Abstract
Motivated by examples from extreme value theory 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 which are implied by stationarity of the underlying time series under minimal assumptions. Of particular interest are the cluster size distributions, where we introduce the two notions of inspected and typical cluster sizes and derive general properties of and connections between them. 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 time series analysis, extreme events—such as floods, heatwaves, or market crashes—rarely occur in isolation. Instead, they tend to cluster in time. Extreme Value Theory (EVT) traditionally uses the extremal index ($\theta$) to quantify this clustering, often informally describing $1/\theta$ as the "mean cluster size." However, this definition is ambiguous: does "mean size" refer to the average size of a cluster containing a fixed observation point (an "inspected" cluster) or the average size of a cluster chosen at random (a "typical" cluster)?
This paper provides a rigorous mathematical framework for "cluster processes" to resolve this ambiguity. By moving beyond restrictive assumptions like "regular variation," the authors derive invariance properties of these clusters using only the stationarity of the underlying data.
Methods and Framework
The authors define a cluster process as the limiting point process of a binary time series, where 1 represents an event (an extreme) and 0 represents a non-event.
Key methodological steps include:
- Limiting Conditional Distribution: The cluster process is defined as the limit of a stationary process conditioned on an event occurring at time zero (Assumption 1.1).
- The Time Change Formula: Theorem 2.1 establishes a general invariance principle, showing that the probability of observing a specific pattern within a cluster is invariant to shifts, provided the shift remains within the cluster's bounds.
- Defining Cluster Sizes: The paper distinguishes between two quantities:
- Inspected Cluster Size ($S^i$): The size of a cluster that contains a fixed observation point.
- Typical Cluster Size ($S^t$): The size of a cluster viewed from its start or chosen "at random" from a sequence of clusters.
Key Insights: The Inspection Paradox in Extremes
A central finding is the application of the inspection paradox to extreme clusters. Just as one is more likely to wait for a bus that is running late, an observer is more likely to "inspect" a cluster that is unusually long. Consequently, the expected size of an inspected cluster $E(S^i)$ is always greater than or equal to the expected size of a typical cluster $E(S^t)$.
Other key insights include:
- The "Uniform Split" Property: In an inspected cluster of size $k$, the probability of being at any specific position (1st, 2nd, ..., $k$-th) is exactly $1/k$.
- Redefining the Extremal Index: The authors prove that the extremal index $\theta$ is exactly $1/E(S^t)$. The common interpretation of $1/\theta$ as the "mean cluster size" is only correct for the typical cluster.
- Universality: These properties are universal consequences of stationarity and the existence of a limit, rather than being tied to specific distributional assumptions.
Key Results
- Theorem 3.3 (Size-Biasing): Establishes the link between the distributions: $P(S^i = k) = k \cdot \theta \cdot P(S^t = k)$. This is the mathematical heart of the inspection paradox.
-
Theorem 4.2 (Expectation Relationship): $$E(S^i) = \frac{E((S^t)^2)}{E(S^t)}$$ This formula demonstrates that the inspected cluster size is inflated by the variance of the typical cluster size.
-
Symmetry: The number of events occurring before and after an inspection point follow the same distribution.
Limitations and Future Work
The theory relies on the existence of a limiting conditional distribution (Assumption 1.1). While standard in EVT, the paper does not explore cases where this limit fails. Furthermore, the framework is currently built for discrete-time series ($t \in \mathbb{Z}$); extending these invariance properties to continuous-time point processes is a natural next step. Finally, while providing the theoretical "ground truth," the paper does not propose new estimators for $\theta$, leaving empirical application for future research.
Summary updated