The Impact of Green Finance on Urban Energy Efficiency: A Double Machine Learning Analysis

Comments 1: I suggest the author add the research gap and research questions in the abstract. This would provide better context for readers.

Response 1: Thank you for this constructive suggestion. We agree that explicitly stating the research gap and research questions significantly improves the clarity and context of the abstract. We have revised the abstract to explicitly mention the limitations of existing research and the two core questions this study aims to answer. You can find this revision on Page 1, Lines 9-15 of the revised manuscript. The revised section in the Abstract is as follows:

“…However, existing research provides limited and often inconsistent evidence on how green finance affects urban energy efficiency, largely due to heterogeneous measurement systems, methodological constraints, and insufficient identification of underlying mechanisms. To address these research gaps, this study investigates two core questions: Does green finance significantly improve urban energy efficiency? If so, what are the specific transmission mechanisms driving this impact? …”

Comments 2: The introduction contains some repetitive sentences, such as the phrase "210 cities in China between 2006 and 2022," which is also mentioned in the abstract and elsewhere. Furthermore, the introduction should conclude with the study's framework or a figure. Finally, adding a brief description of each paper section at the end of the introduction would improve the manuscript's structure and guide the reader.

Response 2: Thank you for this helpful suggestion regarding the structure and flow of the Introduction.

1. We have removed the specific details regarding the sample size and time period ("210 cities in China between 2006 and 2022") from the Introduction to avoid redundancy, as these details are fully described in the Abstract and Data section. We replaced this with a more general description of the dataset. You can find this revision on Page 3, Lines 103-104 of the revised manuscript. The revised Introduction is presented below:

“To achieve these objectives, this study uses a Double Machine Learning (DML) framework to analyze comprehensive panel data of Chinese cities…”

2. We have added a new paragraph at the end of the Introduction to outline the organization of the manuscript. This roadmap guides the reader through the theoretical framework, research design, empirical results, and conclusions. You can find this revision on Pages 3-4, Lines 124-130 of the revised manuscript. The revised Introduction is presented below:

“The remainder of this paper is organized as follows: Section 2 presents the theoretical analysis and research hypotheses. Section 3 details the methodology and mate-rials, including the DML model construction and variable selection. Section 4 provides the results and discussion, covering baseline regressions, robustness checks, mechanism analysis, and heterogeneity. Section 5 presents the conclusions, followed by policy recommendations in Section 6. Finally, Section 7 discusses the limitations and suggests directions for future research.”

Comments 3: Please review Figure 2; its resolution is very low, making the results difficult to see.

Response 3: Thank you for pointing out this issue. We apologize for the poor quality of the original figure. We have redrawn Figure 2 to improve its clarity and readability. The revised figure is now provided at a high resolution (400 DPI) below to ensure that all text and structural details are clearly visible.

Figure 2. Mechanisms of Green Finance’s Impact on Urban Energy Efficiency

Comments 4: The manuscript uses inconsistent reference styles. Please unify them according to the journal's guidelines, as the reference on line 192 is incorrect. All references must be corrected.

Response 4: Thank you for your careful review and for pointing out this formatting oversight. We apologize for the inconsistent citation styles. We have carefully reviewed the entire manuscript and unified all citations according to the journal's guidelines. Specifically, we have converted all "Author-Year" citations (such as the one mentioned on line 192) into the standard numbered format (e.g., [1]). We have also ensured that the order of the reference list corresponds strictly to the order of citation appearance in the text.

Comments 5: Does the "Research Design" section refer to the research methodology and materials? This section is vital but currently unclear and requires revision.

Response 5: Thank you for this vital comment. We acknowledge that the original structure was not sufficiently clear regarding the inclusion of both methodology and materials. To address your concern and improve the logical flow, we have made substantial revisions to Section 3:

1. We have renamed the section "3. Methodology and Materials" to explicitly reflect its scope, covering both the DML framework and the data/variables.

2. We have added an introductory paragraph to outline the organization of the section. You can find this revision on Page 6, Lines 219-224 of the revised manuscript. The revised introductory paragraph is presented below:

“This section systematically presents the research methodology and data materials employed in this study. It is organized into three subsections: Section 3.1 elaborates on the Model Construction, specifically the rationale and specification of the Double Machine Learning (DML) framework; Section 3.2 presents the Variable Setting, including the dependent, independent, and mechanism variables; and Section 3.3 de-scribes the Data Sources and provides descriptive statistics”

3. We have restructured Section 3.1 (Model Construction) into four distinct subsections to present the methodology in a more logical, step-by-step manner. Furthermore, we have significantly refined the content of Section 3.1.1 to provide a rigorous theoretical justification for selecting the Double Machine Learning (DML) framework. You can find this revision on Pages 6-8 of the revised manuscript. The revised content is as follows:

“3.1.1Rationale for Method Selection

To rigorously investigate the impact of green finance on urban energy efficiency, this study adopts the Double Machine Learning (DML) framework. This approach offers significant advantages over conventional causal inference models, particularly in terms of variable selection and estimation accuracy [37]. Urban energy efficiency is a complex system influenced by a wide range of socioeconomic factors. DML is uniquely suited to this high-dimensional context, providing a more accurate assessment than traditional alternatives.

First, DML overcomes the structural limitations of traditional econometric models. Conventional methods rely on pre-specified control variables, which expose research to omitted variable bias due to theoretical gaps or data unavailability. Furthermore, in high-dimensional settings, these models suffer from the "curse of dimensionality" and are highly sensitive to multicollinearity, which can distort coefficient estimates. In contrast, DML utilizes machine learning and regularization techniques to autonomously select optimal subsets of control variables from high-dimensional datasets. This capability effectively mitigates the adverse effects of multicollinearity and excessive controls, thereby minimizing estimation bias and enhancing model reliability.

Second, DML fundamentally distinguishes itself from standard machine learning algorithms (e.g., Random Forests, Support Vector Machines, Gradient Boosting) by prioritizing causal inference over pure prediction. While standard ML algorithms excel at predictive accuracy, they generally fail to yield valid causal interpretations. DML addresses this issue by combining orthogonalization with cross-fitting. This mechanism effectively removes the influence of high-dimensional nuisance parameters on the target parameter, preventing overfitting and ensuring consistent, asymptotically nor-mal estimates of the treatment effect.

Finally, DML is superior in handling complex, non-linear relationships. Traditional linear regressions often fail to capture the dynamic interactions between green finance and energy efficiency, leading to model misspecification. Unlike generic ML models that capture non-linearity but lack causal validity, DML integrates the flexible modeling capabilities of algorithms (such as Random Forests) into a rigorous econometric identification strategy. This integration allows the model to effectively consider non-linearities and complex variable dependencies without sacrificing the unbiased-ness required for robust causal analysis.

3.1.2 Model Specification

…

3.1.3 Estimation Procedure

…

3.1.4 Algorithm Implementation

…”

Comments 6: The discussion section is currently confusing and lacks a clear structure. I recommend reviewing the structure of recently published papers in Sustainability and revising this section to align with the standard format.

Response 6: Thank you for this constructive suggestion. We have carefully reviewed recent publications in Sustainability and standard practices in empirical economic research. We observed that presenting empirical results alongside their theoretical interpretation is a widely accepted structure in the journal. However, we agree that the original title might have confused the section's scope. To address your concern and align with the "Results and Discussion" format without disrupting the logical flow of the empirical tests, we have made the following adjustments:

1. We have renamed Section 4 from "Empirical results and analysis" to "4. Results and Discussion". This explicit title clarifies that the section includes both the statistical findings and the theoretical discussion of mechanisms and heterogeneity, effectively serving the function of a discussion section.

2. We maintained the logical progression: "Baseline Regression-Robustness Checks -Mechanism Discussion-Heterogeneity Discussion". This ensures that the discussion of transmission channels (mechanisms) immediately follows the empirical evidence, providing a cohesive narrative.

Comments 7: The Conclusion and Policy Recommendation sections should be separate, as the current version is unclear. Please revise this section carefully, as it contains numerous errors. I also suggest the author add a "Future Research Directions" section after the Conclusion. This would provide valuable insights for readers and other researchers.

Response 7: Thank you for this crucial guidance. We apologize for the errors and the lack of clarity in the original combined section. Following your recommendation, we have completely restructured the end of the manuscript into three distinct sections and carefully revised the text to correct grammatical errors and improve the logical flow.

1. We have separated the original Section 5 into Section 5 (Conclusions) and Section 6 (Policy Recommendations). Section 5 now focuses strictly on summarizing the research objectives, the DML methodology, and the key empirical findings (impact, mechanisms, and heterogeneity). Section 6 provides actionable policy suggestions based directly on these findings.

2. We have rewritten the introductory paragraph of the Conclusion to be more concise and corrected the grammatical errors noted in the previous version (e.g., refining the statement of research objectives).

3. As suggested, we have added a dedicated Section 7 (Limitations and Future Research Directions), which explicitly discusses the need for context-specific policy mixes, the potential synergy between digital technology and green finance, and the adoption of dynamic modeling approaches with refined indicators to capture long-term innovation impacts. You can find this revision on Pages 16-17 of the revised manuscript. The revised content is as follows:

“5. Conclusion

In the pursuit of high-quality development and carbon neutrality, improving ur-ban energy efficiency is a critical challenge for China. To this end, green finance has emerged as a key policy tool that optimizes resource allocation. To rigorously evaluate its effectiveness and underlying mechanisms, this study employs a Double Machine Learning (DML) framework including panel data from 210 Chinese cities. The key conclusions drawn from the empirical analysis are as follows:

First, green finance significantly boosts urban energy efficiency. This improvement is realized through three distinct mechanisms: (1) strengthening environmental regulation, which promotes green infrastructure and energy-saving technologies while discouraging energy-intensive firms; (2) optimizing industrial structures, which shifts high-energy sectors toward low-carbon models and supports green industry clusters; and (3) incentivizing green innovation, which improves resource allocation and market expectations, thereby encouraging corporate R&D in green technologies.

Second, the impact of green finance exhibits significant regional heterogeneity. The positive effects are more pronounced in non-resource-based cities, regions outside traditional industrial bases, and financially developed areas. These variations are at-tributed to mature market mechanisms, advanced technology adoption capabilities, and greater industrial adaptability in these regions.

Beyond these empirical findings, this study also provides theoretical insights. The application of a Double Machine Learning framework underscores the importance of nonlinear and high-dimensional interactions in understanding how green finance affects urban energy efficiency. This extends existing theories that predominantly assume linear and homogeneous policy effects. In addition, the differentiated impacts observed across city types enrich the literature on urban energy efficiency by demonstrating that structural and institutional conditions fundamentally shape the effectiveness of green financial policies. These contributions help advance a more context-sensitive theoretical understanding of green finance in the urban development domain.

6. Policy Recommendation

Based on the empirical findings, this study proposes the following policy recommendations to enhance the effectiveness of green finance:

First, governments should establish a unified system for green finance standards. Inconsistencies in current standards can hinder funding for low-carbon projects. Therefore, developing a comprehensive certification and regulatory framework for green bonds, loans, and insurance is essential to enhance transparency and investor confidence. Strict entry and exit rules should be enforced to ensure funds effectively support energy conservation and green innovation while restricting financing for high-pollution industries. 

Second, policymakers should implement targeted financial support strategies for key regions. Resource-based cities and old industrial bases face specific barriers to energy transition. Special funds and low-interest loans should be directed toward these areas to facilitate green technology upgrades, low-carbon projects, and sector-specific initiatives such as green mining. Effective use of these funds requires active collaboration between financial institutions and local governments, which drives regional low-carbon transitions.

Third, decision-makers should align green innovation initiatives with industrial transformation goals. Joint government-financial innovation funds should prioritize R&D in energy-saving technologies and clean energy, directing capital toward green credit and bonds. Furthermore, supporting small and medium-sized cities in sectors like green buildings can foster green industry clusters and establish an innovation-driven model, thereby enhancing energy efficiency on a national scale.

7. Limitations and Future Research Directions

While this research deepens insights into the function of green finance in promoting urban energy efficiency, three specific limitations highlight key avenues for future research.

First, regarding generalizability and context, this study relies on Chinese data, and the specific institutional environment may limit the direct transferability of findings to market-based systems. Future research should adopt a comparative perspective, replicating this analysis in other transition economies (e.g., BRICS) to test external validity. Additionally, applying Qualitative Comparative Analysis or Synthetic Control Methods could help disentangle how regional variations in development stages affect the optimal alignment of policy mixes.

Second, the depth of mechanism analysis can be expanded. The current study focuses on macro-level data and traditional financial instruments, leaving the synergy with digital technology and micro-foundations underexplored. Future studies should employ Spatial Durbin Models to test the interaction effects of digital finance. Furthermore, complementing quantitative analysis with qualitative methods, such as semi-structured interviews with financial practitioners and urban decision-makers, would be valuable for validating the specific implementation mechanisms proposed in this paper.

Third, measurement and estimation techniques can be further refined. The use of static patent counts may overlook the quality and time-lagged effects of innovation. Future research should adopt dynamic panel data models and use patent citation metrics to capture long-run impacts. Moreover, incorporating advanced estimators like Causal Forests could improve the identification of heterogeneous treatment effects, providing deeper insights into the causal dynamics of green finance.”

Comments 1: The abstract needs to provide specific conclusion data to indicate the extent of the impact.

Response 1: Thank you for this helpful suggestion. We agree that including specific quantitative results strengthens the abstract by illustrating the magnitude of the impact. We have revised the abstract to explicitly include the specific coefficient from our baseline regression model, which represents the most robust estimate of the impact. You can find this revision on Page 1, Lines 16-18 of the revised manuscript. The revised sentence in the Abstract is as follows:

“... The analysis demonstrates a significant positive impact of green finance on urban energy efficiency, with an estimated coefficient of 0.1910...”

Comments 2: Regional heterogeneity analysis is required, and it should be illustrated using map tools.

Response 2: Thank you for this insightful suggestion regarding visualization. We fully agree that addressing regional differences is crucial. After carefully considering a traditional geographical heterogeneity analysis (e.g., Eastern vs. Western China) illustrated with map tools, we ultimately chose to focus on attribute-based heterogeneity (Resource-Based Cities and Old Industrial Bases) rather than simple geographic divisions. This decision is based on the rationale that the impact of green finance is driven more by industrial structure and resource dependence than by mere geographical location, meaning our current classification provides deeper economic insights than a geographic map could offer. Furthermore, unlike contiguous geographic regions, the cities in our specific categories are geographically scattered across the country, which would result in a fragmented map distribution that fails to convey clear visual patterns. Thus, to ensure scientific rigor and focus on the most relevant economic attributes, we have retained the detailed tabular presentation in Table 7, which accurately quantifies these structural differences.

Comments 3: If the double machine learning method is used, it should be clearly stated in the abstract.

Response 3: Thank you for this suggestion. We agree that highlighting the specific methodology is crucial for the abstract's clarity. Although the method was mentioned in the previous version, we have revised the sentence to make it more prominent and explicit by adding the acronym "(DML)" and refining the phrasing. You can find this revision on Page 1, Line 15 of the revised manuscript. The revised sentence in the Abstract is as follows:

“…Methodologically, this exploration employs a Double Machine Learning (DML) approach to…”

Comments 4: The research design needs to clearly outline the research process and key steps, and explain the rationale for using the double machine learning method.

Response 4: Thank you for this constructive suggestion. To clearly outline the research process and justify our methodological choice, we have systematically restructured Section 3 (Methodology and Materials). Specifically, we subdivided the 'Model Construction' section into four distinct subsections, including 3.1.1 Rationale for Method Selection, 3.1.2 Model Specification, 3.1.3 Estimation Procedure, and 3.1.4 Sample Splitting and Learning Algorithm, to present the research steps in a logical, step-by-step manner.

Furthermore, regarding the rationale, we have significantly refined Section 3.1.1 to provide a rigorous theoretical justification. The revised text explicitly articulates that the Double Machine Learning (DML) method was selected for its superior capability to mitigate omitted variable bias in high-dimensional settings, prioritize causal inference over mere prediction, and capture complex non-linear relationships that traditional models fail to address. We believe these revisions establish a clear roadmap of our research design and convincingly justify our methodological choice. You can find this revision on Pages 6-8 of the revised manuscript. The revised content is as follows:

“3.1.1Rationale for Method Selection

To rigorously investigate the impact of green finance on urban energy efficiency, this study adopts the Double Machine Learning (DML) framework. This approach offers significant advantages over conventional causal inference models, particularly in terms of variable selection and estimation accuracy [37]. Urban energy efficiency is a complex system influenced by a wide range of socioeconomic factors. DML is uniquely suited to this high-dimensional context, providing a more accurate assessment than traditional alternatives.

First, DML overcomes the structural limitations of traditional econometric models. Conventional methods rely on pre-specified control variables, which expose research to omitted variable bias due to theoretical gaps or data unavailability. Furthermore, in high-dimensional settings, these models suffer from the "curse of dimensionality" and are highly sensitive to multicollinearity, which can distort coefficient estimates. In contrast, DML utilizes machine learning and regularization techniques to autonomously select optimal subsets of control variables from high-dimensional datasets. This capability effectively mitigates the adverse effects of multicollinearity and excessive controls, thereby minimizing estimation bias and enhancing model reliability.

Second, DML fundamentally distinguishes itself from standard machine learning algorithms (e.g., Random Forests, Support Vector Machines, Gradient Boosting) by prioritizing causal inference over pure prediction. While standard ML algorithms excel at predictive accuracy, they generally fail to yield valid causal interpretations. DML addresses this issue by combining orthogonalization with cross-fitting. This mechanism effectively removes the influence of high-dimensional nuisance parameters on the target parameter, preventing overfitting and ensuring consistent, asymptotically nor-mal estimates of the treatment effect.

Finally, DML is superior in handling complex, non-linear relationships. Traditional linear regressions often fail to capture the dynamic interactions between green finance and energy efficiency, leading to model misspecification. Unlike generic ML models that capture non-linearity but lack causal validity, DML integrates the flexible modeling capabilities of algorithms (such as Random Forests) into a rigorous econometric identification strategy. This integration allows the model to effectively consider non-linearities and complex variable dependencies without sacrificing the unbiased-ness required for robust causal analysis.

3.1.2 Model Specification

…

3.1.3 Estimation Procedure

…

3.1.4 Algorithm Implementation

…”

Comments 1: The research objectives are not explicitly formulated in a dedicated section but inferred from the introduction (p. 3, rows 81–103), and the research gap is mentioned but not clearly highlighted through a synthetic sentence that explicitly stems from the reviewed literature.

The literature review is well-structured and inventories a considerable number of contributions, although we note a clear predominance of Chinese sources, with few recent international references (partially explained by the research focus on a suggestive national context). The hypotheses are theoretically well-founded (H1, H2a–H2c), yet a clearer presentation of how they derive from the identified gaps would be useful. We therefore recommend clarifying the study’s objectives and including an explicit statement regarding this research’s contribution to filling the existing gap in contemporary literature on this topic.

Response 1: Thank you for this insightful and comprehensive suggestion. We appreciate your positive evaluation of our literature review and theoretical foundation. We fully agree that explicitly formulating the research objectives, highlighting the research gap with a synthetic statement, and broadening the international perspective significantly enhance the manuscript's clarity and academic rigor. To address your concerns comprehensively, we have made the following four major revisions to the Introduction and Theoretical Analysis:

1. We added a synthetic sentence to the Introduction that explicitly summarizes the limitations of existing literature. You can find this revision on Page 3, Lines 89-93 of the revised manuscript. The revised content is as follows:

"…The reviewed literature highlights the following critical research gap: while the correlation between green finance and energy efficiency is acknowledged, the causal transmission mechanisms remain "black-boxed", and traditional linear estimations fail to capture the complex, non-linear dynamics inherent in this relationship…"

2. We inserted a dedicated paragraph to explicitly formulate the study's specific objectives, rather than leaving them to be inferred. You can find this revision on Page 3, Lines 96-102 of the revised manuscript. The revised content is as follows:

“To fill this gap, the primary objective of this study is to rigorously identify the causal impact of green finance on urban energy efficiency. Specifically, this research aims to achieve three sub-objectives: (1) to accurately estimate the net effect of green finance by eliminating potential biases from high-dimensional control variables; (2) to decode the "black box" of transmission mechanisms by verifying the mediating roles of environmental regulation, industrial upgrading, and technological innovation; and (3) to reveal the heterogeneous impacts across different resource endowments and industrial bases.”

3. To address the predominance of Chinese sources and broaden the study's scope, we have incorporated three recent international studies (published in 2024-2025) into the Introduction. These references highlight the global relevance of green finance in enhancing energy efficiency beyond the specific national context. You can find this revision on Page 2, Lines 63-65 and Page3, Lines 76-79 of the revised manuscript. The revised content is as follows:

“... Recent empirical evidence further indicates that green finance accelerates energy transitions across diverse socioeconomic settings, showing distinct threshold and cli-mate-risk moderation effects [8]

Furthermore, recent studies highlight that specific green finance instruments, particularly green bonds, significantly drive renewable energy growth [18], with the digital economy acting as a key moderator in low-carbon transformations [19] ...”

4. We refined the introductory paragraph of Section 2 (Theoretical Analysis). We added a transition statement to explicitly clarify that the theoretical derivation of hypotheses is directly aimed at addressing the "black-boxed" mechanism gap identified in the introduction. You can find this revision on Page 4, Lines 132-135 of the revised manuscript. The revised content is as follows:

"…Addressing the identified research gap regarding the "black-boxed" transmission mechanisms, this section theoretically deduces the specific pathways through which green finance influences energy efficiency…"

Comments 2: The DML method is described in detail (pp. 5–6), with justification for choosing the Random Forest algorithm. However, we believe the explanations are excessively technical for a non-specialist reader interested in the topic and proposed solutions; a synthetic diagram of the estimation steps could be helpful. It would also be useful to include a clearer discussion on model validation and methodological limitations, such as dependence on algorithm parameters.

Response 2: We sincerely thank the reviewer for this constructive suggestion. Although we believe that providing a detailed rationale for the Double Machine Learning (DML) framework is necessary to demonstrate its superiority in handling high-dimensional controls, we acknowledge that the original description was overly technical for a broad audience. Therefore, we have streamlined the technical expression in the methodology section to improve readability. To further enhance accessibility and rigor, we have made the following specific revisions:

1. To assist non-specialist readers, we have added a synthetic diagram (Figure 3) in Section 3.1.3. This figure visually outlines the DML estimation framework, intuitively illustrating how the model filters out confounding effects before causal inference.

Figure 3. Estimation Framework of Double Machine Learning  

2. We inserted a brief, plain-language explanation in Section 3.1.3 describing DML as a "denoising" process, which complements the technical mathematical descriptions and improves the readability of the method section. You can find this revision on Page 7, Lines 272-275 of the revised manuscript. The revised content is as follows:

“In essence, this procedure serves as a 'denoising' process, where machine learning algorithms first remove the confounding effects of control variables, allowing the subsequent linear regression to capture the pure causal effect of green finance on energy efficiency”

3. Regarding model validation and the concern about algorithm parameter dependence, we have clarified in Section 3.1.4 that we employed K-fold cross-validation to adaptively tune the hyperparameters of the Random Forest algorithm. We explicitly state that this procedure minimizes the risk of the model being overly dependent on specific parameter settings, thereby ensuring the robustness and stability of our estimation results. You can find this revision on Pages 7-8, Lines 300-303 of the revised manuscript. The revised content is as follows:

“…To minimize the potential bias from parameter dependence, we employed K-fold cross-validation to adaptively select optimal hyperparameters (e.g., tree depth) for the random forest algorithm, ensuring the robustness of our estimation results.”

Comments 3: The discussions (pp. 12–14) are predominantly descriptive, with few critical comparisons to similar studies that would enrich and better position the paper’s contribution within the international research flow on this theme. Such additions would be welcome.

Response 3: We sincerely thank the reviewer for this valuable suggestion. We fully agree that the Discussion section should not only describe empirical results, but also clearly position our findings within the broader international research landscape. Following the reviewer’s recommendation, we have substantially revised the Discussion components across Sections 4.2, 4.4, and 4.5.

1. We added analytical text that explains how our findings align with or differ from existing studies on green finance and energy efficiency. The revisions highlight the methodological distinctions between the DML framework and commonly used fixed-effects, DID, and spatial econometric models, which help explain why our estimates are more conservative and robust than those reported in earlier research. This addition improves the positioning of our results within the global research landscape and enhances their empirical relevance. You can find this revision on Pages 10-11, Lines 409-417 of the revised manuscript. The revised content is as follows:

“Overall, these findings align with existing evidence that green finance enhances energy efficiency across different socioeconomic settings [9,12]. Compared with prior studies using linear fixed-effects or spatial models, the magnitude of our estimates is more conservative. This is likely due to the DML framework, which flexibly absorbs high-dimensional confounders and avoids functional-form biases commonly present in traditional econometric approaches. Recent international studies similarly note that policy effects may be overstated when nonlinearities and endogenous targeting are not properly addressed [18,24]. Thus, our results provide a more robust benchmark for assessing the true contribution of green finance and help refine the empirical under-standing of its effectiveness.”

2. We expanded the mechanism analysis by comparing our identified transmission pathways with findings from OECD countries and other emerging economies. While previous studies often focus on a single channel, such as innovation or industrial restructuring, our DML-based mediation approach identifies three simultaneous pathways. We further clarify that the stronger role of environmental regulation in China reflects its regulatory-centered green finance framework, which differs from the innovation-driven patterns observed in many other contexts. These additions provide a more internationally grounded interpretation of how green finance affects energy efficiency. You can find this revision on Page 14, Lines 526-534 of the revised manuscript. The revised content is as follows:

“The mechanism findings complement and extend existing research that typically evaluates single channels—such as innovation or industrial upgrading—in isolation. By using a DML-based mediation framework, this study uncovers three simultaneous and mutually reinforcing pathways. Interestingly, the dominant role of environmental regulation contrasts with evidence from Europe and other emerging economies, where innovation effects tend to prevail [18]. This divergence reflects China’s regulatory-centered green finance system, which generates strong compliance incentives. These results emphasize that institutional context shapes how green finance operates, offering a more integrated explanation of the finance–energy efficiency nexus than previously available.”

3. We refined the heterogeneity discussion by comparing our results with studies showing relatively uniform effects across cities in advanced economies. Our findings reveal significantly weaker impacts in resource-based cities and old industrial bases in China, largely due to structural inertia and path dependence. These contextual differences have been insufficiently examined in the international literature, and our results contribute new comparative insights into urban transitions toward higher energy efficiency. You can find this revision on Pages 14-15, Lines 563-571 of the revised manuscript. The revised content is as follows:

“The heterogeneity results further reveal that the effectiveness of green finance varies substantially across urban contexts. While prior research suggests that financial development enhances the environmental impact of green finance [19], few studies assess the combined roles of resource dependence and industrial legacy. Our findings show markedly weaker effects in resource-based and old industrial-based cities, where structural inertia and path dependence constrain the transition toward energy-efficient development. This result contrasts with the more uniform gains observed in OECD urban studies [3]. These patterns highlight that green finance is context-dependent rather than universally effective, contributing new comparative insights into urban energy transition dynamics”

Comments 4: We appreciate the concise nature of the Conclusions, which emphasize the results well; however, we recommend that theoretical implications to be more clearly articulated. In the current configuration, the emphasis falls mainly on policy recommendations. This is not a deficiency per se, but rather indicates a need to balance both theoretical and practical implications. For example, a short discussion (1–2 paragraphs) on this paper’s contribution to green finance theory and the literature on urban energy efficiency would be useful. Limitations are mentioned, but brief considerations regarding possible methodological solutions would enhance their significance.

Response 4: Thank you very much for this thoughtful and constructive suggestion. To improve the clarity and coherence of the manuscript, we reorganized the original concluding section into three independent parts, namely Conclusions, Policy Recommendations, and Limitations and Future Research Directions. This structural refinement allows theoretical insights, policy implications, and future research considerations to be presented more clearly and in a more balanced manner. In response to the reviewer’s specific comment, we made two substantive revisions.

1. We strengthened the Conclusions section by adding a concise paragraph that explicitly articulates the theoretical implications of the study. The new text explains how the Double Machine Learning framework advances theoretical understanding by capturing nonlinear and high-dimensional mechanisms, and how the observed urban heterogeneity contributes to the literature on energy efficiency under differentiated institutional and structural contexts. You can find this revision on Page 16, Lines 614-640 of the revised manuscript. The revised content is as follows:

“5. Conclusion

In the pursuit of high-quality development and carbon neutrality, improving ur-ban energy efficiency is a critical challenge for China. To this end, green finance has emerged as a key policy tool that optimizes resource allocation. To rigorously evaluate its effectiveness and underlying mechanisms, this study employs a Double Machine Learning (DML) framework including panel data from 210 Chinese cities. The key conclusions drawn from the empirical analysis are as follows:

First, green finance significantly boosts urban energy efficiency. This improvement is realized through three distinct mechanisms: (1) strengthening environmental regulation, which promotes green infrastructure and energy-saving technologies while discouraging energy-intensive firms; (2) optimizing industrial structures, which shifts high-energy sectors toward low-carbon models and supports green industry clusters; and (3) incentivizing green innovation, which improves resource allocation and market expectations, thereby encouraging corporate R&D in green technologies.

Second, the impact of green finance exhibits significant regional heterogeneity. The positive effects are more pronounced in non-resource-based cities, regions outside traditional industrial bases, and financially developed areas. These variations are at-tributed to mature market mechanisms, advanced technology adoption capabilities, and greater industrial adaptability in these regions.

Beyond these empirical findings, this study also provides theoretical insights. The application of a Double Machine Learning framework underscores the importance of nonlinear and high-dimensional interactions in understanding how green finance affects urban energy efficiency. This extends existing theories that predominantly assume linear and homogeneous policy effects. In addition, the differentiated impacts observed across city types enrich the literature on urban energy efficiency by demonstrating that structural and institutional conditions fundamentally shape the effectiveness of green financial policies. These contributions help advance a more context-sensitive theoretical understanding of green finance in the urban development domain.”

2. We enhanced the Limitations and Future Research Directions section by incorporating specific methodological solutions for each identified gap. The revised text now outlines promising approaches, including Qualitative Comparative Analysis (QCA), Spatial Durbin Models (SDM), and Causal Forests. This strengthens the significance of the limitations and provides clearer guidance for subsequent research. You can find this revision on Page 17, Lines 663-685 of the revised manuscript. The revised content is as follows:

“7. Limitations and Future Research Directions

While this research deepens insights into the function of green finance in promoting urban energy efficiency, three specific limitations highlight key avenues for future research.

First, regarding generalizability and context, this study relies on Chinese data, and the specific institutional environment may limit the direct transferability of findings to market-based systems. Future research should adopt a comparative perspective, replicating this analysis in other transition economies (e.g., BRICS) to test external validity. Additionally, applying Qualitative Comparative Analysis or Synthetic Control Methods could help disentangle how regional variations in development stages affect the optimal alignment of policy mixes.

Second, the depth of mechanism analysis can be expanded. The current study focuses on macro-level data and traditional financial instruments, leaving the synergy with digital technology and micro-foundations underexplored. Future studies should employ Spatial Durbin Models to test the interaction effects of digital finance. Furthermore, complementing quantitative analysis with qualitative methods, such as semi-structured interviews with financial practitioners and urban decision-makers, would be valuable for validating the specific implementation mechanisms proposed in this paper.

Third, measurement and estimation techniques can be further refined. The use of static patent counts may overlook the quality and time-lagged effects of innovation. Future research should adopt dynamic panel data models and use patent citation metrics to capture long-run impacts. Moreover, incorporating advanced estimators like Causal Forests could improve the identification of heterogeneous treatment effects, providing deeper insights into the causal dynamics of green finance.”

Comments 5: The theoretical model is coherent but overloaded by the inclusion of numerous control variables, which may affect interpretability and the flow of arguments. The degree of generalization is limited to the Chinese context, without discussion of transferability. We therefore recommend a comparative perspective or at least comments on the possibilities of replication in other regions and countries. Among possible future research directions, complementing the analysis with qualitative methods (e.g., interviews with actors in the financial sector, urban decision-makers) could help validate the findings, mechanisms, and the proposed solutions.

Response 5: We sincerely thank the reviewer for these thoughtful comments regarding model specification and future research directions. We have addressed these concerns as follows:

1. We respectfully clarify that the inclusion of a high-dimensional set of control variables is a deliberate methodological choice driven by the Double Machine Learning (DML) framework. Unlike traditional linear models where numerous controls can cause multicollinearity and reduce interpretability, DML is specifically designed to handle high-dimensional data. It utilizes machine learning algorithms (Random Forest in our case) to autonomously "select" relevant confounders and remove their bias. Therefore, retaining a comprehensive set of controls is essential for the DML model to effectively isolate the "pure" causal effect of green finance.

2. We fully accept the reviewer's constructive suggestions regarding the limitations of the Chinese context and the need for qualitative validation. Accordingly, we have substantially revised and streamlined Section 7 (Limitations and Future Research Directions) to explicitly incorporate these points. Specifically, regarding transferability, we acknowledged that findings from China may not directly transfer to market-based systems and added a call for comparative studies in other transition economies to test external validity. Furthermore, we integrated the suggestion to complement quantitative analysis with qualitative methods, such as semi-structured interviews with financial practitioners and urban decision-makers, to validate the specific implementation mechanisms proposed in this paper. You can find this revision on Page 17 of the revised manuscript. The revised text in Section 7 now reads:

"First, regarding generalizability and context... Future research should adopt a comparative perspective, replicating this analysis in other transition economies (e.g., BRICS) to test external validity...

 Second... complementing quantitative analysis with qualitative methods—such as semi-structured interviews with financial practitioners and urban decision-makers—would be valuable for validating the specific implementation mechanisms proposed in this paper."

Comments 6: The text is generally clear, but there are minor language errors: e.g., p. 5: “he influence” instead of “the influence.”

Response 6: We sincerely thank the reviewer for the careful reading and for pointing out this oversight. We have corrected the specific error ("he influence" to "the influence") as suggested. Furthermore, we have thoroughly proofread and polished the entire manuscript to eliminate grammatical errors and ensure the language meets the high standards of the journal.

Comment7:Certain expressions are either complicated (“Population scale (SP) is gauged by the year-end registered population in logarithmic form”), overly technical (the explanation of “orthogonalization” and “Neyman orthogonality”), or vague (“three key limitations emerge that point to future research need”). Where possible, we recommend revising for greater clarity or adapting phrasing (without losing precision and consistency of arguments).

Response7: We sincerely thank the reviewer for the careful attention to linguistic details and agree that certain expressions in the original manuscript were either overly complex or technical. In response, we have revised these sentences to enhance clarity and precision: we simplified the phrasing regarding population scale to "measured as the natural logarithm of..." to align with standard terminology, removed the overly specific term "(Neyman orthogonality)" to avoid technical distraction, and rephrased the transition to future research as "three specific limitations highlight key avenues for..." to make the statement more direct. Furthermore, we have conducted a thorough proofreading of the entire text to ensure a smoother reading flow.

Comment8: Figures lack source indication (especially Fig. 1, which uses external data).

Response8: We sincerely thank the reviewer for this reminder regarding data attribution. We clarify that the data presented in Figure 1 is derived from the China Energy Statistical Yearbook. Accordingly, we have updated the figure caption to explicitly include "Source: China Energy Statistical Yearbook" to ensure transparency and proper citation.

Comment9: Titles of some tables and subsections are correct but overly technical (e.g., “Double Machine Learning with Gradient Boosting” could be simplified), or quite generic, suitable for any similar context, and offer few indications related to the specific case of this research (e.g., “… test results”). We recommend, where possible, modifying them to the context.

Response9: Thank you very much for this helpful suggestion.  We agree that several table titles and subsection headings could be made clearer and more directly connected to the context of this study.  In the revised manuscript, we have refined these titles to improve readability and ensure closer alignment with the research content. More specifically, highly technical expressions were simplified to enhance clarity, and generic labels were rewritten to better reflect the specific analysis conducted in this paper.  These adjustments help strengthen the coherence between the titles and the empirical context of green finance and urban energy efficiency.

Comment10:There are a few footnotes—although their use is not inherently problematic, we recommend avoiding them, as they interrupt reading flow and create confusion regarding their relevance. In our view, most explanations are contextual and of questionable significance (hence their placement in footnotes). We suggest either integrating them into the text (if truly meaningful) or removing them.

Response10: We fully agree with the reviewer’s view that the footnotes interrupted the reading flow. Since these footnotes primarily listed specific names of cities or provinces covered by the exclusion policies, which are details secondary to the main argument, we have decided to remove them entirely. The descriptions retained in the main text provide sufficient context for the reader without the distraction of lengthy lists.

Comment11: In-text citation style is inconsistent: sometimes numbers in square brackets [ ], sometimes author-year in parentheses ( , ). The final list includes 40 sources, but the style is uneven, and some entries lack DOI (where available). Some citations in the text do not clearly appear in the final references list (e.g., Liu et al., 2023 mentioned twice in the text but not identifiable in the final list). We recommend checking completeness and standardizing the style according to the journal’s guidelines.

Response11: We sincerely apologize for the oversight regarding the inconsistency in citation styles and the reference list. In response, we have conducted a comprehensive review of the entire manuscript to ensure strict adherence to the journal’s guidelines. We have standardized all in-text citations to the numbered format consistent with Sustainability’s style, cross-checked every citation against the final list to ensure a one-to-one correspondence, and specifically identified and added the missing "Liu et al., 2024" reference to ensure completeness and accuracy.