Abstract
Skew‐symmetric families of distributions such as the skew‐normal and skew‐t represent supersets of the normal and t distributions, and they exhibit richer classes of extremal behaviour. By defining a non‐stationary skew‐normal process, which allows the easy handling of positive definite, non‐stationary covariance functions, we derive a new family of max‐stable processes – the extremal skew‐t process. This process is a superset of non‐stationary processes that include the stationary extremal‐t processes. We provide the spectral representation and the resulting angular densities of the extremal skew‐t process and illustrate its practical implementation.
Introduction
In fields such as meteorology, finance, and structural engineering, the accurate modeling of extreme events—such as 100-year floods or sudden market crashes—is a critical challenge. Max-stable processes have emerged as the primary mathematical framework for modeling these spatial and temporal extremes. However, traditional models, including the Brown-Resnick and extremal-$t$ processes, often rely on assumptions of symmetry and stationarity. In real-world scenarios, spatial dependence frequently varies by location, and data distributions are often skewed. This paper addresses the gap in existing literature by introducing the extremal-skew-$t$ process, a flexible family of max-stable processes designed to capture non-stationarity and asymmetric dependence.
Methodology
The authors bridge the gap between mathematical rigor and real-world complexity by leveraging skew-symmetric families of distributions, specifically the skew-normal and skew-$t$. The methodology involves several key innovations:
1. Non-central Extended Skew-$t$ Distribution
The research defines a new variant of the skew-$t$ family incorporating a non-centrality parameter. This allows for more nuanced tail behavior than standard symmetric distributions can provide.
2. Non-stationary Skew-Normal Process
A central innovation is the construction of a random process where the "slant" (skewness) is defined as a function of location, $\delta(s)$. This approach induces a non-stationary covariance structure that remains positive definite—overcoming a significant mathematical hurdle in spatial modeling. By allowing the slant to vary across space, the model adapts to localized geographic or environmental factors.
3. Spectral Representation and Inference
Using the fundamental theory of max-stable processes, the authors derive the spectral representation for the extremal-skew-$t$ process. This provides the necessary framework for simulation and statistical inference. For parameter estimation, the paper utilizes composite likelihood methods and derives angular densities on the unit simplex, which are essential for visualizing how variables co-depend during extreme events.
Empirical Validation
The model was validated through both simulation and real-world application:
- Simulated Datasets: The authors generated 500 replicate datasets to verify the accuracy and efficiency of their likelihood estimation methods in bivariate and trivariate settings.
- Oklahoma Wind Speed Data: The model was applied to weekly maximum wind speeds (1996–2012) from four monitoring stations in Oklahoma, USA. This dataset is characterized by significant marginal skewness and non-stationary dependence, making it an ideal test case for the new framework.
Key Insights
- Skewness as a Driver of Non-stationarity: The use of a varying slant function $\delta(s)$ allows the model to adapt to different regions without losing the tractable properties of Gaussian-based processes.
- A Generalization of Existing Models: The extremal-skew-$t$ process acts as a "superset." By setting skewness parameters to zero, the model simplifies to the standard extremal-$t$ process. This ensures that the model is robust; it performs like a standard model for symmetric data but offers "extra gears" for asymmetric data.
- Richer Dependence Classes: Unlike symmetric models, these skew-symmetric families can describe a wider range of tail dependence levels, transitioning smoothly between negative and positive associations.
Results and Discussion
The extremal-skew-$t$ model demonstrated superior performance in the Oklahoma wind speed application, consistently outperforming the standard extremal-$t$ model. This was evidenced by lower Composite Likelihood Information Criterion (CLIC) scores across all station combinations.
Qualitatively, the model proved more robust to outliers. In instances where a single extreme observation (e.g., ~40 km/h) heavily influenced symmetric models, the skew-symmetric model provided a more accurate statistical fit. Furthermore, bivariate return level plots revealed asymmetric contours, reflecting real-world directional dependencies—such as prevailing wind patterns—that symmetric models are mathematically incapable of representing.
Limitations and Future Work
While the extremal-skew-$t$ process offers significant advantages, it is computationally demanding. Scaling the triplewise composite likelihood estimation to very large spatial grids remains a challenge. Future research may explore semi-parametric estimation to increase flexibility without over-parameterization, as well as the development of simpler, more interpretable covariance functions specifically tailored for skew-normal processes.
Summary updated