Balancing Techniques for Advanced Financial Distress Detection Using Artificial Intelligence

Round 1

Reviewer 1 Report

While the paper addresses an important topic and employs advanced techniques, there are several areas where improvements are needed to enhance clarity, comprehensiveness, and impact. Addressing the outlined suggestions will strengthen the overall quality of the paper and contribute to its significance within the field of financial distress detection using artificial intelligence.

Abstract: The abstract provides a succinct overview of the paper's content; however, it would be beneficial to avoid using abbreviations. This ensures clarity and accessibility for readers who may not be familiar with the specific terminology.

Introduction: While the introduction outlines the scope of the research, it falls short in addressing the underlying rationale for the investigation. The authors need to articulate the research gap more explicitly. What previous studies have explored this topic, and what gaps in knowledge do they leave? By elucidating the motivations behind the study, such as the limitations of existing methodologies or the evolving nature of financial markets, the authors can establish the significance of their work more effectively. Additionally, it would be valuable to clarify how this study contributes to the existing body of research. What unique perspectives or methodologies does it offer that distinguish it from previous studies?

Literature Review: The literature review appears to lack completeness, as indicated by the notation of missing tables. Providing these tables would enhance the comprehensiveness of the review and facilitate a deeper understanding of the existing research landscape. Moreover, it is essential to ensure that the literature review covers relevant studies comprehensively, highlighting both the methodologies employed and the findings reported. This will enable readers to grasp the broader context within which the current study is situated.

Methodology: In the methodology section, it is unclear how the time-series data of companies were treated. Providing a detailed explanation of the data preprocessing steps, particularly concerning time-series analysis, is crucial for ensuring the reproducibility and validity of the study's findings.

Discussion: The absence of a discussion section is notable, as it represents a missed opportunity to contextualize the results within the broader literature and to elucidate the study's contributions and implications. The authors should engage in a critical analysis of their findings, comparing them to those of previous studies and identifying any discrepancies or consistencies. Furthermore, they should articulate how their research advances the field, whether by introducing novel methodologies, offering new insights, or addressing limitations of prior work. By providing a comprehensive discussion, the authors can enhance the overall impact and relevance of their study.

Conclusion: The conclusion provides a summary of the study's findings; however, it would be strengthened by reiterating the significance of the research in addressing the identified gap in the literature.

Minor

Author Response

Your comments are very valuable, and we have taken them all into account.

Abstract: we added explanations.

Introduction: We supplemented the introduction to reveal the research gap and the motivations behind the study. Additions are marked in red.

Literature Review:  The tables are revised, the new researchers are added. Additions are marked in red.

Methodology: You can find it in line 335:341

Discussion: We added the discussion section.

Conclusion: We hope you will find the missing part in the discussion and in lines 952-953, 960-961

Reviewer 2 Report

The authors should characterise the market, although they indicate the type of company and the country in which the research is applied, it would be important to consider larger markets that imply greater market efficiency. 

The authors should indicate why the reason for looking at these variables is of importance, many of the bases derive from the decision making itself and not from the market conditions for the application of the techniques they mention. 

The authors should decide what to talk about in terms of having relevant information rather than checks and balances. 

It should be defined what accuracy implies from the point of view of finance and not from the point of view of numbers. 

The graphs look good and could contribute to the research, but should be refocused on what ML techniques can allow.

The manuscript should be focused, and not provide so much different information because in the end it muddies what it wants to show. 

Author Response

1. The authors should characterise the market, although they indicate the type of company and the country in which the research is applied, it would be important to consider larger markets that imply greater market efficiency.

Now we have accessible data from Lithuania SME’s and we based our research based on this information. It was mentioned in all main parts: introduction, data, and conclusion.

2. The authors should indicate why the reason for looking at these variables is of importance, many of the bases derive from the decision making itself and not from the market conditions for the application of the techniques they mention.

Done. You can find it in the introduction, lines: 61:72

3. The authors should decide what to talk about in terms of having relevant information rather than checks and balances.

We understand that the question comes from the appendix part. There was a misunderstanding before the appendix part was given before the reference. Now it is fixed. The appendix shows the features of the data set used in analyses. Because in some cases it is periodic data, its frequency has been shown with checkmarks. However, we left it to machine learning dimensionality reduction algorithms to find out the most relevant information on the FD.

4. It should be defined what accuracy implies from the point of view of finance and not from the point of view of numbers.

Done. You can find it in the discussion part, lines 902-903

5. The graphs look good and could contribute to the research, but should be refocused on what ML techniques can allow.

We understand your question from ML evaluation metrics perspective. We use several ML metrics (see 4.7 section), mainly focusing on AUC score. The max of all analyzed metrics is 1 or 100 proc., which shows that all FD identification cases are classified correctly.

6. The manuscript should be focused, and not provide so much different information because in the end it muddies what it wants to show.

Done. We have written the discussion part, we hope it will help readers to understand our research better. Our main focus is on FD research methodology creation. The usability of different methods helps us to understand which methods are more or less suitable for this precise problem. 

Reviewer 3 Report

In order to improve the quality of this work, some comments have been given as below.

1.In fact, there are lots of algorithms have been published for handling imbalanced data, such as SMOTE, TOMEK-Links, NearMiss and other Cost-Sensitive Training methods including SVM. Authors have to compare their method to these traditional techniques no matter in classificaiton performance and training time.

2.Authors seem to find a optimal combination from limited methods. However, the perfromance of such kind combination heavily depends on the selected training set. In other works, the obtained weights and combination of methods will be different when we employed different training data. Without trial-and-error, authors should provide a algorithm or procedure to make sure readers can obtain optimal solutions.

3.Literature review is sinsufficient to let readers understand the developmemnt of class imbalance problems.

4.In section 3.2, authors merely used binary classifiers. But, the used data set has more than 2 class labels. Authors should use multi-class classifiers. In fact, the used machine learning also could handle multi-class classification problem.

5.RQ1~RQ5 should be moved to introduction to enhance the research gap.

6.From the results, we only prove the selected features of feature selection methods are good for classifying data. How do we know the select feature set is optimal? 

7.From page 36~58, they should be appendix. Please make them clearly or just remove them.

8.It's not a survey paper, 198 references are too many. Please select some important references.

This paper should be edited by an English speaker.

Author Response

1. In this study, we employ data-level approach techniques and plan to expand the research in the future by using algorithm-level techniques. The SMOTE-Tomek link was used in the study; however, the Tomek link was not separately applied. At the beginning of the experiment, we applied the Tomek link; however, the data imbalance changed only to 2%. For this reason, further experiments were not continued due to the remaining significant imbalance.
2. You can find it in the discussion part, lines: 934-947
3. You can find it in lines: 164-171
4. Good idea for further research development. Now, we want our findings to be as much comparable as possible in this area of research.
5. You can find in lines: 96-103
6. It needs further investigation, you can find it in lines: 942,990.
7.  
8. We do not see the relevance of the comment. One of this article's ideas is to help other researchers find relevant information in such fields as financial distress (FD) class condition analyses, feature inclusion in the data set, and use of imbalance techniques in FD field.

Round 2

Reviewer 2 Report

The authors improved the manuscript with the recommendations made, but it would be important for them to adjust the graphs to make them meaningful in the research. 

Author Response

Maybe you could specify which graph is unclear, what it is missing. And how would you recommend changing it?

Reviewer 3 Report

In order to improve the quality of this work, some comments have been given as below.

1.Author presents 5 research questions. But there is not any academic evidence to support these 5 questions are worth to study. Authors needs to add some literature review then make a concluding remark to support these 5 RQs are truly research gap.

2.Since authors present 5 RQs, they have to add a summary table to provide all answers of these RQs. 

3.From cited reference [1], since it's reported that GBoost model with random oversampling (ROS) performs the best, what did authors try to find optimal solutions or combination again?

4.As we know, the performance of combination of feature selection techniques, machine learning approaches, and imbalanced data skills heavily depends on the selected data. Different data set will result in totally different performance. So, it's difficult to make a general conclusion regarding toto find an optimal solution of the combination of feature selection techniques, machine learning approaches, and imbalanced data skills. At least, in conclusion section, authors should mention that the given conclusion merely for this case.

The quality of English is enough to be read.

Author Response

Thank you for your comments and feedback. I appreciate your input and have carefully considered your suggestions. Below are my responses to your comments.

1. It is done. You can find in lines: 161:204
2. You will find in the following lines 1011:1030
3. Researchers analyze not only different methods but also apply them to different datasets. Similarly, here these researchers found an optimal solution from the methods they analyzed for the Polish dataset. Meanwhile, our dataset under analysis is completely different.
4. It is done.  You can find inline 1007:1008

Round 3

Reviewer 2 Report

The authors could have improved some of the graphs to better show the results.

Author Response

Thank you for your comments.

The graphs and their description have been improved. You can find the description in the 363:365 lines.

Graphs based on Table 10 are provided; please refer to fig. 9, 13, 16, and 17.

We chose the AUC metric for comparison because it is the most popular in these studies and is suitable for an unbalanced testing sample, considering both classes equally. For this reason, all the figures are visualized based on AUC, while other comparative metrics are shown in Tables 10 and 11.

Reviewer 3 Report

All of my comments have been fully considered in this version.

Author Response

Thank you :)