general comments



One important piece of information that is not mentioned explicitly enough in the introduction or the abstract is that the main part of the manuscript is about using a statistical model to generate suitable parameters for an existing mechanistic model (referred to DMO20). Reading these sections for the second time, it becomes a bit more clear but it would be beneficial to the reader to describe this more clearly early in the manuscript.



The authors are careful in creating a forecasting scenario in which no summer data is included. However, the entire set of historical data is used to produce the linear regressions, which may include future data w.r.t. year for which the forecast is produced. If I understand the approach correctly, even in the Case 4 setup, the forecast may include information from summer data of the current year, if the current year is "relevant" (as defined in the manuscript). As a result, does the forecasting system produce significantly better results for the "relevant" years compared to other years? In addition to the 4 cases currently included in the manuscript, I would suggest to add a Case 5 that excludes all data from the future (forecasts for the first few years with little data could be skipped) or, alternatively, excludes the data from the current forecast year, even if it is relevant.



Looking at the results in Fig. 2 and 3, it appears as if the forecast-hindcast as well as the forecast-observation comparisons show a pattern in August and September: The forecast appears to overestimate BWDO and HA for low values and underestimate it for high values and this pattern appears to increase in time. The authors already introduce a linear regression for the purpose of bias reduction but apply it only to the June forecast. A similar linear correction could be applied to correct this pattern which appears to increase with lead time. Yet, I am a bit hesitant to recommend such a correction because it, just like the bias reduction, adds a non-mechanistic element to the model.



How difficult would it be to extend the approach presented in this study and estimate August and September values from all data available until then? I am not suggesting that this needs to be done in this manuscript, yet creating successive 2 month forecasts appear a suitable course of action for producing more accurate estimates. This could be mentioned in the discussion.



Overall, while the manuscript is well written it sometimes overestimates the study-specific knowledge of the reader. Including more information explicitly would benefit readers. In some instances, I had to read ahead to answer questions which could have been addressed right away. I have listed some of those instances below.





specific comments



l 8: "Several models" Here it would be helpful to specify what type of model is meant, e.g. "dynamical", "statistical" etc.



l 52: The same group has previously considered different sources of uncertainty in the 3d model, finding that variations in wind forcing had the largest impact on hypoxia estimates (J. P. Mattern, K. Fennel, and M. Dowd (2013), Sensitivity and Uncertainty Analysis of Model Hypoxia Estimates for the Texas-Louisiana Shelf, Journal of Geophysical Research, doi: 10.1002/jgrc.20130).



l 77: After the first read, I am assuming that the DMO20 model has four compartments, an eastern and a western one, each divided into two layers. This could be made a bit more explicit in the text.



l 91: Is there any indication about the cause of this bias? It is nice to have an underlying parsimonious mechanistic model, yet the bias correction introduces a non-mechanistic element. By the way, it would be helpful to mention again that June is the start of the prediction interval and that the bias disappears over the course of several weeks, so that it can be neglected in the following month.



l 110: Is this done for one or multiple years? All years with data? This information is probably given later but it would be useful to mention it here already or even earlier.



l 118: Do the ten most relevant years represent the full time period accurately?



l 144: I may be good to give some examples of the model parameters that contribute to the uncertainty here, so that the reader does not need to consult the DMO20 paper to get this information.



l 150: What if one, multiple, or maybe all relevant years are in the future w.r.t. to the estimated year?



l 156: "Sixteen multiple linear regressions": I assume, the 16 refers to 4 (months) * 2 (rivers) * 2 (discharge, nitrogen loading) but this could be made a bit clearer, or a reference to Table 1 could be added here already. In my opinion, it would be good to clearly state again that there are 4 regressions for each month.



l 185: The "hypoxia model" is DMO20, correct? I would suggest to include this here again.



l 192: Is there a distinction between "forecasted" in this line and "pseudo-forecast" a few lines above? It would be good to stay consistent with the use of "pseudo". Maybe even drop the "pseudo-" prefix after describing that this is the way the word forecast is used in the context of the manuscript.



Fig. 2: Am I correct in assuming that there are 32*30 red dots in the top panels, one for every day in June in the 32 years with data? But if the monitoring cruises are typically in late July, why are there so many red dots in the bottom panels?



Fig. 4: I can only distinguish between 3 shades of gray here, yet the caption suggests there should be 4. Are the uncertainties plotted cumulatively or is the effect of parameter uncertainty generally smaller than that of the riverine and and meteorological inputs?



l 238: "Note that the relative magnitudes of the variance components are somewhat different from the relative magnitudes of the 95% IQR components ...": If the goal here is to say that the relative magnitudes differ because the variance has squared units I think it would be easier for the reader to state this directly, rather than drawing a line from IQR to standard deviation.

This manuscript builds upon previous statistical models for hypoxia in the Gulf of Mexico to make a temporally resolved forecast. This work is potentially valuable for management of the ecosystem’s fisheries and offers some important insights regarding the contributions of different sources of uncertainty to this type of forecast. The manuscript is well written and will be of interest to readers of HESS. I have one substantial concern and a number of comments for the authors to consider.  

This modeling approach is impressive as it handles the different sources of uncertainty without excessive parameterization or an actual process-based model. To do accomplish this, the model relies on use of historical data for the trajectory of summer meteorology, discharge and loading. The authors select the most relevant years’ summer records based on similarity of spring forcings. This selection process at line 115 is explained clearly, but more justification would be helpful as this is potentially influential for the forecast. What are the consequences if this selection step were omitted? What happens to predictive performance if 10 less relevant years are used? Does that lead to degradation of performance?

It appears that this temporally resolved forecast would, in fact, be static once the spring hydrology data are used to identify the most similar years to use for summer forcings. Could this somehow be informed by additional data from after May, perhaps using some information from previous years?

Figure 2. Are the hindcasted data for all of the years, or only those that were matched as most relevant and used for the model?

Figure 4. The the predictive intervals in the shaded region do not appear to increase over the course of the season, which is unexpected based on Figure 3. The authors provide parts of the explanation for this in lines 240-245, but I suggest adding an additional sentence that puts together the 1) disparity in uncertainty between inputs and fitted parameters and 2) change in contribution of inputs over the season.

Table 2 and Line 326. The caution about increasing uncertainties over the season is appropriate, but is not adequately captured in the figures or tables in the main text or supplement. Instead, the decrease in R2 is presented. Although the uncertainties are referenced in the text, it would be helpful to have a table or visual that shows the uncertainty for the pseudo-forecast by month (by that I mean IQR in units of HA or BWDO)

Technical Comments

Table 2. What does the bold styling indicate? It appears to be the highest R2 of the four cases, but this could be explained in the caption

Line 91 - The rationale for this correction is clear and sound. Presumably there is some uncertainty associated with both the predictor and fitted parameters of this regression. Were those carried forward into the forecast? I would expect that uncertainty to be impactful.

Line 135 - Did this occur frequently? Assigning a threshold for performance is appropriate, but consider adding some justification for why this particular threshold was selected.

General Comments:

This is a well written manuscript on an extremely important topic of predicting hypoxia in a water body of global significance. The methods and findings from this study are also applicable to similar water bodies across the world. The novel approach of using data-driven modelling (particularly, Bayesian statistical modelling) makes this a suitable paper for HESS. The results and discussion are generally well written (save one comment about being confused by Table 2). However, I do think that it would be good to see more details in the methodology about the approach taken in this study. This is coming from the perspective of someone who thinks this work is very relevant to the work of watershed planners and managers - who may want more detail on how they could apply similar work themselves.

I understand the reasoning for not detailing the Bayesian Mechanistic Model in this paper (because it was previously published) - however, I do think that more details on the model forumulation are required in Section 2.1 so that this can be a stand-alone paper in its own right. Particularly - what are the key parameters and what are the model equations? Similarly in Section 2.2, readers need to refer to the DMO 20 publication, and to Matli et al. 2018. I do think that these details should be brought across to this paper too.

I also think more emphasis in the methodology needs to be placed on the forecasting method and the regression modelling of June-September Discharge and Loading. I didn't realise until quite later in the paper that Bayesian methods were used in determining these data. This is quite important and could be applied not only in the context presented in the paper, but to other sites. In particular the details I would like to see are: what software/platform was used for the modelling (is this available on github etc?), how did you check convergence of chains?, how many iterations?, was there a burn-in period?, how many models did you produce in the exhaustive search (just to place the scale of the work in context), what are the prior distributions?, how were the assumptions checked?

I would also like to see an assessment or further discussion of how the uncertainty in the forecasts are propogated through the Bayesian Mechanistic Model. How do you account for this? How does the final uncertainty change? More details on how this assessment is conducted would be good to see in the methodology.

Specific Comments:

Lines 60-63: I am a little confused as to what the difference here is between hindcasting and forecasting? Would these need different modelling/simulation strategies? Is the forecasting process the same as the hindcasting process - just using input data representing future conditions? Then how would you validate the success of the forecasting? Perhaps a brief clarification of these points would be useful here.

Line 101: 'geostatistical estimates from Matli et al' - please elaborate on this a bit more: what are the geostatistical estimates of? What does Matli et al. 2018 provide?

Line 101-102: Could you please provide a few more details here too? I assume that this is referring to the estimates of BWDO and HA, but it is not 100% clear. Why is it that the monitoring cruise data have lower uncertainty? Also - how often do these monitoring cruises happen?

Line 321-324: The authors state that the projections of summer riverine inputs improve hypoxia forecasting skill - and refer to Table 2, however even after the explanation in Section 3.2, I am still confused about this table. I would have thought that if the projectings are improving forecasting skill, we should see higher R2 in the Forecasted vs Observed columns compared to the Forecasted vs Hindcasted columns. Or perhaps I am competely misunderstanding the table.

Technical Comments:

 - The authors have used a large number of acronyms throughout the manuscript (e.g., HA, BDOM,  NGoM, MAR). This is perhaps a subjective comment, but I highly suggest that these acronyms are avoided and terms written out in full to make the manuscript more readable.

- Fig S1: I suggest putting this figure in the main manuscript - not all readers are familiar with the Gulf of Mexico

- figure 1 was a very useful figure for me and helped me understand the process - but for those who are not familiar with the different text box shapes - could you please provide a legend either in the figure or in the caption? also perhaps to make it easier for the reader, it might be nice to have the section heading numbers in the figure.