Dear Leonie and co-authors:

Thank you very much for submitting your responses to the reviewer comments. I apologize that it has taken me so long to look at your responses and the entire manuscript in detail. I would like you to go ahead with the edits you describe in your responses, but I also have a few additional comments that I would like you to consider when you revise your manuscript. I have very much enjoyed reading about your innovative study, and I hope my comments will help you to highlight the contribution of your study even more strongly.

Please let me know if you have any questions or would like to discuss some of my points in more detail. I am looking forward to receiving the revised version of your manuscript.

Best wishes
Pascal
Pascal Haegeli, PhD
NHESS Editor
Simon Fraser University, Vancouver BC, Canada

MORE SUBSTANTIAL COMMENTS
L61: Research objectives
In the description of your research objectives, you state “(a) find out if and how real-world behaviour corresponds to online behaviour”. Given that your dataset of GPS tracks is rather small and its representativeness and relationship to the click data is unclear, I would like you to be more precise in the description of your objective. Your dataset does not allow you to draw general conclusions about “real-world behaviour” in backcountry skiing. Instead, you are comparing the clicks of people that use the Skitourenguru for trip planning to GPS tracks that have been uploaded to the site. You can hypothesize what these datasets and your analysis might say about general behaviours in the discussion section, but it is beyond your research question. Please be clear about this throughout the manuscript.

L156: Representation of tracks
Why did you choose average elevation to represent your GPS tracks? Minimum and maximum elevation of a tracks seems more informative as it would allow you to identify the likely minimum temperature that they would experience and relate the track to the avalanche prone area described in the bulletin. I am sure you had your reasons for the mean value, but please explain.

L165: Use of dataset
Studies using different parts of their dataset for different analyses always raises warning signs. You could have included an additional “season” variable in your click model to account for this.

L175 and Figure 4: Normalizing of counts
In my opinion, normalizing unnecessarily obscures the data you are presenting, and it creates an inconsistency between Figures 4 & 8. You can produce the exact same presentation by showing the clicks and counts on two different scales. This allows the reader to see the count numbers. Furthermore, please use the full width of the page for these charts and use gridlines that highlight the weekends. This will allow the reader to see the patterns more clearly. Finally, one should use consistent representations for the same data type. Both of your time series are count data, and they should therefore be presented the same way.

L220: Characterization of regions
Why did you calculate ski route density by dividing the total length of all ski routes in the region by the area of the region. I assume this variable aims to represent the options that backcountry users have. However, your approach makes longer routes count more than shorter routes, which does not seem quite right. Why not using number of routes divided by area? Please also explain the thinking behind your accessibility variable in a little bit more detail.

L295: Location of detailed explanation of presentation
All of the employed methods should be described in the methods section. However, it might be easier for the reader if you moved some of the detailed explanations of the pros and cons of the normalized PD plots to the results section where readers are looking at these charts. You want to present the information where it is most useful for the reader. This detailed description is difficult to understand for the reader in the methods section since they have not seen the charts yet.

L341+: Order of presentation in results section
It seems odd that you talk about the model performance before you have even described the models themselves. I strongly recommend that you describe the models and the influence of the variables first. Once the reader has a sense of this, it is much more meaningful to discuss performance.

L341: Details on model performance
Your section on the model performance is very detailed, and one can get easily get lost in it. While interesting, is all this information really necessary for the addressing the research questions you specified in the introduction? Instead of documenting everything possible, being more selective can help direct readers’ attention to the most important parts of the analysis.

L378 and Fig. 8: Examination of prediction error
It is interesting to look at the time series of the prediction errors. Similar to my earlier comment about Figure 4, I would use the full width of the page to display the information, and I would use the same presentation for the same type of data (counts represented as lines). In addition, however, why are you presenting the errors as weekly and not daily summaries like the original count data? Since weekday/weekend is an important predictor, this presentation obscures any related patterns. Given the structure in the time series of the prediction error for the click model, it seems like there is something at play (most likely weather related) that is not captured by your predictor variables. It would be useful to comment on this in either the results or discussion section.

L412 and Fig. 10: Impact of different variables
I very much enjoyed the discussion of the impact of the different variables. I think that your description of the spatial pattern in the response to increasing danger ratings is interesting, but your current presentation of the effect in Figure 10b is making it difficult to properly see it. In essence, you are saying that there is an interaction between the increase in danger rating and some of the variables that describe the forecast regions (accessibility, wildlife areas, and ski route density) which results in a response that is spatially variable. However, since the change in danger ratings in your scenario is not homogenous across the Alps, it is difficult to separate the two patterns. I think two maps that show the change in response to a consistent change from a) moderate to considerable, and b) considerable to high would be more insightful for showing what is going on. In addition, I think it would be good to flip the color scale to have positive values at the top and negative values at the bottom.

L419+: Discussion
I think your discussion could be much more powerful and insightful. Right now, a considerable amount of text repeats background information explaining the nature of the data (e.g., L438+), and the thinking behind some of the design choices of your study. In addition, you repeat information from the results section and in some circumstances, you introduce additional results (e.g., L446+: discussion of spatial patterns) without further discussing its higher-level meaning. Furthermore, many aspects of the discussion section touch on different limitations of the analysis without ever giving the reader a complete picture
There are lots of interesting insights coming out of your study, and I think that tightening and slightly restructuring the discussion section would help the main results and implications of your study come out more strongly. The current structure of the section (data, methodology, and variables) makes it hard to get the full picture. I suggest the following subsections for the discussion:
• A synthesis of the results that very explicitly answers the research questions outlined at the beginning of the paper.
• A dedicated limitation section that summarizes the limitations from all sources (data, variables, and methodology) to produce a comprehensive picture of that these models can and cannot do.
• A more targeted implications section that focuses on what the results mean for avalanche safety initiatives and research. What important insights and new opportunities does your study create (e.g., highlight areas that warrant extra avalanche forecast resources because people tend to flock there under certain conditions) and what are the steps required to take this research to the next level (e.g., integrating information about the terrain characteristics of the route and tracks to better understand how people change their choices based on the avalanche conditions).

L560+: Conclusion
The conclusion section is very short and does not seem to add much value right now. It would be useful if you could provide a bit of an outlook that looks at next steps for this line of research.

EDITORIAL COMMENTS
L65: Calling Section 2 “Background” instead of “State of the Art” would make the structure of your manuscript more in line with standard scientific papers.

L70 (and others): Mccammon should be spelled McCammon

L73: Surveys and questionnaires are generally analyzed with quantitative methods.

L86: Replace “emergency avalanche beacons” with “avalanche transceivers”

L133: Why are you providing the mean elevation of the Swiss Alps? Is this a meaningful/necessary statistic for your study? The fact that 50% of the area is above 1500m seems mire informative. Furthermore, precision of these numbers seems unnecessarily high. It would be sufficient to say that the Swiss Alps cover roughly 26,000 km2.

L153: “Obfuscated” is not a commonly used term and therefore makes the explanation more difficult to understand. You can easily provide the same explanation without it. E.g., “To protect the privacy of the contributing users, we converted/transformed/simplified the GPS track to …”

L175 (and others): Please avoid using dashes in scientific writing. I know that ChatGPT likes using dashes, but they are really not commonly used in scientific writing.

L186: I would not use the term “human factor” here. This term is generally used to refer to specific challenges in the human decision-making process, and not any non-weather, snowpack, or terrain factor that affects people’s choices. In general, I think it is an overused term, and we should use it in a more precise way.

L273 (and others): As pointed out by Reviewer 1, please do not use brackets like this. It creates hard to read text. Saving a single sentence is not worth making readers read the same sentence several times.

L314 (Fig. 3 and Table 2): I would integrate the correlations presented in Table 2 directly into Fig. 3 and deleted Table 2 to save space. Am I correct in assuming that the correlations presented in Table 2 should be consistent with the correlation values shown for a 1-day time lag in Fig. 3. This is currently not the case. Please explain.

L320 (and others): Since you are not calculating spatial autocorrelations, I would describe the observed patterns in a more qualitative way. Like the tracks are spatially clustered or concentrated. It would also be useful to describe these patterns in more detail (e.g., along main transportation corridors, key tourism destinations). The same applies for the description of the spatial pattern of the click data.

L337: The description that “tracks were more abundant than clicks in some areas” is incorrect because the click dataset is so much larger than the track dataset that I assume it was larger in all regions. Please describe differently.

Fig. 5: Including the two variables into a single color scale makes this chart difficult to read. It would be much easier to have two separate charts, one for click and one for tracks. The color saturation currently also does not properly line up with the magnitudes of differences between the click and track data (i.e., more saturated does not mean bigger difference). You also do not explain what you mean with “correspondence”.

Fig. 7: The labels for the four count categories are not quite right. The way it is presented right now, the top category (e.g., < 294) includes all data points. Is it necessary to group the values into four categories or could the size of the circles just represent the actual values? Would it make sense to turn the color scale upside down to have lower values at the bottom and higher values at the top?

L384: “weekly periods of over- and under predictions …” is not quite right because the periods were actually 3-4 weeks long.

L387: Replace “classification model” with “track model” to make it easier for the reader to relate it to the other information about this model.

Appendix B: I do not think this appendix is necessary. All these equations are well established in the scientific community.

Dear Leonie, Frank, Guenter, and Ross:

Thank you very much for taking the time to respond to all reviewer comments in such detail and revising your paper accordingly. I feel that this version of the manuscript is much improved, and I am happy to accept it for publication. Congratulations on a great paper and contribution to the literature. You are examining an important question in public avalanche safety, and you are going it in a very sophisticated but accessible way.

While I was going through the manuscript, I found a few typos and made suggestions for a few minor editorial changes (see attached file). Please have a look at these suggestions before you submit the final files, but I do not need to see the manuscript again before publication.

I sincerely apologize for taking so long to make a decision on your paper. This has taken way too long, and I generally aim to process manuscripts much more speedily. I am very sorry about that, and I hope it has not caused any issues.

Congratulations again on a great paper.
Best wishes
Pascal

Pascal Haegeli, PhD
NHESS Editor
Simon Fraser University, Vancouver BC, Canada