Second-order refined peaks-over-threshold modelling for heavy-tailed distributions

Jan Beirlant, Elisabeth Joossens, Johan Segers

Abstract

Modelling excesses over a high threshold using the Pareto or generalized Pareto distribution (PD/GPD) is the most popular approach in extreme value statistics. This method typically requires high thresholds in order for the (G)PD to fit well and in such a case applies only to a small upper fraction of the data. The extension of the (G)PD proposed in this paper is able to describe the excess distribution for lower thresholds in case of heavy tailed distributions. This yields a statistical model that can be fitted to a larger portion of the data. Moreover, estimates of tail parameters display stability for a larger range of thresholds. Our findings are supported by asymptotic results, simulations and a case study.

Introduction

In extreme value theory, the "Peaks-Over-Threshold" (POT) method is the standard for modeling rare, high-impact events such as insurance catastrophes or market crashes. Traditionally, excesses over a high threshold are assumed to follow a Generalized Pareto Distribution (GPD). However, practitioners face a significant bias-variance tradeoff: low thresholds lead to model bias because the GPD is only an asymptotic limit, while high thresholds lead to high variance due to data scarcity. This paper introduces the Extended Pareto Distribution (EPD) to bridge this gap, allowing for accurate modeling at lower thresholds by accounting for second-order effects.

The Extended Pareto Distribution (EPD)

The core innovation of this research is the inclusion of a second-order parameter ($\delta$) that explicitly models deviations from a pure power law. Unlike previous methods that used second-order theory merely to correct existing estimators (like the Hill estimator), this approach fits a more flexible distribution directly to the data.

The EPD distribution function incorporates a polynomial-like correction term to the standard Pareto tail. By treating the second-order deviation as an intrinsic part of the model rather than a nuisance, the authors provide a tool capable of estimating not just the tail index ($\gamma$), but also tail probabilities and quantiles (Value-at-Risk) with greater robustness. The model relies on a second-order parameter $\rho$, and the authors demonstrate that even with crude initial estimates of $\rho$, the resulting EPD estimates for the primary tail index $\gamma$ remain stable.

Experimental Framework

The model's performance was evaluated using a variety of heavy-tailed distributions, including Burr, Fréchet, and Student-t distributions. These distributions serve as benchmarks to test how well the EPD captures different "flavors" of tail behavior and how the second-order parameters influence the fit.

Table 1: Summary of distributions and their associated second-order parameters used for model validation.

The table above illustrates the relationship between the primary tail index $\gamma$ and the second-order parameters $\tau$ and $\rho$ for these common distributions. This mathematical context is essential for understanding how the EPD accounts for the specific curvature of different heavy-tailed models.

Key Results and Case Study

Simulation results, particularly for the Student-t distribution (with 4 degrees of freedom), showed that the EPD estimator remains asymptotically unbiased even as more data is included from lower thresholds. This increased data usage leads to a significant reduction in asymptotic variance compared to standard GPD models that require very high thresholds.

The practical utility of the EPD was further demonstrated using the Secura Belgian Re dataset, which contains 371 large automobile insurance claims from 1988–2001. While traditional Hill and GPD estimators fluctuated wildly depending on the chosen threshold, the EPD provided stable estimates for the tail index ($\gamma \approx 0.3$) and the probability of extreme claims (e.g., exceeding €7 million) across a wide range of thresholds.

Conclusion and Limitations

The EPD framework offers a significant advancement in threshold stability for heavy-tailed distributions ($\gamma > 0$). While the current implementation is optimized for heavy tails, it provides a more efficient use of available data than traditional POT methods. Future work may involve extending this framework to light-tailed distributions, though the authors note that the numerical complexity increases significantly in those cases due to more intricate parameter dependencies.

Summary updated

Abstract

Modelling excesses over a high threshold using the Pareto or generalized Pareto distribution (PD/GPD) is the most popular approach in extreme value statistics. This method typically requires high thresholds in order for the (G)PD to fit well and in such a case applies only to a small upper fraction of the data. The extension of the (G)PD proposed in this paper is able to describe the excess distribution for lower thresholds in case of heavy tailed distributions. This yields a statistical model that can be fitted to a larger portion of the data. Moreover, estimates of tail parameters display stability for a larger range of thresholds. Our findings are supported by asymptotic results, simulations and a case study.
BibTeX
@article{beirlant2009second,
  author = {Jan Beirlant and Elisabeth Joossens and Johan Segers},
  title = {{Second-order refined peaks-over-threshold modelling for heavy-tailed distributions}},
  journal = {Journal of Statistical Planning and Inference},
  year = {2009},
  doi = {10.1016/j.jspi.2009.01.006},
}
Figure