Extreme value analysis is frequently employed to estimate environmental loading on offshore structures. However, there are many possible approaches to applying this method. This study adopts a non-asymptotic framework to investigate the statistical behavior of significant wave height ( H s ) extremes across multiple temporal aggregation scales. Six parametric probability distributions were evaluated using four goodness-of-fit criteria, together with different block maxima definitions. The results show that the Anderson–Darling test provides the most robust model selection, due to its higher sensitivity to tail behavior and its ability to produce spatially coherent patterns. The Exponentiated Weibull distribution performs best for high-frequency data, while the traditional Weibull becomes more appropriate as the block size increases, reflecting the smoothing of extremes. The analysis demonstrates that both the choice of goodness-of-fit test and the temporal aggregation significantly influence the selected distribution and the resulting return-level estimates for long return periods. Spatial patterns reveal strong regional variability, indicating that the use of a single distribution or aggregation scale may lead to substantial biases. Therefore, regional heterogeneity and the specific objectives of the analysis must be carefully considered in the statistical modeling of extremes. • Global sensitivity of extreme Hs using WW3 hindcast (1993–2024) • Effect of AD, KS, χ 2 , and MSE on best-fit distribution selection • Anderson–Darling yields more coherent and tail-relevant patterns • Larger blocks shift dominance from Exponentiated Weibull to Weibull • Methods affect 100- and 1000-year return levels across regions
Reis et al. (Tue,) studied this question.