Marginal Risks in Quantitative Analysis

• Marginal risks are quantitative measures defining risk for individual components in multivariate and time-dependent settings, exemplified by MES.
• Estimation employs threshold exceedance and bootstrap methods to ensure consistency and robust inference under heavy-tailed conditions.
• Applications span financial risk management, survival analysis, causal inference, and AI systems, offering actionable insights for model selection and policy design.

Marginal risks encompass a class of risk quantifications, statistical measures, and inferential procedures where risk is quantified for individual components, outcomes, or exposures—often in multivariate, time-dependent, or competing scenarios—either as direct quantities or as decompositions from joint models. In financial mathematics, survival analysis, systemic risk measurement, and AI risk evaluation, marginal risk metrics are instrumental in attributing risk, designing robust inference, and informing policy or model selection. This article presents a comprehensive technical treatment focusing primarily on marginal expected shortfall and its extensions but with general connections to other domains in which marginal risk arises.

1. Marginal Expected Shortfall and Its Time-Lagged Extension

Marginal Expected Shortfall (MES) is a pivotal systemic risk metric assessing the expected loss (or negative return) at a particular institution, conditional on a systemic (market) loss exceeding a high threshold. For losses $(X, Y)$ of an institution and the system, respectively, and for tail probability $p \approx 0$, the MES is defined as

$$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$

where $\mathrm{VaR}_{\mathrm{sys}}(p)$ is the $p$-quantile of $Y$. This formulation captures tail dependence and the institution’s risk conditional on systemic extremes (Liu et al., 7 May 2025).

To incorporate delayed effects—where shocks to the system may propagate over time—Liu, Liu, and Zhao introduced the time-lagged MES (TMES). For a stationary time series $(X_t, Y_t)$ and lag $h \geq 0$,

$$\delta(h) = \lim_{u \to \infty} \mathbb{E}[ X_t \mid Y_{t-h} > u ]$$

This generalizes static MES, enabling assessment of risk propagation dynamics and systemic shock spillovers beyond immediate temporal coincidence (Liu et al., 7 May 2025).

2. Estimation and Asymptotic Behavior of Marginal and Time-Lagged Risk Measures

Estimation of MES and TMES in heavy-tailed and weakly dependent settings is achieved via threshold exceedance and empirical averaging. For TMES, the empirical estimator is

$$\widehat\delta(h) = \frac{m_n}{n} \sum_{t=1}^n X_t\, I_{t-h}$$

where $p \approx 0$0 indicates exceedance above a high threshold $p \approx 0$1, with $p \approx 0$2 and $p \approx 0$3. Under sufficient mixing and anti-clustering, $p \approx 0$4 is consistent: $p \approx 0$5 where $p \approx 0$6 is a function of extremograms and tail auto-covariances. The estimator satisfies a central limit theorem, giving a pre-asymptotic normal limit (Liu et al., 7 May 2025).

For confidence bands, a stationary (Politis–Romano) bootstrap is applied on centered summands $p \approx 0$7, and bootstrap copies of the estimator provide empirical quantiles for construction of simultaneous or pointwise confidence intervals. Bootstrap consistency holds under regularity conditions (Liu et al., 7 May 2025).

Extensive simulation validates these approaches, showing estimator concentration, accurate normal approximations, and nominal bootstrap coverage in both max-moving-average and copula-based time series models.

3. Multivariate and Copula-Based Perspectives on Marginal Tail Risk

In high dimensions or under general dependence structures, marginal risk properties—particularly joint tail probabilities—are governed by the interaction between marginal tail behavior and copula structure.

For Gaussian copula models with regularly varying (heavy-tailed) marginals, the tail decay of risk vectors

$p \approx 0$8

varies depending on both the number of components simultaneously exceeding high thresholds and the geometry of the underlying correlation matrix $p \approx 0$9. Asymptotic estimates involve effective tail indices $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$0, computable via quadratic programming tied to $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$1, and exhibit "hidden regular variation": certain joint exceedances decay more slowly than would be expected under pure Gaussian or independent structure (Das et al., 2023).

Tail probability decay rates for sets where multiple marginals simultaneously exceed thresholds differ markedly from the Gaussian marginal case. In particular, with heavy-tailed marginals and Gaussian copula, for any fixed thresholds,

$$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$2

for large $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$3, highlighting asymptotic tail independence yet structurally distinct joint risk compared to the Gaussian-marginal scenario. The nature of the decay exponent depends intricately on both marginal tail indices and correlation structure (Das et al., 2023).

4. Marginal Risk in Competing Risks, Survival, and Causal Inference

Marginal risks arise in survival analysis and causal inference via cumulative incidence and marginal structural models.

In competing risks, the marginal cumulative incidence function for event type $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$4 is

$$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$5

With right censoring and induced dependent censoring in recurrent event settings, robust nonparametric estimation using inverse-probability-of-censoring weighting (IPCW) enables consistent inference for marginal cause-specific cumulative incidence and hazard functions (Li, 2017). Bootstrapping is used for confidence bands, and the approach detects clinically meaningful differences between risk profiles.

For complex settings such as semi-competing risks, marginal structural models (MSMs) combine inverse probability weighting with Markov and frailty-based illness–death models to define and estimate marginal risk contrasts (risk difference, risk ratio) with explicit causal interpretation under the potential-outcomes framework. Estimation uses weighted Cox regression or EM algorithms with nonparametric bootstrap variance estimation (Zhang et al., 2022).

5. Marginal Contribution and Sensitivity in Aggregate and Robust Risk

In robust risk management, marginal risk often refers to sensitivities or allocations—how much an individual component or exposure contributes to the risk of an aggregate portfolio or functional.

In multi-marginal optimal transport (MMOT) formulations, the worst-case value of a spectral risk measure under only marginal knowledge is characterized as a linear MMOT problem: $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$6 The dual potentials $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$7 arising from the MMOT dual problem quantify the marginal risk contribution of component $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$8 at level $$\mathrm{MES}(p) = \mathbb{E}[ X \mid Y > \mathrm{VaR}_{\mathrm{sys}}(p) ]$$9. In the setting where $\mathrm{VaR}_{\mathrm{sys}}(p)$0 (sum-aggregate), worst-case spectral risk decomposes into the sum of marginal spectral risks, and the corresponding sensitivities recover classic comonotonic additivity (Ennaji et al., 2022).

In distributionally robust frameworks, optimizing over Wasserstein balls around marginal reference measures yields robust bounds for aggregate loss/risk expectations. The dependence on marginal uncertainty radii is continuous and can be computed explicitly via strong duality. A wide range of computational algorithms support these robust marginal risk bounds (Fan et al., 2023).

6. Marginal Risk: Dominance, Decomposition, and Model Uncertainty

Within stochastic modeling for risk management and systemic risk, the decomposition of model risk into marginal and dependence (copula) components is critical. Empirical studies from the Copula-GARCH literature find that model risk attributable to the choice of copula (dependence) dominates that due to marginal models, especially during crises. The mean absolute deviation (mad) of risk forecasts, used as a quantitative measure, confirms that marginal model risk is typically an order of magnitude smaller than dependence (copula) risk (Fritzsch et al., 2021).

Moreover, the practical utility of model confidence set (MCS) procedures highlights that narrowing the set of plausible copulas can further reduce marginal model risk, but such improvements are smaller compared to gains from dependence model selection.

In causal inference and mediation for binary outcomes, explicit relationships between conditional and marginal relative risk parameters are established via recursive log-mean regression graph models. The marginal effect—quantified by the log-marginal relative risk—decomposes into a direct (conditional) effect plus a deviation term, capturing the influence of mediating variables, thus distinguishing marginal risk from conditional or pathway-specific effects (Lupparelli, 2018).

7. Marginal Risk Assessment without Ground Truth in AI Systems

In engineered systems, particularly AI deployment, marginal risk is operationalized as the vector difference in multi-dimensional risk profiles between a new system and a baseline, evaluated without reference to ground-truth outcomes. The MARIA framework formalizes marginal risk as

$\mathrm{VaR}_{\mathrm{sys}}(p)$1

across risk dimensions (performance, reliability, safety, fairness, etc.), and evaluates via relative (ground-truth-free) methods: predictability dominance, capability dominance, and interaction dominance. Metrics such as output stability, agreement rates, or interaction win-rates quantify where AI adoption introduces new risks or delivers reductions, ensuring that the focus is on marginal risk increments rather than absolute risk (Chen et al., 31 Oct 2025).

Computation involves systematic comparative runs, bootstrapped confidence intervals, and rigorous calibration procedures to guard against spurious findings. This methodology offers a rigorous, scalable approach for pre-deployment marginal risk evaluation in high-stakes AI systems.

---

In summary, marginal risks and their rigorous quantification, estimation, and decomposition arise throughout advanced quantitative risk management, systemic risk, causal inference, and complex system evaluation. Methodologies span threshold-based estimation for time-lagged dependence, robust optimization under distributional uncertainty, deconvolution of model risk, causal mediation, and novel frameworks for ground-truth-free operational assessment, with the technical toolkit leveraging extreme value theory, copula analysis, optimal transport, and advanced statistical inference (Liu et al., 7 May 2025, Das et al., 2023, Li, 2017, Ennaji et al., 2022, Fan et al., 2023, Fritzsch et al., 2021, Chen et al., 31 Oct 2025).