In this study, the authors evaluated the performance of the quantitative precipitation forecasts (QPFs) by a high-resolution CRM during the mei-yu seasons of Taiwan in 2012-2014, using categorical statistics. The results showed that the QPF skill is better for larger precipitation events, and improved compared to previous results. In addition, case analysis indicates that the strength of the high-resolution CRM lies in an improved ability to capture smaller scale processes for the phase-locked rainfall systems. These findings verify that the high-resolution CRM has good potential application in actual QPF during the mei-yu seasons of Taiwan. However, some major issues need to be clarified.

General comments:

1. It needs to give more explanation for the novelty of this study. As mentioned in the introduction, the purpose of this study is to clarify the dependency property in categorical scores of QPF and whether the skill of the high-resolution CRM is better than those in previous studies, although the studied object is changed from typhoons to mei-yu systems. This purpose has been basically fulfilled by W15 and W16. Therefore, they should not be considered as the novelty of this study, unless the study can prove the CReSS is sensitive to different weather systems. However, from the conclusions, the higher-resolution CReSS primarily improved the forecast skill of the phased-locked topographic rainfall, as it better resolves the terrain and related small scale processes, which means the improvement of QPF caused by this CRM is not attributed to a better capture of the evolution of mei-yu front.
2. Regard the QPF skill of the CReSS on different categories of rainfall events, this study shows a better QPF skill for larger rainfall events. However, this phenomenon may also happen for other high-resolution models, as a higher resolution permits the model to capture more small scale processes to improve convection development, and thus, more rainfall production. To a certain extent, this can be indicated by Figure 3 which shows that the QPF skill of the “All” category (the black lines) has a smaller success ratio (about false alarm) than those of large rainfall categories (A and A plus; the orange and red lines) for high rainfall thresholds (such as larger than 100 mm). It means that the high-resolution CReSS not only produces larger rainfall for large rainfall events, which leads to a higher TS scores, but also produces larger rainfall for small rainfall events, which leads to a smaller success ratio. Thus, this study needs to clarify more about the advantage of the CReSS model, apart from the resolution.
3. As discussed in section 4, the QPF error is also attributed to the forecast error on the evolution of mei-yu front. This error may lead to a worse QPF, as the location and timing of large rainfall can be completely incorrect. What is the cause of this error? Is it related to the boundary condition or the domain processes simulated by the CReSS? The answer of this issue can clarify the novelty of this study, as it is about the QPF associated with the mei-yu front.
4. The comparative analysis or verification is based on a key indicator, the TS score. However, how large the value of TS score could be defined as skillful or a good skill? The study mentioned that when TS is larger than 0.15 it can be indicated “some predictive skill” (in line 228). Is there any objective definition or reference from operational prediction to support that?

Specific comments:

1. There are many places in the article that are not clearly expressed or improper use of vocabulary, which require major revisions. For examples (not exhausted), the sentences in lines 27-28 (“weaker events”->”smaller rainfall events”?), 35-40 (“where”->“when”?), 68 (“to hit”-> “that hit”?), 77 (“or event magnitude”->“or rainfall magnitude”?), 89 (“whether …” ?), 110-114 (“which are also run …”-> “which are applied for the model run …”?), 117-118 (“doubled the resolution”->“increase the resolution”?), 125-130 (“used include …”-> “used for QPF verification include …”?).
2. The manuscript uses too many abbreviations, which makes the readers hard to get the meaning of the sentences conveniently. Please delete the abbreviations which appear not frequently in the manuscript.
3. Figure 3: Please explain more why after the rainfall events have been categorized into different rainfall magnitude events (A-D), different rainfall thresholds are still needed for each magnitude event.
4. Figure 5: Why not put the CReSS results along with these model results for comparison? Are these models at a resolution of 5 km? If so, the comparison in the TS scores between the CReSS and these models is discounted, as their resolution are different.
5. Lines 556-557: Did these previous results come from forecasts of an equal resolution (2.5 km)?

Review to “Evaluation of Mei-yu Heavy-Rainfall Quantitative Precipitation Forecasts in Taiwan by A Cloud-Resolving Model for Three Seasons of 2012–2014” from Chung-Chieh Wang et al.

This paper evaluates the performance of a convection-permitting model (the Cloud-Resolving Storm Simulator; CReSS) in simulating heavy precipitation over Taiwan during three mei-yu seasons in 2012, 2013 and 2014. The simulations are validated against rain gauges, radar data and NCEP analyses. A well-chosen classification criteria for the events is suggested, dividing the data into segments depending on whether at least 10% of the gauges showed heavy, moderate, some or no rain. For the validation, the authors employ several verification metrics based on contingency tables, namely Threat Score (TS), Probability of Detection (POD), False Alarm Ratio (FAR) and Bias Score (BS).

Although the methods and findings of the manuscript are not totally ground-breaking or unknown to the scientific community, the study presents a compelling analysis of Quantitative Precipitation Forecasts (QPFs) validation. Provided the evaluated model works at a convection permitting resolution and the investigation region is prone to suffering heavy precipitation events with high impact, the manuscript can be relevant to scientists working on the topic as well as stake-holders dealing with the impacts of such events. Therefore I believe there is interest in its publishing although it requires major revisions.

General Comments

1. The conclusions in Section 5 (also in the Abstract) should be further elaborated. There is no need in these sections to write again the numbers of TS, FAR, etc. (e.g. lines 554 to 555 or 563 to 564). By doing so the main conclusions of the paper are hidden e.g. that QPFs are clearly improved by the use of convection permitting in heavy precipitation situations. Please, rewrite the conclusions and the Abstract to clearly point out the findings of the paper.

2. It is not demonstrated that the CReSS model does a better job in orographic precipitation (phase locked) situations compared to transient systems as its stated in the conclusions (Lines 573 to 574). I agree that presumably, predictability is larger in those situations, due to 1) the fact that stationary systems have larger intrinsic predictability and 2) the better representation of the model orography. However the paper does not demonstrate this aspect. The concept “predictability” is lightly used and no quantification is provided (see for instance Hochman et al., 2021 where intrinsic predictability is quantified using dimension and persistence metrics for the systems). Instead only one case is shown (09-10 Jun) and two more cases are mentioned (L480) but no results are provided. From the case shown (09-10) the larger predictability of orographic precipitation is assumed by the fact that the location of the precipitating front (convective line) is not well represented but the TS scores are high (TS=0.4). This is not sufficient proof. It is advised that, given the type of information and analysis provided, the paper focuses on the “accuracy” of the simulation avoiding the analysis on “predictability”.

Specific Comments

Title and Abstract (L18): It is highly possible that the reader is not familiar with what the Mei-Yu season is and when it occurs. Please include this information.

L20-L21, L26-27, L71-73, L79-80: A perfect forecast would show a TS of 1. Why are TS values close to 0.1 a good result then? To support this statement either provide the information about the skill for that score or provide the TS values of the “past results and 5-km models”. Since at this point of the paper the TS has not yet been defined please also include the information that a perfect forecast has a TS=1.

L88. The “dependency property” regarding the link between large events and improvement of the QPFs has not yet been explicitly explained. If I understood correctly these are defined in the papers W15, W16. A brief explanation of what this is, is required here.

L89: The subobjective “further evaluate the model QPFs for larger and extreme events” should be part of the purpose 1) .

L100-101 and L113: What do you mean by “CReSS needs no nesting”. Dynamical downscales always need initial and boundary conditions from other, coarser, model”.

L104-109: More information is needed about the parametrizations used. You need to explicitly mention the shallow convection parameterization scheme and the turbulence scheme. Also referring to Table 1 and W15 and W16.

Table 1: How is the turbulence closure treated? Some models use a TKE 1D parametrization, others use a 3D, etc. What is the case in your simulations?

L228: Why is TS > 0.15 the threshold for predictive skill? Please explain, also if this information comes from previous literature provide the corresponding references.  

L297-303: The statement “In model forecasts, when errors grow and the evolution deviates, the chance to become less rainy (not as favourable) is higher than more rainy, more so in forecasts made earlier at longer ranges.” needs demonstration, either from literature or results.

Figure5: Could the authors elaborate on why are the scores lower for the events during June 2013 (either Day 1 or 2)? This aspect should also be included in the manuscript.

Figure 7: In some panels, it seems as though the scores (TS or BS) are better on the third day of the forecast (day 3, blue line) than for the previous 2 days. Could you please explain this behaviour?

Figure 12: The understanding of the Figure, its caption and explanation provided in the text is incomprehensible. Please rewrite. Why is the number of appearances of the different weather types, proof of the better model performance? Besides, large precipitation totals are usually linked to large-scale systems rather than localized convection, this is already known to the scientific community.

Summary and Concluding Remarks: Please recap the aim of the paper, main steps carried out and briefly the relevance of the study before enumerating the conclusions.

L556: Please, again, include what do you refer to by “compared to previous results”.

L564: When you mention “larger groups” are you referring to events with large coverage? Please reword.

L573: The statement “and such QPFs with high hit rates are clearly very useful for hazard mitigation.” is not documented by your investigation. This was not shown in the paper. Delete.

Writing Comments

L17: What does in real-time mean? Please consider deleting.

L33: Should read: “Quantitative Precipitation Forecasting (QPF)…”

L37: Should read: “… mainly during two periods …”

L63-64: Delete “While the scores at the CWB will be compared with our results later,” and “obviously much”.

L68-69: Model resolutions of 2.2 km are already being used in operational centres, for example de German Weather Service not only in research.

L78: Change “more rain” by “the larger the rain”.

L84: Wang (2015) has already been defined as W15. Please correct.

L85-86: Delete “For mei-yu rainfall in Taiwan, we are certainly keen to find out how this CRM performs, especially for the extreme events.”

L90-91: Delete “To answer these questions above are our objectives”.

L113-114: Delete “which are also run four times a day, each out to 72 h (now 78 h).”

L116: Change: “highly dictated” by “forced”.

Table 1: Provide the grid spacing of the topography in km as well.

L135-136: Include this information about the relevance of the study in the abstract.

L189.: Include the information that BS=1 implies no biases and that BS>(<)1 implies overestimation (underestimation) of the events.

L218-236: Why is the explanation of Fig. 4 before the results concerning Figure 3?

L276: Rephrase “…under-prediction for low…”

L295: The word “serious is not appropriate in this context”

L447: Reword: “… June are compared…”

L496: Rephrase: “… sizes, as shown in Fig. 10…” and “… as an example…” by illustrated.

L580: The sentence “it is also recommended that such events should be examined with caution and proper classification. ” does not bring any information and its colloquial.

Reference

Hochman, A., Scher, S., Quinting, J., Pinto, J. G., and Messori, G.: A new view of heat wave dynamics and predictability over the eastern Mediterranean, Earth Syst. Dynam., 12, 133–149, https://doi.org/10.5194/esd-12-133-2021, 2021.