Moody’s Ratings Statistical Forecasting for Industrial and Retail Firms

Round 1

Reviewer 1 Report

Thank you for the opportunity to review the paper Moody’s Rating Statistical Forecasting for Industrial and Retail Firms. Reviewing and testing existing rating models, and proposing new ones tends to be an important topic both for academics as well as practitioners. Following you can find some remarks concerning your paper:

Introduction/Literature review: In the introduction the author/s give a clear sense of what the main aim or purpose of the paper is – and that is to give several alternatives for rating model for Moody's long-term companies’ ratings for industrial and retailing companies using Artificial Intelligence methods (neural networks). Literature review is done systematically, moreover author/s give a thorough overview of rating forecasting literature.

Data and Method: In the paper author/s use neural networks. I find that this part of the paper needs to be improved in the sense that the methodology needs to be described in the more detail. You state that you are using MLP neural networks and that is that, furthermore you state that you used SPSS and mention that some parameters were changed, network layers and neurons, however, you do not state which ones. As for evaluating the success of the network you do not state what are you using (I figure it was a classification) but this should be stated in the paper. Additionally, you could explain in more detail following - as I see in your paper you have a lot of classes and that is why the accuracy of the models is low. Further, table 7 (matrix of confusion), shows that some classes are very poorly represented so maybe you could explain it a bit more. Selected sample using a random selection in Bloomberg corporate database is adequately used for this research.

Discussion/ Conclusion: The conclusion is done correctly. What I would like to see is that author/s reflect what the contributions of the research are in more detail. What is the new knowledge? Way is this model important/better. You need to apostrophize your contribution to the field.

I wish you a lot of success in your future work.

Author Response

Methodology:

As suggested a brief description of the MLP neural network has been added with the description of the topology adopted. When forecasting ratings, effectively it is a classification problem, as the results in table 7 are presented, as the ordered classes are discrete. It has been clarified in the manuscript. Some classes include a small number of companies. This is natural, as most industrial firms are rated in the ranges between and Caa1. There are just a small number top rated or in the lower Caa classes; the mortality in the later is quite high. As referred in table 3, the number of companies top rated or at the bottom of the scale are scarce; this happens in the whole population of firms, and, consequently in the sample; also, when dividing it in training and test sets, this situation is more evident. To attain the proposed model, several alternative topologies were selected, with one or two hidden layers, and different number of neurons in these. In addition, several previous normalizations in the input layer were used and different activation functions (the hyperbolic tangent in the hidden layer was finally adopted); 80% of the firms were used to train the network, and the remaining 20% out of sample cases were used for validation purposes. SPSS package disposes of several parameterization alternatives, which can be useful in dealing with different topologies and specifications, such as selecting the number of hidden layers (one in the proposed network), and limiting the number of hidden neurons.

Discussion/Conclusion:

Some additional comments have been included, and the reference of a classification in all rating classes, as a novelty on previous works, where the ratings are aggregated in a low number of classes (prime/no prime, or 3 to 5 classes). Of course, when the objective is to forecast the precise rating of a company several problems arouse; there is a need of more abundant statistical information, and, the situation in both tails of the rating scale is complicated by the fact that there are just a few companies top rated, and, on the other end, there is a high mortality of firms, complicating the follow up of companies several years. As a practical result, with the proposed methodology, and having access to statistical information, such as the provided (as a cost) by Bloomberg, it is possible to assess the reliability of ratings provided by CRA's, or to estimate a rating for a new company without having to support the costs of the rating process.

Reviewer 2 Report

Dear Authors

This is a very interesting article. In my opinion, changes should be made according to the structure of articles published in mdpi. The abstract should include the purpose, methodology and conclusions. The 'Methodology' section should describe what constituted the research matter and its analytical methods used. In addition, the research questions should be formulated, the limitations and future directions of the research.

Author Response

Abstract:

It has been expanded including some remarks about the objective, methodology and conclusions. Regarding the structure of the articles in MDPI, in the Economies journal, papers tend to include abstracts without a separation in these points as such, differing from other journals, and this why it has the structure presented.

Objectives: research questions and limitations:

A brief description of the MLP neural network has been added with the description of the topology adopted. When forecasting ratings, effectively it is a classification problem on discrete ordered classes, and the results in table 7 are presented as such. It has been clarified in the manuscript. Some classes include a small number of companies. This is natural, as most industrial firms are rated in the ranges between and Caa1. There are just a small number top rated or in the lower Caa classes; the mortality in the later is quite high. As referred in table 3, the number of companies top rated or at the bottom of the scale are scarce; this happens in the whole population of firms, and, consequently in the sample; also, when dividing it in training and test sets, this situation is more evident. To attain the proposed model, several alternative topologies were selected, with one or two hidden layers, and different number of neurons in these. In addition, several previous normalizations in the input layer were used and different activation functions (the hyperbolic tangent in the hidden layer was finally adopted); 80% of the firms were used to train the network, and the remaining 20% out of sample cases were used for validation purposes. SPSS package disposes of several parameterization alternatives, which can be useful in dealing with different topologies and specifications, such as selecting the number of hidden layers (one in the proposed network), and limiting the number of hidden neurons.

Discussion/Conclusion:

Some additional comments have been included, and the reference of a classification in all rating classes, as a novelty on previous works, where the ratings are aggregated in a low number of classes (prime/no prime, or 3 to 5 classes). Of course, when the objective is to forecast the precise rating of a company several problems arouse; there is a need of more abundant statistical information, and, the situation in both tails of the rating scale is complicated by the fact that there are just a few companies top rated, and, on the other end, there is a high mortality of firms, complicating the follow up of companies several years. As a practical result, with the proposed methodology, and having access to statistical information, such as the provided (as a cost) by Bloomberg, it is possible to assess the reliability of ratings provided by CRA's, or to estimate a rating for a new company without having to support the costs of the rating process. Some comment about the limitations in the proposed models for forecasting is included, and about the need to be able to develop further modelling.

Reviewer 3 Report

The paper under title “Moody’s Rating Statistical Forecasting for Industrial and Retail Firms”, deals with a very interesting topic on how to create a credit worthiness model (similar to Moody’s) using only publicly available data. The study is using industrial and retail firms as the sample. The study could be of interest to an international audience, however it lacks a theoretical background and a clear contribution to the literature, which make them unsuitable for publication in its current form. So, I urge the author(s) to consider the following comments and revise-resubmit the paper.

1. At first the study needs to be re-edited and checked for syntax and grammatical mistakes. Misspellings are also a lot throughout the paper.
2. On the line 22 author(s) refer to the “variability of the global situation”. What do they mean by that?
3. Author(s) need to explain and support the focus on the industrial and retail sectors. Why they have been chosen, and what are their specific features warranting their examination. Connected with that is the fact that the introduction lacks a clear contribution of the study to existing literature.
4. Also, author(s) refer to the case of opinion shopping (where the CRA is paid by the one who they evaluate) but without any theoretical background regarding this phenomenon. I think the following studies (even though they are indicative) they can help autho(s) to create a sound theoretical framework:

-Newton, N.J., Persellin, J.S., Wang, D., Wilkins, M.S. Internal control opinion shopping and audit market competition (2016) Accounting Review, 91 (2), pp. 603-623. DOI: 10.2308/accr-51149

-Chen, F., Peng, S., Xue, S., Yang, Z., Ye, F. Do Audit Clients Successfully Engage in Opinion Shopping? Partner-Level Evidence (2016) Journal of Accounting Research, 54 (1), pp. 79-112.

DOI: 10.1111/1475-679X.12097

-Goh, L., Gupta, A. Executive compensation, compensation consultants, and shopping for opinion: Evidence from the United Kingdom (2010) Journal of Accounting, Auditing and Finance, 25 (4), pp. 607-643. DOI: 10.1177/0148558X1002500407

-Lu, T. Does opinion shopping impair auditor independence and audit quality? (2006) Journal of Accounting Research, 44 (3), pp. 561-583. DOI: 10.1111/j.1475-679X.2006.00211.x

-Lennox, C. Do companies successfully engage in opinion-shopping? Evidence from the UK

(2000) Journal of Accounting and Economics, 29 (3), pp. 321-337. DOI: 10.1016/S0165-4101(00)00025-2

5. On the literature review section it will be useful for the reader to include a final paragraph on the novelty of your research design (or method proposed) relative to other studies discussed previously on this section.
6. On the research design section author(s) need to be more specific on the methodology they will follow. It is very descriptive without the reader to understand the steps that have been followed for producing the results. On section 5 also author(s) mention the MLP method, but they need to elaborate further on the specifics of this method and how it works.
7. Moreover, all variables used on page 9 (lines 314-324) need to be supported by previous studies. So relevant citations need to be added there.
8. Figure 4 needs to be explained, there is no explanation, and it is confusing by itself.
9. Author(s) need to add some policy implications of the study, limitations and directions for future research on the conclusion part.

Author Response

Syntax and spelling:

Has been checked and corrected

Sentence confusing:

The sentence about the ‘variability of the global situation’ has been changed to express the changing environment in its financial an economic aspects

 

Support on the sectors analyzed:

The two sectors show a distribution of ratings quite similar. The focus could have been limited to industry, but it was convenient to increase the sample size, as the forecast was extended to all rating classes, and in some categories there are nor many companies. Comparing different sectors, their distribution shows different patterns of variation, but in these two, without having to use an additional dummy variable. Additional references have been added.

About opinion shopping:

This is a topic quite wide, and that needs additional information not used in this case, as it was not the central topic. It comes afloat when trying to forecast ratings, as there is a large literature about it; when Fitch entered the rating market, in theory this additional agent could have increased the competition, and affect prices; but the results were an inflation of rates, consequence of the opinion shopping. The opinion shopping of firms when choosing a CRA or an auditor has been treated in the literature; Newton et al. (2016) try to measure this phenomenon, and its occurrence, particulary in competitive audit markets. Chen et al. (2016) detect an increase of favorable audit opinions after switching auditors, while Lu (2006) finds that neither the predecessor nor the successor consultant are affected by these changes. Several indices to quantify the inflation of ratings have been proposed, and, it is Moody’s where this phenomenon is less present. Some comment has been added with the suggested bibliography.

About the novelty:

In most previous works, ratings are predicted on an aggregate scale (prime, not-prime, or in three to five classes. Here, it is the full range of rates that are forecasted, and using non-linear data-based methods, such as neural networks. Also in the results this topic is specified.

Methodology:

As suggested a brief description of the MLP neural network has been added with the description of the topology adopted. When forecasting ratings, effectively it is a classification problem, as the results in table 7 are presented, as the ordered classes are discrete. It has been clarified in the manuscript. Some classes include a small number of companies. This is natural, as most industrial firms are rated in the ranges between and Caa1. There are just a small number top rated or in the lower Caa classes; the mortality in the later is quite high. As referred in table 3, the number of companies top rated or at the bottom of the scale are scarce; this happens in the whole population of firms, and, consequently in the sample; also, when dividing it in training and test sets, this situation is more evident. To attain the proposed model, several alternative topologies were selected, with one or two hidden layers, and different number of neurons in these. In addition, several previous normalizations in the input layer were used and different activation functions (the hyperbolic tangent in the hidden layer was finally adopted); 80% of the firms were used to train the network, and the remaining 20% out of sample cases were used for validation purposes. SPSS package disposes of several parameterization alternatives, which can be useful in dealing with different topologies and specifications, such as selecting the number of hidden layers (one in the proposed network), and limiting the number of hidden neurons.

Variables used:

Many of the variables used are referenced on section three, when commenting the literature, and linking these to several previous publications

Figure 4 explained

Several explanations have been added, and the reference to the figure (that was not as is should have been), changed.

Policy implications, limitations, future directions:

Some applications and directions of development have been included in the last section. Comments have been included, and the reference of a classification in all rating classes, as a novelty on previous works, where the ratings are aggregated in a low number of classes (prime/no prime, or 3 to 5 classes). Of course, when the objective is to forecast the precise rating of a company several problems arouse; there is a need of more abundant statistical information, and, the situation in both tails of the rating scale is complicated by the fact that there are just a few companies top rated, and, on the other end, there is a high mortality of firms, complicating the follow up of companies several years. Some ratios on public accounts can be used as additional information. As a practical result, with the proposed methodology, and having access to statistical information, such as the provided (as a cost) by Bloomberg, it is possible to assess the reliability of ratings provided by CRA's, or to estimate a rating for a new company without having to support the costs of the rating process. Some comment about the limitations in the proposed models for forecasting is included, and about the need to be able to develop further modelling.

Round 2

Reviewer 3 Report

I would like to thank the author(s) for their effort to accommodate all comments raised. I believe the paper has improved materially and author(s) addressed all issues mentioned on the review process. I think the paper could be published in "Economies" after a general check on potential syntax or editing mistakes on the text.

Author Response

Dear Reviewer;

Thank you for your comments and advice to improve the quality of our work. After a double check on language syntax and edition mistakes we have corrected them.