Spectral Risk Measures¶

Spectral risk measures generalize Value-at-Risk (VaR) and Expected Shortfall (ES) by applying weights across the quantile distribution. They provide a coherent, theoretically principled framework for risk measurement that captures risk preferences through a spectrum of weights on tail outcomes.

Key Concepts¶

Coherent Risk Measures A risk measure $\rho: \mathcal{L} \to \mathbb{R}$ is coherent if it satisfies: 1. Monotonicity: $X \leq Y$ ⟹ $\rho(X) \leq \rho(Y)$ 2. Subadditivity: $\rho(X + Y) \leq \rho(X) + \rho(Y)$ (diversification benefit) 3. Positive homogeneity: $\rho(\lambda X) = \lambda \rho(X)$ for $\lambda > 0$ (scaling) 4. Translation invariance: $\rho(X + c) = \rho(X) - c$ (cash adds linearly)

VaR violates subadditivity (not coherent). Expected Shortfall is coherent.

Spectral Risk Measure Definition A spectral risk measure applies weight function $\phi(\alpha)$ across quantiles:

\[\rho_\phi(X) = \int_0^1 \phi(\alpha) q_\alpha(X) d\alpha\]

where $q_\alpha(X) = \text{VaR}_\alpha(X)$ is the $\alpha$-quantile, and: - $\phi(\alpha) \geq 0$ (non-negative weights) - $\int_0^1 \phi(\alpha) d\alpha = 1$ (weights sum to 1) - $\phi$ non-decreasing in $\alpha$ (higher weight on worse outcomes)

Intuition The weight function $\phi(\alpha)$ encodes: - How much emphasis to place on each tail region - Risk aversion profile: flat $\phi$ = uniform weight, increasing $\phi$ = conservative - $\phi(\alpha) = \delta(\alpha - \alpha_0)$: reduces to $\text{VaR}_{\alpha_0}$ - $\phi(\alpha) = \frac{1}{1-\alpha_0}$ for $\alpha \geq \alpha_0$: gives Expected Shortfall

Special Cases

Measure	Weight Function	Interpretation
Value-at-Risk	$\delta(\alpha - 0.95)$	Single quantile
Expected Shortfall	$\phi(\alpha) = \frac{1}{1-\alpha_0} \mathbf{1}_{\alpha \geq \alpha_0}$	Average of worst 5%
AveVar	$\phi(\alpha) = \frac{2}{1-\alpha_0}\mathbf{1}_{\alpha \geq \alpha_0}$	Linear increasing weights
Omega ratio	Specific choice for return/risk tradeoff	Gain/loss focus

Kusuoka Representation Kusuoka showed any spectral risk measure can be expressed as:

\[\rho_\phi(X) = \max_{p \in \mathcal{P}} \mathbb{E}_p[-X]\]

where the maximum is over distributions consistent with marginal constraints. This connects spectral measures to robust optimization.

Parametric Approaches For common distributions, spectral measures have closed forms: - Normal distribution: $\rho_\phi(X) = \mu + \sigma \int_0^1 \phi(\alpha) \Phi^{-1}(\alpha) d\alpha$ - Student-t: leads to heavier tail weighting - Mixture distributions: captures multi-modal risk (e.g., normal + jumps)

Advantages Over Standard Measures 1. Flexibility: adjust risk aversion without changing measure type 2. Theoretical soundness: coherence axioms satisfied 3. Tail sensitivity: directly incorporates tail behavior through weights 4. Portfolio optimization: convex optimization framework available 5. Regulatory acceptance: ES is mandated in Basel III, spectral measures studied for advanced approaches

Practical Implementation Computing spectral risk measures: 1. Estimate quantile function $q_\alpha$ from data or model 2. Choose weight function $\phi$ reflecting risk preferences 3. Numerically integrate: $\rho_\phi(X) = \int_0^1 \phi(\alpha) q_\alpha(X) d\alpha$ 4. Validate through backtesting and stress testing

Limitations - Weight function choice not uniquely determined by data - Requires stable quantile estimation (challenging in tails) - Computational cost vs. VaR (which is just a single quantile) - Interpretation less intuitive than "worst 5% loss"

Risk Management Practice

Spectral measures provide: - Theoretically principled risk aggregation - Flexibility to match institutional risk tolerance - Framework for incorporating expert judgment - Bridge between academic rigor and practical needs

Exercises¶

Exercise 1. A spectral risk measure is defined as $\rho_\phi(X) = \int_0^1 \text{VaR}_u(X)\,\phi(u)\,du$ where the weight function $\phi$ satisfies $\phi \ge 0$, $\int_0^1 \phi(u)\,du = 1$, and $\phi$ is non-decreasing. Show that ES at level $\alpha$ is a spectral risk measure by identifying its weight function $\phi(u) = \frac{1}{1-\alpha}\mathbf{1}_{\{u \ge \alpha\}}$.

Solution to Exercise 1

We must show that ES at level $\alpha$ can be written in the form $\rho_\phi(X) = \int_0^1 \text{VaR}_u(X)\,\phi(u)\,du$ with a valid spectrum $\phi$.

Claim: The weight function for $\text{ES}_\alpha$ is:

\[ \phi_\alpha(u) = \frac{1}{1-\alpha}\mathbf{1}_{\{u \ge \alpha\}} \]

Verification of the three conditions:

Non-negativity: $\phi_\alpha(u) = \frac{1}{1-\alpha} > 0$ for $u \ge \alpha$ and $\phi_\alpha(u) = 0$ for $u < \alpha$. Hence $\phi_\alpha(u) \ge 0$ for all $u \in [0,1]$.
Normalization:

\[ \int_0^1 \phi_\alpha(u)\,du = \int_\alpha^1 \frac{1}{1-\alpha}\,du = \frac{1}{1-\alpha}(1-\alpha) = 1 \checkmark \]

Non-decreasing: $\phi_\alpha$ is a step function that jumps from 0 to $\frac{1}{1-\alpha}$ at $u = \alpha$ and remains constant thereafter. It is non-decreasing.

Verification that the spectral measure equals ES:

\[ \rho_{\phi_\alpha}(X) = \int_0^1 \text{VaR}_u(X)\,\phi_\alpha(u)\,du = \int_\alpha^1 \text{VaR}_u(X) \cdot \frac{1}{1-\alpha}\,du = \frac{1}{1-\alpha}\int_\alpha^1 \text{VaR}_u(X)\,du \]

This is precisely the integral representation of $\text{ES}_\alpha(X)$. Therefore, ES at level $\alpha$ is a spectral risk measure with the weight function $\phi_\alpha(u) = \frac{1}{1-\alpha}\mathbf{1}_{\{u \ge \alpha\}}$.

Exercise 2. Explain why the non-decreasing condition on the weight function $\phi$ is necessary for the spectral risk measure to be coherent. What economic property does this condition encode (hint: aversion to tail risk)?

Solution to Exercise 2

Why the non-decreasing condition is necessary for coherence:

The non-decreasing condition on $\phi$ ensures that the spectral risk measure satisfies subadditivity, which is the key axiom separating coherent from non-coherent risk measures.

Economic property encoded: The non-decreasing condition encodes risk aversion in the sense that worse outcomes (higher quantiles in the loss distribution) receive at least as much weight as better outcomes. Formally:

If $u_1 < u_2$, then $\text{VaR}_{u_1}(X) \le \text{VaR}_{u_2}(X)$ (higher quantile levels correspond to larger losses).
The condition $\phi(u_1) \le \phi(u_2)$ means the risk measure puts more weight on these larger losses.
An agent who weights extreme losses at least as heavily as moderate losses is displaying aversion to tail risk.

What goes wrong without the non-decreasing condition:

If $\phi$ is allowed to decrease, one could construct a weight function that puts high weight on moderate losses but low weight on extreme losses. Such a measure would:

Fail subadditivity: Acerbi (2002) proved that a spectral risk measure is coherent if and only if $\phi$ is non-decreasing. A decreasing $\phi$ produces a spectral measure that violates subadditivity.
Ignore tail risk: By down-weighting the extreme tail, the measure underestimates the risk of rare catastrophic events.
Penalize diversification: Without subadditivity, combining portfolios could appear to increase risk, contradicting the fundamental principle of diversification.

Connection to the Dirac delta (VaR): The extreme case where $\phi(u) = \delta(u - \alpha)$ puts all weight on a single quantile and zero weight everywhere else. This is "infinitely non-monotone" in a distributional sense and corresponds to VaR, which indeed fails subadditivity.

Exercise 3. The exponential spectral risk measure uses $\phi(u) = \frac{\gamma e^{\gamma u}}{e^\gamma - 1}$ for risk aversion parameter $\gamma > 0$. Compute $\phi(0)$ and $\phi(1)$ for $\gamma = 2$. Explain how increasing $\gamma$ shifts more weight toward the tail of the loss distribution.

Solution to Exercise 3

The exponential spectral weight function is:

\[ \phi(u) = \frac{\gamma e^{\gamma u}}{e^\gamma - 1} \]

Computing $\phi(0)$ and $\phi(1)$ for $\gamma = 2$:

\[ \phi(0) = \frac{2 \cdot e^{2 \cdot 0}}{e^2 - 1} = \frac{2 \cdot 1}{e^2 - 1} = \frac{2}{7.389 - 1} = \frac{2}{6.389} \approx 0.313 \]

\[ \phi(1) = \frac{2 \cdot e^{2 \cdot 1}}{e^2 - 1} = \frac{2 \cdot e^2}{e^2 - 1} = \frac{2 \times 7.389}{6.389} = \frac{14.778}{6.389} \approx 2.313 \]

Ratio: $\phi(1)/\phi(0) = e^2 \approx 7.389$. The weight at the worst quantile ($u=1$) is about 7.4 times the weight at the best quantile ($u=0$).

Verification of normalization:

\[ \int_0^1 \phi(u)\,du = \frac{\gamma}{e^\gamma - 1}\int_0^1 e^{\gamma u}\,du = \frac{\gamma}{e^\gamma - 1} \cdot \frac{e^\gamma - 1}{\gamma} = 1 \checkmark \]

Effect of increasing $\gamma$:

As $\gamma$ increases:

$\phi(0) = \frac{\gamma}{e^\gamma - 1} \to 0$ (the weight on the best outcomes vanishes)
$\phi(1) = \frac{\gamma e^\gamma}{e^\gamma - 1} \to \gamma$ (the weight on the worst outcomes grows)
The ratio $\phi(1)/\phi(0) = e^\gamma$ grows exponentially

This means increasing $\gamma$ concentrates more weight on the extreme tail of the loss distribution, making the risk measure more conservative. In the limit:

As $\gamma \to 0$: $\phi(u) \to 1$ uniformly, giving the mean $\mathbb{E}[X]$ (risk-neutral).
As $\gamma \to \infty$: the weight concentrates near $u = 1$, approaching the worst-case loss (maximum loss).

Exercise 4. VaR at level $\alpha$ can be written as $\text{VaR}_\alpha(X) = \int_0^1 \text{VaR}_u(X)\,\delta(u - \alpha)\,du$, where $\delta$ is the Dirac delta. Explain why this weight function is not non-decreasing and hence VaR is not a spectral risk measure. What property of coherence does VaR consequently lack?

Solution to Exercise 4

VaR as a spectral-like integral:

Formally, $\text{VaR}_\alpha(X)$ can be written as:

\[ \text{VaR}_\alpha(X) = \int_0^1 \text{VaR}_u(X)\,\delta(u - \alpha)\,du \]

where $\delta(u - \alpha)$ is the Dirac delta function centered at $\alpha$.

Why this "weight function" fails the non-decreasing condition:

The Dirac delta $\delta(u - \alpha)$ is not a non-decreasing function. In fact, it is not even a proper function -- it is a distribution (generalized function) that is zero everywhere except at $u = \alpha$ where it is "infinite." Considering any smooth approximation (e.g., a narrow bump function centered at $\alpha$):

The approximation increases as $u$ approaches $\alpha$ from below
Then decreases as $u$ moves past $\alpha$

This violating the non-decreasing requirement is fundamental: the weight function rises to a peak at $\alpha$ and then drops back to zero. It puts zero weight on outcomes worse than the $\alpha$-quantile and zero weight on outcomes better than the $\alpha$-quantile, concentrating all weight on a single point.

Consequence for coherence:

Since VaR's weight function is not non-decreasing, VaR is not a spectral risk measure. By Acerbi's theorem, only spectral risk measures with non-decreasing weight functions are coherent. Therefore, VaR lacks subadditivity.

Intuition: VaR ignores the severity of losses beyond the $\alpha$-quantile. By putting zero weight on outcomes worse than $\text{VaR}_\alpha$, it throws away information about tail severity. A spectral measure with a proper non-decreasing $\phi$ must give these extreme outcomes at least as much weight as the $\alpha$-quantile, ensuring that tail risk is captured and subadditivity is preserved.

Exercise 5. For a portfolio loss $X \sim N(0, \sigma^2)$, compute the spectral risk measure with the exponential weight function from Exercise 3. Express the result as a function of $\sigma$ and $\gamma$. How does it compare to $\text{ES}_{0.95}$?

Solution to Exercise 5

For $X \sim N(0, \sigma^2)$, the quantile function is $\text{VaR}_u(X) = \sigma \Phi^{-1}(u)$.

Spectral risk measure with exponential weights:

\[ \rho_\phi(X) = \int_0^1 \phi(u) \,\text{VaR}_u(X)\,du = \int_0^1 \frac{\gamma e^{\gamma u}}{e^\gamma - 1} \cdot \sigma \Phi^{-1}(u)\,du \]

\[ = \frac{\gamma \sigma}{e^\gamma - 1} \int_0^1 e^{\gamma u} \Phi^{-1}(u)\,du \]

Evaluating the integral: Let $z = \Phi^{-1}(u)$, so $u = \Phi(z)$ and $du = \phi(z)\,dz$:

\[ \int_0^1 e^{\gamma u} \Phi^{-1}(u)\,du = \int_{-\infty}^{\infty} z \, e^{\gamma \Phi(z)} \phi(z)\,dz \]

This integral does not have a simple closed form in general. However, we can use the moment generating function approach. For the normal distribution, we can use integration by parts or the identity:

\[ \int_0^1 e^{\gamma u} \Phi^{-1}(u)\,du = \frac{e^\gamma}{\gamma}\left[1 - e^{-\gamma}\right] \cdot m(\gamma) \]

A more direct approach uses the result that for $X \sim N(0, \sigma^2)$:

\[ \rho_\phi(X) = \sigma \cdot c(\gamma) \]

where $c(\gamma) = \frac{\gamma}{e^\gamma - 1}\int_0^1 e^{\gamma u}\Phi^{-1}(u)\,du$ is a constant depending only on $\gamma$. Numerical evaluation gives:

For $\gamma = 1$: $c(1) \approx 0.725$
For $\gamma = 2$: $c(2) \approx 1.166$
For $\gamma = 5$: $c(5) \approx 1.868$
For $\gamma = 10$: $c(10) \approx 2.197$

Result:

\[ \rho_\phi(X) = \sigma \cdot c(\gamma) \]

Comparison with $\text{ES}_{0.95}$:

For the normal distribution:

\[ \text{ES}_{0.95} = \sigma \cdot \frac{\phi(1.645)}{0.05} = \sigma \cdot \frac{0.1031}{0.05} = 2.063\sigma \]

For moderate $\gamma$ (e.g., $\gamma = 2$), $\rho_\phi \approx 1.166\sigma < 2.063\sigma = \text{ES}_{0.95}$. The exponential spectral measure is less conservative because it distributes weight across all quantiles, not just the top 5%.
For large $\gamma$ (e.g., $\gamma \ge 8$), $\rho_\phi$ approaches and may exceed $\text{ES}_{0.95}$ as the exponential weight increasingly concentrates on the extreme tail.
The exponential spectral measure is a smooth risk measure that transitions gradually from low to high tail weighting, whereas $\text{ES}_{0.95}$ has a sharp cutoff at the 95th percentile.

Exercise 6. Discuss the practical challenges of implementing spectral risk measures in a trading desk risk system. How would you estimate a spectral risk measure from historical P&L data? What are the advantages over simply using ES at a fixed confidence level?

Solution to Exercise 6

Practical challenges of implementing spectral risk measures:

Choice of weight function: The spectrum $\phi$ is not uniquely determined by the data. Unlike VaR (which requires only a confidence level) or ES (confidence level), a spectral measure requires selecting an entire function. This introduces a subjective element:
- What functional form should $\phi$ take (exponential, power, piecewise linear)?
- How should the risk aversion parameter(s) be calibrated?
- Different desks or business units may have different risk preferences.
Quantile estimation in the tails: Spectral measures require estimating the full quantile function $q_\alpha(X)$ for $\alpha$ across $[0,1]$. Tail quantiles are notoriously difficult to estimate accurately from limited data. With 500 daily observations, the 99.5th percentile estimate relies on only 2--3 data points.
Computational cost: Unlike VaR (a single quantile) or ES (average of a few tail observations), spectral measures require numerical integration over the entire quantile function weighted by $\phi$. For large portfolios with Monte Carlo pricing, this adds computational burden.
Communication and governance: VaR and ES are widely understood by traders, senior management, and regulators. A spectral measure with an exponential or power-law weight function is harder to explain and may face resistance in governance processes.
Regulatory acceptance: Basel III mandates ES, not general spectral measures. Using a spectral measure internally requires maintaining ES for regulatory reporting while potentially using a different measure for internal risk management.

Estimating a spectral risk measure from historical P&L data:

Given $n$ historical losses $L_1, \ldots, L_n$:

Sort the losses: $L_{(1)} \le L_{(2)} \le \cdots \le L_{(n)}$
Assign quantile levels: $u_i = (i - 0.5)/n$ for $i = 1, \ldots, n$ (midpoint convention)
Compute the discrete approximation:

\[ \hat{\rho}_\phi = \frac{1}{n}\sum_{i=1}^{n} \phi(u_i) \, L_{(i)} \]

This is a Riemann sum approximation to the integral $\int_0^1 \phi(u)\,q_u(X)\,du$.

Alternatively, using trapezoidal or Simpson's rule for better accuracy, or kernel-smoothed quantile estimates for smoother integration.

Advantages over ES at a fixed confidence level:

Flexible risk aversion: A spectral measure allows the institution to express its specific risk preferences through the shape of $\phi$. A bank with higher aversion to extreme tail events can use a steeply increasing $\phi$, while one primarily concerned with moderate losses can use a more gradual weight function.
Smooth tail weighting: ES has a discontinuity in its weight function at $\alpha$ (jumping from 0 to $\frac{1}{1-\alpha}$). Spectral measures with smooth $\phi$ (like the exponential) produce risk measures that vary continuously with changes in the loss distribution, avoiding the sensitivity to the exact choice of $\alpha$.
Information about the full tail: ES treats all tail losses above $\text{VaR}_\alpha$ equally. A spectral measure can assign progressively more weight to more extreme losses, better reflecting the economic reality that a $1B loss is not just twice as bad as a $500M loss.
Unified framework: Different confidence-level ES measures (ES$_{0.95}$, ES$_{0.975}$, ES$_{0.99}$) are all special cases of spectral measures. A single spectral measure can simultaneously capture information that would otherwise require reporting multiple ES levels.

Measure	Weight Function	Interpretation
Value-at-Risk	\(\delta(\alpha - 0.95)\)	Single quantile
Expected Shortfall	\(\phi(\alpha) = \frac{1}{1-\alpha_0} \mathbf{1}_{\alpha \geq \alpha_0}\)	Average of worst 5%
AveVar	\(\phi(\alpha) = \frac{2}{1-\alpha_0}\mathbf{1}_{\alpha \geq \alpha_0}\)	Linear increasing weights
Omega ratio	Specific choice for return/risk tradeoff	Gain/loss focus