This is a very interesting study with three explicit objectives: 1) assessing the dependence between four main flood drivers along the coastline of CONUS (surge, wave, rainfall and river discharge), 2) assessing the seasonal patterns of such dependence, and 3) comparing tail dependence detected from observational records and from the reanalysis data. Not only the idea is very interesting, but also the research design is robust and alternative approaches (i.e. POT vs block-maxima and rank vs tail dependance) have been well-discussed. This study and its outputs are very helpful for coastal compound flood hazard assessment and management, as address a very important challenge that compound flooding community is dealing with it. Appropriate characterization of exceedance probability (which is instrumental to risk assessment) is a difficult task when multiple variables are involved and such studies help the flood risk modelers in at least two ways. First, they will now where there is no significant dependence between flooding variables and make the simplifying assumption of independence between variables. Second, in case a significant dependence is detected what type of numerical modeling scheme can provide the most relevant information about the nonlinear interactions between interacting variables. The manuscript is very well-written and results are nicely presented. Thus, I recommend publication of this manuscript in HESS after a minor revision. My comments are mainly about necessary justification/explanations and some editorial. Please, see the details below.

My main concern here is the time-scale of hydrologic variables. While wave and surge are obtained and analyzed on an hourly basis, accumulated daily precipitation depth is taken as a representative of extreme pluvial flooding driver (L167). I doubt that such characteristic of rainfall can be a good representative of extreme precipitation contributing to compound coastal flooding events. This is a relatively long time window that somehow dampens the density of rain storm and so provides irrelevant information for compound coastal flood modeling. Please, explain if such selection was due to lack of data or methodological reasons? Anyways, I didn’t find it explicitly discussed in your manuscript. Also, for the two-way selection scheme, I guess 10 days is too long to justify coincidence/concurrence. It’s been justified with regard to the lag-time of river flow, but I doubt in any of the selected systems the transition time of the peak river flow from the upper tributaries to the outlet would be more than 5-6 days, at most. Do you have any reference on this? I am aware of other studies with comparable window length (i.e. 7 days), but “river flow lag time" might not be necessarily sufficient to justify such long sampling window length. Also, this is not a similar process between all cases. For example, such time window with justification from river discharge lag time has nothing to do with sampling wave and surge that have occurred up to 10 days apart. Moreover, even if it justifies the 10 days lag time in Q-S case, it does not automatically justify pairing an extreme surge with a river flow that has happened up to 10 days later. Such unified 10day time window needs a much stronger justification.

I understand the value of information presented in Figures 5, 7, 9, but I don’t see much value in the way that these are currently presented. There are simply too much information on a single chart. I suggest either breaking these down to subpanels (a, b, c, …) that separately plots the pairs, or removing these charts (as kind of similar information is already presented in the heatmaps). Moreover, I found the dashed lines simply meaningless here. Both rank and tail dependence measures used here assess the nonlinear dependencies and I don’t understand what a linear regression means here? Please, explain if I am missing an important meaning behind these linear regression practice.

L142: WAM has not been defined yet.

L224: The equation does not seem right to me. There should be a typo in it.

L305-321: Better to clearly explain earlier that numbers inside parenthesis show the number of sites.

L351: Looks like R here represents the Pearson corr coeff. (As mentioned in L441 and L413). If so please clarify first time appreas in the manuscript.

L475-479: The sentence is too long. Please, consider breaking it down.

L556: These citations are not necessary (better to say useful) at the end of conclusions.

Great work!

Congrats to the authors for putting together an interesting and generally well-written manuscript. The study uses observational and model-based data to characterize the pair-wise dependence between various flood drivers (storm surge, waves, rainfall, river discharge) for sites across the entire CONUS and for different seasons. The authors then compare model-based dependence against the observed dependence structures between flood drivers to investigate the ability of models to represent multi-hazard dependence. I recommend publication after addressing some points I raise below.

Main Points:

• There are significant differences between the Kendall tau and tail dependence for the seasonal analysis. In particular, the kendall tau analysis suggests that the correlation is higher during the tropical season compared to extra tropical season (for the majority of locations and pairs). But the tail dependence analysis basically shows the exact opposite (figure 8 note all the blue boxes compared to the prevalence of red boxes in figure 6). The authors mention that some differences could be due to sampling differences and autocorrelation of the time series used for the tail dependence analysis, but I don’t think this fully explains the large discrepancy between the two methods. Moreover, the authors do not comment on how to interpret these differences. Which metric should we use to quantify hazard dependence?
• I didn’t see the KL divergence results presented anywhere in the manuscript. As the authors point out the KL metric takes into account the entire dependence structure rather than just the tail dependence, so I was expecting to see/read the KL values for different locations and pairs. I think this needs to be added to the manuscript
• I don’t think the discussion could be organized better. Since one of the objectives/novelties of this work is to investigate hazard correlation across the entire CONUS, the discussion should be used to highlight regional trends. I suggest devoting one paragraph for each region, highlighting the main hazard correlations (and the meteorological reasons for the correlation), and then describing which season multi-hazard events are more likely to occur and why. This should be followed by a paragraph discussing the differences between the tail dependence and kendall tau analysis (see the first bullet). Then I would have a paragraph describing the comparison of the model with the observations. There are a lot of results presented in this paper so the discussion really needs to summarize everything in an organized manner.
• I would also suggest adding some discussion about the limitations of using a bivariate approach for this analysis. For example, along the west coast of Florida, there is significant correlation between almost all the hazard pairs (Figure 3), suggesting that there may also be a significant threat for three or even all four of the hazards to occur simultaneously.

Specific Comments:

Abstract:

Line 33-34: change “elderlies” to “the elderly” and “the poor” to “low income communities”

Section 3.1: It’s not clear to me why two different dependence measures are implemented in this work (Kendall tau and tail dependence). Both measures aim to investigate dependence between hazard extremes (either indirectly by selecting the maximum value per year or by directly utilizing the 90^th percentile threshold). What is the benefit of using two different metrics? What insight does the Kendall tau provide compared to the tail dependence (and vice versa)? The motivation for using these metrics needs to be stated explicitly here.

Figure 3: I suggest breaking up the East Coast region into two regions: Southeast and Northeast. It’s clear from Figure 3 that the Southeast region (up to North Carolina) shows similarities with the Gulf Coast region. In both regions the occurrence of TC events likely causes significant dependence between surge/discharge and surge/precipitation. However, moving to the northeast there are relatively few sites with significant S_P/P_S or S_Q/Q_S correlation. Therefore, I think it makes more sense to talk about the Southeast and Northeast rather than the entire East Coast as a whole.

Section 4.2:

It doesn’t really make sense to talk about a tropical cyclone season for the West coast. I think it’s valid to look at the seasonal differences between dependence values for locations in the West coast, but the authors should provide some justification for why June-November vs December-May still makes sense since TCs do not impact the West coast.

Figure 5: Again, I think it would be more interesting to show the southeast and northeast regions separately.

Section 4.3:

Where exactly were the KL results presented? The authors mention that the KL divergence provides information about whether the dependence structure is significantly different between observations and model, but I did not find any figure/table presenting the KL results.

Figure 10: The caption here is not correct. The caption refers to differences between seasons but I believe this figure is supposed to be showing the difference between reanalysis and observations

Figure 12: Caption incorrect again.

Line 480-481: I don’t understand this sentence: “not all extreme precipitation events can occur with moderate or high discharge, but extreme discharge events can occur with moderate or high precipitation”. I think you mean that high precipitation can occur in the absence of high river discharge, but high river discharge usually occurs simultaneously with high precipitation.