TL;DR
This paper introduces "Bregman superquantiles," a novel risk measure that extends the concept of superquantiles by incorporating Bregman divergences. This new measure offers improved robustness and statistical properties, particularly for heavy-tailed distributions, and the authors provide methods for its estimation and analyze its asymptotic behavior.
What Problem Does This Paper Address?
Traditional risk measures like quantiles have limitations: they are not subadditive (meaning the risk of a sum of assets can be greater than the sum of individual risks) and don't fully capture tail risk. Superquantiles address some of these issues by being subadditive and providing more information about the tail of a distribution. However, classical superquantiles can still suffer from poor statistical properties, especially with heavy-tailed distributions. This paper aims to develop a more flexible and robust risk measure by generalizing the superquantile concept using Bregman divergences, which allow for a broader range of distance metrics beyond the standard Euclidean one.
Methods
The core contribution is the definition of the Bregman superquantile. This is achieved by:
- Bregman Divergence: Utilizing a Bregman divergence, $d_\gamma(x, x') = \gamma(x) - \gamma(x') - \gamma'(x')(x - x')$, which measures dissimilarity between real numbers based on a strictly convex function $\gamma$.
- Bregman Mean: Defining a Bregman mean, which is analogous to the classical expectation but uses the Bregman divergence for minimization.
- Bregman Superquantile Definition: Defining the Bregman superquantile $Q_\alpha^\gamma$ as a transformation of the Bregman mean of $\gamma'(X)$ conditional on $X$ being above its $\alpha$-quantile. This can be interpreted as changing the scale of the random variable using $\gamma'$, calculating the mean in that scaled space, and then transforming back.
- Coherence Axioms: Investigating whether Bregman superquantiles satisfy the axioms of a coherent risk measure (constant invariance, positive homogeneity, subadditivity, monotonicity, and closeness).
- Monte Carlo Estimation: Developing a plug-in estimator for the Bregman superquantile based on sample data.
- Asymptotic Analysis: Studying the consistency and asymptotic normality of the proposed estimator under various regularity conditions on $\gamma$ and the distribution's quantile function.
Datasets
Not discussed in the paper. The paper uses simulated data for numerical experiments and a real-world dataset for an application example.
Key Insights and Ideas
- Flexibility through $\gamma$: The choice of the strictly convex function $\gamma$ in the Bregman divergence allows for tailoring the risk measure to specific problems. Different $\gamma$ functions lead to different "distances" and thus different risk sensitivities.
- Improved Robustness: Bregman superquantiles can offer better statistical properties than classical superquantiles, especially for heavy-tailed distributions. For instance, they can remain finite and consistent even when the classical superquantile is not.
- Scale Transformation Interpretation: The Bregman superquantile can be intuitively understood as a risk measure obtained by transforming the data's scale using $\gamma'$, calculating a mean in that transformed space, and then transforming back. This provides a new perspective on risk aggregation.
- Coherence is Achievable: Under certain conditions on $\gamma$ (e.g., concavity and sub-additivity of $\gamma'$), Bregman superquantiles can satisfy the desirable axioms of coherent risk measures.
- Practical Estimation: The paper provides a concrete Monte Carlo estimator and analyzes its asymptotic behavior, making the theoretical concept practically applicable.
Key Results
- The paper defines the Bregman superquantile and demonstrates that it can be a coherent measure of risk under specific conditions on the underlying convex function $\gamma$.
- A Monte Carlo estimator for the Bregman superquantile is proposed and shown to be consistent and asymptotically normal under certain regularity assumptions.
- Numerical simulations on exponential and Pareto distributions show that Bregman superquantiles (especially geometric and harmonic ones) exhibit better statistical properties (consistency and asymptotic normality) than classical superquantiles, particularly for heavy-tailed Pareto distributions.
- An application to real data from a nuclear safety exercise illustrates the practical relevance of these measures.
Limitations and Future Work
- Assumptions on $\gamma$: The coherence and asymptotic properties of the Bregman superquantile depend on specific assumptions about the convexity and differentiability of $\gamma$.
- Choice of $\gamma$: While the flexibility of choosing $\gamma$ is an advantage, selecting the "optimal" $\gamma$ for a given problem might require further investigation.
- Future Work: The authors suggest exploring the link between Bregman superquantiles and capital allocation in economic theory, as well as applying these concepts to other areas like physical component reliability assessment. They also mention the need for more efficient estimators beyond Monte Carlo methods for computationally intensive problems.
Summary updated