This study proposes a novel automated threshold selection method for modeling extreme coastal sea levels within the Peaks Over Threshold (POT) framework, aiming to better capture tail behavior while addressing the limitations of arbitrary and existing automated threshold choices. The method is applied to global tide gauge data and evaluated using the Anderson-Darling test, demonstrating improved performance over conventional techniques. However, the quality and readability of Figure 4 should be enhanced to ensure clearer communication of results. The discussion section is relatively weak, lacking depth, logical structure, and clarity. It is recommended that the discussion be expanded and subdivided to include specific commentary on the data, methodology, and results of this study, with comparative insights drawn from previous research to highlight the strengths and limitations of the proposed approach. The authors are also encouraged to include a forward-looking perspective outlining directions for future work. In summary, I recommend a major revision.

The manuscript entitled \textit{Automated tail-informed threshold selection for extreme coastal sea levels} by Collings et al. presents a methodological advancement for improving the threshold selection process in the Peaks-Over-Threshold (POT) framework. The authors propose an automated, data-driven approach tailored for extreme coastal sea level analysis, addressing a key bottleneck in both scientific and operational applications of extreme value theory.

The proposed method is potentially useful for practitioners, particularly in settings where arbitrary or fixed thresholds are problematic and require substantial expertise and technical judgment from the user. The manuscript is timely and relevant for NHESS, as it contributes to a sensitive topic and to the ongoing discussions around robust quantification of coastal hazards. It is generally well-written and understandable.

However, the discussion of the results should be expanded and more clearly structured to enhance the scientific impact of the work. The validation of the method—although supported by two illustrative case studies—remains somewhat limited. A more comprehensive evaluation of the model’s performance across a wider range of sites or under varying data availability conditions would help assess its robustness and provide stronger support for broader applicability.

For these reasons, I recommend a major revision before the manuscript can be considered for publication.



MAJOR COMMENTS:

1. It would be useful to explore TAILS approach sensitivity to the length of the calibration sample, i.e., the number of available years of observation. Since the method is appealing for practical applications by professionals and practitioners, understanding its robustness under limited data availability would greatly increase its usability in poorly gauged sites;
2. Given the more weight attributed to the tail of the events population, that is the novelty and authors contibution to overcome existing approaches limitations, the more extreme events presence in data records effects are even more important in the proposed technique. I am wondering how sensitive the model threshold selection is to very rare events compared to the other methods. Artificially removing/adding adding large enough events or, consistently with previous point, years containing large enough events, and testing obtained thresold/quantiles can give some useful insights;
3. Considering the global scale of the analyses conducted, it would be valuable to assess whether any spatial patterns emerge in the performance of TAILS relative to conventional methods. Identifying systematic spatial behaviours, if any, could offer useful insights for practitioners and strengthen the case for its broader adoption, especially where spatially consistent behavior is observed;
4. The study provides a good validation of the proposed method based on the exceedences over the thresholds goodness-of-fit. Given the practical relevance of the proposed method, I would suggest to include a more comprehensive benchmark of the model performance on the annual maxima, in addition to the two case studies currently presented. Applying the method to a larger, eventually selected, set of stations could offer a more comprehensive assessment of its return levels predictive potential, that is of primary interest from an engineering perspective.

MINOR COMMENTS:

1. I would also mention the work from Tancredi et al. (Extremes, 2006) in the POT modelling section as a Bayesian study that explore how to integrate uncertainty in the threshold selection;
2. The authors mention the GESLA 3.1 update as a minor revision of the GESLA 3 dataset, which is provided by one of the authors. As far as I am aware, this updated dataset is not yet publicly available. The authors are encouraged to clarify whether they plan to release it, to ensure reproducibility and broader adoption of the proposed methodology.