Goulart et al. provide an interesting analysis of soybean failures in the US as modeled using Random Forest in the present climate and in future climates. The research is well conducted and uses appropriate methods. My main concern is not with the methods or results, but with the interpretation of those results. The manuscript is well written and the figures describe the data well. I thank the authors for presenting the research in such a complete and coherent manner.

Major comments

I rather like your use of Random Forest and you apply the model rigorously. Your use of shuffling data is also clever and provides a nice analysis structure, but your interpretation of the results could be improved. See my two comments below

• The interpretation of the compound factor becoming less important needs contextualization. I don’t think it’s quite correct to say that it becomes less important, but only that it becomes more dependent on precipitation. I think this relates to your test of importance being to shuffle the data in an attempt to break the correlation structure as compared to the ordered data. But in a warming climate, when all years are hot years, the dependence of joint hot-dry extremes becomes dependent mostly on rainfall, making the joint failure and shuffled failure years similar to one another at high return periods. I think it is misleading then to say that the compound nature of these events becomes less important, only that their occurrence depends primarily on precipitation anomalies.
• L 429-438 - There is no real evidence (from figure 7) that precipitation contributes to the increase in soybean failures overall. There is only evidence that it contributes in a secondary way to truly extreme events, such as 2012 (Figure 10), although its influence is much smaller than that of temperature

• There is an unintended consequence of using a threshold to define failures that leads to the conclusion that joint dependence will become less important in the future. When you set a threshold to identify “failure” years arising from hot years, you lose the sensitivity of yields to warming above that threshold. For example, in the future even if all years meet the criteria for hot years, drought years may be exceptionally hot. This joint temperature-precipitation correlation structure would show up if you were modeling yields as a continuous variable but is lost once you have converted your dependent variable to a binary (e.g. no discrimination between “failure” and “exceptional failure”).
• In particular, your analysis cannot say that “the correlation between contributing meteorological variables becomes less relevant” (L392-393). You can only say that it becomes less relevant for your chosen yield threshold. If you were to up your threshold definition of “failure”, it would once again become important.
• It is not true that “non-critical values of precipitation […] can lead to failures” (L394). We can see in Figure 7 that in fact all of the 3C failures still occur at very low levels of precipitation, but it is true that now low levels of precipitation lead to crop failures at higher rates. My qualm here is really with the interpretation of the precipitation not DTR as you already discuss the limitations of the model with respect to DTR well in the discussion section.

Minor comments

Does EPIC-IIASA include changing management in the historical run or is it static?

Why do there only appear to be nine values in your partial dependence plots (Figure 4)? Or am I misinterpreting the ticks at the bottom of the plots on the x-axis?

Figure C7 - not required, but I’d suggest using different colors. It’s nearly impossible to distinguish the differences in the PDFs at the top of the graph using orange/red and yellow

L 355 - I’d qualify that the change in temperature is much greater than the change in precipitation

L288-289 - It is not true that you see an increase in extreme dry years. The two blue PDFs are nearly identical, and the difference seems marginal. You certainly see an increase in the failure rate at the low-end of the rainfall distribution though (e.g. “Failure 3C” PDF larger than “Failure PD” PDF). And because heat is increasing you will still see increasingly frequent joint warm and dry conditions.

Figure 7: change the axis labels to be interpretable to the reader instead of being the variable names used in the code

L 282 -283 (and relevant to 390-392) - The interpretation that compound weather extremes become less important in a warming world because extreme temperature alone drives failures is not necessarily true. It may be (and probably is) rather that compound weather extremes become more likely in your shuffled data when you uniformly warm all years. For example, if every year is an exceptionally hot year then your crop failures depend almost exclusively on rainfall, making the joint failure and shuffled failure years similar to one another at high return periods.  In fact, we can see that all of the 3C failures still occur at very low levels of precipitation in Figure 7, but it is true that now low levels of precipitation lead to crop failures at higher rates.

L 255 - I don’t think the data support a robust increase of failure probability at higher precipitation levels unless you are willing to put equal weight on the corresponding dips in higher maximum temperatures leading to lower rates of failure and higher diurnal temperature ranges leading to lower rates of failure. I’d at least caution the reader that this may be noise rather than signal in the data.

L243 - define DTR at first use in the text. While it’s defined in the table it is not defined before first use in the text

L159-161 - what does assigning weights mean in this case? Was the dependent variable weighted? Are you running RF on binary data instead of crop yields generally? If so, justify this decision.

- Also, why does it matter if the failure observations are less frequent?

L158 - need to better define for the reader what a “data split” and a “shuffle” is in your cross validation procedure

- Described on L190

L100 - note that soybean yield data is available for this entire period going back to 1900 from USDA, but it is difficult to remove the management and technology changes trend. So there is still reason to use crop models and the work provides a useful complement to observation-based analyses.

L 37: “the majority of climatic shocks are compound events” - is this true?

Goulart et al. analyse weather conditions leading to soybean failure in the Midwest US using crop model data, training of a random forest model, and analysing a particular historical event and possible analogous events in the present day and future. The paper is well-written and clearly describes the approaches used and the results, and I enjoyed reading it. I have some comments particularly around the justification of the data used and the storyline approach, which I have noted below.

Specific comments

• I think the reasoning behind using a crop model rather than observations is sound, however you haven’t shown whether the model is reliable at modelling soybean yields in the US. Could you provide any references, or analysis (if there is observed data available), to show whether the crop model provides realistic results for soybean production in the region of interest and can reproduce some of the important impacts being considered in the paper (e.g. does the model represent the impact of hot and dry conditions on soybean plants well?)? It would then be relevant to reflect in the discussion section on how the model-based results are likely to compare to real-world crop failures.
• I assume all meteorological variables were considered at monthly scale for the analysis, but I don’t think this is explicitly stated anywhere. Is there justification for using this timescale that could be provided in the text? Did you consider shorter timescale events that may also cause crops to fail but would be averaged out when looking at monthly timescales (e.g. a short very dry period), or if there are periods of less than a month in the crop lifecycle when the soybean may be more vulnerable to particular weather conditions?
• Given that the paper is titled “a storyline approach”, I think the storyline analysis is quite limited. For the 2012 season, only the meteorological variable values are presented and the probability of failure calculated. Can you include more insight into the storyline of events, e.g. how did the weather impact the crop via changes to the soil moisture? How did the high DTR affect the crop? This information would help to provide a much more complete chain of events for the 2012 season, otherwise it is unclear to me how it is a storyline approach being used to identify the drivers of the event. I think using the impact analogues as well as the event analogues is a very interesting approach and, as you note in the discussion, the impact perspective is likely to be of more interest to society. However it would be interesting to include analysis of how the storyline chains of events differ (compared to the 2012 season and to each other) in these cases leading up to similar impacts. There is some relevant discussion already in section 4 but describing the range of plausible storylines more explicitly would help to incorporate more of a storyline approach in the paper.
• In Figure 5a there is a pronounced jump in the observed data at return period of around 6 years. Are you able to provide any insight into why this might be?

Technical corrections

• Line 26: It seems strange to refer to Figure 2 in the text before Figure 1. Also this figure shows the mean yield rather than anything related to low yield in 2012. Combining information from Figure 9 might be relevant here to show the yield anomaly.
• Figure 2: Where does this mean yield data come from (is it from the model or is it observed data)?
• Line 28-30: Sentence starting “On a global level…” – I think there is a missing word or grammar check needed on this sentence
• Line 30: Change “Extremes” to “Extreme”
• Line 44-45: changing “to link” to “in linking” and “to explain” to “in explaining” would improve the readability here
• Line 166: Correct spelling of Matthews
• Line 184: I assume RF is the abbreviation for Random Forest but this needs clarifying
• Table C1: Need to define what the codes for the variables mean in this table or use their full names
• Line 234: Do you mean high correlations with the yield or between the variables?
• Figure 4: It would be easier for the reader to interpret the figures if you could change the axis labels to the full names of the variables. Also it is not clear what the small lines along the x axis correspond to.
• Line 267: Is the observed data mentioned here the yield data from EPIC-IIASA, rather than real-world observations?
• Figure 9a): Are these yield anomalies from the model or is other observed data used?
• Line 374: Do you mean Figure 10d here?

Review of “Weather-induced crop failure events under climate change: a storyline approach”

This study explored climate impacts on soybean crop failures in the Midwestern United States under present and future warming scenarios using the random forest model and storyline approach with model simulated crop data and CRU climate data. The findings suggest that the failures are likely to increase with global warming, changes in both univariate and multivariate climate drivers of failure decrease the importance of compound factor, and impact-analogues are significantly increased under global warming compared to event analogues. The manuscript is well in line with the scope of the journal Earth System Dynamics. While the random forest model is not new, it is novel to use that to analyze climate impacts on crop failures. Also, it is very interesting to use storyline approach to study this issue. The findings are relevant. I think, in general, the paper is publishable after revisions. While the statistical analysis is done rigorously, the data may be improved, and method description and the interpretation of the results and findings also need to be improved. Here I provide specific comments as follows.

Major comments:

1. It is reasonable to use model simulated crop data to obtain longer time series for model training. However, the study does not show the performance of the simulated data compared to observations. Since the study region has accessible observational yield data over long period, it should demonstrate the simulated crop data can well represent observations. Otherwise, it may require a bias correction to improve the quality of the simulated crop data.  
2. It can be simple to average all the data over the region, but will discard many useful informaiton. It is arbitrary to state “local scale of impacts are not meaningful for national and global implications”. Using all the data points within the study region can increase the sample size by taking into account of the within region variability and increase the robustness of the model.
3. It is not clear how the random forest model was trained and validated. While a description of the evaluation metrics is provided, it is not clear how the data was split as training data and validation data. The validation data needs to be excluded as unseen data when is used for model training.
4. All the results require thorough interpretations such as the mechanisms and biophysical meaning besides just reporting the data.
5. Based on the previous literature, the increase in DTR may actually benefit yields, which is contradictory with the finding from this study. Please explain.

Minor comments:

1. The approach formulated in this study consists of three parts. The storyline approach is part 3 of the approach (part c) and address part of the research objectives. However, based on the title, storyline approach seems the key message of this study. I think the title needs to be changed since the storyline approach seems not the major focus of this study.
2. Little result is presented in the abstract. It can be more informative to shorten the background and methods descriptions and include more results.
3. Line 55: it is not clear to say “crop failures are compound events”. Do you mean crop failures can be consequences of compound events? It requires a clear definition on compound events.
4. Lines 74-75: “…, a natural next step is to …” It may be not so natural. I suggest the authors better explain their motivations and why such study is needed.
5. The last paragraph of the introduction mostly described the method/experimental outline. The majority of this part should be better move to the method section as a general statement of the method prior to detailed descriptions. The introduction section should focus on explaining why the research is needed, is novel, and stating the scope, research questions/objectives.
6. Line 119: “… a similar 5-year period …” Please clarify what are the similar 5-year period at 2C and 3C.
7. Line 194: Please interpret what is compound factor besides providing the equation.
8. Lines 196-197: the definition of return period is confusing. Based on Figures 5 and 6, return period is not a one to one relationship with inverse of the failure probability, how it can be calculated as the inverse of the failure probability? It needs to be clarified.
9. It is not clear which month of data was used as candidate meteorological input prior to eliminating and combining monthly data in July and August.
10. Variable names and acronyms are inconsistent in the text and tables and do not follow standard practice.
11. Figure 3b: Is Pearson’s correlation performed on detrended or categorical yield?
12. I would suggest only providing half of the correlation matrix. To limit redundancy
13. Figure 4: I am assuming that the ticks on the bottom of the partial dependency plot represent the percentiles of the data. If they are being included, they need to be explained.
14. Ln 390 decreasing compound factor. Is this just an artifact of choosing DTR.
15. Ln 374 Figure 10d?