Tail Conditional Expectation

• Tail Conditional Expectation is a coherent risk measure that averages losses beyond a specified quantile, clearly defining tail risk.
• It leverages integral and conditional expectation methods, ensuring properties like monotonicity, subadditivity, and positive homogeneity for robust risk evaluation.
• Extensions to multivariate, copula-based, and covariate-conditioned frameworks enable practical applications in risk capital allocation, insurance, and reinsurance.

Tail Conditional Expectation (TCE), also widely known as Conditional Tail Expectation (CTE) or Expected Shortfall in risk management literature, is a fundamental risk measure in probability, actuarial science, quantitative finance, and statistical theory of extremes. TCE quantifies the average loss or value that exceeds a specified quantile (threshold) of a loss distribution, capturing tail risk beyond classical Value-at-Risk (VaR) and exhibiting coherence properties essential for robust risk assessment.

1. Definition and Basic Properties

For a real-valued random variable $X$ with cumulative distribution function $F$ and a threshold $t$ or confidence level $\alpha$, the Tail Conditional Expectation at level $t$ or $\alpha$ is

$\mathrm{TCE}_t(X) = \mathbb{E}[X \mid X \geq t]$

or for continuous $F$ and $\alpha \in (0,1)$,

$$\mathrm{TCE}_\alpha(X) = \mathbb{E}[X \mid X > \mathrm{VaR}_\alpha(X)] = \frac{1}{1-\alpha}\int_\alpha^1 \mathrm{VaR}_u(X)\,du$$

where $$\mathrm{VaR}_\alpha(X) = \inf\{x : F(x) \geq \alpha\}$$ denotes the Value-at-Risk at level $\alpha$ (Laidi et al., 2020, Brahimi, 2012, Mohammed et al., 2021, Najafabadi et al., 2017). For discrete distributions, the corresponding sum-formula applies.

TCE admits several equivalent representations, including integral, conditional expectation, and infimum forms strictly within the support beyond the VaR threshold (Laidi et al., 2020, Mohammed et al., 2021).

Coherence

TCE is a coherent risk measure: it is monotone, translation-invariant, subadditive, and positively homogeneous. Formally, for fixed $\alpha$, these properties are:

• Monotonicity: $X \leq Y$ (a.s.) $\Rightarrow$ $\mathrm{TCE}_\alpha(X) \leq \mathrm{TCE}_\alpha(Y)$
• Translation invariance: $$\mathrm{TCE}_\alpha(X + c) = \mathrm{TCE}_\alpha(X) + c$$
• Positive homogeneity: $$\mathrm{TCE}_\alpha(\lambda X) = \lambda \mathrm{TCE}_\alpha(X),\,\forall \lambda > 0$$
• Subadditivity: $$\mathrm{TCE}_\alpha(X+Y) \leq \mathrm{TCE}_\alpha(X) + \mathrm{TCE}_\alpha(Y)$$ (Mohammed et al., 2021, Brahimi, 2012)

2. Explicit Formulas and Special Cases

TCE can be computed exactly in many parametric cases. For location-scale mixture of elliptical distributions $Y = \mu + \Theta \beta + \Theta \sigma X$ with independent mixing variable $\Theta$ and $X$ standard elliptical, TCE takes the explicit form: $$\mathrm{TCE}_\alpha(Y) = \mu + \mathbb{E}[\Theta]\,\beta + \sigma\,\mathbb{E}[\Theta] \frac{G_1(z_\alpha^2)}{\bar F_Z(z_\alpha)}$$ where $z_\alpha$ standardizes to the base distribution, $G_1$ is a tail generator, and $\bar F_Z$ denotes tail probability (Zuo et al., 2020). Closed-form results for the Normal, Student-$t$, logistic, and Laplace cases follow as specializations.

For binomial and Poisson random variables, sharp upper bounds for TCE are available:

• If $X \sim \mathrm{Bin}(n,p)$ with $np < k \leq n$,

$$\mathbb{E}[X \mid X \geq k] \leq k + \frac{(n-k)p}{k-np+p}$$

• If $Y \sim \mathrm{Poi}(\mu)$, $k \geq \mu$,

$$\mathbb{E}[Y \mid Y \geq k] \leq k + \frac{\mu}{k+1-\mu}$$

These bounds provide input for sharp non-asymptotic lower tail estimates (Pelekis, 2016).

3. Extensions: Multivariate and Copula-Based TCE

Tail Conditional Expectation extends naturally to multivariate and conditional settings.

Multivariate TCE (MTCE)

For a random vector $\mathbf{X}$ in $\mathbb{R}^d$ and marginal VaRs at level vector $\mathbf{q}$,

$$\operatorname{MTCE}_{\mathbf{q}}(\mathbf{X}) = \mathbb{E}[\mathbf{X} \mid X_i > \mathrm{VaR}_{q_i}(X_i) \;\forall i]$$

where explicit formulas can be given for generalized skew-elliptical laws (Zuo et al., 2022, Zuo et al., 2020).

Copula Conditional Tail Expectation (CCTE)

For a bivariate loss vector $(X_1, X_2)$, CCTE captures the expected loss of $X_1$ conditional on both $X_1$ and $X_2$ exceeding respective quantiles. Using copula $C$,

$$\mathrm{CCTE}_{X_1}(s;t) = \frac{1}{\widehat C(1-s,1-t)} \int_s^1 F_1^{-1}(u)[1 - C_u(u, t)] du$$

where $\widehat C$ is the survival copula, and $C_u(u, v) = \partial C(u, v)/\partial u$ (Brahimi, 2012). When $X_1, X_2$ are independent, this reduces to the univariate CTE; under positive quadrant dependence, the CCTE dominates the univariate CTE, highlighting the influence of tail dependence.

Covariate Conditional Tail Expectation (CCTE, Depth-Based)

For a pair $(Y, \mathbf{X})$, with a statistical depth function $D$ over $\mathbf{X}$, the CCTE at level $\alpha$ is

$$\mathrm{CCTE}_{D,\alpha}(Y \mid \mathbf{X}) = \mathbb{E}[Y \mid \mathbf{X} \in \{x : D(x, P_{\mathbf{X}}) \leq \alpha\}]$$

with plug-in estimators constructed from empirical depth lower-level sets (Elisabeth et al., 2021).

4. Statistical Estimation for Heavy-Tailed and Dependent Risks

For heavy-tailed loss distributions with regularly varying right tails (index $\alpha > 1$), standard quantile-based estimators of CTE suffer from bias. Improved estimation utilizes bias-reduced high quantile estimators, such as the Li–Peng–Yang approach, integrating second-order regular variation information for substantially reduced bias and mean squared error in the extreme tail (Laidi et al., 2020).

In dependent risk settings, explicit formulas for TCE of maxima, minima, and sums under parametric copula models (e.g., FGM, Archimedean) and under various marginal laws (exponential, Pareto) demonstrate how dependence inflates or attenuates TCE, especially in aggregate and minimum risks (Thapa et al., 2021).

Table: Key Estimation Features for Heavy-Tailed CTE

Method	Assumptions	Main Benefit
Classical (Weissman)	First-order RV	Simplicity, poor bias
Li–Peng–Yang, Hall Model	Second-order RV	Dramatically reduced bias

Here, RV denotes regular variation.

5. Applications in Risk Management and Insurance

TCE is widely used for regulatory capital computation and allocation, insurance product pricing, and optimal reinsurance design.

Reinsurance

Optimal stop-loss and multi-layer reinsurance contracts under the CTE criterion minimize the insurer's tail risk. Multi-layer contracts retain the same CTE as the classical stop-loss but split the ceded risk into multiple layers, optimizing secondary criteria such as variance trade-off, expected ceded loss, or utility-based objectives. The space of CTE-optimal reinsurance treaties is thus highly non-unique, requiring further criteria for practical specification (Najafabadi et al., 2017).

Risk Capital Allocation

CTE-based risk capital allocation (RCA) assigns capital to business units according to their expected losses in extreme aggregate loss events. Mathematically, this allocation is $A_i = \mathbb{E}[X_i \mid S_X > s_\alpha]$, with $s_\alpha = \mathrm{VaR}_\alpha(S_X)$ and $S_X$ the portfolio loss (Mohammed et al., 2021). However, these allocations align with profit-maximizing allocations (i.e., $A_i$ proportional to $\mathbb{E}[X_i/S_X | S_X > s_\alpha]$) only under restrictive conditions—such as identical partial size-biased distributions or independent gamma marginals with common scale. For most dependence structures, regulatory CTE allocations diverge from economic allocations.

6. Theoretical Significance and Tail Probability Bounds

TCE plays a critical role in deriving non-asymptotic lower bounds for tail probabilities of discrete laws. Upper bounds on TCE for binomial and Poisson random variables, combined with recursive identities linking TCE and tail probabilities, yield constructive lower bounds for right tails. These results can, in many parameter regimes, outperform classical results such as Chernoff–Hoeffding or Ash's bounds, especially for moderate $n, k$ (Pelekis, 2016). This approach is essential in reliability, rare-event estimation, and robust risk quantification.

7. Numerical and Empirical Aspects

Simulation studies across parametric and heavy-tailed models validate both the accuracy and efficiency of bias-reduced CTE estimators over classical estimators. For depth-based CCTE, empirical convergence rates scale as $n^{-1/2}$ under standard moment and smoothness conditions, with error controlled by the accuracy in estimating underlying depth level sets (Laidi et al., 2020, Elisabeth et al., 2021). In copula-based settings, empirical analyses demonstrate the sensitivity of tail risk measures to dependence structure, copula family, and threshold selection (Brahimi, 2012, Thapa et al., 2021).

---

In summary, Tail Conditional Expectation provides a robust, coherent, and statistically rich framework for tail risk measurement, with extensive developments in estimation, theoretical bounds, and practical applications in finance, insurance, and statistics. Its generalizations to multivariate, covariate-conditioned, and dependent frameworks ensure relevance in modern risk aggregation and allocation contexts.