Marginal expected shortfall risk measure for time series

Yuri Goegebeur, Armelle Guillou, Jing QinStatistical 2025 · DOI

Abstract

This paper addresses the estimation of the Marginal Expected Shortfall (MES), defined as $\mathbb{E}(X_h | Y_0 > U_Y(1/p))$, at extreme levels within the context of strictly stationary $\beta$-mixing time series. Unlike traditional models that assume independent observations, this framework accounts for temporal dependence and allows for differing tail behaviors between variables. We propose a Weissman-type estimator and establish its weak convergence using empirical process arguments and the cluster method. The methodology is extended to real-valued time series via positive/negative part decomposition and supported by a multiplier block bootstrap procedure. The performance is validated through simulation and applied to river flow data for flood risk assessment.

Introduction

In risk management, quantifying the impact of a catastrophic event in one variable on another is critical. The Marginal Expected Shortfall (MES) measures the expected loss of variable $X$ given that variable $Y$ has exceeded a high quantile. While well-studied for independent and identically distributed (i.i.d.) data, real-world systems like financial markets or hydrological networks exhibit significant temporal dependence. A shock to an upstream river station, for instance, may take time to propagate downstream, necessitating a model that accounts for time lags ($h$) and serial correlation.

Existing MES estimation methods often suffer from two primary limitations: they assume independence, which underestimates risk in clustered events, and they often require $X$ and $Y$ to share the same extreme value index (tail thickness). This research bridges these gaps by developing a robust statistical framework for dependent time series with flexible tail indices.

Methodology

Extreme Value Framework

The approach is rooted in Extreme Value Theory (EVT), specifically targeting "Pareto-type" distributions where the probability of extreme events decays polynomially. The tail behavior is characterized by the extreme value index $\gamma$, which determines the "heaviness" of the distribution's tail.

Estimator Construction

We utilize a Weissman-type estimator to extrapolate risk measures beyond the range of observed data. This is essential because extreme events are, by definition, rare. The construction involves:

  1. Hill Estimator: Used to estimate the tail indices $\gamma_X$ and $\gamma_Y$ for the respective marginal distributions.
  2. Cluster Method: To handle the "clustering" of extremes common in time series, we adopt the framework of Drees and Rootzén (2010). This allows us to establish the mathematical weak convergence of the estimator to a Gaussian process.
  3. Multiplier Block Bootstrap: Given the complexity of the asymptotic variance in dependent structures, we implement a bootstrap procedure that divides data into blocks. This preserves local dependence while allowing for practical uncertainty estimation and confidence interval construction.

Experimental Results

Simulation Studies

The model was tested using two primary synthetic setups:

  • Model 1: A Markovian time series utilizing a Gumbel copula with Burr and Fréchet marginals to test non-linear dependence.
  • Model 2: An AR(1) process with multivariate $t$ innovations to evaluate performance under linear dependence.

Results across 1,000 replications demonstrated that the estimator remains stable even when extrapolating to very extreme levels (e.g., $p = 1/n$). However, the authors noted an increase in bias as the time lag $h$ increases, suggesting a trade-off between lead time and predictive accuracy.

Case Study: River Flow Analysis

The methodology was applied to daily river flow rates ($m^3/s$) from the River Coquet in the UK, comparing an upstream station (Rothbury) and a downstream station (Morwick). Both stations exhibited tail indices of approximately 0.4. The model successfully quantified the expected surge at Morwick one day after an extreme flow event at Rothbury, providing a concrete tool for regional flood risk assessment.

Discussion and Limitations

Tail Index Constraints

A key theoretical constraint is that $\gamma_X$ must be less than $1/2$ for certain convergence properties to hold. While this covers many heavy-tailed applications, it excludes "super-heavy" tails where variance becomes infinite.

Practical Implementation

The choice of block size ($r_n$) for the bootstrap remains a heuristic challenge. Future work could focus on data-driven methods for optimal block selection to improve the precision of the multiplier bootstrap.

Stationarity

The current framework assumes strict stationarity. In the era of climate change, where the definition of an "extreme" event is shifting, extending this model to non-stationary time series represents a vital next step for environmental risk modeling.

Conclusion

This paper provides a rigorous and flexible framework for estimating MES in dependent time series. By decoupling tail indices and accounting for temporal clusters, the proposed estimator offers a more accurate reflection of risk in complex, interconnected systems than traditional i.i.d.-based models.

Summary updated

Abstract

Abstract We consider the estimation of the marginal expected shortfall $${\mathbb {E}}\left( X_h | Y_0&gt;U_Y(1/p)\right) $$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>E</mml:mi> <mml:mfenced> <mml:msub> <mml:mi>X</mml:mi> <mml:mi>h</mml:mi> </mml:msub> <mml:mrow> <mml:mo>|</mml:mo> </mml:mrow> <mml:msub> <mml:mi>Y</mml:mi> <mml:mn>0</mml:mn> </mml:msub> <mml:mo>&gt;</mml:mo> <mml:msub> <mml:mi>U</mml:mi> <mml:mi>Y</mml:mi> </mml:msub> <mml:mrow> <mml:mo>(</mml:mo> <mml:mn>1</mml:mn> <mml:mo>/</mml:mo> <mml:mi>p</mml:mi> <mml:mo>)</mml:mo> </mml:mrow> </mml:mfenced> </mml:mrow> </mml:math> at extreme levels, when $$((X_t, Y_t))_{t\in {\mathbb {Z}}}$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msub> <mml:mrow> <mml:mo>(</mml:mo> <mml:mrow> <mml:mo>(</mml:mo> <mml:msub> <mml:mi>X</mml:mi> <mml:mi>t</mml:mi> </mml:msub> <mml:mo>,</mml:mo> <mml:msub> <mml:mi>Y</mml:mi> <mml:mi>t</mml:mi> </mml:msub> <mml:mo>)</mml:mo> </mml:mrow> <mml:mo>)</mml:mo> </mml:mrow> <mml:mrow> <mml:mi>t</mml:mi> <mml:mo>∈</mml:mo> <mml:mi>Z</mml:mi> </mml:mrow> </mml:msub> </mml:math> is a strictly stationary $$\beta -$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>β</mml:mi> <mml:mo>-</mml:mo> </mml:mrow> </mml:math> mixing time series with marginal distributions of Pareto-type, $$U_Y$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msub> <mml:mi>U</mml:mi> <mml:mi>Y</mml:mi> </mml:msub> </mml:math> is the tail quantile function associated to $$Y_t$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:msub> <mml:mi>Y</mml:mi> <mml:mi>t</mml:mi> </mml:msub> </mml:math> , h is a positive integer and $$p\in (0, 1)$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>p</mml:mi> <mml:mo>∈</mml:mo> <mml:mo>(</mml:mo> <mml:mn>0</mml:mn> <mml:mo>,</mml:mo> <mml:mn>1</mml:mn> <mml:mo>)</mml:mo> </mml:mrow> </mml:math> is such that $$p\rightarrow 0$$ <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML"> <mml:mrow> <mml:mi>p</mml:mi> <mml:mo>→</mml:mo> <mml:mn>0</mml:mn> </mml:mrow> </mml:math> . We propose an estimator for this risk measure based on a Weissman-type construction. First, in case of a non-negative time series, we establish the weak convergence of our estimator by using empirical processes arguments combined with the cluster method of Drees and Rootzén (2010). Then, we extend our result to the case of real-valued time series by using the decomposition of the original time series into the positive and negative parts, and we also propose a bootstrap procedure. The performance of our estimator is illustrated on a simulation experiment. Finally, the method is applied on river flow data.
BibTeX
@inproceedings{goegebeur2025marginal,
  author = {Yuri Goegebeur and Armelle Guillou and Jing Qin},
  title = {{Marginal expected shortfall risk measure for time series}},
  booktitle = {Statistical},
  year = {2025},
  doi = {10.1007/s11203-025-09334-9},
}
Figure