss and finish with the largest positive loss.
Accumulate the weights. Keep each probability with its own loss and combine weights at ties.
Read the first crossing. Return the loss in the first row whose cumulative weight reaches or exceeds the requested confidence.

For \(n\) equally likely observations and the inverse empirical distribution convention, the ascending rank is \(k=\lceil n\alpha\rceil\), counting from one. With 100 observations and 99% confidence, rank 99 is the second-largest loss. If a question specifies interpolation between order statistics, apply its stated convention: sample quantiles from different conventions can differ even on identical data. With unequal weights, ranking the rows still orders severity, but counting rows no longer measures probability.

In Table 2, the 95th percentile is USD 4 million. The cumulative weight jumps from 90% to 98%, so 95% lies inside that jump. Raising confidence to 97% leaves VaR unchanged. Raising it to 99% moves VaR to USD 14 million. For a fixed distribution, increasing confidence can keep the threshold constant or raise it. A calculated decline signals that another input, sample or convention changed.

Reading a P&L percentile. With a continuous distribution and unique quantiles, \(q_{\alpha}(L)=-q_{1-\alpha}(\Pi)\), where \(q\) denotes a quantile. A first-percentile P&L of minus USD 7 million gives 99% loss VaR of USD 7 million. At a discrete probability boundary, negation reverses endpoint conventions. Converting observations to losses before sorting gives the specified loss quantile directly.

Worked case

A desk holds USD 80 million and models 200 equally likely one-day simple total returns. Its three worst returns, from worst upward, are −6.0%, −3.5% and −2.0%. Yesterday's realized loss was USD 0.7 million. The committee requests 99% absolute one-day VaR using the inverse empirical distribution.

The opening value converts the scenario returns into currency. The three largest losses are USD 4.8 million, USD 2.8 million and USD 1.6 million. In ascending loss order they occupy ranks 200, 199 and 198. The requested rank is \(\lceil200(0.99)\rceil=198\). The reported VaR is therefore \(80(0.020)=1.6\) million dollars. Two scenarios are strictly worse, representing \(2/200=1\%\) of the distribution. USD 4.8 million would report the sample maximum. Yesterday's realized loss measures one historical outcome and does not enter this forecast.

A Checkpoint

A desk has 1,000 equally weighted scenarios. Its lowest 985 signed losses are below USD 3 million, the next 12 equal USD 3 million and the last three exceed USD 3 million. Which pair gives 99% VaR and its strict exceedance probability?

  1. USD 3 million and 0.3%
  2. USD 3 million and 1.0%
  3. USD 3 million and 1.5%
  4. The largest loss and 0.1%
Answer: A. Rank 990 lies among the ties, and only three of the 1,000 losses exceed the cutoff. B uses the nominal tail probability even though the cumulative probability at the cutoff overshoots 99%.

1.1.1.4 Interpret confidence, reference value and holding period

A report of USD 6 million one-day 99% absolute VaR says the model assigns at least 99% probability to a one-day loss of USD 6 million or less. In a continuous model that attains exactly 99% at the cutoff, the remaining 1% is the probability of a strictly larger loss. It does not specify when that loss will occur. In particular, a breach today does not use up the rest of the month's allowance.

For \(N\) calibrated continuous forecasts, each with exceedance probability \(1-\alpha\), the expected number of exceptions is \(N(1-\alpha)\). Over 250 trading days, 99% daily VaR gives an expected count of 2.5. Actual counts are integers. Linearity of expectation gives the mean count without an independence assumption. Independence matters for a binomial count distribution and for multiplying daily probabilities to calculate the chance of at least one exception. An exception count assesses frequency; a separate record of the losses is needed to assess severity.

Table 3 distinguishes report fields that can otherwise make two VaR numbers appear comparable.

Table 3. Reading the specifications on a VaR report
Report fieldExampleWhat it determines
Holding periodTen trading daysThe time over which the modeled P&L accrues.
Lookback windowTwo years of observationsThe market history used to estimate shocks or model inputs.
Confidence97.5%The cumulative loss probability defining the cutoff.
Reference valueOpening wealth or expected ending wealthWhether the loss is measured in absolute or mean-relative terms.
Portfolio assumptionToday's holdings held fixedWhich exposures are revalued across the scenarios.

Absolute and mean-relative VaR. Absolute VaR measures the loss from opening wealth. Mean-relative VaR measures the shortfall from expected wealth. If expected P&L is positive, the expected ending wealth is higher than opening wealth, so the distance from that reference to an adverse outcome is larger. In signed loss units, \(\mathrm{VaR}_{\alpha}(L-\mu_L)=\mathrm{VaR}_{\alpha}(L)-\mu_L\). For expected P&L of USD 0.4 million and absolute VaR of USD 3 million, mean-relative VaR is USD 3.4 million. A negative absolute VaR is mathematically possible when even the specified adverse quantile corresponds to a gain. A reporting floor of zero would be a separate convention.

Changing the horizon. For independent, identically distributed normal daily currency losses that add across days, let \(m\) be daily mean loss, \(s\) daily loss standard deviation and \(z_{\alpha}\) the standard normal quantile. Over \(h\) trading days, \(\mathrm{VaR}_{\alpha,h}=hm+z_{\alpha}s\sqrt{h}\). The mean accumulates linearly and the dispersion grows with the square root of time. Multiplying the entire daily absolute VaR by \(\sqrt{h}\) is exact here when the mean is zero. Independent increments alone give variance scaling when variances are equal and finite; they do not generally give exact quantile scaling.

For example, take zero mean, daily standard deviation of USD 1 million and a 99% normal quantile rounded to 2.326. Four-day VaR is \(2.326(1)\sqrt{4}=4.652\) million dollars. Fig 2 lets you change daily volatility while keeping those assumptions fixed.

Fig 2. Holding period and normal loss VaR
Key point: Four independent trading days double the zero-mean normal VaR relative to one day.
{"x":{"label":"Holding period (trading days)","min":1,"max":16,"step":1},"y":{"label":"99% VaR (USD millions)","min":0,"max":20},"curves":[{"label":"Zero-mean normal VaR","expr":"2.326*s*sqrt(x)"}],"sliders":[{"name":"s","label":"Daily loss standard deviation (USD millions)","min":0.5,"max":2,"step":0.1,"value":1}],"readouts":[{"label":"Four-day VaR (USD millions)","expr":"2.326*s*sqrt(4)","digits":3}],"try":"Predict how four-day VaR changes when daily volatility doubles, then raise the slider from 1 to 2 to check.","say":"The curve rises more slowly as the holding period grows. Doubling daily volatility doubles its height at every horizon.","prompt":"a live graph of 99% VaR (USD millions) against holding period (trading days). The curve shows zero-mean normal VaR. The curve rises more slowly as the holding period grows. Doubling daily volatility doubles its height at every horizon. On the page, a slider changes the daily loss standard deviation (USD millions), and the curve redraws as it moves. It also displays four-day VaR (USD millions). Try this: Predict how four-day VaR changes when daily volatility doubles, then raise the slider from 1 to 2 to check."}

Positive serial correlation adds positive covariance terms to aggregate variance. Changing volatility makes a single daily variance unrepresentative of later days. A desk that rebalances during the interval or holds options whose sensitivities change with the underlying price also needs a horizon model consistent with those exposures. Ten-day VaR concerns the aggregate ten-day loss; a temporary loss halfway through the interval can be larger than the loss at the end.

A Checkpoint

Daily losses are independent and identically distributed normal variables with a mean loss of USD 0.1 million. The one-day 99% mean-relative VaR is USD 2 million. What is the four-day absolute VaR?

  1. USD 4.0 million
  2. USD 4.2 million
  3. USD 4.4 million
  4. USD 8.4 million
Answer: C. With the same daily distribution, four-day dispersion contributes USD 4 million and four days of mean loss add USD 0.4 million. B scales the entire one-day absolute VaR, including its mean, by two.

B Distinguish forecast coverage from estimation precision

A desk estimates one-day 99% VaR at USD 6 million and gives a 95% confidence interval of USD 5 million to USD 8 million for that estimate. The 99% selects the population loss quantile the desk wants to estimate. The 95% describes the interval procedure: under its assumptions, repeated samples produce intervals containing the true quantile about 95% of the time. The interval endpoints do not define a band for tomorrow's realized loss.

For independent daily exceptions with common probability \(p\), the probability of at least one exception over \(N\) days is \(1-(1-p)^N\). At a daily exception probability of 1%, the chance across 100 independent trading days is \(1-0.99^{100}\approx63.4\%\). The expected count is one, yet the chance of seeing no exception is about 36.6%. This calculation concerns the sequence of daily forecasts. A 100-day terminal loss requires its own distribution.

Table 4 gives the one-sided normal multipliers for three frequently compared coverage levels.

Table 4. One-sided normal loss cutoffs
VaR confidenceUpper-tail probabilityStandard deviations above mean loss
95%5%1.645
97.5%2.5%1.960
99%1%2.326

With zero mean and USD 2 million daily loss volatility, 95% normal VaR is approximately USD 3.29 million. The familiar multiplier 1.960 leaves 2.5% in the upper tail because a central 95% interval splits its excluded probability between two tails. To change confidence under the normal model, rescale the dispersion component by the ratio of the two multipliers and retain the same horizon mean.

1.1.1.5 Average the worst probability mass

VaR identifies where the tail begins. To measure severity, give every unit of probability in the worst \(1-\alpha\) share equal weight, then calculate its average loss. For an integrable loss distribution the general definition is \[\mathrm{ES}_{\alpha}=\frac{1}{1-\alpha}\int_{\alpha}^{1}q_u(L)\,du,\] where \(u\) is cumulative probability and \(q_u(L)\) is the corresponding loss quantile in currency. The integral adds losses over a probability interval; division by the interval width returns a currency amount. Every included loss is at least the cutoff, so ES is at least VaR for the same portfolio, horizon and confidence. Equality holds if the whole selected tail has the cutoff loss.

For a continuous distribution with the intended tail probability above the cutoff, ES equals the conditional mean loss beyond VaR. A discrete distribution requires one extra decision. Let \(v\) denote VaR, \(t=1-\alpha\) the target tail probability, and \(b=P(L>v)\) the mass strictly above the cutoff. Include all outcomes strictly above \(v\), then add exactly \(t-b\) of probability at \(v\). This gives

\[\mathrm{ES}_{\alpha}=\frac{\sum_{i:L_i>v}p_i L_i+(t-b)v}{t}.\]

Worked case

The trading book in Table 2 has a USD 15 million stress limit. The committee wants its 95% ES and the change in ES if the dislocation loss increases to USD 24 million with all state probabilities fixed. The limit is expressed in dollars but supplies neither a scenario loss nor a probability.

The worst 5% contains the entire 2% dislocation state. To reach 5%, take 3 percentage points of the 8% state at USD 4 million. The original ES is \([0.02(14)+0.03(4)]/0.05=8\) million dollars. This lies between the USD 4 million boundary loss and the USD 14 million maximum. USD 14 million alone averages only the worst 2%. Including every outcome at the cutoff would average the worst 10%.

After the shock revision, the 95% cutoff remains USD 4 million because 98% of outcomes are still at or below it. ES rises by \((0.02/0.05)(24-14)=4\) million dollars, to USD 12 million. The 2% state carries 40% of the weight inside the selected tail. That conditional weight explains why a USD 10 million deterioration moves ES by USD 4 million.

Fig 3 varies the severe loss while holding its probability and the USD 4 million boundary fixed. For this distribution, if \(d\) is the dislocation loss in USD millions, \(\mathrm{ES}_{0.95}=[0.02d+0.03(4)]/0.05\).

Fig 3. Tail severity at a fixed quantile
Key point: The slope of ES equals the severe state's share of the selected tail.
{"x":{"label":"Dislocation loss (USD millions)","min":4,"max":30},"y":{"label":"One-day risk (USD millions)","min":0,"max":18},"curves":[{"label":"95% expected shortfall","expr":"(0.02*x+0.03*4)/0.05"},{"label":"95% value at risk","expr":"4"}],"sliders":[],"readouts":[],"try":"Predict the rise in expected shortfall for a dislocation loss of 24 instead of 14, then compare the two points on the rising line.","say":"The value-at-risk line stays horizontal at four million dollars. The expected-shortfall line rises by forty cents per additional dollar of dislocation loss.","prompt":"a live graph of one-day risk (USD millions) against dislocation loss (USD millions). It has 2 curves: 95% expected shortfall and 95% value at risk. The value-at-risk line stays horizontal at four million dollars. The expected-shortfall line rises by forty cents per additional dollar of dislocation loss. Try this: Predict the rise in expected shortfall for a dislocation loss of 24 instead of 14, then compare the two points on the rising line."}

A Separate the loss beyond the cutoff from total tail loss

Let \(v\) be VaR and define the positive excess as \((L-v)^+=\max(L-v,0)\). Its unconditional expectation is the average amount beyond the cutoff per forecast period, counting zero on every day that stays within the cutoff. For integrable losses, \(\mathrm{ES}_{\alpha}=v+E[(L-v)^+]/(1-\alpha)\). This identity also holds when some probability sits exactly at VaR.

In Table 2, a desk absorbs the first USD 4 million of loss, and a hypothetical protection contract covers all loss above that amount. The expected daily payout is \(0.02(14-4)=0.20\) million dollars. At 95% confidence, the excess averaged over the full 5% tail is \(0.20/0.05=4\) million dollars. On days that strictly breach the cutoff, the average payout is USD 10 million. The denominator distinguishes the three quantities.

1.1.1.6 Choose the measure for the risk decision

A manager asking how much a desk usually earns needs information from the whole distribution. A manager asking how severe a bad week could be needs to specify whether the question concerns a percentile, an average of adverse outcomes, or a particular market event. A stress test can reprice the book under a named shock even when a reliable probability for that shock is unavailable.

Fig 4 distinguishes four requests that can all arrive under the heading of portfolio loss.

Fig 4. Matching a risk request to its measure
Key point: A specified market shock determines a scenario loss without requiring a percentile.
flowchart TD
A[Risk committee request] --> B{Named market shock?}
B -->|Yes| C[Reprice under shock]
B -->|No| D{Which outcomes matter?}
D -->|All states| E[Expected signed loss]
D -->|Coverage threshold| F[Value at risk]
D -->|Worst fraction| G[Expected shortfall]

Aggregation can reverse a VaR comparison. Suppose two books each lose USD 10 million with probability 4% and otherwise break even. Their loss events are mutually exclusive. Each book has 95% VaR of zero because it breaks even 96% of the time. Together they lose USD 10 million with probability 8% and break even only 92% of the time. Combined 95% VaR is USD 10 million, exceeding the sum of the stand-alone VaRs. The portfolio has gained sufficient probability of a loss to cross the percentile boundary, although the two books never lose together.

Subadditivity means \(\rho(L_A+L_B)\le\rho(L_A)+\rho(L_B)\), where \(\rho\) is a risk measure and \(L_A,L_B\) are same-horizon losses. ES satisfies this property. In the example, each stand-alone 95% ES is \([0.04(10)+0.01(0)]/0.05=8\) million dollars. Combined ES is USD 10 million because the entire worst 5% consists of USD 10 million losses. The sum of stand-alone ES values gives an upper bound; it does not specify the combined value. The joint distribution supplies that value.

Precision of tail estimates. In 250 equally weighted observations, a 99% ES estimate draws on only 2.5 observations' worth of probability. One unusually large loss can therefore materially change the estimate. Higher confidence moves the calculation farther into the tail while reducing the amount of observed tail evidence. The largest historical loss is a sample outcome, not an economic ceiling.

See also. Parametric Portfolio VaR Estimation, Historical Simulation VaR Estimation and Monte Carlo Market Risk Estimation develop alternative ways to obtain a distribution. Expected Shortfall and Coherent Measures and Risk Measure Estimation Uncertainty develop the severity and precision questions. VaR Mapping Across Risk Factors, Extreme Value Tail Estimation and Correlation Estimation and Breakdown address exposures, extremes and joint outcomes.

A Check how a reported measure responds to a change in the book

Table 5 gives three tests that both absolute VaR and ES pass when confidence, horizon and scenario weights stay fixed. Together with subadditivity, these are the four defining properties of a coherent risk measure under the signed-loss convention.

Table 5. Risk-measure checks under matched scenarios
PropertyChange to scenario lossesRequired response
Translation invarianceAdd a certain loss of USD 2 million in every state.Absolute VaR and ES each rise by USD 2 million.
Positive homogeneityScale every signed loss by 1.5.VaR and ES each become 1.5 times their previous value.
MonotonicityReduce the loss in every state or leave it unchanged.Neither VaR nor ES can rise.

A fixed funding expense shifts every loss and its mean by the same amount. Absolute VaR rises by the expense, while mean-relative VaR stays unchanged because centering removes the shift. This difference identifies which report should move when a certain expense is incorporated.

Positive homogeneity applies to a nonnegative multiple of the entire loss variable. Reversing a trade swaps gains and losses and requires reading the opposite tail. Increasing a position can also change liquidation costs disproportionately. In that case the scenario losses themselves no longer scale with the position size.

For monotonicity, compare the two books in each market state. A hedge that reduces the worst loss but introduces a loss elsewhere does not establish state-by-state dominance. Conversely, a hedge that improves every adverse state can leave a percentile unchanged if the outcome that sets the cutoff remains unchanged.

B Recall check

Give the convention or quantity that resolves each question.

Which cash flow must be removed when it increases account value without earning a return?
An external capital contribution.
What happens to standard deviation when P&L is converted to signed loss?
It stays the same because squared deviations are unchanged.
Which ascending observation gives 95% inverse empirical VaR among 80 equally weighted losses?
The 76th observation.
What fills the gap when strict exceedances occupy less than the ES tail probability?
Use enough probability mass at the VaR boundary to complete the tail.
What portfolio information is lost when position losses are sorted independently?
Which outcomes occur together in each market state.
COMMON MISTAKE

Common mistakes

Table 6 gives the detail in each problem that changes the calculation.

Table 6. Similar-looking calculations with different targets
Tempting choiceWhy it looks plausibleDetail that settles it
Take the upper P&L percentileThe requested confidence is high.P&L ranks gains highest; adverse P&L occupies the lower tail.
Average every loss tied at VaREvery tied loss is equally adverse.The specified tail has a fixed probability budget.
Multiply ten-day VaR by ten to obtain 100-day VaRThe new horizon is ten times longer.Mean and dispersion have different time scaling, even under independent normal increments.
Use the scenario count as its weightHistorical simulation often weights dates equally.Explicit scenario probabilities govern the cumulative distribution.
Read the worst observed loss as the portfolio's maximum possible lossIt occupies the last row in the sample.A historical sample lists realized shocks only.
QUIZ

Self-check quiz

Question 1

Using Table 2, calculate the trading book's one-day 97.5% ES.

  1. USD 4 million
  2. USD 8 million
  3. USD 14 million
  4. USD 12 million

Question 2

A portfolio opens at EUR 50 million. Its owner contributes EUR 3 million. It closes at EUR 52.4 million after paying EUR 0.2 million of investment income to the owner. What is the signed economic loss?

  1. EUR −2.6 million
  2. EUR 0.4 million
  3. EUR 0.6 million
  4. EUR 0.8 million

Question 3

A model reports calibrated, continuous one-day 98% VaR on each of 150 trading days. The risk manager has no evidence about independence. Which statement follows?

  1. The expected number of exceptions is three.
  2. The probability of no exceptions is 98% raised to the 150th power.
  3. Exactly three daily losses will exceed their forecasts.
  4. The mean loss on exception days equals the reported VaR.

Question 4

A desk reports a one-week 99% mean-relative VaR of GBP 5 million and expected one-week P&L of GBP 0.7 million. Which absolute VaR uses the same distribution?

  1. GBP 5.7 million
  2. GBP 5.0 million
  3. GBP 4.3 million
  4. GBP −4.3 million

Question 5

Two desks each have a 3% probability of losing USD 6 million and otherwise break even. Their loss events cannot coincide. Which pair gives the combined 95% VaR and 95% ES?

  1. USD 0 and USD 3.6 million
  2. USD 0 and USD 6 million
  3. USD 6 million and USD 7.2 million
  4. USD 6 million and USD 6 million
ANSWERS

Answer key

Question 1 - Answer: D. The worst 2.5% comprises 2% at USD 14 million and 0.5% at USD 4 million, giving \([0.02(14)+0.005(4)]/0.025=12\) million dollars. A gives the cutoff, B uses the 95% tail, and C includes only strict exceedances.

Question 2 - Answer: B. Economic P&L is \(52.4-50+0.2-3=-0.4\) million euros, giving a positive signed loss of EUR 0.4 million. A counts contributed capital as earnings, C omits distributed income, and D subtracts that income from P&L.

Question 3 - Answer: A. Adding the 150 daily exception probabilities gives \(150(0.02)=3\) expected exceptions. B requires independence to multiply the daily probabilities, C treats an expectation as a scheduled count, and D substitutes the threshold for average tail severity.

Question 4 - Answer: C. Expected loss is GBP −0.7 million. Adding expected loss back to mean-relative VaR gives \(5-0.7=4.3\) million pounds. A reverses the adjustment, B assumes a zero mean, and D reports the sign of adverse P&L instead of signed loss.

Question 5 - Answer: D. The portfolio breaks even with probability 94%, so the 95th percentile lies at USD 6 million and the entire worst 5% has that loss. A gives one desk's measures, B retains a stand-alone zero quantile, and C adds stand-alone ES estimates despite the mutually exclusive losses.

RAPID

Rapid review

  • Positive economic P&L → negative signed loss.
  • Portfolio percentile → aggregate within joint scenarios before ranking.
  • Unequal weights → locate the first cumulative probability that reaches confidence.
  • Ties at the cutoff → strict exceptions can occupy less than the nominal tail.
  • Expected exception count → multiply forecast count by calibrated exceedance probability.
  • Mean-relative VaR → subtract mean signed loss from absolute VaR.
  • Independent normal daily losses → scale the mean linearly and dispersion by the square root of time.
  • Discrete ES → fill the worst probability fraction exactly, including a partial boundary weight when needed.
  • Stress loss → revalue the portfolio under the named shock.
  • Log total return → exponentiate to recover the closing-to-opening wealth ratio.
  • Foreign holding → multiply local and exchange-rate gross returns in the reporting-currency quote.
  • One-sided 95% normal VaR → use 1.645 standard deviations above mean loss.
  • At least one independent breach → subtract the probability of zero breaches from one.
  • Certain expense in every state → absolute VaR rises by the expense and mean-relative VaR stays fixed.

References

  • 2026 FRM Study Guide, Part II. Global Association of Risk Professionals, 2026.
  • 2026 FRM Learning Objectives, Part II. Global Association of Risk Professionals, 2026.
End of your free lesson

Keep going with the full FRM Part II Exam Prep Course

You just studied the way members do, in the same lesson interface. Full access unlocks every lesson, the question bank and flashcards for a full year.

Up next with full access1.1.2 - Parametric Portfolio VaR Estimation

One payment per program