The article presents a comparison of regionalization methods to estimate the calibration parameter of the WaterGAP3 GHM model in ungauged catchments worldwide.

The article presents relevant work, and it is well written. Although, in my opinion, the fact that the methodology relies completely for data and evaluation in the model rest validity to analysis and findings of the results. Nevertheless, if these constraints and their implications are better analyzed and addressed, the paper is publishable.

General comments:

In several parts of the manuscript, the authors claim they are using new regionalization methods, and that is misleading, they are applying regionalization methods (that are not new) for the first time in the context of the WaterGap model. Please clarify throughout the text.

The implementation of regionalization methods is not very well explained, many assumptions and choices are made without proper justification (see specific comments). These may have important impacts in the performance of each regionalization method, unfairly favoring one over another.

I have some reservations with the fact that all regionalization is made within the WaterGap model, e.g., the use of model input data as basin descriptors, for example “soil storage” coming from a look-up table, does not seem to be a proper descriptor of a basin; the selection of these descriptors based on the correlation with the calibrated parameter; the evaluation of each method by comparing regionalized gamma versus calibrated gamma (if I understood correctly);  and also, to use as benchmark the resulting parameter distribution of a previous version of the model at a different resolution and regionalization method.

Specific comments:

Line 131, explain the standard calibration approach and sufficient performance indicators referred here.

Line 148-149, explain what do you mean with the need to define sensitive parameters for a better calibration of WaterGAP3 if the model only has one calibration parameter.

Line 162-163: I would say you are minimizing redundant information, but not avoiding it, some of the descriptors are correlated.

Line 173 to 177: To use the calibrated parameter to select descriptors comes with many drawbacks, as you explain here, why then not to choose another more independent variable to define meaningful basin descriptors?

Line 177: More than a complex relationship, a very uncertain one, if there is one at all.

Table 1: Define IG(&) in the Table

Line 210 to 214: Please clarify, the evaluation is done by comparing regionalized gamma with calibrated gamma? Or model predictions? Why a MAE value of 2.1?

Line 220 to 222: Please explain better the “tunning approach”, what are gamma 1 and gamma 2?

Line 252, why three clusters?

Line 259, explain why the “highly flexible version” uses 162 clusters?

Line 305, re-write to clarify.

Line 306 to 308, re-write to clarify the meaning of the sentence. I believe that it has to do more with the relevance of the descriptors included in explaining catchment behavior, or in this case calibrated parameter gamma.

Line 317-317: Change “However” for “Therefore”?, if I understand the intention of the sentence correctly.

Line 356-357, the “high flexible” k-means results are not in Figure 4? They need to be added.

Line 358 to 361: I do not understand this statement, why and how a regionalization method will “extract” information form the descriptors. Pagliero et al. (2019) is not correctly cited in the context of this statement.

Figure 5d: I have reservations to the validity of comparing global distribution of gamma for WG2 and WG3 given they have different resolutions. Please comment whether this is an issue for this comparison.

Figure 6, what was the reason to choose the year 1989 for this comparison? Seems arbitrary. Same for year 2010 in Figure 7. Why not use an average year, or the driest/wetter year of the period.

Line 453-454. This statement contradicts itself. Basin descriptors selected for regionalization are of crucial importance because they need to be selected due to the amount of information they contain.

Line 459: use a more academic term instead of “blurring”.

This manuscript investigated the ability of parameter transfer to ungauged basins in global hydrological models based on the calibration experiments of WaterGAP3. The combination of traditional and novel regionalization approaches revealed that machine learning-based methods may be too flexible for regionalizing. However, some findings are not convincing and need more discussions or elaboration, some terms need better definitions, and the figures need improvements. Therefore, I recommend a major revision. Please see below for my detailed comments.

Major comments:

1) Title. The term “Regionalization and its impact” is too vague to present the content of this manuscript. Elaborate the titles based on the novelty of the presented work. 

2) The range of gamma values is too narrow. The cluster of values around the maximum range of gamma implies that the higher values seem appropriate but are excluded from the candidates. Please recalibrate gamma values using wider ranges, and present the scores in the accuracy of predicted streamflow in terms of score index (e.g., NSE).

3) The authors assessed the impact of regionalization only from the mean absolute errors in gamma values but neglecting the ability of parameters to represent the river discharges. The parameter transferred from one basin to another should reasonably represent river discharges (accepted behavior) at least in the donor basin. The success of the parameter transfer should be assessed by the accuracies of simulated river discharges in the transferred (ungauged) basins with the parameter. Missing information on the accuracies in the simulated river discharges (both in donor and transferred basins) impedes the reliability of this work.

Minor comments:

1) L 174. The usage of 'heavy-tailed' distribution seemed wrong. In probability distribution, heavy-tailed distributions have heavier tails than the exponential distribution. The authors may intend different meanings, but the use of 'heavy-tail' here would misleading.

2) L102. Srmax values are not presented in the manuscript. Maybe it is based on the look-up tables of this model. Please provide it as many researches indicated that soil storage is an important parameter for GHM.

3) L259. Explain why 162 groups were used.