THE GOLD STANDARD OF JUDGEMENT · SPECIALIST STUDY
Read EVaR as a bound, an optimisation and a modelling commitment
Entropic Value at Risk turns an exponential-moment bound into a coherent tail-risk measure. It can be computed as an infimum over a positive scale and interpreted through a relative-entropy ambiguity set. Those strengths come with a sharp condition: the relevant exponential moment must exist and be estimated responsibly. EVaR is therefore not a superior number by default; it is a transparent commitment to a tail model and confidence level.
This guide belongs to the Judgement subject guide, the Gold Standard of Education, and the How X Works learning library. Choose a reading route, work the examples, then use the transfer checks to test whether the method survives a changed context.
Choose a reading route
Build the formula · chapters 1–3
Work the numbers · chapters 4–5
Compare risk measures · chapters 6–7
Estimate and stress · chapters 8–9
Govern the decision · chapters 10
The whole problem at a glance
| Object | Definition or role | Main question |
|---|---|---|
| Loss L | Higher values are worse | Is the sign convention explicit? |
| Moment-generating function | M_L(t)=E[exp(tL)] | Is it finite for some t>0? |
| Confidence level | 1−α, with α a small tail probability | Which rare-event probability is being bounded? |
| EVaR | inf over t>0 of [ln M_L(t)−ln α]/t | How sensitive is the result to tails and estimation? |
| Dual set | Distributions within a KL-divergence radius | Which alternative laws are considered plausible? |
Different publications use α for the confidence level or for the tail probability. This article states the convention each time; software users must verify theirs.
CHAPTER 1 OF 10 · Build the formula
1. Fix the loss sign and confidence convention
Core idea. Risk measures become confusing when profit, return, cost and loss signs are mixed. Let L denote loss so larger values are worse. Let α be the allowed upper-tail probability, such as 0.05 for a 95% confidence statement. State both conventions beside every formula and output.
Method. Translate source data into one loss variable, document units and horizon, choose α from a decision requirement, and verify software parameterization. Test the sign with a simple scenario: adding a sure loss should increase the reported risk by the same amount under cash-invariant measures.
Worked example. A portfolio return of −8% becomes a loss of +8 percentage points. If software expects returns but the analyst supplies losses, the tail can reverse. A unit test with a constant +2 loss should move EVaR upward by 2.
Changed-context transfer. Move from one-day monetary loss to annual percentage loss. The distribution, dependence and units change, so the same α does not make the outputs directly comparable. Rebuild the model rather than rescaling mechanically.
Limits. Confidence level is not a probability that the risk number is correct. It indexes a tail bound under the model. Parameter uncertainty and sampling confidence require separate analysis.
Misconception to avoid. EVaR at 95% means losses cannot exceed the number with 95% certainty. It is a model-based bound or risk measure, not a guarantee about a single future observation.
Practical evidence check
- Write a one-line contract containing variable, sign, unit, horizon, α convention, data window and whether the model is parametric or empirical.
- Return to chapter contents
CHAPTER 2 OF 10 · Build the formula
2. Derive the Chernoff bound from exponential moments
Core idea. For t>0, Markov’s inequality applied to exp(tL) gives P(L≥z) ≤ exp(−tz)E[exp(tL)]. Rearranging a target probability α yields z ≥ [ln E(exp(tL))−ln α]/t. Minimising over t selects the tightest bound in this exponential family.
Method. Check that the moment-generating function is finite for at least one positive t. Evaluate the log MGF rather than the raw exponential where possible. Treat t as an optimisation variable with units reciprocal to loss, and preserve the domain over which the model is valid.
Worked example. For the two-outcome loss, M_L(t)=0.99+0.01e^(100t). Each positive t yields a valid Chernoff-derived threshold under the model. The smallest objective value is EVaR for the chosen α.
Changed-context transfer. Replace the bounded loss with a lognormal-type heavy loss whose MGF is infinite for every t>0. The derivation cannot deliver a finite EVaR even if a finite quantile exists.
Limits. A sample MGF is always finite for finite observations, which can hide population nonexistence. Empirical computation does not prove the theoretical tail condition.
Misconception to avoid. EVaR is just the expected exponential loss. The risk measure optimises a transformed log moment divided by t; the exponential moment is an ingredient, not the final scale.
Practical evidence check
- For a proposed distribution, state the interval of t for which the MGF exists and verify the numerical optimiser never leaves it.
- Return to chapter contents
CHAPTER 3 OF 10 · Build the formula
3. Read EVaR as an infimum over risk sensitivity
Core idea. The optimisation balances two terms. Larger t emphasises severe losses in the log MGF; smaller t magnifies the confidence penalty −ln α through division by t. The minimiser selects the exponential tilt that makes the probability bound tightest within the model.
Method. Define f(t)=[log M_L(t)−log α]/t for t>0, examine limits near zero and the domain boundary, then solve a one-dimensional convex problem when assumptions permit. Use stable gradients or bracketing and verify the result with neighbouring t values.
Worked example. In a normal loss model with mean μ and variance σ², log M_L(t)=μt+σ²t²/2. Minimising gives EVaR=μ+σ√(−2ln α). The formula displays how confidence and volatility interact.
Changed-context transfer. Change the normal model to a mixture with rare high variance. The closed form disappears, and the optimum can move toward tail components. Numerical EVaR becomes more sensitive to the mixture assumptions.
Limits. An infimum may occur near a numerical boundary, and floating-point overflow can create a false minimum. Report solver diagnostics, domain and sensitivity.
Misconception to avoid. The optimizing t is a universal risk-aversion parameter. It depends on the distribution, unit, horizon and α; it is not automatically a stable preference trait.
Practical evidence check
- Recompute EVaR after multiplying all losses by ten. Confirm the risk scales by ten and the optimizer t scales inversely.
- Return to chapter contents
CHAPTER 4 OF 10 · Work the numbers
4. Work a discrete example without hiding optimisation
Core idea. Discrete losses make the calculation transparent because the MGF is a weighted sum of exponentials. They also show why a single feasible t gives only an upper value of the objective, not the minimum that defines EVaR.
Method. Tabulate outcomes and probabilities, compute log-sum-exp for several t values, bracket the minimum, solve, and compare the reported objective with a dense grid. Preserve enough precision to avoid rounding away the optimum.
Worked example. For losses 0 with probability .99 and 100 with probability .01, α=.05. At t=.02, the objective is about 154.5. Trying t=.04 gives a different bound. A solver must locate the minimum; quoting the first calculation as EVaR is wrong.
Changed-context transfer. Change the severe loss from 100 to 1,000 while keeping its probability. The exponential term grows dramatically and the optimum changes. This is the intended tail sensitivity, but it also exposes estimation fragility.
Limits. Probabilities near zero and large losses can overflow in ordinary arithmetic. Use log-sum-exp, scaled variables and high-precision checks for consequential work.
Misconception to avoid. EVaR must lie inside the observed loss range. A Chernoff bound can exceed the maximum observed sample loss when the empirical distribution or confidence penalty warrants it; population models can extend further.
Practical evidence check
- Implement the example two ways—direct exponentials and log-sum-exp—then increase the loss until the direct version fails and confirm the stable version remains coherent.
- Return to chapter contents
CHAPTER 5 OF 10 · Work the numbers
5. Understand coherence, convexity and translation
Core idea. EVaR is a coherent risk measure under its domain: it is monotone, cash-invariant, positively homogeneous and subadditive. These properties support aggregation and convex optimisation, but they do not guarantee that the fitted distribution or chosen confidence level is appropriate.
Method. Test properties with controlled inputs. Add a constant c and confirm risk rises by c; multiply nonnegative losses by λ and confirm scaling; compare ordered losses for monotonicity; and verify diversification under a jointly modelled dependence structure.
Worked example. Two units have dependent losses. Summing their separately estimated EVaRs can be more conservative than modelling the joint total, while subadditivity says the coherent measure of the sum should not exceed the sum of individual risks under a common probability model.
Changed-context transfer. If the units were estimated with inconsistent scenarios or horizons, the algebraic property does not rescue the aggregation. Harmonise the joint model before invoking coherence.
Limits. Subadditivity can encourage false confidence when common shocks, tail dependence or model boundaries are omitted. Coherence is a mathematical consistency property, not a data-quality certificate.
Misconception to avoid. Coherent means objectively correct. It means the measure satisfies specific axioms; those axioms are useful but not the only criteria for a decision.
Practical evidence check
- Run property tests on the exact production implementation, including constants, rescaling and a small dependent two-loss example.
- Return to chapter contents
Worked two-outcome loss
| Input | Value | Role |
|---|---|---|
| Loss outcomes | 0 with probability 0.99; 100 with probability 0.01 | A rare severe loss |
| Tail probability α | 0.05 | Corresponds to 95% confidence under this convention |
| MGF | 0.99 + 0.01 exp(100t) | Feeds the EVaR objective |
| Objective at t=0.02 | [ln(0.99+0.01e²)−ln(0.05)]/0.02 ≈ 154.5 | One feasible upper value, not the optimum |
| EVaR | Minimum over t>0 | Requires numerical optimisation |
The numbers are illustrative. Optimisation, units, distribution and α convention must be checked before comparison.
CHAPTER 6 OF 10 · Compare risk measures
6. Compare EVaR with VaR and Expected Shortfall
Core idea. VaR names a quantile, Expected Shortfall averages losses in the tail, and EVaR gives the tightest exponential-moment Chernoff bound. Under matched definitions and regularity, EVaR is generally at least as conservative as Expected Shortfall, which is at least as conservative as VaR.
Method. Compute all three on the same loss variable, horizon, α convention and distribution. Add scenario losses and report tail sample size. Explain which decision each measure supports rather than ranking numbers without purpose.
Worked example. A bounded discrete loss has 95% VaR near zero because the severe event probability is 1%. Expected Shortfall incorporates part or all of the severe outcome depending on convention, while EVaR responds through the exponential moment and may be larger.
Changed-context transfer. Raise the severe-event probability above 5%. VaR now jumps to the severe loss, while Expected Shortfall and EVaR move more continuously. The change shows why measure behaviour depends on the distribution around α.
Limits. Ordering can be misreported when one source uses α as confidence and another as tail probability. At atoms, Expected Shortfall conventions also require care.
Misconception to avoid. EVaR replaces stress testing because it is more conservative. A scalar risk measure cannot expose scenario mechanisms, liquidity, contagion or model failure.
Practical evidence check
- Create a comparison sheet that displays formula convention, horizon, α, estimate, uncertainty, worst scenarios and decision threshold for every measure.
- Return to chapter contents
CHAPTER 7 OF 10 · Compare risk measures
7. Use the KL-divergence dual to see model ambiguity
Core idea. EVaR has a dual interpretation as the worst expected loss over probability distributions within a relative-entropy, or Kullback–Leibler, neighbourhood of the reference law. The confidence level determines the divergence budget. This connects tail risk to distributional robustness.
Method. State the reference distribution P and ambiguity set Q with D_KL(Q||P) bounded by a radius tied to −ln α. Identify support assumptions: KL divergence penalises placing mass where P assigns none, so the empirical reference can impose a hidden support boundary.
Worked example. For the two-outcome model, the worst alternative law shifts more probability toward the 100 loss while paying a KL penalty. The optimiser’s exponential tilt describes that reweighting. EVaR is the resulting worst expected loss within the allowed information distance.
Changed-context transfer. Add a disaster loss absent from the reference support. A KL ball around an empirical distribution may be unable to assign it mass. The dual looks robust yet misses the scenario entirely; explicit stress augmentation is needed.
Limits. KL divergence is asymmetric and sensitive to reference support. Other ambiguity sets, such as Wasserstein balls, express different ideas about plausible model error.
Misconception to avoid. The KL dual protects against every unknown distribution. It protects against a defined neighbourhood of a chosen reference, not unmodelled support or arbitrary structural breaks.
Practical evidence check
- List every severe scenario outside the reference support and decide whether to add it, use another ambiguity set or disclose the blind spot.
- Return to chapter contents
CHAPTER 8 OF 10 · Estimate and stress
8. Estimate log moments with numerical and statistical discipline
Core idea. EVaR depends on exponential weighting, so a few large observations can dominate. Naive sample estimates can be unstable, overflow numerically and understate unseen tails. Estimation design is part of the judgement, not a technical afterthought.
Method. Use log-sum-exp, unit scaling, out-of-sample validation, block or dependence-aware resampling, influence diagnostics and confidence intervals. Compare empirical, parametric and semiparametric models. Report how the minimising t and EVaR change across plausible specifications.
Worked example. A five-year daily sample has one extreme loss. Removing it lowers EVaR sharply, while VaR barely moves. That influence is informative: it shows the result is tail-driven and requires scrutiny of both data validity and tail plausibility.
Changed-context transfer. Extend the sample into a different regime. Pooling may dilute a new tail; fitting only the new regime reduces sample size. Present both models and the decision consequences instead of choosing silently.
Limits. Bootstrap methods can fail under heavy tails or dependence. Parametric models can look stable because they impose the tail. No interval compensates for an implausible distribution.
Misconception to avoid. More decimal places indicate a better EVaR estimate. Precision in printing is unrelated to model and sampling uncertainty.
Practical evidence check
- Publish an influence chart, solver diagnostics and at least three model variants; require the decision to be robust or explicitly conditional on the variant.
- Return to chapter contents
Risk-measure comparison
| Measure | Tail information | Typical strength | Typical limit |
|---|---|---|---|
| VaR | One quantile | Simple threshold interpretation | Ignores severity beyond the quantile and is not coherent in general |
| CVaR / Expected Shortfall | Average loss beyond a quantile under continuity conventions | Direct tail-severity interpretation and coherence | Estimation can be unstable with little tail data |
| EVaR | Entire exponential moment through a Chernoff bound | Coherent, convex and linked to KL robustness | May be conservative and can be infinite for heavy-tailed laws |
| Stress test | Specified scenarios | Transparent causal challenge | Does not assign a complete probability law by itself |
Under common regularity and matched conventions, EVaR is at least as conservative as CVaR, which is at least as conservative as VaR. Check definitions before using the ordering.
CHAPTER 9 OF 10 · Estimate and stress
9. Recognise heavy tails and MGF nonexistence
Core idea. Some important loss distributions have no finite positive MGF. In those cases population EVaR is infinite or outside the model domain, even though finite samples yield a number. That is a substantive warning about exponential-tail assumptions.
Method. Test theoretical MGF existence for the chosen family, inspect empirical tail plots and exceedances, compare tail-index estimates, and stress truncated or tempered alternatives. Never declare finiteness from a successful numerical routine alone.
Worked example. A Pareto loss with a heavy upper tail has finite quantiles but infinite positive exponential moments. The empirical plug-in EVaR is finite because the sample maximum truncates reality. As more severe observations arrive, it can rise without stabilising.
Changed-context transfer. Cap losses through an enforceable contract. The bounded exposure may now have a finite MGF, but counterparty failure or legal non-enforcement can reopen the tail. Model the cap’s credibility as a scenario.
Limits. Tail diagnostics have high uncertainty with few extremes. State the range of plausible tail behaviour and avoid using EVaR as the sole acceptance criterion where finiteness is doubtful.
Misconception to avoid. If software returns a finite number, EVaR exists. Software has evaluated the sample or fitted model; existence is a property of the assumed population law and domain.
Practical evidence check
- For each loss class, document whether exposure is physically or contractually bounded, whether the MGF exists, and what observation would invalidate that belief.
- Return to chapter contents
CHAPTER 10 OF 10 · Govern the decision
10. Turn the number into a governed decision
Core idea. A risk measure has value only when connected to appetite, limits, alternatives and actions. EVaR’s conservatism can support robust decisions, but unexamined conservatism may immobilise useful activity or hide distributional burdens.
Method. Define the decision threshold, escalation rule, model owner, independent review, update frequency and override process. Pair EVaR with Expected Shortfall, scenarios, liquidity or service impacts, and explain who benefits or bears the risk.
Worked example. A programme uses EVaR to size a reserve. The committee also reviews cash-flow stress, stakeholder harm and model sensitivity. If EVaR breaches its limit only under one disputed tail model, the decision record states that dependency and selects a reversible action.
Changed-context transfer. Transfer to public-service capacity. Loss may be unmet demand rather than money. Units, ethics and risk-sharing change, yet the same discipline—define loss, horizon, α, tail model and action—still applies.
Limits. Optimising against one coherent measure can encourage gaming of the model boundary. Independent scenarios, outcome monitoring and periodic loss-definition review remain necessary.
Misconception to avoid. A larger EVaR always means a worse decision. Scale, benefit, reversibility, risk bearer and available safeguards matter; the measure describes one dimension of loss.
Practical evidence check
- For the recommended option, write the threshold, all companion metrics, three model sensitivities, affected groups, review date and trigger that would reverse the choice.
- Return to chapter contents
Derive the normal-model closed form carefully
For L normally distributed with mean μ and variance σ², the log MGF is μt+σ²t²/2. Substituting produces μ+σ²t/2−ln(α)/t.
Differentiate for t>0: σ²/2+ln(α)/t²=0. Because ln(α)<0, the minimiser is √(−2ln α)/σ when σ>0.
Substitution gives μ+σ√(−2ln α). The result is cash-invariant and positively homogeneous. As α shrinks, the confidence penalty grows through the square root of the log.
The elegance is conditional on normality. Real loss tails may be skewed, bounded, mixed or heavy. Use the formula as a transparent benchmark, not a universal shortcut.
Audit a production EVaR calculation
Confirm loss sign, unit, horizon, valuation date, α convention, data lineage and treatment of missing or revised observations.
Review distribution choice, dependence model, MGF domain, support, parameter estimation, regime handling and scenario augmentation. Preserve rejected models and reasons.
Inspect log-sum-exp implementation, optimizer bounds, gradients, convergence, scaling, reproducibility and property-based tests. Recompute a small case independently.
Trace the output to its limit, decision, approver, override and monitoring rule. A correct number with no governed action is not a complete control.
Use sensitivity surfaces instead of one confidence level
Compute EVaR across a justified grid of α values, horizons and model specifications. Plot or tabulate the risk and minimising t, not just the selected headline value.
Look for cliffs where small changes create large reserve or ranking shifts. Those are decision fragilities that deserve scenario review and reversible action.
Compare with Expected Shortfall and a severe stress at each grid point. Divergence among measures can reveal whether probability near the quantile, average tail severity or exponential extremes drive the choice.
Document which region reflects policy and which is exploration. Do not choose α after observing which value makes a preferred option pass.
Communicate tail risk without false certainty
Lead with the decision question and loss definition, then give EVaR with units, horizon and α. Avoid presenting a bare number.
State the model, data period, tail condition and dominant observations. Explain whether the MGF is theoretically finite and how results change under plausible alternatives.
Translate the output into a concrete but bounded statement: it is a model-based coherent risk measure and Chernoff bound, not a guaranteed maximum loss.
End with actions, safeguards and review triggers. Readers need to know what the measure changes, not merely how sophisticated it is.
Questions readers ask
Is EVaR the same as exponential utility?
- They share exponential moments, but EVaR is a coherent risk measure obtained by optimizing over the exponential scale for a confidence level. Exponential utility fixes a preference parameter.
Why can EVaR exceed Expected Shortfall?
- It uses the tightest Chernoff bound based on the full exponential moment and is generally more conservative under matched conventions.
Can EVaR be infinite?
- Yes. If the positive moment-generating function does not exist for the loss distribution, finite population EVaR may not exist.
Does a finite empirical EVaR prove the population value is finite?
- No. A finite sample is bounded and its empirical MGF is finite, even when the population MGF is infinite.
What does the KL dual add?
- It interprets EVaR as a worst expected loss over a relative-entropy neighbourhood of the reference distribution, exposing a distributionally robust meaning and its support limits.
Should EVaR replace VaR and Expected Shortfall?
- No. Use measures together with scenarios and decision context. Each reveals different aspects and assumptions.
Practise, verify and continue
For losses 0, 10 and 50 with probabilities .90, .08 and .02, choose a tail probability α, compute the log MGF on a grid of positive t, locate the approximate minimum and compare the result with VaR and Expected Shortfall under stated conventions.
Replace the bounded loss with a Pareto tail and then with a credible contractual cap. Explain when population EVaR becomes infinite, when it becomes finite, and which legal or counterparty scenario could invalidate the cap.
Related eduKate reading
Conditional Value at Risk, Expected Shortfall and coherent risk limits
Spectral risk measures, quantile weighting and tail preferences
Shortfall risk measures, acceptance sets and loss functions
Distributionally robust optimisation, ambiguity and out-of-sample guarantees
Expectiles, asymmetric loss and tail-sensitive decisions
Authoritative sources and scope
Ahmadi-Javid, Entropic Value-at-Risk: A New Coherent Risk Measure
Portfolio optimisation with Entropic Value at Risk
Delbaen, remarks on the structure of the entropic value-at-risk
Chernoff, a measure of asymptotic efficiency for tests based on sums of observations
The formulas and examples are educational and use simplified distributions. This article is not financial, investment, actuarial or regulatory advice. Consequential use requires domain validation, independent model review and jurisdiction-specific requirements. Before deployment, reproduce a benchmark calculation independently, confirm the moment-generating-function domain, stress unseen support, compare EVaR with Expected Shortfall and scenario losses, and record the decision rule. Reviewers should be able to see which observations dominate the exponential moment, how the optimiser behaved, and whether the recommendation survives plausible changes in confidence level, horizon, dependence and tail family. Where the decision affects people, also examine who bears the modeled loss, who bears the cost of conservatism, and which important harms the scalar loss variable leaves outside the model. Retain a dated model inventory, validation owner, limitation register and next review trigger so later users can distinguish an approved use from an attractive reuse outside the evidence boundary.