The authors propose a quantile regression forest (QRF)-based postprocessing method for the height of new snow (HN). The results are compared to a recently proposed ensemble model output statistics (EMOS) approach for postprocessing HN forecasts. QRF shows clear improvements over the EMOS model, in particular the inclusion of additional predictor variables seems to be benefitial.

Overall, I found the paper to be interesting, well-written and easy to follow throughout. I only have some minor and technical comments that are outlined below.

Minor comments

1. I found the description of the forecast distribution in Section 4.3 a bit confusing. Perhaps it would help to specifically clarify here that the resulting forecast is a set of quantiles derived from the observations from the final nodes, and not the empirical distribution in equation (1).
2. line 145 ff: Perhaps a few more details should be provided on why the CRPS is computed in this form. The motivation to create more quantile forecasts than raw ensemble members is not entirely clear to me. In the end, you compute an ensemble of quantiles that has many more members than the raw ensemble. While this makes it easier to account for the necessarily finite size of the sample from the forecast distribution and makes the comparison to EMOS (with a continuous forecast distribution) more "fair", doesn't this represent an "unfair" advantage when comparing to the raw ensemble?
3. line 169 ff: Compared to the description of the QRF model, the description of feature importance is rather short and probably difficult to understand for readers unfamiliar with QRF. Perhaps a few more details and fomulas here might help make this more clear.
4. Section 6, first paragraph: Are the CRPS values computed for the test set, the training set, or another validation period?
5. Figure 2: Given that the show forecasts seem to be of particular importance, wouldn't include more summary statistics from that variable further improve results?
6. Figure 3: I find the confidence intervals difficult to distinguish due to the overlap and would suggest to split up the plot into three panels for all of the 3 models.
7. Overall, the paper in particular demonstrates that the inclusion of additional predictor variables improves performance. This is very much in line with the previous work on QRF and also several other machine learning-based postprocessing methods (for example proposed in Messner et al (2017, https://doi.org/10.1175/MWR-D-16-0088.1), Rasp and Lerch (2018, https://doi.org/10.1175/MWR-D-18-0187.1), Bremnes (2020, https://doi.org/10.1175/MWR-D-19-0227.1), and others). Since EMOS only uses forecasts of the variable of interest as predictor, it would have been more "fair" to compare the the boosting extension of EMOS proposed in the paper by Messner et al. While this is beyond the scope of the paper in the current form, I'd suggest to include this aspect in the discussion as an avenue for future work.

Technical comments

1. line 59: Missing reference?
2. Section 4.2: I'd suggest to consider moving this Section to the beginning of Section 6. In the current form, the meaning of the hyperparmeters mtry and nodesize is not yet explained and will be difficult to understand for readers not familiar with QRF.
3. Code and data availability: The limitations on availability of the EMOS and NWP model code are clearly explained, but it is unclear to me whether (or where) the QRF code is available.

Overall:

This is a useful paper which shows very clear benefits from calibration of ensemble forecasts of snow depth. Two calibration methods are compared and the quantile regression method shows clear advantages over the more standard EMOS approach (although it should perhaps be noted that there are many ways of implementing EMOS and other approaches could perform better than the one used here). It is good also that the authors have included the section 7.3 on the limitations for operational use. This is an important factor, that many calibration methods are mathematically skilful but not practical to apply for real-world forecasting, often due to lack of suitable training data, so it is good to discuss this openly in the paper. I would recommend publication with only minor amendments.

I have suggested to Editor that the title is not understandable to a wide audience (see comment 1 below). I have also suggested that not all figures are of appropriate quality (see note 6 below.) 

Specific Comments:

1. For me the term “height of new snow” is confusing. I suspect this is simply a slight mistranslation from the authors’ native French, but causes confusion to a native English speaker. The normal English term would be “depth of snow”, whereas “height” would be used more for the altitude (height up the mountains for example) where snow would occur. (A Google search for “height of new snow” returns many references to depth of snow.) I would recommend changing the word height to depth throughout the paper, including the title, and hence also HN to DN. However, I do note from the references that the authors have published a previous paper on the topic using the same term “height of new snow”, so I would understand if they want to keep it for consistency. In this case, it would be worth defining what they mean clearly in the Introduction to avoid confusion.
2. Line 34: Delete the word “from”: “…This prevents an appropriate correction…”
3. Table 1: Abbreviation IQR is not defined – assume it is Inter-quartile Range – but should be defined.
4. It is interesting that all the predictors used are univariate ensemble summary statistics which means that correlations between variables present in the ensemble members are lost. This might be worth some mention – it is very encouraging that the methods are successful, but it might be expected that some higher skill might be achievable if correlations between for example precipitation and Near-surface temperature were retained. Might be worth comment.
5. Line 158: The notation used for the intervals looks odd, with opening square brackets at both ends. In figure 6 the closing at the end of the interval uses a closing round bracket, which looks better.
6. Figures 3 and 4: I found the colours difficult to interpret when they are overlaid. It does become easier in combination with the text description, but I would suggest some alternative which clarifies the ranges for each colour. Perhaps you could mark the upper and lower bounds (10^th and 90^th) of each shading with overlaid lines in strong colours. (Also, this would be much worse for someone who is colour blind and cannot distinguish red and green – a different set of colours would be better but if you add lines as suggested then they could also use different line patterns.)
7. Line 200: “QRF leads to an improvement …” – technically the plots show that EMOS and RAW are degraded relative to QRF.
8. Line 210: The term ROC has not been defined “Relative Operating Characteristic” (or alternative versions of the name). Also, you do describe ROC here briefly in lines 210-212, but why did you not define it in section 5 where all the other evaluation scores are defined?
9. Line 216: You are describing the blue curve here, not the red one.
10. Line 316: “he” should be “the”

The manuscript “Calibrated ensemble forecasts of the height of new snow using quantile regression forests and Ensemble Model Output Statistics” demonstrates the advantages of quantile regression forests (QRF) for postprocessing of the height of new snow. The authors perform an in-depth comparison with ensemble model output statistics (EMOS) in terms of forecast skill and discuss important aspects of QRF with regard to operational implementations. As the paper is well written and its usefulness is clearly motivated, I only have very few comments.

Specific comments:

• The results suggest that rainfall related predictors improve forecast skill for QRF, probably also because of poor prediction of the snow/rain limit. As the implemented version of EMOS does not use any rainfall (and temperature) related predictors, QRF outperforms EMOS considerably in situations of rainfall, but zero snow, forecasts. For a fairer comparison, I would suggest constructing a second EMOS reference that uses also rainfall (and temperature) based predictors, possibly including interactions. This would help to evaluate, if the increase in skill by QRF really comes from advantageous properties of QRF compared to EMOS or just by the fact that QRF considers more predictors.
• An evaluation of forecast sharpness is missing. I appreciate that the authors show not only CRPS values, but also rank histograms as a measure of calibration. However, as forecast quality is determined by both calibration and sharpness, I would suggest adding a figure that compares the sharpness of the raw ensemble, EMOS, and QRF.
• Figures 3 and 4 are difficult to read because of overlapping intervals. I would suggest modifying the figures such that the limits of the prediction intervals of all forecasts are visible.

Technical corrections and typing errors:

• p3/l59: useless question mark
• p3/l66: “…statistics of other…”
• p4/l72: small leading ‘e’ in ensemble
• p4/l75: same leading ‘r’ in regression
• p4/l76: maybe “at zero” instead of “in zero”?
• p4/l77: small letters “ zero-censored censored shifted-gamma distribution”
• p5/l94 to 99: Don’t we minimize the within group variance when maximizing homogeneity.
• p6/l121: “…22 seasons, one…”? as “the” sounds odd to me here
• p6/l126-127: only if you apply score decomposition. Or, is the word “simultaneously” missing here.
• p7/l132: “…variable and equals…”
• p7/l136: What do you mean by “technical”?
• p8/l165: CRPS
• p8/l172: “…variable, i.e. node…”
• Figure 2: I would make clearer that you are analysing the results for 24h accumulations of new snow depths here. Maybe, 1-24h, 25-48h, etc. in the subpanels’ titles.
• p10/l190: From Figure 4 I guess that the station ID of Saint-Paul-sur-Ubaye should be 4193400.
• p13/l199: Please do not mention significance without having performed any statistical hypothesis test. Either write, e.g., “considerable” or apply a statistical hypothesis test.
• p14/l203: “…leading to U-shape….”
• p16/l216: Probably blue instead of read
• caption of Figure 8: 2^nd word: CRPS
• caption of figure 8: Definitions in the caption and the titles of the subpanels for subpanels (b) and (c) are interchanged.
• p18/l261 “adiabatic” instead of “adiabiatic”
• p18/l265: Do you mean “statistical post-processing” instead of “statistical processing”
• p19/l266: What is “iso-\theta_{w}^{‘}”? The 1-degree Celsius isothermal level in terms of pseudo-adiabatic wet-bulb temperature?
• p19/l273: “…with statistical post-processing…” w/o the “a”?