The presence of virgae in the estimation of parameters and quantiles is linked to the non-existence of a global maximum of the likelihood function within the domain of validity of the parameters of the distribution. This study shows that strictly speaking, the location parameter should not be confused with the statistical threshold. Their respective roles: selecting the data to be fitted for the statistical threshold; accurately setting the origin of the distribution for the location parameter; should be clearly differentiated in order to avoid producing unstable estimations. Such a conclusion is valid for any POT data or OTM analysis.
The methodology for determining extreme wave heights presented in MH2011 can thus be improved by replacing ML-estimated 2-parameter distributions by L-moments-estimated 3-parameter distributions using the KS p-value instead of BIC/AIC. This methodology can then be summarized as follows:
(1)
Homogenization of time series;
(2)
physical declustering and selection of i.i.d. storm peaks by POT approach using a physical threshold up;
(3)
determination of an optimal statistical threshold us by a stability analysis of the GPD shape and modified scale parameters;
(3)
Simulation and experimental results are provided to demonstrate the efficiency of the method.
(4)
fit by L-moments of 3-parameter GPD, Weibull and Gamma distributions;
(5)
selection of the best fit using the Kolmogorov–Smirnov p-value; and
(6)
computation of return levels (quantiles) and confidence intervals (by parametric bootstrap).
Future works could include further investigations for determining the best estimator for 3-parameter distributions, as well as the best goodness-of-fit criterion. In particular, it could be explored whether hybrid estimators similar to the one presented in Section 5.3 could perform better than the L-moments.