layout: default title: Coherent Risk Measures and Average Value at Risk ——————————————————-
Coherent Risk Measures and Average Value at Risk
Sukrit Mittal Franklin Templeton Investments
Outline
- Why coherence?
- Axioms of coherent risk measures
- Quantiles revisited
- Average Value at Risk (AVaR)
- Representation of AVaR
- AVaR vs VaR
- AVaR in the Black–Scholes model
- Coherence proofs
- Exercises
1. Why Coherence?
VaR answered a practical question.
It failed a theoretical one.
Specifically:
Diversification should not increase risk.
Coherent risk measures enforce this principle.
From Heuristics to Axioms
Risk is not just a number.
It is a functional:
\[\rho : L \to \mathbb{R}\]mapping random losses to capital requirements.
Axioms discipline this mapping.
2. Axioms of Coherent Risk Measures
A risk measure $\rho(L)$ is coherent if it satisfies:
-
Monotonicity If $L_1 \le L_2$, then $\rho(L_1) \le \rho(L_2)$
-
Translation invariance $\rho(L + c) = \rho(L) + c$
-
Positive homogeneity $\rho(\lambda L) = \lambda \rho(L)$ for $\lambda > 0$
-
Subadditivity $\rho(L_1 + L_2) \le \rho(L_1) + \rho(L_2)$
These axioms formalize diversification.
Economic Meaning
- Monotonicity: worse losses imply more risk
- Translation: cash reduces risk one-for-one
- Homogeneity: scaling portfolios scales risk
- Subadditivity: diversification helps
If one axiom fails, interpretation collapses.
3. Quantiles Revisited
Let $L$ denote loss.
Define VaR again:
\[\text{VaR}_\alpha(L) = \inf {x : \mathbb{P}(L \le x) \ge \alpha }\]VaR is a quantile.
Quantiles are not averages.
4. Average Value at Risk (AVaR)
The Average Value at Risk at level $\alpha$ is defined as:
\[\text{AVaR}*\alpha(L) = \frac{1}{1-\alpha} \int*\alpha^1 \text{VaR}_u(L) , du\]AVaR averages the worst losses.
This single change fixes VaR.
Interpretation
AVaR answers:
Given that things are bad, how bad are they on average?
It measures tail severity.
This is what risk managers actually want.
5. Representation of AVaR
An equivalent formulation:
\[\text{AVaR}*\alpha(L) = \inf*{m \in \mathbb{R}} \left{ m + \frac{1}{1-\alpha} \mathbb{E}[(L - m)^+] \right}\]This is a convex optimization problem.
This representation is fundamental.
Why This Representation Matters
- Convexity becomes explicit
- Numerical computation becomes tractable
- Dual representations emerge naturally
VaR has none of these properties.
6. AVaR vs VaR
Key contrasts:
- VaR: quantile, ignores tail shape
- AVaR: tail average, tail-sensitive
VaR may penalize diversification.
AVaR never does.
This difference is structural.
7. AVaR in the Black–Scholes Model
Assume log-returns are normal.
Loss $L$ is log-normal.
AVaR admits closed-form expressions.
The math is heavier—but honest.
Normal Loss Approximation
If losses are normal:
\[L \sim \mathcal{N}(\mu_L, \sigma_L^2)\]Then:
\[\text{AVaR}*\alpha = \mu_L + \sigma_L \frac{\varphi(z*\alpha)}{1-\alpha}\]where $\varphi$ is the standard normal density.
Interpretation
Compared to VaR:
- Same quantile cutoff
- Larger risk value
AVaR prices tail thickness.
Markets ignore it at their peril.
8. Coherence of AVaR
Theorem
AVaR is a coherent risk measure.
Proof outline:
- Monotonicity: inherited from expectation
- Translation invariance: linearity
- Positive homogeneity: scaling inside expectation
- Subadditivity: convexity of $(x)^+$ and Jensen’s inequality
Each axiom can be proven rigorously.
Why VaR Fails Subadditivity
VaR focuses on a single quantile.
Combining distributions can move mass.
Averages smooth.
Quantiles jump.
This is not a technicality.
9. Exercises
Exercise 1
Show that AVaR is monotone and translation invariant.
Exercise 2
Derive the infimum representation of AVaR.
Explain why convexity matters.
Exercise 3
Compare VaR and AVaR numerically for a heavy-tailed distribution.
Discuss implications for regulation.
Final Takeaways
- Risk measures must respect diversification
- VaR fails coherence
- AVaR repairs this failure
- Coherence is not philosophy—it is structure
Next: optimization under coherent risk.
Where finance meets convex analysis.