Prediction of Financial Distress Risk for Green Enterprises from the Perspective of Climate Resilience
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsManuscript ID: systems-4398911
Title: Prediction of Debt Default Risk for Green Enterprises from the Perspective of Climate Resilience
Summaries: The study focuses on improving debt default risk prediction for green enterprises, which face unique climate-related risks not captured by traditional financial models. It introduces a non-financial “climate resilience indicator” built from financial data and annual reports using NLP. Results show this indicator significantly improves prediction accuracy, has both direct and indirect effects on default risk through financial factors, and enhances traditional models, offering a better framework for early warning of financial risk in green firms.
The manuscript has been reviewed; however, there are some suggestions to revise the manuscript carefully.
The abstract has a major problem because it does not clearly cover the main objective of the study. It reads more like an introduction or a general explanation rather than a complete abstract. Please carefully revise it and write a concise abstract of around 170–200 words that clearly presents the research purpose, methodology, key findings, and main contributions.
The introduction is well explained in detail; however, I found that some references are missing where the authors discuss previous studies and related research details. Please add appropriate citations, especially in the sections around lines 82, 93, and 113, where references are currently missing. You can include the following sources because they are closely related to the authors’ discussion: (https://www.mdpi.com/2071-1050/17/15/7094 ), (https://www.inderscienceonline.com/doi/abs/10.1504/MEJM.2024.135152 ), If other studies are more closely related to the topic, please identify and include them in the introduction section to strengthen the literature review.
The Literature Review section is too brief and feels rushed. It would be better to expand it with some additional content and include more relevant studies to strengthen the discussion and avoid overlooking important prior research.
Please rewrite the Methods section. It is not clear what data sources were used or what kind of studies were included.
Please add a description to every table.
Please expand the discussion section, as it is incomplete. Include a more detailed analysis and discussion of the relevant studies.
Please check the Conclusion section also.
Author Response
Thanks for your encouragement. We have enclosed the responses to the editors and referees’ comments, point by point, in the following section. We include our changes in our revised manuscript. A list of all the changes we made to the manuscript is listed in “Response to reviewer 1.docx”.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsThe paper examines whether a firm-level climate resilience indicator extracted from annual report text can improve the prediction of debt default risk for Chinese listed green enterprises. The study combines annual report text with financial indicators for A-share firms in green industries over 2015–2024, uses ST status as the outcome, and compares several machine-learning models, with XGBoost reported as the best performer.
The topic is important, and the paper could be of interest to researchers working on climate finance, corporate disclosure, credit risk, and applied machine learning. It may also be relevant to policymakers and practitioners if climate-related narratives in annual reports provide useful early-warning information. At the same time, several central parts of the research design and interpretation need substantial revision before the paper’s contribution can be assessed with confidence.
My first major concern is the outcome variable. The manuscript is presented as a study of debt default, but the coded event is whether a firm receives ST status. These are not the same outcome. ST status reflects financial abnormality and regulatory treatment; it is not a direct measure of default in the usual credit-risk sense. This is more than a matter of terminology because the title, abstract, stated contribution, and policy implications all rely on the concept of default. If the authors can obtain direct debt events, such as bond defaults, loan nonpayment, missed interest, restructuring, or another credit outcome, they should use them. If they cannot, the paper should be renamed and reframed throughout as a study of financial distress prediction rather than debt default prediction.
My second major concern is the timing of the prediction exercise and the risk of information leakage. The paper randomly divides firm-year observations into training and test samples. In a panel setting, this can place observations for the same firm in both samples in nearby years and may overstate out-of-sample performance. The concern is greater because the manuscript appears to conduct feature screening and compression before the final evaluation. If that is the case, the test sample is not a clean holdout. The analysis should be redesigned as a true forecasting exercise. Annual report information from year t should be used to predict an outcome in year t+1 or later, and model selection, feature selection, dimensionality reduction, and tuning should all be carried out within the training folds.
My third major concern is the construct validity of the text-based measure. The manuscript defines climate resilience as the product of climate-word intensity and sentiment in climate-related paragraphs. At present, the paper does not establish convincingly that this measure captures resilience. More climate-related language may reflect stronger resilience, but it may also reflect greater exposure, heavier disclosure requirements, discussion prompted by financial stress, or a strategic choice to emphasize climate issues. Positive language may likewise reflect tone management rather than underlying capacity to withstand climate shocks. The current measure may therefore combine several different concepts.
The authors need stronger validation, for example by comparing the measure with a hand-coded subsample, known climate events, external disclosure scores, or future real outcomes. The paper should also explain more clearly how the measure relates to established text-based research in finance. Li (2010) shows how forward-looking language in corporate filings can be measured carefully, while Hassan et al. (2019) provide an example of linking a narrative risk measure to economically meaningful outcomes. Mugerman, Sade, and Shayo (2014) are also very useful for the paper’s positioning because their study shows how institutional context and non-fundamental channels can shape financial decisions. That perspective is relevant when interpreting a disclosure-based measure as a signal rather than as a purely mechanical word count.
My fourth major concern is that the paper moves too quickly from prediction to mechanism. SHAP values and partial dependence plots are useful for describing how a fitted model uses the observed variables, but they do not identify a transmission channel. The manuscript currently relies on these tools to support a chain from climate resilience to financial conditions and then to debt default. That conclusion is too strong for the evidence presented. The same issue applies to the discussion of transmission effects. If the authors want to make a mechanism claim, they need a design with explicit timing, careful controls, and a method suited to mediation or path analysis. Otherwise, this part of the paper should be presented more cautiously as descriptive evidence about model behavior and associations in the data.
My fifth major concern is the econometric reporting in the nonlinear and interaction analyses. The manuscript states that it uses polynomial logit models, but the tables report R-squared, residual sum of squares, and F-statistics, which are not the standard outputs for logit estimation. This makes it unclear what models were actually estimated. The result for the log specification is especially hard to reconcile with the rest of the table. These analyses should be re-estimated and reported in a standard form, with the coefficients, standard errors, and fit measures appropriate for nonlinear binary-response models, together with a clear description of each specification.
My sixth major concern is model evaluation. Several models fit the training data almost perfectly, which raises a clear concern about overfitting. In this setting, accuracy, precision, recall, and F1 are not sufficient on their own. Because the outcome is imbalanced, the paper should also report ROC-AUC, precision-recall performance, calibration, and sensitivity to the classification threshold. It should test whether the improvement from adding the climate variable is statistically meaningful rather than relying only on raw differences in performance. A comparison with a simple benchmark model would also help readers judge the economic value of the more complex methods.
My seventh major concern is the sample design. The paper combines several green industries that differ in business models, policy exposure, capital intensity, and financing conditions. The authors should explain more clearly why these firms belong in one prediction sample and should examine the extent to which the results differ across groups. The regression analysis would also benefit from year effects, industry effects, and a clearer set of firm characteristics that are relevant to distress risk. More generally, the paper presents a debt-default application, while the data are firm-year observations with a broad distress label. This mismatch should be addressed directly.
My eighth major concern is the paper’s positioning for a broad economics and finance audience. In its current form, the manuscript reads mainly as a comparison of prediction models followed by several additional analyses. I believe the contribution would be clearer if the paper focused on one general question: whether climate-related narratives in mandatory filings contain incremental information about future credit deterioration beyond standard accounting ratios. That question can interest readers beyond a narrow green-finance setting. To answer it persuasively, however, the paper needs a more appropriate outcome, a clean forecasting design, stronger validation of the text measure, and a firm distinction between predictive evidence and causal explanation.
These concerns are substantial, but they are also concrete.
References
Hassan, T. A., Hollander, S., van Lent, L., & Tahoun, A. (2019). Firm-level political risk: Measurement and effects. Quarterly Journal of Economics, 134(4), 2135–2202. https://doi.org/10.1093/qje/qjz021
Li, F. (2010). The information content of forward-looking statements in corporate filings—A naïve Bayesian machine learning approach. Journal of Accounting Research, 48(5), 1049–1102. https://doi.org/10.1111/j.1475-679X.2010.00382.x
Mugerman, Y., Sade, O., & Shayo, M. (2014). Long term savings decisions: Financial reform, peer effects and ethnicity. Journal of Economic Behavior & Organization, 106, 235–253. https://doi.org/10.1016/j.jebo.2014.07.002
Author Response
Thanks for your encouragement. We have enclosed the responses to the editors and referees’ comments, point by point, in the following section. We include our changes in our revised manuscript. A list of all the changes we made to the manuscript is listed in “Response to reviewer 2.docx”
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThank you for the revised manuscript, and futher I have no comments
Good Luck.
