Composite Risk Measures

• Composite risk measures are methods that combine various risk quantifiers (e.g., VaR, ES, ML) to capture complex tail behaviors and provide a comprehensive risk profile.
• They employ methodologies such as vector-valued aggregation, scalar functions, and nested approaches to improve sensitivity, robustness, and informational content in risk assessments.
• They facilitate more robust regulatory reporting and decision-making by incorporating deviation measures and addressing the limitations of single risk metrics.

A composite risk measure is a functional or procedure designed to aggregate multiple aspects of uncertainty, loss, or variability into a single or vector-valued summary, blending information from diverse risk quantifiers. Such measures fundamentally extend classical risk quantification (e.g., Value-at-Risk, Expected Shortfall) by combining several risk statistics—either in parallel, through aggregation, or sequentially, via nested or coupled operations—to achieve improved sensitivity, robustness, or information content, particularly for tail-risk modeling and regulatory reporting. Composite risk frameworks enable regulators, insurers, and asset managers to capture and communicate complex tail behaviors that would remain undetected under any single classical risk measure.

1. Formal Definitions and Main Classes

Composite risk measures are constructed by systematically combining basic risk measures or functionals. Core classes include:

• Vector-valued composites: Collect multiple risk measures (e.g., $\VaR_{0.95}$, $\ES_{0.99}$, ML) as a vector to form an information-rich fingerprint of a risk profile (Gu/'egan et al., 2011).
• Scalar-aggregated composites: Apply an aggregation function (e.g., weighted sum, supremum, or a nonlinear coupling) to the individual components of a vector (e.g., $f(\VaR_{0.95}, \ES_{0.99}, \ML)$) (Righi, 2015, Alexander et al., 2024).
• Compositions of risk and deviation measures: Add a deviation functional to a risk measure, e.g., $\rho(X) + \beta D(X)$, with suitable axiomatic control (e.g., Limitedness) to ensure coherence (Righi, 2015, Han et al., 2023).
• Coupled risk measures: Construct as $H(L_1(F), L_2(F))$, for $H$ bivariate and $L_i$ (possibly different) classical risk measures, typical forms include ratios, convex combinations, etc. (Necir et al., 2011).
• Nested or composite functionals: Compose risk functionals in a layered manner, for example, by calculating a risk measure on an outer distribution applied to the statistic of an inner risk measure (e.g., $\mu(\rho_F(H(x,\xi)))$) (Qian et al., 2015, Dentcheva et al., 2015).

Key requirements for the building blocks and their aggregation include monotonicity, translation invariance, convexity, and, in the coherent case, subadditivity and positive homogeneity.

2. Composite Risk Measures as Information Constraints

Every risk measure reported by a financial institution acts as an information constraint on the (unknown) loss distribution. A collection of risk measures $R = (\VaR_{0.95}, \VaR_{0.99}, \ES_{0.95}, \ES_{0.99}, \ML)$ restricts the admissible set of distributions by enforcing pointwise and integral conditions (Gu/'egan et al., 2011). Formally, the intersection

$$\mathcal{F}_\mathcal{R} = \{ F : \rho_i(F) = r_i, \; \forall i \}$$

becomes the information set. The more constraints, the less ambiguity remains.

However, even with five canonical measures (e.g., $\ES_{0.99}$0, $\ES_{0.99}$1, $\ES_{0.99}$2, $\ES_{0.99}$3, ML), the tail distribution is still underdetermined in the most general setting, but ambiguity is drastically reduced in practice (Gu/'egan et al., 2011, Guégan et al., 2011).

3. Construction Principles and Mathematical Properties

Risk-Deviation Sums and the Limitedness Axiom

The sum of a risk measure and a deviation measure, $\ES_{0.99}$4, produces a coherent risk measure if and only if the Limitedness axiom is enforced:

$\ES_{0.99}$5

This ensures the composite functional does not exceed the worst-case loss. Fatou-continuity and law-invariance yield representation results, including dual forms and Kusuoka-type integral representations (Righi, 2015).

Examples include:

• Mean + semi-deviation (coherent)
• Loss-deviation generated by $\ES_{0.99}$6 (coherent when $\ES_{0.99}$7)
• Mean + standard deviation (not coherent, fails Limitedness)

Mean-Deviation and Monotonic Models

Monotonic mean–deviation risk measures are of the form

$\ES_{0.99}$8

and are monetary and SSD-consistent. Convexity is equivalent to $\ES_{0.99}$9 being convex, and coherence to $f(\VaR_{0.95}, \ES_{0.99}, \ML)$0 linear (Han et al., 2023).

Adjusted and Aggregated Profiles

Adjusted risk measures generalize the adjusted Expected Shortfall by defining, for a target risk profile $f(\VaR_{0.95}, \ES_{0.99}, \ML)$1 and a family $f(\VaR_{0.95}, \ES_{0.99}, \ML)$2,

$f(\VaR_{0.95}, \ES_{0.99}, \ML)$3

Coherence is characterized by the jump structure of $f(\VaR_{0.95}, \ES_{0.99}, \ML)$4; only one nonzero jump is allowed for positive homogeneity (Alexander et al., 2024).

Nested/CRM Functionals

The CRM (Composite Risk Measure) framework formalizes the nested application of risk measures for decision making:

$f(\VaR_{0.95}, \ES_{0.99}, \ML)$5

where the inner measure $f(\VaR_{0.95}, \ES_{0.99}, \ML)$6 accounts for risk under a fixed parameter set, and the outer measure $f(\VaR_{0.95}, \ES_{0.99}, \ML)$7 incorporates uncertainty in the parameters themselves. This framework includes stochastic programming, robust optimization, and DRO as special cases (Qian et al., 2015).

Law-Invariant and Quasi-Convex Aggregations

Risk measures constructed via acceptance sets parameterized by probability–loss functions (e.g., $f(\VaR_{0.95}, \ES_{0.99}, \ML)$8 controlling risk aversion to size) furnish composite, law-invariant, and quasi-convex risk measures. This generalizes $f(\VaR_{0.95}, \ES_{0.99}, \ML)$9 to balance loss probabilities and sizes according to $\rho(X) + \beta D(X)$0 (Frittelli et al., 2012).

4. Special Classes and Examples

Composite Type	Construction	Key Reference
Vector-valued	$\rho(X) + \beta D(X)$1	(Gu/'egan et al., 2011, Guégan et al., 2011)
Summed risk–deviation	$\rho(X) + \beta D(X)$2	(Righi, 2015, Han et al., 2023)
Coupled functionals	$\rho(X) + \beta D(X)$3	(Necir et al., 2011)
Adjusted-family	$\rho(X) + \beta D(X)$4	(Alexander et al., 2024)
Nested/composite	$\rho(X) + \beta D(X)$5	(Qian et al., 2015, Dentcheva et al., 2015)
Multivariate OCE	Min over allocations $\rho(X) + \beta D(X)$6	(Kaakai et al., 2022)
Fréchet-based	Barycentric aggregation	(Papayiannis et al., 2022)

Special instances:

• Adjusted ES, SCRM, CRM, and AERM provide substantial flexibility for regulatory and practical capital requirement settings (Alexander et al., 2024).
• Coupled risks such as ratios of L-functionals accommodate heavy-tail settings and are empirically estimable, with explicit asymptotic theory for high quantiles (Necir et al., 2011).

5. Statistical Estimation and Asymptotics

Composite risk measures, whether of the nested expectation type or coupled L-functional type, admit plug-in estimators and central limit theory under suitable regularity:

• For composite functionals $\rho(X) + \beta D(X)$7, the plug-in estimator is

$\rho(X) + \beta D(X)$8

and $\rho(X) + \beta D(X)$9 converges to a normal law with explicit variance (Dentcheva et al., 2015).

• For coupled functionals in heavy-tail regimes, mixed scaling (e.g., $H(L_1(F), L_2(F))$0) and bias–variance trade-offs are required, employing tail index estimators (Hill, Weissman) and Brownian bridge approximations (Necir et al., 2011).

Empirical findings suggest composite/adjusted measures maintain robust performance across volatility regimes when calibrated with appropriate benchmarks and improve tail sensitivity relative to singly parameterized risk measures (Alexander et al., 2024).

6. Implementation and Optimization

Composite measures remain tractable under convexity assumptions:

• Convexity of $H(L_1(F), L_2(F))$1 and $H(L_1(F), L_2(F))$2 allows convex programming techniques for optimization (Qian et al., 2015, Han et al., 2023).
• Stochastic approximation algorithms enable scalable estimation of multivariate OCEs with provable convergence and central limit theorems for error quantification (Kaakai et al., 2022).
• Sample-average approximation (SAA) techniques are standard for empirical estimation, with dimension and sample size scaling empirically documented in portfolio contexts (Qian et al., 2015).

Dual representations facilitate robust scenario analysis, capital allocation, and sensitivity calculations via adversarial distributions or stress scenarios (Alexander et al., 2024, Kaakai et al., 2022).

7. Regulatory and Practical Implications

Composite risk measures address well-documented limitations of classical risk measures in regulatory, insurance, and asset management contexts:

• Single measures (e.g., $H(L_1(F), L_2(F))$3 alone) fail to identify pathological or “mosaic” tail phenomena; pairs or triples are also insufficient for full characterization (Gu/'egan et al., 2011, Guégan et al., 2011).
• The combination of multiple measures (five or more) is strongly advocated for regulatory reporting, with practical recommendation for $H(L_1(F), L_2(F))$4, $H(L_1(F), L_2(F))$5, $H(L_1(F), L_2(F))$6, $H(L_1(F), L_2(F))$7, and ML (Gu/'egan et al., 2011, Guégan et al., 2011).
• Flexible adjustment via target profiles (as in adjusted ES or SCRM) allows tailoring capital requirements to specific loss environments and regime shifts (Alexander et al., 2024).
• In insurance, composite (Fréchet-type) risk measures robustly aggregate model uncertainty and expert heterogeneity, enabling stable allocation and pricing (Papayiannis et al., 2022).
• In portfolio optimization, mean–deviation and CRM frameworks offer less conservative, decision-adaptive distribution sets, improving out-of-sample performance (Qian et al., 2015, Han et al., 2023).

Composite risk measures thus furnish a scientific, mathematically justified path toward more robust and informative risk assessment and management, especially when facing model uncertainty, regime changes, or regulatory scrutiny.