A Hybrid XGBoost-MLP Model for Credit Risk Assessment on Digital Supply Chain Finance

Round 1

Reviewer 1 Report

The authors demonstrate the superior performance of an XGBoost-multilayer perceptron model for predicting credit risk (bank defaults by Chinese companies).  Feature selection occurs in a first stage with the XGBoost model, followed by credit risk assessment by the MLP model.  Features include not only traditional risk features, but also digital features related to the company’s use of technology. 

Strengths of the paper include its completeness in comparing MLP model with a suite of other classification methods for credit risk analysis, and the inclusion of partial dependence plots to test the influence of each explanatory variable to the model.

Some specific comments below:

• The Introduction is quite long and contains an extensive but excellent literature review/history of the development of SCF and DSCF. Would this better serve the reader as a separate section? 
• The paper is also very long. I am not sure whether the full descriptions of the machine learning methods are needed for this audience.  Perhaps a short description about how the methods work, and where they are best applied, with references to sources for the mathematics.  Also, can the discussion about the different thresholds also be shortened?
• Another error statistic that is becoming more widely used and useful in classification in the Mathews Correlation Coefficient. It has advantages over F1 and accuracy in binary classification.  This could also be computed.
• The major weakness of the paper, as the authors themselves point out, is the limited dataset of Chinese companies. This is understandable.  However, given this, more information about how these datapoints/companies were selected would be helpful.  For example, what is the “Main Board” for someone not familiar with the Chinese stock market? Also, some characterization of the companies, other than SME or core enterprises, would be helpful.  Are they all private companies?  To derive usefulness from these results for a decision-maker, this information is important.
• This brings me to my last comment. The focus of the paper is on model building and the strengths and weaknesses of the various models and analytical tools in assessing credit risk. However, given the effort (and number of pages) given to feature selection, I was expecting more discussion which contributing features were most important, and what those results mean for assessing credit risk in practice.  Specifically, what does the paper conclude in summary for decision-makers who must make lending decisions about which factors are most important?  The conclusions only state that feature selection is found to be important, and that digital features are found to be important.  Some summary is needed about which features are most valuable in assessing risk, especially given the significant work to assess feature important at the various thresholds.
• The paper needs a more complete editing. Some of the phrases did not make sense in standard English. 

Line 55 – derivate?

Line 201 – use of adjective “wealthy”?

Line 212, 517-  “DSCF exists gap”?

Line 285 – Two Figure 1’s.   Should line 285 reference be to Figure 2?

Line 382 – maps, not map

Line 519 – Missing word?

And so forth….

Author Response

Dear reviewer, 

Thank you for your previous comments and advice. Those comments are all valuable and helpful for revising and improving our paper. We have revised the manuscript accordingly. The main corrections in the paper and the responses to the reviewer’s comments are as follow:

Regarding the overlong introduction, we have shortened the introduction and created the new section 2 of the literature review which include two subsections: 2.1 Background of DSCF(Line 101-174) and 2.2 Machine Learning and Credit Risk models (Line 175-289).

Regarding the oversize full paper, we simplified the full description of the machine learning method into summarised description with reference (see Line 380-441) and made the derivation of the XGBoost method more compact (see Line 329-365).  Further, we removed the redundant robustness results and left the most significant robustness result of Test=0.1. We have added the explanation that we would provide extra results if required in Line 707.

Regarding the additional error statistic method, we computed the score of the Matthew Correlation Coefficient. The updated description of the Matthew Correlation Coefficient is located in Line 520-524.

In addition, the updated MCC score could be found in Table 4 (Line 552), Table 5 (Line 563), Table 6 (Line 605), Table 7 (Line 684).

Regarding the incomplete provision of dataset information, we provide more basic information and reference in the footnote (see Line 451-458).  In addition, we updated the characterisation of selected SMEs in Line 464.

Regarding the lack of discussion of contributions, we added some intuitive conclusions and recommendations for different decision-makers. (See Line 723-768)

In addition, we proofread and revised the following typo and unstandard words and format:

Original Line 55- “derivate"

Revised Line 104-"The role of inter-firm coordination and facilitation through new supply chain enterprises, leading to the derivation of a supply chain production model."

Original Line 201- “wealthy”

Revised Line 56- "The motivation of this paper is driven by three aspects: Firstly, the model for credit risk assessment is various and ambivalent."

Original Line 212, 517- “DSCF exists gap”

Revised Line 67, 472- “there are gaps in the research on DSCF”

Original Line 285- Figures misleading 

Revised Line 292- Updated 

Original Line 382- “map”

Revised Line 382- Updated “maps”

Original Line 519- Missing word

Revised Line 474- "Thirdly, there are few data treatments in the existing literature that focus on feature selection."

Finally, thank you again for your suggestions. According to your suggestions, we have made corresponding changes and adjustments in the manuscript. The amendment to the question chapter can be found in the new version.

Regards,

Yixuan

Reviewer 2 Report

High-quality article. Very well developed methodically and methodologically. The authors have made every effort to explore the topic discussed. A well-prepared literature review as well as a presentation of the results. The whole thing is well perceived by the potential reader. It is very important after all. It is also important that the verification of the model is carried out extensively, on a large sample, which confirms its effectiveness. It should be emphasized that the authors are experts in the subject and specialists in this field.
Congratulations on your good publication.
If I can :) I would suggest better explain the novelty and significance of your findings in conclusions.
Short and to the point: well done :).

Author Response

Dear reviewer, 

Thank you for your summary. We really appreciate your effort in reviewing our manuscript. Those comments are all valuable and encouraged to our future works. We have carefully provided the modifications following these comments and suggestions as follows: 

Regarding the lack of novelty and contributions in conclusion, we have listed some intuitive and constructive recommendations for different decision-makers (see Line 747-768) and we also added novelty description about the digital supply chain features in Line 723-746.

Finally, thank you again for your positive comments and suggestions.

Regards,

Yixuan