Invariance properties of limiting point processes and applications to clusters of extremes

Anja Janßen, Johan Segers

Abstract

Motivated by examples from extreme value theory, but without using the theory of regularly varying time series or any assumptions about the marginal distribution, we introduce the general notion of a cluster process as a limiting point process of returns of a certain event in a time series. We explore general invariance properties of cluster processes that are implied by stationarity of the underlying time series. Of particular interest in applications are the cluster size distributions, and we derive general properties and interconnections between the size of an inspected and a typical cluster. While the extremal index commonly used in extreme value theory is often interpreted as the inverse of a “mean cluster size”, we point out that this only holds true for the expected value of the typical cluster size, caused by an effect very similar to the inspection paradox in renewal theory.

Introduction

In the analysis of stationary time series, extreme events—such as financial crashes, environmental disasters, or network traffic spikes—rarely occur in isolation. Instead, they exhibit a tendency to "cluster" in time. In classical Extreme Value Theory (EVT), this clustering behavior is quantified by the extremal index ($\theta \in [0, 1]$). A common heuristic in the field is to interpret $1/\theta$ as the "mean cluster size."

However, this interpretation masks a significant statistical nuance. If an observer identifies a cluster because an extreme event occurred at a specific time (e.g., "now"), they are effectively "inspecting" the process. Much like the inspection paradox in renewal theory—where one is more likely to wait for a bus during a long inter-arrival interval—an observer is more likely to land within a large cluster than a small one. This paper provides a rigorous framework to distinguish between "inspected" and "typical" clusters, relying solely on the assumption of stationarity rather than specific distributional tails.

Theoretical Framework

The authors move away from standard EVT frameworks that require heavy-tailed assumptions like regular variation. Instead, they define a cluster process through the lens of limiting point processes. Consider a stationary time series $X_t$. By defining a high threshold $u$, we can generate a binary indicator process $I_t = \mathbb{1}\{X_t > u\}$, where $I_t = 1$ denotes an exceedance.

Illustration of a time series $X_t$ and its corresponding exceedance indicator process $I_t$ relative to a threshold $u$.

The cluster process is defined as the weak limit of these binary processes, conditioned on an event occurring at the origin ($t=0$). By applying Time Change Formulas and tools from Palm theory, the authors derive a fundamental invariance principle (Theorem 2.1). This principle describes how the probability of observing a specific pattern of events shifts as the reference point moves within a cluster.

The Inspection Paradox in Extremes

A central contribution of this work is the formalization of the inspected cluster size ($S^i$) versus the typical cluster size ($S^t$).

  • Typical Cluster ($S^t$): The size of a cluster as viewed from its natural beginning.
  • Inspected Cluster ($S^i$): The size of the cluster that contains a randomly chosen exceedance (e.g., at $t=0$).

The paper proves that the inspected cluster is stochastically larger than the typical cluster. Specifically, the relationship between their expectations is given by: $$E[S^i] = \frac{E[(S^t)^2]}{E[S^t]}$$ This confirms that $E[S^i] \geq E[S^t]$, with equality holding only if the cluster size is constant. Furthermore, the authors demonstrate a Uniform Split Property: given the total size of an inspected cluster, the number of events occurring before the inspection point is uniformly distributed.

Key Results and Identities

The research clarifies the long-standing relationship between the extremal index and cluster size.

  • Theorem 4.2: The authors establish that $\theta = P(S^i < \infty) / E[S^t]$. In cases where clusters are almost surely finite, this simplifies to the elegant identity $\theta = 1/E[S^t]$.
  • Distributional Link: Theorem 3.3 provides the mathematical machinery to derive the distribution of the inspected cluster size from the typical one, and vice versa, showing they are intrinsically linked by stationarity.

These results were validated using mathematical models such as Moving Maxima Processes (where cluster structure is fixed) and Urn Models (where risk persists for random durations, leading to geometric cluster sizes).

Conclusion and Future Directions

By stripping away the requirement for regular variation, this paper shows that the core properties of clusters in extremes are fundamental consequences of stationarity. The "inspection paradox" highlights a systematic bias: averaging the sizes of clusters that contain observed exceedances will overestimate the "typical" cluster size.

While the paper is theoretical, it poses a significant question for practitioners: should $\theta$ be estimated using "blocks" (which target typical clusters) or "runs" (which may be subject to inspection bias)? Future work may extend these invariance properties to "marked" point processes, incorporating the magnitude of extreme events alongside their timing.

Summary updated

Abstract

Abstract Motivated by examples from extreme value theory, but without using the theory of regularly varying time series or any assumptions about the marginal distribution, we introduce the general notion of a cluster process as a limiting point process of returns of a certain event in a time series. We explore general invariance properties of cluster processes that are implied by stationarity of the underlying time series. Of particular interest in applications are the cluster size distributions, and we derive general properties and interconnections between the size of an inspected and a typical cluster. While the extremal index commonly used in extreme value theory is often interpreted as the inverse of a “mean cluster size”, we point out that this only holds true for the expected value of the typical cluster size, caused by an effect very similar to the inspection paradox in renewal theory.
BibTeX
@inproceedings{janen2024invariance,
  author = {Anja Janßen and Johan Segers},
  title = {{Invariance properties of limiting point processes and applications to clusters of extremes}},
  booktitle = {Dependence},
  year = {2024},
  doi = {10.1515/demo-2023-0109},
}

Loading related content...

Figure