The paper is well structured while the content is very dense and properly concise. The paper gives an important contribute to the analysis of factors affecting the FF curve by performing a deep analysis of the effects produced by models structure, model parameters, interaction between model parameters and model structure, initial conditions (in terms of water content) and precipitation events. In my opinion the paper can be accepted subject to minor revision.

I suggest the authors prepare a flux diagram describing their workflow, i.e. all the steps of their procedure. Indeed, there is a lot of attention on how the available data are used and this, in the end, slightly obscures the logic and the sequence of the steps. In this (or these) diagram(s) the authors should highlight the deep meaning of each step independently of the way the available data are used for their quantification.  

It is not completely clear to me how the rainfall events are generated starting from the precipitation frequency curve. I understand that simulations are performed with a time step of one day and thus precipitations are generated with this time step. However, in the case of Island Park, a two-day precipitation event is generated from the frequency curve and I do not understand if the total of precipitation in two days is generated or if the event is scanned at daily level.

Overall, I prefer continuous simulations than event simulations: in fact, in this way the “natural” combination between rainfall periods and flood periods is obtained without any artificial combination between initial conditions and rainfall events. Instead of using the regional analysis, why don’t the authors have set up a rainfall model (Neyman-Scott or Poisson model), performed long simulations and extracted annual maxima from them?

Incidentally, Hashimi -> Hashemi

Review of HESS-2021-49: “Identifying Sensitivities in Flood Frequency Analyses using a Stochastic Hydrologic Modeling System” by Newman et al.

The authors present a sensitivity study examining the relatively contributions to flood quantiles from precipitation, initial conditions, model structure and parameters, and meteorological sequencing for two watersheds in the western US. The results are interesting and useful, while the manuscript is well-written. Like reviewer #1, I think that only minor revisions are needed. Similarly, like reviewer #1, I think occasionally found it difficult to understand exactly what was done or why. I’ll point out those issues that I noticed, but I agree with reviewer #1 that an overall workflow diagram might be helpful if done with care.

Specific comments:

I agree with the authors’ discussion of AEP equality assumptions, but the potential problems don’t end there. Even the assumption that precipitation annual maxima—which are the values used here and in most studies—are the drivers of streamflow annual maxima is not really correct. In Yu et al. (2019; specifically, Table 3 in that paper), we found that you need to get into 200+ year return periods before that assumption is really reliable, at least for the midsized midwestern watershed we looked at. Clearly, this is less of an issue for really big floods.

L68-70: It seems like some element is missing from this sentence. “higher sensitivity…” higher than what?

Section 3.1: I think you need to provide more explanation on how you used the total probability theorem. I *think* I understand what you did, but the reader shouldn’t have to guess. Out of curiosity, I’m wondering if that approach would be valid when using distributed models. With lumped models (which I assume the authors are using here, but I’m not actually sure; see below), a bigger rainfall event combined with a higher IC will result in a higher peak than a smaller rainfall combined with a drier IC. But with distributed models, that is only true in general but not universally due to routing effects—I’ve seen cases where this isn’t.

Section 3.4: It would have been nice to know how important this assumption of picking a few (high) ICs is, as opposed to letting the ICs vary more widely. My particular concern is that to some degree or another, your rainfall quantiles are probably based in part on some events (probably some big ones!) that are outside of this Feb-July (Altus) and Apr-June (Island Park) periods. The Colorado 2013 floods are a good example of this. I suspect that there is some degree of misrepresentation of the relative importance of ICs and precipitation for this reason.

Section 3.5: I found the explanation of spatial precipitation structure to be unclear-both how it was done, and why it was done. In the latter case, my confusion stems from the lack of description of the models’ spatial discretization (or maybe I missed that somewhere).

Section 3.6: I found this section difficult to follow, and didn’t totally understand what was being done.

Section 3.7: While neither Peleg et al. (2017) for I (in Zhu et al. 2018) examined model structure, we did use ANOVA (in Zhu et al.) or something like ANOVA (in Peleg et al.) to examine the roles of other things (ICs, for one) in FFA. I won’t be offended if you don’t, but you may consider whether those prior studies’ findings provide relevant contrasts with your work.

L304: Usage of “overrepresentation” is unclear.

L304 and more generally: given all the moving parts here, some section referencing would help, as well as a bit more precision with terminology. For example, “KGE interval metric-based calibration”-it took me a minute to figure out what you were talking about. You mean calibration based on peak flows, right? Furthermore, referring back to Section 3.3 (e.g. “(see Section 3.3)”) would help the reader the track down the relevant details they might have missed or forgotten. This section referencing would help in a number of other places too.

Section 5: It would be nice to know if the “shapes” of the flood frequency curves are driven by the shapes of the precip frequency curves, which aren’t shown.

L365 and around there: I struggled with this paragraph, in part because I didn’t understand the descriptions in Section 3.6. Also, this is another good place to refer back to earlier sections/descriptions.

L376: “Dry to historical meteorological sequence”-I found this wording confusing

L390: You could refer back to the first mention that you’re analyzing different streamflow timescales

L428: consider replacing “across” with “between”

References:

Yu, G., D. B. Wright, Z. Zhu, C. Smith, and K. D. Holman. “Process-Based Flood Frequency Analysis in an Agricultural  Watershed Exhibiting Nonstationary Flood Seasonality.” Hydrol. Earth Syst. Sci. 23, no. 5 (May 7, 2019): 2225–43. https://doi.org/10.5194/hess-23-2225-2019.

Zhu, Zhihua, Daniel B. Wright, and Guo Yu. “The Impact of Rainfall Space‐Time Structure in Flood Frequency Analysis.” Water Resources Research 54, no. 11 (2018): 8983–98. https://doi.org/10.1029/2018WR023550.

Peleg, N., F. Blumensaat, P. Molnar, S. Fatichi, and P. Burlando. “Partitioning the Impacts of Spatial and Climatological Rainfall Variability in Urban Drainage Modeling.” Hydrol. Earth Syst. Sci. 21, no. 3 (March 14, 2017): 1559–72. https://doi.org/10.5194/hess-21-1559-2017.

General comments :

This paper presents a comprehensive study of the uncertainties of a Flood Frequency analysis method based on stochastic simulation. The different sources of uncertainty are distributed between the structure of the model, the estimation of the parameters, the initial conditions and the inputs (rainfall). Overall, this paper is well written and presents significant and comprehensive scientific results. As the other reviewers I will recommend minor corrections, mainly to clarify some points about the tools used, described only by publications. Clarifications on the methodology would allow a better understanding of certain points detailed below.

Specific comments :

Section 3.1: As the other reviewers, I think that a diagram presenting the workflow would help in understanding the different steps and tools used.

L134: specify if the modeling is lumped or distributed, knowing that the input data for the calibration looks distributed (in Newman et al, 2015). Clarify how you calibrate the hydrological models from an ensemble of historical meteorology (is the ensemble related to ground data interpolation uncertainties?). If the modeling is global, are the ensemble really very different sets (spatial mean)?

L137: Clarify if "two event sequence possibilities" are in fact two periods/seasons? And which ones?

L141: We ask ourselves the question of 11 parameters sets: how are they obtained ? This is not from the sets of historical meteorology since there are 100 used. (ok, the answer is in line 328... but only 10 sets for the Altus basin)

L220: why not take the whole distribution of initial conditions (IC) and only the strongest initial conditions. This does not allow to associate dry CI with heavy rainfall, and it reduces the impact of the uncertainties related to IC (for the current frequencies at least).

L250: Rainfall events are in fact total rainfall generated from a regional probability law. Is it a limitation of the method, to simulate two-day events to generate extreme flood flows? As the time step of the hydrological modeling is daily, how are the rains distributed over the two days simulated? Moreover, can you explain how the problem of changing from a point rainfall to a basin rainfall is solved (how the areal-reduction factor is taken into account).

L250-252 and 260-262: it should be better explained how you use historical precipitation events that are equal to the basin-average magnitudes sampled from the frequency curve, especially for extremes events.

L309: How is the calibration performance assessed. There do not appear to be any validation procedures to verify this (as LOO procedure for instance). The graphs in Figure 4 presented calibration results. It is therefore difficult to judge the performance of the different models, which also depends on their robustness. It is better to show validation results.

Specific comments : Figure 1: as it is a theoretical explanation diagram, put theoretical curves (straight lines for example ?) because we have the impression that it is a result.