A risk spectrum makes the weight placed on each part of the loss distribution visible
Spectral risk measures average loss quantiles with a non-negative, normalised and tail-emphasising weight function, extending expected shortfall while exposing how decision makers value progressively worse outcomes.
Expected shortfall gives equal weight to losses inside a selected worst tail. Some decisions need a smoother or differently shaped emphasis across quantiles. A spectral risk measure supplies that weighting explicitly, allowing moderate losses and extreme losses to receive different importance.
The flexibility is useful and dangerous. A spectrum can be coherent yet poorly elicited, unstable in a thin tail or inappropriate for irreversible non-monetary harms. This guide treats the weighting function as a decision assumption that must be tested, not as a hidden technical setting.
Chapter contents
- 1. Define the loss distribution and quantiles
- 2. Represent risk attitude with a spectrum
- 3. Check the admissibility conditions
- 4. See expected shortfall as a spectral building block
- 5. Understand law invariance and comonotonic additivity
- 6. Worked example: compare two ordered loss profiles
- 7. Estimate a spectral measure from finite samples
- 8. Optimise decisions under spectral risk
- 9. Elicit or choose the spectrum responsibly
- 10. Quantify sampling and model uncertainty
- 11. Compare spectral risk with CVaR, utility and robust models
- 12. Govern implementation, communication and revision
Choose a reading route
- Route 1: Build the measure — losses, quantiles and spectra
- Route 2: Connect the theory — coherence and expected-shortfall mixtures
- Route 3: Work a calculation — discrete weights and interpretation
- Route 4: Choose and test — preferences, estimation and optimisation
- Route 5: Compare and govern — alternatives, limits and revision
How to use this guide
Start with the chapters that match your immediate decision, then return to the worked example and audit chapters. Each chapter separates a model from the evidence needed to use it responsibly. A claim is not strengthened merely by being technical; it is strengthened when assumptions, observations, alternatives and limits can be inspected.
| Layer | Question | Minimum record |
|---|---|---|
| Concept | What exactly is being represented? | Definitions and boundary |
| Method | How will the reasoning be carried out? | Steps, inputs and assumptions |
| Transfer | Would it work in a changed setting? | A deliberately altered case |
| Audit | What could prove it wrong? | Evidence, limits and revision trigger |
Chapter 1. Define the loss distribution and quantiles
The idea
A spectral measure starts with a random loss, a horizon and its quantile function. Sign, units, population and probability model must be explicit before weights are applied.
A disciplined method
Define positive loss, horizon, portfolio or system boundary, valuation rule and probability distribution; state the quantile convention for discrete data.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which ordered outcome is worse and which probability model orders it. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A service plan has ten scenario losses measured as unmet-demand hours, sorted from smallest to largest before risk weights are applied.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change the horizon and rebuild the distribution instead of rescaling the old spectrum mechanically.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Quantiles inherit missing scenarios, dependence errors and structural change.
Evidence to retain: loss definition, horizon, scenarios, probabilities, order statistics and quantile rule
Common misconception: The spectral measure belongs to the decision option without reference to a model.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 2. Represent risk attitude with a spectrum
The idea
The spectrum is a weighting function across cumulative probabilities. Under a loss convention, greater weight on worse quantiles expresses increasing concern toward the tail.
A disciplined method
Plot the function, check its direction under the chosen quantile ordering, normalise total weight to one and explain the decision meaning of its shape.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which probability regions receive more decision weight. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A flat spectrum gives every quantile equal weight; a rising spectrum places progressively more weight on severe losses.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Reverse the loss ordering and verify the monotonicity condition is translated correctly.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: A mathematical weight function does not capture every ethical or operational preference.
Evidence to retain: formula, plot, normalisation, ordering convention and interpretation
Common misconception: A steeper tail weight objectively means better risk management.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 3. Check the admissibility conditions
The idea
A coherent spectral risk measure uses a non-negative, normalised and monotonic risk-aversion spectrum under the stated convention. These conditions support monotonicity, translation invariance, homogeneity and subadditivity.
A disciplined method
Test non-negativity, integrate or sum weights to one, verify monotonicity and numerically check coherence properties on simple positions.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the selected weights define the claimed coherent measure. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A proposed spectrum has one negative interval to reward moderate losses; it fails admissibility even though the final number looks plausible.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Approximate a smooth spectrum on a coarse grid and verify discretisation preserves valid ordered weights.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Coherence is conditional on definitions and does not validate the loss model or objective.
Evidence to retain: condition tests, numerical integration, discretisation and property checks
Common misconception: Any weighted average of quantiles is coherent.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 4. See expected shortfall as a spectral building block
The idea
Expected shortfall applies constant weight across a worst tail and zero elsewhere. Spectral measures can be constructed as mixtures of expected shortfalls at different confidence levels.
A disciplined method
Write the target spectrum, approximate it with weighted expected-shortfall components and verify the mixture reproduces quantile weights.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which tail regions each component emphasises. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A measure blends 90-percent and 99-percent expected shortfall, giving broad tail concern plus extra emphasis on extremes.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Shift mixture weight toward the 99-percent component and inspect sensitivity to sparse observations.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Mixture interpretation depends on regularity and convention; extreme components can be unstable.
Evidence to retain: component levels, weights, reconstructed spectrum and numerical error
Common misconception: A spectral measure is unrelated to expected shortfall because its weights vary.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 5. Understand law invariance and comonotonic additivity
The idea
Spectral measures depend on the loss distribution rather than labels attached to scenarios. They are also additive for comonotonic risks that worsen together under the same underlying ordering.
A disciplined method
Test relabelled scenarios for invariance, identify comonotonic examples and avoid interpreting additivity as independence or universal diversification.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether two risks share the same adverse ordering. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Two losses both rise with the same heat-severity index; their comonotonic sum receives no diversification benefit from offsetting ranks.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Break the shared ordering and compare the combined spectral risk.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Distribution-only measures can ignore path, source, rights and causal structure.
Evidence to retain: joint scenarios, ranks, dependence structure, additivity check and interpretation
Common misconception: Law invariance means context and causation never matter to a decision.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 6. Worked example: compare two ordered loss profiles
The idea
A discrete example shows how two decisions with similar means and expected shortfall can separate under a more tail-sensitive spectrum.
A disciplined method
Sort eight equally likely losses for each plan, assign increasing weights that sum to one, multiply and sum, then compare the contribution of each quantile.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which severe outcomes drive the ranking. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
Plan A has several moderate losses and one very large loss; Plan B has higher middle losses but a lower maximum. A steep spectrum favours B more strongly than a flat spectrum.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Flatten the spectrum and show when the ranking reverses.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Eight scenarios are illustrative and cannot establish a production preference or tail model.
Evidence to retain: ordered losses, weights, products, total, contribution plot and alternative spectra
Common misconception: The option with the lower spectral value is best under every reasonable preference.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
| Quantile region | Flat spectrum | Tail-steep spectrum |
|---|---|---|
| Lower losses | Material weight | Small weight |
| Middle losses | Material weight | Moderate weight |
| Severe tail | Material weight | Large weight |
| Decision effect | Mean-like balance | Tail-dominated trade-off |
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 7. Estimate a spectral measure from finite samples
The idea
Empirical calculation sorts observations and combines order statistics with discrete weights. Unequal probabilities and ties require careful integration rather than convenient rank labels.
A disciplined method
Preserve scenario probabilities, integrate the spectrum over each quantile interval, allocate tied mass consistently and report effective weight concentrated in the observed tail.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is how continuous weights map onto finite observations. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
One scenario spans a wide probability interval and receives the integrated spectrum over that interval instead of one point weight.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Replace equal probabilities with importance-sampling weights and recompute interval contributions.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: High tail concentration can leave the estimate driven by very few observations.
Evidence to retain: sample, probabilities, sorting rule, interval weights, ties and effective tail size
Common misconception: Assigning the spectrum value at each rank is always an unbiased estimator.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 8. Optimise decisions under spectral risk
The idea
Spectral objectives can be optimised through formulations related to weighted expected shortfalls, avoiding a sorting operation inside some convex models.
A disciplined method
Select an admissible spectrum, build an equivalent finite mixture or auxiliary formulation, verify against direct calculation and sweep preference parameters.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the feasible loss model preserves convexity and valid weights. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A capacity plan minimises a spectral shortage measure subject to cost and service constraints; direct and optimisation calculations agree on test portfolios.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add integer facility decisions and reassess solver guarantees and computational burden.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Nonlinear, discrete or endogenous distributions can make optimisation difficult or nonconvex.
Evidence to retain: formulation, solver status, equivalence check, parameter sweep and feasibility audit
Common misconception: A coherent spectral objective guarantees an easy optimisation problem.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 9. Elicit or choose the spectrum responsibly
The idea
The spectrum encodes relative concern across loss quantiles. It can be derived from an explicit family and parameters, policy limits or structured comparisons, but cannot be inferred from a slogan such as risk averse.
A disciplined method
Use understandable trade-offs, test implied weights, compare stakeholder views, document authority and reject shapes that violate admissibility or ethical constraints.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whose risk attitude the spectrum represents. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A board compares one-unit reductions at the 90th and 99th loss quantiles; responses constrain a family of increasing spectra rather than selecting arbitrary points.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Repeat with service users who bear the losses and report conflicting spectra instead of averaging away responsibility.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: People struggle with tail probabilities, preferences can be context-dependent and aggregation is normative.
Evidence to retain: elicitation questions, participants, fitted family, conflicts, approvals and revision date
Common misconception: There is one objective spectrum for a risk-averse organisation.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 10. Quantify sampling and model uncertainty
The idea
A flexible tail-weight function can amplify estimation error. Confidence intervals, resampling and alternative scenario models should accompany the point estimate.
A disciplined method
Bootstrap with caution, compare parametric and empirical tails, stress dependence and extremes, and plot ranking stability across samples and spectra.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is whether the decision survives plausible uncertainty in losses and weights. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A plan ranks first under the central estimate but switches in 35 percent of resamples when the steepest tail weights are used.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Add a plausible unobserved stress scenario and measure its contribution.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Resampling cannot create structural breaks absent from data, and parametric tails can be misspecified.
Evidence to retain: data window, resamples, model alternatives, stress cases, intervals and rank stability
Common misconception: A smooth spectrum produces a stable risk estimate.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 11. Compare spectral risk with CVaR, utility and robust models
The idea
Spectral measures expose quantile weights; expected utility models wealth-dependent preferences; CVaR fixes one tail average; robust models protect against distributional ambiguity. They solve different problems.
A disciplined method
Match the method to the decision, compute several measures on the same scenarios and identify which assumption causes ranking differences.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is which risk concept is decision-relevant. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A repeated financial loss uses spectral comparison, while an irreversible safety limit remains a hard constraint and ambiguity is tested with a robust model.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Change the outcome from money to rights or health and reassess whether numerical aggregation is legitimate.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: No scalar measure captures all distributional, causal and ethical concerns.
Evidence to retain: comparison matrix, rankings, assumptions, excluded harms and decision rationale
Common misconception: Spectral flexibility subsumes every useful risk approach.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Chapter 12. Govern implementation, communication and revision
The idea
A spectral result should travel with the loss definition, spectrum, horizon, model, uncertainty and contribution chart. Independent review should detect sign or ordering errors.
A disciplined method
Version formulas and code, test simple cases, disclose weights visually, run challenger spectra, define override authority and trigger review after regime or objective change.
A useful working record makes the sequence visible: first state the question and unit of analysis; next record the observations and their provenance; then apply the stated rule; finally compare the result with plausible alternatives. For this chapter, the decisive choice is what event invalidates the selected spectrum or distribution. Keeping that choice explicit prevents a later reader from mistaking a convenient convention for a fact about the world.
Worked example
A new legal service standard converts part of the tail from a preference to a hard limit; the model is redesigned rather than merely steepened.
Work the example twice. On the first pass, follow the stated assumptions. On the second, reverse one assumption, remove one piece of evidence and ask whether the conclusion, ranking or boundary changes. This small stress test matters because polished outputs can hide sensitivity. The objective is not to force certainty; it is to locate the evidence on which the conclusion actually turns.
Changed-context transfer check
Communicate the result to nontechnical decision makers using contribution plots without hiding assumptions.
A transfer succeeds only when the structure of the problem survives the change. Match roles, dependencies and measurement conditions rather than surface vocabulary. Record what was preserved, what changed, and why the method still applies—or why it does not. If the new context changes the data-generating process or the people bearing the burden, re-specify the model instead of importing the old answer.
Limits, evidence and revision
Limit: Governance cannot turn unmodelled harms into valid weights.
Evidence to retain: model inventory, code tests, spectrum approval, challenger results, overrides and triggers
Common misconception: Publishing the final number is enough for accountability.
A gold-standard conclusion is therefore conditional. It should name the observation period, excluded cases, uncertainty and the event that would trigger review. Negative findings and unresolved contradictions belong in the record. They protect future readers from repeating a hidden assumption and make correction possible without reconstructing the entire analysis from memory.
Practical evidence check
Can another reader recover the input, reproduce the intermediate decision and identify at least one observation that would change the result? If not, the chapter has produced an explanation but not yet an auditable finding.
Field workbook
Create eight ordered loss scenarios for two decisions. Calculate a flat-weight average, a worst-quarter expected shortfall and one admissible increasing spectrum. Plot contributions, flatten and steepen the spectrum, then explain any ranking reversal.
- Write the decision question in one sentence and name the decision owner.
- List observations separately from interpretations, with dates and provenance.
- Record at least two plausible alternatives before choosing a preferred account.
- Run the changed-context test and note which assumptions fail.
- State the limit, confidence and next review trigger in plain language.
| Layer | Evidence | Failure signal |
|---|---|---|
| Loss model | Horizon, units and scenarios | Hidden sign or boundary |
| Spectrum | Non-negative, normalised, monotonic | Invalid weight shape |
| Estimate | Probability-interval weights | Thin-tail overconfidence |
| Decision | Sensitivity and authority | Preference hidden in code |
Continue through the verified ecosystem
- Conditional Value at Risk — expected shortfall and tail-loss limits
- Distributionally robust optimisation — distributional ambiguity and guarantees
- Chance-constrained optimisation — violation probabilities and reliability
- Multistage stochastic programming — scenario trees and recourse
- SMAA — missing weights and robust rank regions
Frequently asked questions
What is a spectral risk measure?
A weighted average of loss quantiles whose admissible weight function places greater concern on worse outcomes.
Is expected shortfall spectral?
Yes. It is a special case with equal weight across a selected worst tail and zero weight elsewhere.
What makes a spectrum coherent?
Under the stated convention, the weights are non-negative, normalised and ordered to give no less weight to worse losses.
Can a spectrum be estimated from preferences?
It can be constrained through structured elicitation, but tail probabilities and normative aggregation make this difficult.
Why test several spectra?
A decision may be driven by a hidden weighting choice; sensitivity shows whether the ranking survives plausible alternatives.
Does coherence prove the model is safe?
No. Coherence is an axiomatic property and does not validate scenarios, data or omitted harms.
Authoritative sources and further reading
- Acerbi: Risk Aversion and Coherent Risk Measures — spectral representation theorem
- Acerbi and Simonetti: Portfolio Optimization with Spectral Measures of Risk — optimisation formulation
- Acerbi and Tasche: Expected Shortfall — coherent tail-risk foundation
- Shapiro: On Kusuoka Representation of Law-Invariant Risk Measures — mixtures of Average Value at Risk
- Artzner et al.: Coherent Measures of Risk — coherence axioms
This guide explains a quantitative risk framework. It is not financial advice and does not validate the chosen losses, spectrum, distribution, stakeholder authority or use for irreversible harms.