Abstract
This paper investigates the estimation of the tail index $\alpha$ based on independent and identically distributed (i.i.d.) multivariate observations. We focus on "standard" heavy-tailed distributions where Pareto-like marginals share a common tail index. By establishing a multivariate central limit theorem (CLT) for a random vector of local Hill estimators, we introduce the BEAR (Best Empirical AggRegation) estimator. This estimator utilizes a proper convex linear combination of local estimates to minimize the asymptotic mean squared error (AMSE). Theoretical results are supported by simulation experiments and an application to real-world industrial pressure data.
Introduction
In risk assessment and extreme value theory, the "tail index" ($\alpha$) is the critical parameter governing the behavior of heavy-tailed distributions. These distributions characterize phenomena where extreme events—such as financial market crashes, catastrophic insurance claims, or massive structural loads—occur more frequently than predicted by Gaussian models.
While univariate tail estimation is a mature field, many modern applications involve multivariate data or random fields where multiple variables share a common tail index but exhibit complex dependencies. For instance, sensors distributed across a ship's hull or various stocks within a single sector often share tail characteristics. The standard approach—estimating $\alpha$ at a single location or using a simple arithmetic mean of local estimates—often fails to account for the dependence structure between variables, leading to suboptimal precision. This paper addresses this gap by proposing an optimal aggregation strategy.
Methodology
1. Multivariate Central Limit Theorem
The foundation of our approach is the establishment of a multivariate CLT for local Hill estimators. We consider a vector of estimators where each component corresponds to a Hill estimator calculated at a specific location or for a specific variable. Under mild conditions, we prove that this vector is asymptotically Gaussian. This result allows us to characterize the covariance structure of the estimators, which is essential for understanding how local errors propagate through the system.
2. The BEAR Estimator
We define the BEAR (Best Empirical AggRegation) estimator as a weighted linear combination of these local estimators: $$\hat{\alpha}_{BEAR} = \sum_{j=1}^d \lambda_j \hat{\alpha}_j$$ where $\lambda_j$ are weights that sum to one. The goal is to select weights $\lambda$ that minimize the Asymptotic Mean Squared Error (AMSE).
3. Quadratic Optimization
Finding the optimal weights involves solving a quadratic programming problem. By estimating the asymptotic covariance matrix ($\Gamma$) from the data, we can determine the specific convex combination that provides the most reliable global estimate. This method is flexible; it allows for different numbers of extreme observations ($k$) at different locations and can be adapted to other tail estimators beyond the Hill method, such as Moment or Ratio estimators.
Empirical Validation
Simulation Studies
The BEAR estimator was tested against synthetic datasets using a variety of heavy-tailed distributions, including Student-t, Generalized Pareto (GPD), Fréchet, and Burr distributions. To simulate realistic multivariate dependence, these distributions were linked using Gumbel and t-copulas.
Across various dimensions (from 2x2 to 4x4 grids), the BEAR estimator consistently outperformed the simple average. In large-sample scenarios ($n=25,000$), the BEAR method reduced the AMSE by 34% to 53% compared to a standard unweighted average. These results demonstrate that accounting for the dependence structure significantly enhances estimation precision.
Real-World Application: Sloshing Data
To demonstrate industrial relevance, we applied the BEAR estimator to "sloshing" data provided by GTT (Gaztransport & Technigaz). The dataset tracks extreme pressure events recorded by a 3x3 array of sensors on a scale-model LNG (liquefied natural gas) carrier tank.
With nearly 50,000 observations, the BEAR estimator yielded a tail index of $\alpha \approx 3.6$. Crucially, the estimated AMSE for the BEAR estimator was 0.7, compared to 1.1 for the simple average. This reduction in uncertainty provides safety engineers with a more reliable basis for structural design and risk mitigation.
Discussion and Limitations
While the BEAR estimator provides a significant leap in precision for multivariate tail estimation, certain challenges remain:
- Bias vs. Variance: Our current optimization focuses on minimizing variance through optimal weighting. Future work is needed to address the estimation and reduction of the asymptotic bias inherent in tail estimators.
- Sample Size Sensitivity: The method requires a sufficient sample size (typically $n > 1000$) to accurately estimate the covariance matrix. In very small datasets, the noise in the covariance estimate may diminish the gains of the BEAR approach.
- Parameter Selection: Choosing the optimal number of extreme values ($k$) for the local Hill plots remains a non-trivial task. While automated heuristics were used in this study, further refinement of these selection criteria would improve the estimator's robustness.
Conclusion
The BEAR estimator offers a theoretically grounded and practically effective method for pooling information across multivariate heavy-tailed data. By treating local sensors or variables as a "committee" of estimators and weighting them based on their mutual dependence, we can achieve substantially more accurate risk assessments in complex systems.
Summary updated