Comments on: Modelling Flood Losses to Microbusinesses in Ho Chi Minh City, Vietnam by Buch et al. (2024-2340)

This paper aims to identify the drivers of flood losses in microbusinesses by employing a Conditional Random Forest on survey data collected from microbusinesses in Ho Chi Minh City in Vietnam. Based on the drivers identified, probabilistic loss models (non-parametric Bayesian Networks) were developed using a combination of data-driven and expert based model formulation. The transferability of this models was assessed by applying data from a different city to evaluate their broader applicability.

I have read the paper with great interest, and the main objective addressed by the manuscript is within the scope of the journal. Nevertheless, major revisions are necessary to make a few points clearer and I recommend accepting it only after these revisions.

Major comments:

I believe the paper could benefit from separating the results from the discussion to enhance clarity. As someone without extensive expertise in ML algorithms, I found it challenging at time to connect the information in the figures and tables with the text. For instance, on pages 10 and 11, Figure 3 is only referenced once, and while the authors discuss correlations among variables, they do not always provide specific numerical values from the figures. The paper contains a substantial amount of results, which makes it difficult to easily connect the text with the accompanying figures and tables. Please, check the figures axes names. Sometime you start with capital letter and sometimes with small

Additionally, I recommend simplifying certain figures (e.g. Figure 5) or providing more detailed descriptions of them in the text. Given that this journal focuses on natural hazards research and is accessed by readers who may not be experts in ML algorithms, a clearer structure with separate results and discussion would aid.

In the methodology section, I suggest that the authors provide further clarification in certain areas. For example, in the abstract, content losses are reported 317 and business interruption losses as 361. However, in the section presenting the data is noted that 250 responses were collected resulting in 397 loss records in the HCMC and for Can Tho, responses were received from 373 microbusiness, of which 313 provided information on losses. It is unclear how these numbers were derived and calculated. So, please check the numbers and a more detailed explanation of the methods used would be helpful.

Additionally, presenting the equations for the error formulas mentioned in lines 147–150 would enhance clarity. There are also methods referenced in the results that are not described in the methodology section. For instance, cumulative distribution functions (CDFs) are discussed in line 357, yet they are not explained in the methods. Including these details would improve transparency and ensure a more complete understanding of the approach used.

In the introduction, the authors place emphasis on the case study to motivate the analysis. It may be beneficial to move the detailed description of the case study to a separate section, allowing the introduction to focus more directly on the research gaps. This would help to clearly establish the broader motivation and context for the study before delving into the specifics of the case study.

The discussion and conclusion sections could be enhanced by further exploring how the findings may be utilized by other experts and their implications for flood risk management. Expanding on these aspects would clarify the broader relevance of the outputs.

The manuscript addresses the topic of flood losses to microbusinesses in low- and middle-income countries and introduces an application of Conditional Random Forests for feature selection and Bayesian Networks to model flood losses, bridging gaps in existing methods that primarily cater to larger firms or macro-level analyses. In addition, it provides empirical evidence for key drivers of flood losses (e.g., water depth, building age, monthly revenue) and evaluates the models’ transferability to another city (Can Tho), highlighting its potential for broader regional application. The approach is good in general; however, the manuscript should provide more details for readers to understand and reproduce this approach in other case studies.

1. Provide more explanation about machine learning techniques of “Bayesian Network” and “Conditional Random Forest” and briefly explain why these methods suit the study.

2. Case study selection: The manuscript justifies the choice of Ho Chi Minh City and Can Tho as case studies based on their high flood risk and economic importance. However, additional clarification is recommended (e.g., explain how these cities represent other flood-prone urban areas in Vietnam or Southeast Asia).

3. Defining sample size: The manuscript mentions the number of surveyed microbusinesses but lacks a detailed rationale for the sample size determination (What statistical considerations or sampling techniques were used to determine the sample size?).

4. Ethical application for surveys: Ethical considerations for conducting surveys need to be explicitly addressed to ensure transparency and compliance with research standards.

5. A flowchart summarizing the methodological workflow (from data collection to modeling and validation) can improve clarity for readers unfamiliar with machine learning or Bayesian methods.

6. A GIS map that shows locations of surveyed microbusinesses can provide clear context and can be combined with flood hazard maps to give an overview of hazard and impact, which can be suitable for visualization and dissemination.