A Sliding Blocks Estimator for the Extremal Index

Christian Y. Robert, Johan Segers, Christopher A. T. Ferro

Abstract

In extreme value statistics for stationary sequences, blocks estimators are usually constructed by using disjoint blocks because exceedances over high thresholds of different blocks can be assumed asymptotically independent. In this paper we focus on the estimation of the extremal index which measures the degree of clustering of extremes. We consider disjoint and sliding blocks estimators and compare their asymptotic properties. In particular we show that the sliding blocks estimator is more efficient than the disjoint version and has a smaller asymptotic bias. Moreover we propose a method to reduce its bias when considering sufficiently large block sizes.

Introduction

In Extreme Value Theory (EVT), understanding how extreme events cluster in time is as critical as measuring their magnitude. This tendency for extremes to "clump" is quantified by the extremal index ($\theta$). If $\theta = 1$, extreme events occur independently; if $\theta < 1$, they cluster, with the average cluster size being approximately $1/\theta$.

Traditionally, researchers estimate $\theta$ using the "blocks method," which involves partitioning a time series into chunks and counting threshold exceedances. To simplify statistical proofs, these blocks are usually disjoint (non-overlapping). However, this paper argues that this traditional approach ignores valuable information at block boundaries. By using sliding (overlapping) blocks, we can create a more robust and efficient estimator.

Methods and Estimators

The paper compares two primary approaches to estimation:

  1. Disjoint Blocks Estimator: Partitions $n$ observations into $k$ non-overlapping blocks of length $r$.
  2. Sliding Blocks Estimator: Slides a window of length $r$ through the sample, yielding $n - r + 1$ overlapping blocks.

The authors derive the asymptotic variance and covariance for both, proving that the sliding blocks estimator is essentially the "most efficient convex combination" of all possible disjoint block partitions. To handle finite-sample errors, they also introduce a bias-reduction technique based on a second-order Taylor expansion.

Efficiency and Performance

The core mathematical insight of the paper is that the loss of independence in sliding blocks is actually a benefit. Even though overlapping blocks are dependent, the sliding estimator captures more local information about clusters, leading to a more stable result with lower variance.

Comparison of asymptotic variance and relative efficiency between disjoint and sliding blocks estimators.

As illustrated in the figures above, the sliding blocks estimator consistently exhibits lower asymptotic variance than the disjoint version. Furthermore, the relative efficiency of the disjoint estimator drops significantly as clustering becomes more intense (indicated by smaller values of $\alpha = \theta\tau$). In Monte Carlo simulations using Max Auto-Regressive (MAR) and Moving Maximum processes, the sliding blocks estimator consistently outperformed the disjoint version across all tested $\theta$ values.

Real-World Application: The FTSE 100

The researchers applied these estimators to the FTSE 100 index (daily log returns from 1968 to 2001) to study the clustering of market crashes. Both estimators stabilized around $\theta \approx 1/3$, suggesting that extreme market drops tend to occur in clusters of roughly three days. However, the sliding blocks version, combined with the authors' optimization parameters, produced much narrower 90% confidence intervals, providing a significantly more precise risk assessment for financial volatility.

Conclusion

Moving from disjoint to sliding blocks provides a clear statistical advantage in estimating the extremal index. By leveraging overlapping data, the sliding blocks estimator achieves higher efficiency and lower bias, offering a more reliable tool for analyzing risks in stationary sequences. While the method requires careful tuning of thresholds and block sizes, the resulting gains in precision make it a superior choice for practical extreme value analysis.

Summary updated

Abstract

In extreme value statistics for stationary sequences, blocks estimators are usually constructed by using disjoint blocks because exceedances over high thresholds of different blocks can be assumed asymptotically independent. In this paper we focus on the estimation of the extremal index which measures the degree of clustering of extremes. We consider disjoint and sliding blocks estimators and compare their asymptotic properties. In particular we show that the sliding blocks estimator is more efficient than the disjoint version and has a smaller asymptotic bias. Moreover we propose a method to reduce its bias when considering sufficiently large block sizes.
BibTeX
@inproceedings{robert2009sliding,
  author = {Christian Y. Robert and Johan Segers and Christopher A. T. Ferro},
  title = {{A Sliding Blocks Estimator for the Extremal Index}},
  booktitle = {Electronic},
  year = {2009},
  doi = {10.1214/08-ejs345},
}

Loading related content...

Figure