Green credit risk assessment and prediction in China’s construction industry: based on the optuna-catboost model Authors: Zenan Qin, Meitong Meng, Dongbao Li, Junwei Li Published: 2026 ABSTRACT Green credit, as a new credit model, not only shares the risk characteristics of traditional credit but also incorporates environmental risk factors into its analysis, making the assessment more complex. This article employs a modified KMV model to evaluate the risk of green credit in the construction industry. Utilizing Mann-Whitney U test and a feature ranking method based on learning models for indicator selection, the final indicators are made more objective and scientific. Subsequently, a green credit risk assessment system suitable for the Chinese construction industry is proposed. Finally, the Optuna-CatBoost (OptCAT) algorithm is applied to predict the risk of green credit in the construction industry, aiming to provide a methodological reference for the industry to mitigate green credit risks. Empirical results indicate that the modified KMV model provides an accurate measurement of green credit risk in the Chinese construction industry. Compared to default parameter (DP), random search (RS), grid search (GS), bayesian optimization (BO), and Hyperopt, Optuna demonstrates better optimization effects. When compared to support vector machine (SVM), random forest (RF), XGBoost, and LightGBM models, OptCAT exhibits the optimal predictive performance. Introduction The Equator Principles are a set of international guidelines for environmental and social risk management in project finance, proposed by the International Finance Corporation and ABN AMRO Bank in 2003. They are designed to assess and manage environmental and social risks associated with project finance, requiring financial institutions to comprehensively consider the potential impacts of projects on the ecological environment, local communities, and other stakeholders during project financing decision-making. Financial institutions should only provide financing support to projects that meet environmental and social standards2. The Equator Principles have become an industry norm for international project finance, with more than 100 financial institutions in over 40 countries worldwide having adopted them. Since the Equator Principles became a new standard for international project financing, it incorporates the impact of projects on the environment and society into comprehensive assessments, emphasizing corporate environmental and social responsibility. Financial institutions have gradually shifted from the previous profit-maximizing operating philosophy to one that focuses on the coordinated development of the economy and the environment. The non-performing loan ratio is an indicator used to assess the quality of bank assets1. In the current situation of high non-performing loan ratios in commercial banks, the occurrence of green credit risks will have a significant impact. Green projects often involve large investments, long durations, lack of collateral, and insufficient internalization of green benefits, leading to lower returns on risk-adjusted green loans. Therefore, potential green credit risks should be given attention. It is necessary to assess and predict green credit. Green credit risk refers to the default risk faced by financial institutions when providing green credit to enterprises, which is caused by factors such as poor corporate environmental performance, insufficient environmental protection investment, or environmental compliance issues. Different from traditional credit risk, green credit risk incorporates environmental factors into the risk assessment system, and these environmental factors may affect corporate financial performance through multiple channels. On the one hand, stringent environmental protection policies may expose high-pollution enterprises to risks such as fines and production suspension for rectification; on the other hand, good environmental performance may bring enterprises policy preferences, improved market reputation, and enhanced long-term competitiveness. The construction industry is one of the focal points for global carbon emissions. In recent years, government agencies and regulatory bodies have identified green building as a key area for financial support. Therefore, the innovation in this article is primarily manifested in the following aspects: Firstly, the article evaluates the risk of green credit in the Chinese construction industry, which is a novel research focus. Secondly, environmental factors are taken into account, and a modified KMV model is employed to assess green credit risk. Thirdly, the article uses the OptCAT algorithm to predict green credit risk in the construction industry. This not only expands the application scope of the OptCAT algorithm but also enriches the methods for predicting green credit risk in the construction industry. Green credit risk assessment With the growing awareness of environmental issues and social responsibility, sustainable finance has been vigorously promoted. While commercial banks leverage economic mechanisms to engage in green credit activities, there is an increasing need to focus on environmental risks and establish robust preventive systems and management. Lundgren and Catasús,16 propose a rigorous evaluation of the environmental risks of lending enterprises before deciding whether to grant loans. This approach allows commercial banks to actively respond to government calls for environmental protection, elevating the importance of environmental responsibility while mitigating the green credit risks associated with lending enterprises. Jeucken (2001) suggests that a bank’s sustainability is closely related to environmental factors, reputation, risk assessment, and more. He argues that if companies prioritize environmental protection, they are likely to increase investments in areas such as environmental protection and technology improvement, leading to higher costs and ultimately an increase in the risk of loan defaults. Therefore, managing environmental risks in banks needs to start at the source. Additionally, he believes that banks should implement energy-saving measures internally to reduce green credit risks. Labatt and White,13 analyze how financial institutions, including banks, should conduct financial assessments regarding environmental risks. They express their views on the relationship between environmental protection and financial innovation, asserting that the banking industry should continuously improve environmental risk prevention systems to minimize losses. Weber,23 points out that most European banks have successfully integrated environmental risk management concepts and sustainable development principles into their operations, effectively avoiding environmental risks and achieving sustainability to some extent. Samuel,18 applies concepts from modern portfolio theory to evaluate portfolio risks associated with loans, using new risk measurement tools to assess derivatives of credit risks. Currently, some literature has assessed the risks of green credit using more traditional measurement methods such as KMV. However, environmental risk factors have not been considered. Green credit is the core product of China’s green finance system, which has distinct local characteristics in top-level institutional design, policy incentive mechanisms, and practical application scenarios. It differs from similar products such as green loans in the international market and still faces many problems in actual promotion. It requires continuous improvement and further research. The risk management of green credit is an indispensable component for banks engaging in green credit business. Compared with traditional credit risk, green credit risk has a more complex composition and involves many factors, including environmental policy risk, environmental cost internalization risk, green project characteristic risk, environmental information asymmetry risk, environmental risk spillover risk, etc. Therefore, the risk indicator system also needs to take environmental factors into account. Traditional credit risk assessment methods have some shortcomings: (1) Credit-metrics Model: Parameter estimation for the Credit-metrics model is ineffective due to the large and complex parameter scale. Additionally, historical data is challenging to obtain given its extensive time span. (2) Expert Analysis: Expert analysis suffers from significant subjectivity, leading to less objective results. (3) Neural Network Model: The shortcomings of the neural network model include randomness, requiring a considerable amount of time for debugging, resulting in poor practicality and operability. (4) Logistic Model: The logistic model relies on a substantial amount of sample data for fitting and discrimination. (5) Analytic Hierarchy Process (AHP): AHP requires judging the consistency of matrices, but it is largely influenced by subjective factors, making it less objective for assessing corporate environmental risks. Furthermore, in traditional KMV models, statistical analysis methods cannot accurately calculate the impact of equity value volatility. Therefore, this study utilizes the GARCH(1,1) model with generalized conditional heteroskedasticity for computation. To highlight the differences from traditional credit risk assessments, this paper considers environmental factors. It employs a modified KMV model to measure default distance (DD) and expected default frequency (EDF). Green credit risk prediction In the face of increasingly complex products, the methods for predicting green credit risks also need to be updated. Machine learning methods play a significant role in improving the efficiency and accuracy of credit risk predictions, and both domestic and international research in this area has made progress. Machine learning methods such as decision trees (DT), neural networks, and support vector machine (SVM) have demonstrated good performance in empirical studies. Li and Wu,14 applied a BP neural network to establish a credit risk assessment model, and empirical evidence indicated the feasibility of this model in practical applications. Fu and Liu,5 proposed using a recursive algorithm to train an RBF neural network. Empirical research showed that this method possesses strong learning capability and good generalization ability, highlighting the model’s effective learning. Farquad,3 introduced a method that combines principal component analysis with SVM. The approach involves dimensionality reduction using principal components followed by non-classification using SVM. Empirical results demonstrated its superior performance compared to logistic regression. Harris,7 utilized CSVM to build a credit assessment model, and empirical results indicated that CSVM outperformed SVM, enhancing the model’s performance. García (2015) conducted a study on credit ratings and corporate bankruptcies from 2000 to 2013, evaluating 1-NN, C4.5, SVM, and MLP using two databases and different sample partitioning methods. Based on accuracy, SVM’s performance was judged to be superior to other algorithms in both databases. Wei et al.,24 proposed a least squares SVM with a hybrid kernel, addressing the sparsity and robustness issues of traditional least squares SVM. Experimental results showed that the optimized model could obtain a small number of features and enhance generalization ability. Szwabe and Misiorek,22 introduced a DT model for credit decision-making. They evaluated several methods for classifying loan applications, providing individual DT as output. Ensemble is a method of combining individual learners to enhance the model’s generalization ability, and it has been widely applied in the field of machine learning for assessing credit risks. Xia,25 developed an ensemble method that is both parallelizable and sequential in the domain of credit risk assessment. This provides the model with greater flexibility and simplifies the model by eliminating redundant variables. Singh,20 compared the performance of base learners and ensemble learning in the field of credit scoring, and empirical evidence showed that the classification results of ensemble learning were significantly superior to those of single algorithms. Feng,4 proposed a dynamic weighted ensemble credit scoring method, using a Markov chain to construct a dynamically weighted trainable combiner. Empirical results indicated that this model had higher accuracy and efficiency in classification compared to static ensemble models. Sivasankar et al.,21 utilized an ensemble CRP model to predict the credit risk of financial institutions. They introduced an enhanced ensemble method called Weighted Adjusted Boosting Ensemble Method (WABEM), which employed feature selection (FS) techniques based on rough sets (RS) and preprocessing based on balance and regression, referred to as RS-RFS-WABEM. Regression was used to fill missing values in records, improving the performance of the CRP model. Currently, ensemble learning methods have not been widely applied in the prediction of green credit risk. The CatBoost algorithm, introduced relatively recently, builds on many existing excellent algorithms. Despite its short existence, the CatBoost algorithm has gained widespread application across various domains due to its high performance, good robustness, and the need for minimal hyperparameter tuning. This includes applications in areas such as driving style recognition15,17, diabetes prediction12; Jaiswal and Gupta, 2022), and more. In this study, the CatBoost model was selected to predict green credit risk in the construction industry, contributing to the enrichment of research in the field of green credit risk. Methodology Indicator selection and data preprocessing Following the principles of authenticity, comprehensiveness, scientificity, and availability, this paper initially constructs an indicator system from four major categories, namely, debt-paying ability, operational capability, profitability, and development capability, encompassing a total of 114 financial indicators for listed companies in the Chinese construction industry. Additionally, it includes four environmental indicators based on the ESG rating system. This study selected four environmental indicators under the Huazheng ESG Rating system (https://www.chindices.com/), which are formulated centering on the core regulatory requirements for the green development of the construction industry. The ESG score provides a comprehensive evaluation of enterprises’ overall performance in three dimensions, namely environment, society and governance, with a higher score indicating better comprehensive ESG performance. The Environmental score (E score) is dedicated to assessing enterprises’ environmental performance, covering five sub-dimensions with respective weightings: carbon emission management (30%), resource use efficiency (25%), pollution prevention and control (20%), environmental management systems (15%), and biodiversity conservation (10%). While falling under the social dimension, the Social score (S score) includes indicators closely linked to the environment in the construction industry, such as community relations, employee health and safety, and product liability. The Governance score (G score) evaluates enterprises’ environmental governance capabilities, including the transparency of environmental information disclosure, the organizational structure for environmental risk management, and the formulation and implementation of environmental goals. The research data cover Chinese A-share listed companies in the construction industry, and all indicators are measured on a 100-point scale (0-100 points). To ensure temporal consistency, we adopted the ESG scores of the previous year to predict the credit risk of the current year. This lag processing is consistent with the operational logic in practical business, where banks use historical environmental performance to assess current risks. Other data in this study were sourced from the CSMAR Database (https://data.csmar.com/) and the Wind Database (https://www.wind.com.cn/), with partial missing data manually supplemented using the annual reports of listed construction enterprises. Firstly, a preliminary screening was conducted on the initial 118 financial indicators, resulting in the retention of 58 financial indicators after eliminating those with too much missing data and high similarity. Mann-Whitney U test, proposed by H.B. Mann and D.R. Whitney in 1947, was then performed. As shown in Table 1 and 47 indicators exhibited significant differences in means, indicating that green credit risk can be highly revealed through these indicators. Therefore, these 47 indicators are temporarily retained as the initial dataset, suggesting their effectiveness in predicting green credit risk. Table 1The results of Mann-Whitney U test.IndicatorsMann-Whitney U testIndicatorsMann-Whitney U testESG6.570***(0.0000)Net Profit Growth Rate0.137(0.8913)E2.615***(0.0089)Accounts Receivable Turnover4.292***(0.0000)S3.033***(0.0024)Days Sales Outstanding− 4.292***(0.0000)G6.701***(0.0000)Inventory Turnover− 1.191***(0.0037)Quick Ratio1.886*(0.0593)Days Inventory Outstanding1.188***(0.0048)Cash Ratio5.245***(0.0000)Accounts Payable Turnover3.954***(0.0001)Operating Working Capital5.658***(0.0000)Operating Working Capital Turnover− 0.139(0.8896)Net Cash Flow from Operating Activities / Current Liabilities1.688*(0.0914)Cash and Cash Equivalents Turnover− 4.353***(0.0000)Debt-to-Equity Ratio− 2.645***(0.0082)Fixed Assets to Revenue Ratio0.479***(0.0019)Tangible Asset Interest-Bearing Debt Ratio− 0.017***(0.0064)Fixed Asset Turnover− 0.479***(0.0019)Equity Multiplier− 0.577(0.5641)Non-current Asset Turnover3.490***(0.0005)Equity Ratio− 0.576(0.5646)Capital Intensity− 4.264***(0.0000)Long-term Debt to Operating Working Capital Ratio2.672***(0.0075)Total Asset Turnover4.264***(0.0000)Profit Before Tax and Interest / Total Liabilities4.149***(0.0000)Equity Turnover2.206**(0.0274)Net Cash Flow from Operating Activities / Total Liabilities1.602(0.1091)Return on Assets4.612***(0.0000)Net Cash Flow from Operating Activities / Interest-Bearing Debt1.365(0.1721)Fixed Asset Net Profit Margin3.827***(0.0001)Tangible Net Worth Debt Ratio− 1.941*(0.0522)Earnings Before Interest and Taxes5.833***(0.0000)Fixed Charge Coverage Ratio3.365***(0.0008)Net Profit to Net Profit Margin Ratio2.665***(0.0077)Capital Preservation and Appreciation Rate2.942***(0.0033)Net Profit Margin to EBIT Ratio0.085***(0.0020)Fixed Asset Growth Rate2.698***(0.0070)EBIT to Total Assets Ratio4.612***(0.0000)Total Asset Growth Rate4.499***(0.0000)Total Operating Cost Ratio− 3.909***(0.0001)Operating Revenue Growth Rate0.566(0.5714)Sales Expense Ratio− 1.383(0.1667)Operating Total Revenue Growth Rate1.321***(0.0066)Management Expense Ratio− 3.993***(0.0001)Sales Expense Growth Rate1.451(0.1468)Financial Expense Ratio− 3.369***(0.0008)Management Expense Growth Rate2.455**(0.0141)Cost-to-Income Ratio4.350***(0.0000)Accrued Items0.213***(0.0085)Asset Impairment Loss to Operating Income4.165***(0.0000)Sustainable Growth Rate3.184**(0.0015)Cash to Net Profit Ratio1.874*(0.0610)Return on Equity Growth Rate0.231(0.8169)Return on Comprehensive Income Attributable to Parent Company4.208***(0.0000)Net Profit Growth Rate0.339(0.7349)Earnings Before Interest, Taxes, Depreciation, and Amortization (EBITDA) Margin4.142***(0.0000) Table 1The results of Mann-Whitney U test.IndicatorsMann-Whitney U testIndicatorsMann-Whitney U testESG6.570***(0.0000)Net Profit Growth Rate0.137(0.8913)E2.615***(0.0089)Accounts Receivable Turnover4.292***(0.0000)S3.033***(0.0024)Days Sales Outstanding− 4.292***(0.0000)G6.701***(0.0000)Inventory Turnover− 1.191***(0.0037)Quick Ratio1.886*(0.0593)Days Inventory Outstanding1.188***(0.0048)Cash Ratio5.245***(0.0000)Accounts Payable Turnover3.954***(0.0001)Operating Working Capital5.658***(0.0000)Operating Working Capital Turnover− 0.139(0.8896)Net Cash Flow from Operating Activities / Current Liabilities1.688*(0.0914)Cash and Cash Equivalents Turnover− 4.353***(0.0000)Debt-to-Equity Ratio− 2.645***(0.0082)Fixed Assets to Revenue Ratio0.479***(0.0019)Tangible Asset Interest-Bearing Debt Ratio− 0.017***(0.0064)Fixed Asset Turnover− 0.479***(0.0019)Equity Multiplier− 0.577(0.5641)Non-current Asset Turnover3.490***(0.0005)Equity Ratio− 0.576(0.5646)Capital Intensity− 4.264***(0.0000)Long-term Debt to Operating Working Capital Ratio2.672***(0.0075)Total Asset Turnover4.264***(0.0000)Profit Before Tax and Interest / Total Liabilities4.149***(0.0000)Equity Turnover2.206**(0.0274)Net Cash Flow from Operating Activities / Total Liabilities1.602(0.1091)Return on Assets4.612***(0.0000)Net Cash Flow from Operating Activities / Interest-Bearing Debt1.365(0.1721)Fixed Asset Net Profit Margin3.827***(0.0001)Tangible Net Worth Debt Ratio− 1.941*(0.0522)Earnings Before Interest and Taxes5.833***(0.0000)Fixed Charge Coverage Ratio3.365***(0.0008)Net Profit to Net Profit Margin Ratio2.665***(0.0077)Capital Preservation and Appreciation Rate2.942***(0.0033)Net Profit Margin to EBIT Ratio0.085***(0.0020)Fixed Asset Growth Rate2.698***(0.0070)EBIT to Total Assets Ratio4.612***(0.0000)Total Asset Growth Rate4.499***(0.0000)Total Operating Cost Ratio− 3.909***(0.0001)Operating Revenue Growth Rate0.566(0.5714)Sales Expense Ratio− 1.383(0.1667)Operating Total Revenue Growth Rate1.321***(0.0066)Management Expense Ratio− 3.993***(0.0001)Sales Expense Growth Rate1.451(0.1468)Financial Expense Ratio− 3.369***(0.0008)Management Expense Growth Rate2.455**(0.0141)Cost-to-Income Ratio4.350***(0.0000)Accrued Items0.213***(0.0085)Asset Impairment Loss to Operating Income4.165***(0.0000)Sustainable Growth Rate3.184**(0.0015)Cash to Net Profit Ratio1.874*(0.0610)Return on Equity Growth Rate0.231(0.8169)Return on Comprehensive Income Attributable to Parent Company4.208***(0.0000)Net Profit Growth Rate0.339(0.7349)Earnings Before Interest, Taxes, Depreciation, and Amortization (EBITDA) Margin4.142***(0.0000) The results of Mann-Whitney U test. 6.570*** (0.0000) 0.137 (0.8913) 2.615*** (0.0089) 4.292*** (0.0000) 3.033*** (0.0024) − 4.292*** (0.0000) 6.701*** (0.0000) − 1.191*** (0.0037) 1.886* (0.0593) 1.188*** (0.0048) 5.245*** (0.0000) 3.954*** (0.0001) 5.658*** (0.0000) − 0.139 (0.8896) 1.688* (0.0914) − 4.353*** (0.0000) − 2.645*** (0.0082) 0.479*** (0.0019) − 0.017*** (0.0064) − 0.479*** (0.0019) − 0.577 (0.5641) 3.490*** (0.0005) − 0.576 (0.5646) − 4.264*** (0.0000) 2.672*** (0.0075) 4.264*** (0.0000) 4.149*** (0.0000) 2.206** (0.0274) 1.602 (0.1091) 4.612*** (0.0000) 1.365 (0.1721) 3.827*** (0.0001) − 1.941* (0.0522) 5.833*** (0.0000) 3.365*** (0.0008) 2.665*** (0.0077) 2.942*** (0.0033) 0.085*** (0.0020) 2.698*** (0.0070) 4.612*** (0.0000) 4.499*** (0.0000) − 3.909*** (0.0001) 0.566 (0.5714) − 1.383 (0.1667) 1.321*** (0.0066) − 3.993*** (0.0001) 1.451 (0.1468) − 3.369*** (0.0008) 2.455** (0.0141) 4.350*** (0.0000) 0.213*** (0.0085) 4.165*** (0.0000) 3.184** (0.0015) 1.874* (0.0610) 0.231 (0.8169) 4.208*** (0.0000) 0.339 (0.7349) 4.142*** (0.0000) In order to reduce the number of feature variables and enhance the model’s generalization ability, feature selection is generally required. One of the criteria for feature selection is to choose feature variables that have a high correlation with the target variable. Common methods include mutual information, PCA, and others. However, these methods have drawbacks such as weak interpretability and poor correlation, which may not fully meet the requirements of the current experiment. With the development of machine learning technology, feature ranking methods based on learning models are well-suited to address this issue. The basic idea is to directly use the final model to establish a predictive model for each indicator and the dependent variable. When selecting a model, if the relationship between the indicator and the dependent variable is nonlinear, tree-based methods (DT, RF), or extended linear models can be chosen. Because tree-based methods can model nonlinear relationships effectively and require less parameter tuning, they are commonly used. It is essential to be cautious about the depth of the trees in tree-based feature selection methods to prevent overfitting. In addition, the use of cross-validation is another method, scoring the indicators through such training to obtain correlation and then retraining the final model. This article chooses the feature ranking based on learning models in embedded feature selection, and the learning model selected is the CatBoost model. The CatBoost method is employed to rank the importance of indicators, and features with an importance score greater than 0.2 are retained. Due to the large number of indicators, this article only lists the importance scores of indicators that need to be excluded, as shown in Table 2. In the end, out of 114 indicators, 27 indicators are selected to construct the necessary indicator system for this article, as shown in Table 3. Table 2Importance score of excluded indicators.IndicatorsImportance ScoreAsset-to-Equity Ratio0.1723Days Sales Outstanding0.1362Fixed Asset Turnover Ratio0.1149Days Inventory Outstanding0.0972Earnings Before Interest and Taxes to Total Assets Ratio0.0888Net Cash Flow from Operating Activities to Current Liabilities Ratio0.0846Return on Assets0.0731Cash to Earnings Ratio0.0651Fixed Charge Coverage Ratio0.0646Operating Working Capital0.0605Earnings Before Interest and Taxes0.0486Growth Rate of Operating Expenses0.0389Cash Ratio0.0314Total Operating Costs Ratio0.0308Equity Turnover Ratio0.0234Fixed Asset Net Profit Margin0.0183Cost-to-Income Ratio0.0178Net Profit to EBIT Ratio0.0133Total Asset Turnover Ratio0.0098Non-Current Asset Turnover Ratio0.0076 Table 2Importance score of excluded indicators.IndicatorsImportance ScoreAsset-to-Equity Ratio0.1723Days Sales Outstanding0.1362Fixed Asset Turnover Ratio0.1149Days Inventory Outstanding0.0972Earnings Before Interest and Taxes to Total Assets Ratio0.0888Net Cash Flow from Operating Activities to Current Liabilities Ratio0.0846Return on Assets0.0731Cash to Earnings Ratio0.0651Fixed Charge Coverage Ratio0.0646Operating Working Capital0.0605Earnings Before Interest and Taxes0.0486Growth Rate of Operating Expenses0.0389Cash Ratio0.0314Total Operating Costs Ratio0.0308Equity Turnover Ratio0.0234Fixed Asset Net Profit Margin0.0183Cost-to-Income Ratio0.0178Net Profit to EBIT Ratio0.0133Total Asset Turnover Ratio0.0098Non-Current Asset Turnover Ratio0.0076 Importance score of excluded indicators. Table 3Descriptive statistics of the final indicator system.Primary indicatorsSecondary indicatorsSymbolMaxMinMeanStd.Environmental Indicatorsesg scoreESG87.7146.4171.455.818e scoreE79.6429.4659.387.590s scoreS94.9723.5473.249.868g scoreG92.5133.6576.477.383Finan-cial indexSolvency capacityQuick RatioS19.1450.1071.1110.676Tangible Assets to Interest-Bearing Debt RatioS20.9870.0000.2220.150Long-Term Debt to Operating Working Capital RatioS3384.7-61.382.15119.36EBITDA to Total Liabilities RatioS41.388-4.0360.0750.202Tangible Net Worth Debt RatioS560.650.1033.8495.747Development capacityCapital Preservation and Appreciation RateD111.50-9.3841.2150.758Fixed Asset Growth RateD261.90-0.8810.4573.024Total Asset Growth RateD35.419-0.4770.2100.388Operating Revenue Growth RateD427.08-0.8330.2151.133Accrued ItemsD525.3714.2219.861.791Sustainable Growth RateD66.077-10.200.0360.434Operating capacityAccounts Receivable Turnover RatioO1833.80.0787.68643.93Inventory Turnover RatioO220,5990.24343.25674.3Accounts Payable Turnover RatioO318.920.0822.2701.622Cash and Cash Equivalents Turnover RatioO4437.60.1737.21018.32Fixed Assets to Revenue RatioO53.1960.0000.1250.201Capital IntensityO639.510.5102.3402.212Profitability capacityProfit Before Tax to EBIT RatioP133.60-22.820.7811.643Expense-to-Sales RatioP21.0060.0090.0570.068Financial Expense RatioP32.323-0.1920.0230.095Cost-to-Income RatioP40.757-5.1820.0150.307Asset Impairment Loss to Operating Revenue RatioP52.057-2.847-0.0010.164Earnings Before Interest and Taxes MarginP60.982-7.8300.0030.451 Table 3Descriptive statistics of the final indicator system.Primary indicatorsSecondary indicatorsSymbolMaxMinMeanStd.Environmental Indicatorsesg scoreESG87.7146.4171.455.818e scoreE79.6429.4659.387.590s scoreS94.9723.5473.249.868g scoreG92.5133.6576.477.383Finan-cial indexSolvency capacityQuick RatioS19.1450.1071.1110.676Tangible Assets to Interest-Bearing Debt RatioS20.9870.0000.2220.150Long-Term Debt to Operating Working Capital RatioS3384.7-61.382.15119.36EBITDA to Total Liabilities RatioS41.388-4.0360.0750.202Tangible Net Worth Debt RatioS560.650.1033.8495.747Development capacityCapital Preservation and Appreciation RateD111.50-9.3841.2150.758Fixed Asset Growth RateD261.90-0.8810.4573.024Total Asset Growth RateD35.419-0.4770.2100.388Operating Revenue Growth RateD427.08-0.8330.2151.133Accrued ItemsD525.3714.2219.861.791Sustainable Growth RateD66.077-10.200.0360.434Operating capacityAccounts Receivable Turnover RatioO1833.80.0787.68643.93Inventory Turnover RatioO220,5990.24343.25674.3Accounts Payable Turnover RatioO318.920.0822.2701.622Cash and Cash Equivalents Turnover RatioO4437.60.1737.21018.32Fixed Assets to Revenue RatioO53.1960.0000.1250.201Capital IntensityO639.510.5102.3402.212Profitability capacityProfit Before Tax to EBIT RatioP133.60-22.820.7811.643Expense-to-Sales RatioP21.0060.0090.0570.068Financial Expense RatioP32.323-0.1920.0230.095Cost-to-Income RatioP40.757-5.1820.0150.307Asset Impairment Loss to Operating Revenue RatioP52.057-2.847-0.0010.164Earnings Before Interest and Taxes MarginP60.982-7.8300.0030.451 Descriptive statistics of the final indicator system. Modified KMV model Core Modifications of the Revised KMV Model: To address the limitations of the traditional KMV model (namely, the assumption that stock returns follow a normal distribution and the use of simple historical volatility to estimate the volatility of equity value), this study employs the GARCH (1,1) model to capture the leptokurtic, fat-tailed characteristics and volatility clustering of stock returns, thereby enabling a more accurate estimation of equity value volatility. Meanwhile, in response to the requirements for assessing green credit risk in the construction industry, environmental indicators are integrated into the subsequent risk prediction framework. This forms a green credit risk assessment system that combines the quantification of default risk using core financial parameters with the supplementation of risk characteristics via environmental indicators. In this study, environmental indicators are not directly involved in the calculation of KMV model parameters such as asset value, default point, DD, and EDF. Instead, they serve as features together with DD and EDF output from the KMV model for the subsequent training of the OptCAT model. KMV Corporation believes that a company’s default depends on the company’s asset value. There will be no default when the debt value is less than the asset value, and a default event will occur when the debt value is greater than the asset value. At the same time, KMV Corporation considers the company’s default threshold to be the company’s debt value and further calculates the EDF through the correspondence between DD and the probability of default risk. The parameters involved in the model include equity market value E, asset value VA, debt value D, risk-free rate r, debt maturity T, equity value volatility σE, and asset value volatility σA. Among these, variables D, r, and T are known, VA and σA is unknown. Since the premise assumes that the company’s asset value follows a normal distribution, the asset value VA and its volatility σA can be calculated from the equity market value E and its volatility.Equity market value E, the calculation formula is as Eq. (1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{E}} = {\mathrm{SP}} \times {\mathrm{NS}}$$\end{document} where SP represents the stock price, and NS is the number of stocks.Equity value volatility σE, the calculation formula is as Eqs. (2)–(5).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i} = \ln \left( {\frac{{S_{{i + 1}} }}{{S_{i} }}} \right)$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu } = \sum\limits_{{i = 1}}^{n} {\mu _{i} }$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma = \sqrt {\frac{{\sum\limits_{{i = 1}}^{n} {\left( {\mu _{i} + \bar{\mu }} \right)^{2} } }}{{n - 1}}}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma { \times } \sqrt N$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i}$$\end{document} represents the daily return of the stock, Si and Si+1 is the closing price of the stock on the i and i + 1 day, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu }$$\end{document} denotes the average return of the stock over a specific time period, σ is the daily volatility of the stock’s returns, σE represents the annual volatility of the stock’s value, and N is the actual number of trading days for the stock in the current year.Debt value D, the calculation formula is as Eq. (6).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}} = {\mathrm{SL}} + {\mathrm{LL}} N$$\end{document}where SL represents the short-term liabilities, and LL is the long-term liabilities.Asset value VA and and its volatility σA, the calculation formula is as Eq. (7).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma _{A} N\left( {d_{1} } \right)\frac{{V_{A} }}{E}$$\end{document}Default distance (DD).Corporate debt is divided into short-term liabilities and long-term liabilities based on the duration. When the value of corporate assets is less than the liabilities, the debtor may default, and when the value of corporate assets exceeds the liabilities, the debtor will repay the debt according to the agreement. There is a critical point, namely the default point, between default and scheduled debt repayment. Therefore, the default point is inevitably the sum of short-term liabilities and a certain proportion of long-term liabilities. The calculation formula is as Eqs. (8)–(9):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DP}} = {\mathrm{SD}} + {\mathrm{q}} \times {\mathrm{LD}}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DD}} = \left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) - {\mathrm{DP}}} \right]/\left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) \times \sigma _{{\mathrm{A}}} } \right]$$\end{document}where SD represents short-term liabilities, LD represents long-term liabilities, θ is the long-term debt weight coefficient (default point coefficient) and E(VA) represents the value of the enterprise’s assets when the debt matures.For the value of θ, the original KMV Corporation recommended θ = 0.5 based on empirical data from US listed companies, meaning the default point equals short-term liabilities plus 50% of long-term liabilities. This study retains this mainstream setting.Expected default frequency (EDF).The EDF refers to the probability of a company facing default risk within a specific period. The calculation formula is as Eq. (10):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{EDF}} = {\mathrm{1}} - {\mathrm{N}}\left( {{\mathrm{DD}}} \right)$$\end{document} Equity market value E, the calculation formula is as Eq. (1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{E}} = {\mathrm{SP}} \times {\mathrm{NS}}$$\end{document} where SP represents the stock price, and NS is the number of stocks. Equity market value E, the calculation formula is as Eq. (1).1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{E}} = {\mathrm{SP}} \times {\mathrm{NS}}$$\end{document} where SP represents the stock price, and NS is the number of stocks. Equity value volatility σE, the calculation formula is as Eqs. (2)–(5).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i} = \ln \left( {\frac{{S_{{i + 1}} }}{{S_{i} }}} \right)$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu } = \sum\limits_{{i = 1}}^{n} {\mu _{i} }$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma = \sqrt {\frac{{\sum\limits_{{i = 1}}^{n} {\left( {\mu _{i} + \bar{\mu }} \right)^{2} } }}{{n - 1}}}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma { \times } \sqrt N$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i}$$\end{document} represents the daily return of the stock, Si and Si+1 is the closing price of the stock on the i and i + 1 day, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu }$$\end{document} denotes the average return of the stock over a specific time period, σ is the daily volatility of the stock’s returns, σE represents the annual volatility of the stock’s value, and N is the actual number of trading days for the stock in the current year. Equity value volatility σE, the calculation formula is as Eqs. (2)–(5).2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i} = \ln \left( {\frac{{S_{{i + 1}} }}{{S_{i} }}} \right)$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu } = \sum\limits_{{i = 1}}^{n} {\mu _{i} }$$\end{document}4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma = \sqrt {\frac{{\sum\limits_{{i = 1}}^{n} {\left( {\mu _{i} + \bar{\mu }} \right)^{2} } }}{{n - 1}}}$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma { \times } \sqrt N$$\end{document} where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mu _{i}$$\end{document} represents the daily return of the stock, Si and Si+1 is the closing price of the stock on the i and i + 1 day, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\bar{\mu }$$\end{document} denotes the average return of the stock over a specific time period, σ is the daily volatility of the stock’s returns, σE represents the annual volatility of the stock’s value, and N is the actual number of trading days for the stock in the current year. Debt value D, the calculation formula is as Eq. (6).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}} = {\mathrm{SL}} + {\mathrm{LL}} N$$\end{document}where SL represents the short-term liabilities, and LL is the long-term liabilities. Debt value D, the calculation formula is as Eq. (6).6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{D}} = {\mathrm{SL}} + {\mathrm{LL}} N$$\end{document}where SL represents the short-term liabilities, and LL is the long-term liabilities. Asset value VA and and its volatility σA, the calculation formula is as Eq. (7).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma _{A} N\left( {d_{1} } \right)\frac{{V_{A} }}{E}$$\end{document} Asset value VA and and its volatility σA, the calculation formula is as Eq. (7).7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _{E} = \sigma _{A} N\left( {d_{1} } \right)\frac{{V_{A} }}{E}$$\end{document} Default distance (DD).Corporate debt is divided into short-term liabilities and long-term liabilities based on the duration. When the value of corporate assets is less than the liabilities, the debtor may default, and when the value of corporate assets exceeds the liabilities, the debtor will repay the debt according to the agreement. There is a critical point, namely the default point, between default and scheduled debt repayment. Therefore, the default point is inevitably the sum of short-term liabilities and a certain proportion of long-term liabilities. The calculation formula is as Eqs. (8)–(9):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DP}} = {\mathrm{SD}} + {\mathrm{q}} \times {\mathrm{LD}}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DD}} = \left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) - {\mathrm{DP}}} \right]/\left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) \times \sigma _{{\mathrm{A}}} } \right]$$\end{document}where SD represents short-term liabilities, LD represents long-term liabilities, θ is the long-term debt weight coefficient (default point coefficient) and E(VA) represents the value of the enterprise’s assets when the debt matures.For the value of θ, the original KMV Corporation recommended θ = 0.5 based on empirical data from US listed companies, meaning the default point equals short-term liabilities plus 50% of long-term liabilities. This study retains this mainstream setting. Default distance (DD). Corporate debt is divided into short-term liabilities and long-term liabilities based on the duration. When the value of corporate assets is less than the liabilities, the debtor may default, and when the value of corporate assets exceeds the liabilities, the debtor will repay the debt according to the agreement. There is a critical point, namely the default point, between default and scheduled debt repayment. Therefore, the default point is inevitably the sum of short-term liabilities and a certain proportion of long-term liabilities. The calculation formula is as Eqs. (8)–(9):8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DP}} = {\mathrm{SD}} + {\mathrm{q}} \times {\mathrm{LD}}$$\end{document}9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{DD}} = \left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) - {\mathrm{DP}}} \right]/\left[ {{\mathrm{E}}\left( {{\mathrm{V}}_{{\mathrm{A}}} } \right) \times \sigma _{{\mathrm{A}}} } \right]$$\end{document}where SD represents short-term liabilities, LD represents long-term liabilities, θ is the long-term debt weight coefficient (default point coefficient) and E(VA) represents the value of the enterprise’s assets when the debt matures. For the value of θ, the original KMV Corporation recommended θ = 0.5 based on empirical data from US listed companies, meaning the default point equals short-term liabilities plus 50% of long-term liabilities. This study retains this mainstream setting. Expected default frequency (EDF).The EDF refers to the probability of a company facing default risk within a specific period. The calculation formula is as Eq. (10):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{EDF}} = {\mathrm{1}} - {\mathrm{N}}\left( {{\mathrm{DD}}} \right)$$\end{document} Expected default frequency (EDF). The EDF refers to the probability of a company facing default risk within a specific period. The calculation formula is as Eq. (10):10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\mathrm{EDF}} = {\mathrm{1}} - {\mathrm{N}}\left( {{\mathrm{DD}}} \right)$$\end{document} KMV model assumes that stock returns follow a normal distribution and uses the logarithmic return method to calculate the equity value volatility of a company. However, in reality, stock returns do not strictly follow a normal distribution but often exhibit a leptokurtic and fat-tailed distribution. In other words, stock return rates demonstrate volatility clustering (ARCH effect). The GARCH model is an extension of the ARCH model proposed by Bollerslev. This model fits well with financial time series data and effectively models the variance of errors after analyzing the heteroscedasticity of the data. This enables the model to better capture the dynamic volatility characteristics of the data. GARCH(1,1) refers to an autoregressive model of order 1. In academic research in the field of economic and financial studies in China, GARCH(1,1) is often used to estimate the volatility of stock markets. Therefore, this study also adopts GARCH(1,1) and utilizes Eviews 13 software to calculate the equity value volatility. The formula for the GARCH(1,1) model is as Eq. (11):11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{gathered} \sigma _{t}^{2} = \alpha _{0} + \alpha _{1} \alpha _{{t - 1}}^{2} + \beta \sigma _{{t - 1}}^{2} \hfill \\ 0 \le \alpha _{1} ,\beta \le 0,\left( {\alpha _{1} + \beta } \right) < 1 \hfill \\ \end{gathered}$$\end{document} where α represents the coefficient of return, and β represents the lag coefficient. The step-by-step algorithm for calculating the DD and EDF of construction enterprises using the revised KMV model is as follows:Step 1: Data collection. Obtain the daily closing stock price, number of tradable shares, short-term liabilities, long-term liabilities of listed construction enterprises, and the risk-free interest rate for the same period at time t-1.Step 2: Calculation of equity market value. Calculate the equity market value E of the enterprise at the end of period t-1 according to Eq. (1).Step 3: Estimation of equity value volatility using GARCH (1,1). Compute the daily stock return according to Eq. (2), construct the GARCH (1,1) model shown in Eq. (11), estimate the conditional variance, and calculate the annual volatility of equity value σE.Step 4: Calculation of debt value and default point. Calculate the debt value and default point according to Eq. (6) and Eq. (8), respectively.Step 5: Solving for asset value and asset volatility. With the σE inputs, solve for the enterprise’s asset market value Vt and annual asset value volatility σV using the Newton-Raphson iteration method based on the Black-Scholes equations in Eq. (7).Step 6: Calculation of DD and EDF. Calculate DD and EDF sequentially according to Eq. (9) and Eq. (10), yielding the quantitative green credit default risk indicators of the enterprise for period t. Step 1: Data collection. Obtain the daily closing stock price, number of tradable shares, short-term liabilities, long-term liabilities of listed construction enterprises, and the risk-free interest rate for the same period at time t-1. Step 1: Data collection. Obtain the daily closing stock price, number of tradable shares, short-term liabilities, long-term liabilities of listed construction enterprises, and the risk-free interest rate for the same period at time t-1. Step 2: Calculation of equity market value. Calculate the equity market value E of the enterprise at the end of period t-1 according to Eq. (1). Step 2: Calculation of equity market value. Calculate the equity market value E of the enterprise at the end of period t-1 according to Eq. (1). Step 3: Estimation of equity value volatility using GARCH (1,1). Compute the daily stock return according to Eq. (2), construct the GARCH (1,1) model shown in Eq. (11), estimate the conditional variance, and calculate the annual volatility of equity value σE. Step 3: Estimation of equity value volatility using GARCH (1,1). Compute the daily stock return according to Eq. (2), construct the GARCH (1,1) model shown in Eq. (11), estimate the conditional variance, and calculate the annual volatility of equity value σE. Step 4: Calculation of debt value and default point. Calculate the debt value and default point according to Eq. (6) and Eq. (8), respectively. Step 4: Calculation of debt value and default point. Calculate the debt value and default point according to Eq. (6) and Eq. (8), respectively. Step 5: Solving for asset value and asset volatility. With the σE inputs, solve for the enterprise’s asset market value Vt and annual asset value volatility σV using the Newton-Raphson iteration method based on the Black-Scholes equations in Eq. (7). Step 5: Solving for asset value and asset volatility. With the σE inputs, solve for the enterprise’s asset market value Vt and annual asset value volatility σV using the Newton-Raphson iteration method based on the Black-Scholes equations in Eq. (7). Step 6: Calculation of DD and EDF. Calculate DD and EDF sequentially according to Eq. (9) and Eq. (10), yielding the quantitative green credit default risk indicators of the enterprise for period t. Step 6: Calculation of DD and EDF. Calculate DD and EDF sequentially according to Eq. (9) and Eq. (10), yielding the quantitative green credit default risk indicators of the enterprise for period t. Green credit risk prediction model CatBoost algorithm CatBoost algorithm is a machine learning method based on Gradient Boosting Decision Trees (GBDT). It was introduced in 2017 by engineers from the Russian company Yandex. CatBoost can be used for classification or regression of numerical data types, ranking, binary and multi-class classification, and multitasking. The name CatBoost comes from the combination of “Category” and “Boosting,” indicating that the algorithm is designed for handling categorical data. In the algorithm, numerical features of the target statistics are used to replace categorical features to achieve the desired estimation. CatBoost algorithm innovatively proposes a method for handling categorical features as numerical features. It adopts a ranking boosting method to handle noise and avoid gradient bias issues. Additionally, it uses fully symmetric trees as base classifiers during the training process. The following will provide a detailed explanation of these innovative points. Categorical features Categorical features Categorical features In machine learning, categorical features are often challenging to handle. Poor handling of these features can significantly impact the predictive performance of the model. Categorical features may lack comparability between categories, making them unsuitable for direct use in binary decision trees. The most common approach to dealing with categorical features is to preprocess them by converting them into numerical values, such as 0, 1, 2, for each category in each example. CatBoost algorithm, based on traditional boosting algorithms, employs one-hot encoding during the preprocessing stage to convert categorical features into numerical features. However, one-hot encoding generates a large number of new features, and if the number of feature values is too large, it can lead to the curse of dimensionality. When dealing with categorical features in gradient boosting algorithms, a common approach is to use the average value of the corresponding labels for the categorical features. In decision tree models, the Greedy Target-based Statistics method is often used, where the average value of the labels serves as the criterion for node splitting. The expression for this method is given by Eq. (12):12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_{ik}}=\frac{{\sum\limits_{{j=1}}^{n} {\left( {{x_{jk}}={x_{ik}}} \right){y_j}} }}{{\sum\limits_{{j=1}}^{n} {\left( {{x_{jk}}={x_{ik}}} \right)} }}$$\end{document} However, in practice, this statistical strategy has a significant drawback. Simply replacing categorical features with the average value of the labels can result in a certain degree of loss of information carried by the categories. Additionally, if there are structural and distribution differences between the training and testing data, this dissimilarity can lead to the problem of covariate shift. To address the shortcomings of the aforementioned statistical strategy, a smoothing prior distribution term P is introduced. This term utilizes the target variable to estimate the expected values for each category, aiming to reduce the impact of noise. The expression for this approach is provided in Eq. (13):13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${x_{ik}}=\frac{{\sum\limits_{{j=1}}^{n} {\left( {{x_{jk}}={x_{ik}}} \right){y_i}+\alpha P} }}{{\sum\limits_{{j=1}}^{n} {\left( {{x_{jk}}={x_{ik}}} \right)+\alpha } }}$$\end{document}where α is a weight coefficient greater than 0. P serves as the smoothing prior denoting the target mean in the dataset. Through this step, it can mitigate the impact of low-frequency sample means on the model, reduce noise in the data, but may potentially lead to overfitting issues in the model. CatBoost also proposes an efficient improvement method based on Ordered Target Statistics, which relies on the principle of ordering, where the target variable statistics for each sample depend only on the observed history. To implement this idea on offline data, the CatBoost algorithm introduces “artificial time” by randomly shuffling the data. For each sample, its target variable statistics are calculated using the data before it, taking \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${D_k}=D$$\end{document} as a test example. See Eq. (14):14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{gathered} {D_k}=\left\{ {{x_j}:\sigma \left( j \right)<\sigma \left( k \right)} \right\}, \hfill \\ {x_{ik}}=\frac{{\sum\limits_{{{x_j} \in {D_k}}} {\left( {{x_{jk}}={x_{ik}}} \right){y_i}+\alpha P} }}{{\sum\limits_{{{x_j} \in {D_k}}} {\left( {{x_{jk}}={x_{ik}}} \right)+\alpha } }} \hfill \\ \end{gathered}$$\end{document} The CatBoost algorithm uses different permutations in different gradient boosting steps. When constructing a new split point in the current decision tree, the CatBoost algorithm performs the first split without considering any combinations of categorical features. However, in the next split, the CatBoost algorithm takes into account both the combinations of categorical features in the current decision tree and all categorical features in the dataset. (2)Gradient bias (2)Gradient bias Gradient bias Overfitting issues exist in all classical boosting algorithms. When estimating the gradient at each step, the same data points from the current model are used, causing the estimated gradient distribution in any domain of the feature space to deviate from the true distribution of gradients in that domain. This leads to overfitting, known as gradient bias. To address the overfitting problem, the CatBoost algorithm introduces some improvements based on classical gradient boosting. Here is a brief description: Let’s assume that the model constructed after building the i-th tree is denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F^i}$$\end{document}, and the gradient values on the k-th training sample after building the i-th tree are denoted as \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g^i}\left( {{X_k},{Y_k}} \right)$$\end{document}. To make the gradient \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${g^i}\left( {{X_k},{Y_k}} \right)$$\end{document} unbiased with respect to model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F^i}$$\end{document}, it is necessary to train model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_k}$$\end{document} without observing \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F^i}$$\end{document}. In order to obtain unbiased gradients for all training samples, training \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${F^i}$$\end{document} cannot use any observed values. The innovation of the CatBoost algorithm lies in training a separate model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_k}$$\end{document} for each sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M_k}$$\end{document}, and model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M_k}$$\end{document} never uses the gradient estimates of sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_k}$$\end{document} for updates. The gradient of sample \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${M_k}$$\end{document} is estimated using model \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${X_k}$$\end{document}, and this estimate is used to score the generated trees. (3)Prediction shift (3)Prediction shift Prediction shift Prediction shift is a common issue in all algorithms based on the GBDT framework, caused by gradient bias. In the GBDT algorithm framework, each iteration uses the same sample dataset to obtain the gradient of the current model through the loss function, training and obtaining the base learner. This leads to gradient estimation bias, affecting the predictive performance of the model. The CatBoost algorithm uses the Ordered Boosting method instead of traditional gradient estimation methods, effectively reducing the bias in gradient estimation and thereby improving the model’s generalization ability. The CatBoost algorithm optimizes the gradient boosting decision tree method by using ordered boosting to ensure that all datasets can be used for training and learning. It reduces overfitting in training by adjusting the learning rate. Unlike other gradient boosting algorithms such as XGBoost, CatBoost builds symmetric trees, using the same conditions to split the leaves of the previous tree in each step. This balanced tree structure facilitates efficient CPU implementation, reduces prediction time, and regularizes the model structure to prevent overfitting. Based on the innovative features of the CatBoost algorithm described above, several major advantages can be summarized. Firstly, CatBoost algorithm exhibits excellent performance and ranks among the top machine learning algorithms. Secondly, it addresses the issues of gradient bias and prediction shift present in traditional GBDT algorithm frameworks, enhancing the algorithm’s generalization and prediction capabilities. Additionally, CatBoost’s handling of categorical features allows for good predictive performance even with low-dimensional features. Finally, CatBoost provides out-of-the-box GPU support, enabling a significant reduction in training time when using GPUs. Hyperparameter optimization Optuna is an open-source hyperparameter optimization framework proposed by Akiba et al. in 2019. It helps users automatically adjust hyperparameters of machine learning models to achieve optimal performance. Optuna framework has three main features: first, it can implement distributed parallel optimization. Second, the framework can prune experiments that show no hope of improvement. Finally, Optuna supports various hyperparameter optimization algorithms, including Bayesian optimization, random search, gradient-based optimization, etc. It helps users quickly find the best hyperparameter combinations, supporting multiple hyperparameter types such as integers, floats, booleans, categorical values, etc. Optuna also offers various visualization tools to help users visualize the hyperparameter optimization process, better understand the model’s behavior, and efficiently adjust hyperparameters for optimal performance. The core idea of the Optuna framework is an optimized transformation based on the Parzen optimizer’s concept in Bayesian optimization algorithms. The Bayesian optimization algorithm relies on the idea that the combination of hyperparameter values to be tested in the next step depends on the model’s historical data. The Optuna framework selects the configuration space for hyperparameter combinations based on historical data and conducts hyperparameter search within this configuration space. The result obtained from each search is used to update the configuration space and inform the next search. In a simplest Optuna optimization program, there are three fundamental concepts: the objective function, a trial, and a study. The objective function is responsible for defining the function to be optimized and specifying the range of hyperparameters. A trial corresponds to a single execution of the objective function, and a study manages the optimization process, deciding how optimization is performed, the total number of trials, recording trial results, and more. The optimization process for hyperparameters in Optuna follows the following steps:Step 1: Initialize the objective function, defining the types of parameters and the search space for parameters.Step 2: Initialize the study, setting the optimization method and parameters for a single trial, such as the maximum number of iterations.Step 3: Begin the hyperparameter tuning process.Step 4: When the maximum number of iterations is reached, output the optimal hyperparameters and the value of the objective function. Step 1: Initialize the objective function, defining the types of parameters and the search space for parameters. Step 1: Initialize the objective function, defining the types of parameters and the search space for parameters. Step 2: Initialize the study, setting the optimization method and parameters for a single trial, such as the maximum number of iterations. Step 2: Initialize the study, setting the optimization method and parameters for a single trial, such as the maximum number of iterations. Step 3: Begin the hyperparameter tuning process. Step 3: Begin the hyperparameter tuning process. Step 4: When the maximum number of iterations is reached, output the optimal hyperparameters and the value of the objective function. Step 4: When the maximum number of iterations is reached, output the optimal hyperparameters and the value of the objective function. The Optuna framework maintains good stability and can perform hyperparameter optimization in complex spaces. It can predict parameter ranges with low termination probability, allowing more computing power to be used for regions with higher likelihood, significantly improving search efficiency. Therefore, with advantages such as high optimization efficiency and stability, Optuna has become one of the most widely used frameworks for automatic hyperparameter optimization. Benchmark models selection rationale To comprehensively and objectively evaluate the performance of the proposed OptCAT model in green credit risk prediction, this study selects four representative machine learning algorithms as benchmark models: SVM, RF, XGBoost, and LightGBM. The selection of these specific models is based on three core rationales that ensure the rigor, comparability, and validity of the comparative experiments. Firstly, these models are the most widely recognized and universally adopted benchmarks in the field of credit risk evaluation, with extensive validation in both academic research and industrial practice. SVM has long been a classic method for credit scoring due to its strong nonlinear fitting ability and stability in small-sample scenarios9. RF, as a representative bagging ensemble algorithm, is widely used for its excellent robustness and anti-overfitting capability19. XGBoost and LightGBM, as state-of-the-art gradient boosting tree models, have become the de facto standard for financial risk prediction tasks in recent years26. By using these universally accepted benchmarks, our experimental results can be directly compared with existing studies on green credit risk prediction, ensuring the academic comparability of our findings. Secondly, the selected benchmark models cover three distinct machine learning paradigms, allowing us to verify the superiority of the OptCAT model from multiple dimensions: (1) SVM: Represents traditional statistical learning methods, which excel at handling high-dimensional mixed data and provide a baseline for comparing with modern ensemble methods. (2)RF: Represents the bagging ensemble learning paradigm, which reduces prediction variance through parallel training of multiple decision trees. Comparing with RF helps us demonstrate the advantages of gradient boosting (the paradigm adopted by CatBoost) in sequential error correction for risk prediction tasks. (3)XGBoost and LightGBM: Belong to the same GBDT family as CatBoost, representing the two most mature and widely used variants of GBDT. Direct comparison with these two models can isolate and highlight two key contributions of our study: ①the performance improvement brought by Optuna-based hyperparameter optimization; and ② CatBoost’s inherent advantages in handling categorical features and mitigating gradient bias, which are particularly relevant for green credit risk assessment involving environmental indicators. Thirdly, our green credit risk prediction task has two key characteristics: (1) a mixed dataset containing both numerical financial indicators and categorical environmental indicators; and (2) a relatively limited sample size (949 firm-year observations) with imbalanced default samples. The selected benchmark models are all proven to perform well under these conditions. We intentionally excluded models that are unsuitable for this study: ①Simple models such as logistic regression and single decision trees were excluded because their predictive performance is significantly inferior to ensemble methods, as widely documented in previous credit risk literature. ②Deep learning models (e.g., neural networks) were excluded because they require large amounts of training data to achieve optimal performance and are prone to overfitting on our relatively small sample size, which would lead to unfair comparisons. Significance of environmental indicators in the context of china’s green credit policy Since China issued the Green Credit Guidelines in 2012, environmental risk has been formally incorporated into the credit risk management framework of commercial banks. Following the proposal of the “Dual Carbon” goal in 2020, the construction industry, an important source of carbon emissions, has become a key sector for green finance support and supervision. In accordance with the Special Statistical System for Green Loans issued by the People’s Bank of China, construction enterprises applying for green loans must meet the criteria specified in the Catalogue of Green Bond Supported Projects, covering green buildings, ultra-low energy consumption buildings, renewable energy applications in construction, and other related fields. In this study, we link environmental indicators to green credit risk through two approaches. First, among the 119 listed construction enterprises in the sample, 43 publicly disclosed that they had obtained green loans or issued green bonds during the sample period. Second, regulatory requirements on green credit risk issued by financial authorities clearly state that “credit shall not be granted or existing credit lines shall be reduced for clients with non-compliant environmental and social performance.” The ST designation used in this study is highly associated with operational difficulties caused by environmental violations. Therefore, by integrating ESG environmental indicators into the credit risk assessment system, this study not only complies with the requirements of Chinese regulators for green credit risk evaluation but also reflects the real transmission mechanism of environmental risks faced by the construction industry under the “Dual Carbon” strategy. Experimental design Experimental steps This study adopts a two-stage green credit risk prediction framework of “KMV-based risk quantification + OptCAT regression prediction + ST label verification”. The core prediction task is the continuous regression prediction of the EDF, with mean squared error (MSE), root mean squared error (RMSE), and mean absolute error (MAE) used to evaluate the regression performance of the model. Meanwhile, the ST enterprise label is utilized to verify the practical risk effectiveness of the EDF prediction results. The specific experimental steps are as follows:Step 1: Indicator screening and dataset construction. A total of 117 indicators from 2009 to 2022 were preprocessed. Through the Mann-Whitney U test and the feature ranking method based on CatBoost, 27 core features were finally obtained as the input features of the OptCAT model. All indicators were processed with a one-period lag, that is, the indicator data of year t-1 were used to provide a basis for the green credit risk prediction of year t.Step 2: Risk quantification of the revised KMV model. Based on the stock trading data and financial data of year t-1, the DD and EDF of year t were calculated. The EDF of year t was taken as the target output of the OptCAT model, and the environmental and financial indicators of year t-1 were taken as the input features of the OptCAT model, completing the dataset matching of “input features-target variables”.Step 3: Construction and prediction of the OptCAT model. Optuna was combined with CatBoost to construct the OptCAT regression prediction model. The matched dataset was divided into a training set and a test set for model training and testing. Meanwhile, the ST and non-ST labels of enterprises from 2021 to 2022 were used to verify the ability of the predicted EDF values to distinguish the actual credit risks of enterprises. Step 1: Indicator screening and dataset construction. A total of 117 indicators from 2009 to 2022 were preprocessed. Through the Mann-Whitney U test and the feature ranking method based on CatBoost, 27 core features were finally obtained as the input features of the OptCAT model. All indicators were processed with a one-period lag, that is, the indicator data of year t-1 were used to provide a basis for the green credit risk prediction of year t. Step 1: Indicator screening and dataset construction. A total of 117 indicators from 2009 to 2022 were preprocessed. Through the Mann-Whitney U test and the feature ranking method based on CatBoost, 27 core features were finally obtained as the input features of the OptCAT model. All indicators were processed with a one-period lag, that is, the indicator data of year t-1 were used to provide a basis for the green credit risk prediction of year t. Step 2: Risk quantification of the revised KMV model. Based on the stock trading data and financial data of year t-1, the DD and EDF of year t were calculated. The EDF of year t was taken as the target output of the OptCAT model, and the environmental and financial indicators of year t-1 were taken as the input features of the OptCAT model, completing the dataset matching of “input features-target variables”. Step 2: Risk quantification of the revised KMV model. Based on the stock trading data and financial data of year t-1, the DD and EDF of year t were calculated. The EDF of year t was taken as the target output of the OptCAT model, and the environmental and financial indicators of year t-1 were taken as the input features of the OptCAT model, completing the dataset matching of “input features-target variables”. Step 3: Construction and prediction of the OptCAT model. Optuna was combined with CatBoost to construct the OptCAT regression prediction model. The matched dataset was divided into a training set and a test set for model training and testing. Meanwhile, the ST and non-ST labels of enterprises from 2021 to 2022 were used to verify the ability of the predicted EDF values to distinguish the actual credit risks of enterprises. Step 3: Construction and prediction of the OptCAT model. Optuna was combined with CatBoost to construct the OptCAT regression prediction model. The matched dataset was divided into a training set and a test set for model training and testing. Meanwhile, the ST and non-ST labels of enterprises from 2021 to 2022 were used to verify the ability of the predicted EDF values to distinguish the actual credit risks of enterprises. Sample selection and division This study selected 119 enterprises in China’s construction industry from 2009 to 2022 as research samples. ST and *ST companies were defined as ST samples, while normally operating companies were defined as non-ST samples. The total number of samples was 949, divided into a training set and a test set. Data from 2009 to 2020 were assigned to the training set, while data from 2021 to 2022 were assigned to the test set. The training set consisted of 730 samples, including 80 ST samples, and the test set comprised 219 samples, including 22 ST samples. Due to the 10.96% proportion of the “ST group”, it is evident that the imbalance in sample sizes across different categories poses a problem. To address this issue, the Borderline SMOTE algorithm was employed. This study strictly adheres to the no-data-leakage principle of machine learning data processing. The Borderline SMOTE oversampling operation is independently performed only on the training set. Specifically, after completing the time-based division of the training set and test set, oversampling is only applied to the minority class (ST samples) in the training set until the ratio of ST to non-ST samples in the training set reaches 1:1. For the test set, this study completely retains its original sample distribution characteristics without any resampling, data balancing, or other processing. The model trained on the training set is directly applied to the original test set to truly simulate the unbalanced data scenario of the model in actual credit decision-making, ensuring the practical application value of the prediction results. Introduced by Han et al., in 20058, this oversampling algorithm is an improvement on the SMOTE algorithm. The specific steps are as follows:Step 1: Calculate the Euclidean distance between each minority class sample and other samples to obtain the k-nearest neighbors for each point.Step 2: Based on a partitioning criterion, classify minority class samples into three categories: Danger, Safe, and Noise.Step 3: Combine any minority class samples within the Danger region with randomly selected minority class samples within k-nearest neighbors, and use linear interpolation formula \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{new}} = x_{i} + rand\left( {0,1} \right) \cdot \left| {x_{i} - \hat{x}_{i} } \right|$$\end{document} to synthesize new minority class samples until the ratio of normal and ST samples is 1:1, obtaining a balanced dataset and improving the class distribution of the samples to achieve better results. Step 1: Calculate the Euclidean distance between each minority class sample and other samples to obtain the k-nearest neighbors for each point. Step 1: Calculate the Euclidean distance between each minority class sample and other samples to obtain the k-nearest neighbors for each point. Step 2: Based on a partitioning criterion, classify minority class samples into three categories: Danger, Safe, and Noise. Step 2: Based on a partitioning criterion, classify minority class samples into three categories: Danger, Safe, and Noise. Step 3: Combine any minority class samples within the Danger region with randomly selected minority class samples within k-nearest neighbors, and use linear interpolation formula \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{new}} = x_{i} + rand\left( {0,1} \right) \cdot \left| {x_{i} - \hat{x}_{i} } \right|$$\end{document} to synthesize new minority class samples until the ratio of normal and ST samples is 1:1, obtaining a balanced dataset and improving the class distribution of the samples to achieve better results. Step 3: Combine any minority class samples within the Danger region with randomly selected minority class samples within k-nearest neighbors, and use linear interpolation formula \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x_{{new}} = x_{i} + rand\left( {0,1} \right) \cdot \left| {x_{i} - \hat{x}_{i} } \right|$$\end{document} to synthesize new minority class samples until the ratio of normal and ST samples is 1:1, obtaining a balanced dataset and improving the class distribution of the samples to achieve better results. This study strictly follows the time-alignment rule of “1-period lagged input, current-period output and validation” throughout the empirical process. The core framework is as follows: environmental indicators, financial indicators, stock data, and financial data from year t-1 are used to calculate the EDF for year t based on a modified KMV model. Meanwhile, environmental and financial indicators from year t-1 serve as inputs to the OptCAT model to predict the EDF for year t. The predicted EDF for year t is then jointly validated against the firm’s actual special treatment (ST) and non-ST status in year t. The specific time-matching scheme is detailed below and presented in Table 4:Data lagging procedure. Consistent with the setting in the original literature, environmental indicators adopt the ratings from year t-1. Financial indicators, stock data, and financial data are all based on annual observations of year t-1.EDF calculation. EDF for 2009 is estimated using stock and financial data from 2008; EDF for 2010 is computed using data from 2009, and so on. This procedure yields a continuous EDF series covering 2009–2022.Model training and testing. The training set spans 2009–2020, matching input features from 2008 to 2019 with target EDF values from 2009 to 2020. The test set covers 2021–2022, matching input features from 2020 to 2021 with target EDF values from 2021 to 2022.ST status validation. EDF values from 2009 to 2022 are matched with firms’ actual ST and non-ST status in the corresponding year t to verify the discriminative power of EDF for corporate default risk. The 2021–2022 period is used for out-of-sample validation of the test set. Furthermore, to ensure the validity of the EDF estimate for 2009, this study additionally collects pre-data for 2008, guaranteeing the rigor of time alignment. Data lagging procedure. Consistent with the setting in the original literature, environmental indicators adopt the ratings from year t-1. Financial indicators, stock data, and financial data are all based on annual observations of year t-1. Data lagging procedure. Consistent with the setting in the original literature, environmental indicators adopt the ratings from year t-1. Financial indicators, stock data, and financial data are all based on annual observations of year t-1. EDF calculation. EDF for 2009 is estimated using stock and financial data from 2008; EDF for 2010 is computed using data from 2009, and so on. This procedure yields a continuous EDF series covering 2009–2022. EDF calculation. EDF for 2009 is estimated using stock and financial data from 2008; EDF for 2010 is computed using data from 2009, and so on. This procedure yields a continuous EDF series covering 2009–2022. Model training and testing. The training set spans 2009–2020, matching input features from 2008 to 2019 with target EDF values from 2009 to 2020. The test set covers 2021–2022, matching input features from 2020 to 2021 with target EDF values from 2021 to 2022. Model training and testing. The training set spans 2009–2020, matching input features from 2008 to 2019 with target EDF values from 2009 to 2020. The test set covers 2021–2022, matching input features from 2020 to 2021 with target EDF values from 2021 to 2022. ST status validation. EDF values from 2009 to 2022 are matched with firms’ actual ST and non-ST status in the corresponding year t to verify the discriminative power of EDF for corporate default risk. The 2021–2022 period is used for out-of-sample validation of the test set. Furthermore, to ensure the validity of the EDF estimate for 2009, this study additionally collects pre-data for 2008, guaranteeing the rigor of time alignment. ST status validation. EDF values from 2009 to 2022 are matched with firms’ actual ST and non-ST status in the corresponding year t to verify the discriminative power of EDF for corporate default risk. The 2021–2022 period is used for out-of-sample validation of the test set. Furthermore, to ensure the validity of the EDF estimate for 2009, this study additionally collects pre-data for 2008, guaranteeing the rigor of time alignment. Table 4Data time matching relationship.Data typeTime dimensionSpecific purposeEnvironmental indicatorst−1The input features of the OptCAT model, together with the financial indicators for year t-1, form the predicted feature set.Financial indicatorst−1The input features of the OptCAT model, together with the environmental indicators for year t-1, form the predicted feature set.Stock datat−1Used to revise the KMV model for calculating DD and EDF in year t.KMV model outputtEDF is the target output of the OptCAT model; DD is an auxiliary input feature for the OptCAT model.OptCAT model prediction resultstPredicting EDF value for year t based on t-1 characteristics.ST/Non ST LabelstUsed to verify the actual discriminative ability of the EDF prediction results for year t, i.e. the t-1 indicator predicts the ST state for year t. Table 4Data time matching relationship.Data typeTime dimensionSpecific purposeEnvironmental indicatorst−1The input features of the OptCAT model, together with the financial indicators for year t-1, form the predicted feature set.Financial indicatorst−1The input features of the OptCAT model, together with the environmental indicators for year t-1, form the predicted feature set.Stock datat−1Used to revise the KMV model for calculating DD and EDF in year t.KMV model outputtEDF is the target output of the OptCAT model; DD is an auxiliary input feature for the OptCAT model.OptCAT model prediction resultstPredicting EDF value for year t based on t-1 characteristics.ST/Non ST LabelstUsed to verify the actual discriminative ability of the EDF prediction results for year t, i.e. the t-1 indicator predicts the ST state for year t. Data time matching relationship. Evaluation of model performance This article evaluates the performance of the model using three indicators: MSE, RMSE, and MAE. The specific formula can be found in Eqs. (15)–(17):15\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MSE=\frac{1}{n}\sum\limits_{{i=1}}^{n} {{{\left[ {{y_i} - f\left( {{x_i}} \right)} \right]}^2}}$$\end{document}16\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$RMSE=\sqrt {\frac{1}{n}\sum\limits_{{i=1}}^{n} {{{\left[ {{y_i} - f\left( {{x_i}} \right)} \right]}^2}} }$$\end{document}17\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$MAE=\frac{1}{n}\sum\limits_{{i=1}}^{n} {\left| {{y_i} - f\left( {{x_i}} \right)} \right|}$$\end{document} Furthermore, the calculation formulas for classification metrics including AUC, accuracy, precision, recall, and F1-score are given in Eqs. (18)-(21):18\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Accuracy=\frac{{TP+TN}}{{TP+TN+FP+FN}}$$\end{document}19\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Precision=\frac{{TP}}{{TP+FP}}$$\end{document}20\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$Recall=\frac{{TP}}{{TP+FN}}$$\end{document}21\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F1=2 \times \frac{{Accuracy \times Recall}}{{Accuracy+Recall}}$$\end{document}where TP denotes true positives (actual ST and predicted ST), TN denotes true negatives (actual non-ST and predicted non-ST), FP denotes false positives (actual non-ST but predicted ST), and FN denotes false negatives (actual ST but predicted non-ST). The closer the AUC value is to 1, the stronger the discriminative ability of the model. Parameter settings This article mainly optimizes the following parameters, selecting default parameters such as default parameter (DP), random search (RS), grid search (GS), bayesian optimization (BO), Hyperopt, and Optuna for comparison. The final set parameters are shown in Table 5. Next, this article selects traditional SVM, RF, XGBoost, LightGBM, and OptCAT models for comparison. The selected parameters are shown in Table 6. To ensure absolute fairness in the comparison of hyperparameter optimization algorithms, this study adopts two strict control measures: Firstly, all six algorithms use exactly the same search space for the five core CatBoost hyperparameters (learning_rate, depth, iterations, border_count, l2_leaf_reg), as specified in Table 5. Secondly, all optimization algorithms are allocated an identical maximum trial budget of 1000 iterations. For RS, 1000 independent random sampling trials are conducted as specified. However, due to the discrete nature of the search space, some parameter combinations are sampled multiple times. To optimize computational efficiency, repeated parameter combinations are not re-evaluated; instead, their previously computed performance results are directly reused. For GS, which exhaustively evaluates all discrete parameter combinations in the search space, the algorithm will automatically terminate early after traversing all unique grid points if the total number of combinations is less than 1000, without performing redundant repeated trials. For BO, Hyperopt, and Optuna, which are sequential model-based optimization algorithms, exactly 1000 independent iterative trials are executed as specified. Table 5Set parameters of the different optimization algorithms.AlgorithmsDPRSGSBOHyperoptOptunalearning_rate0.030.300.100.030.570.09depth686699iterations1000900300400899531border_count128280200140218422l2_leaf_reg311060116567number of trials/10001000100010001000Parameter search range/learning_rate∈[0.01,1]; depth∈[1,10]; iterations∈[100,1000]; border_count∈[32,512]; l2_leaf_reg∈[1,200] Table 5Set parameters of the different optimization algorithms.AlgorithmsDPRSGSBOHyperoptOptunalearning_rate0.030.300.100.030.570.09depth686699iterations1000900300400899531border_count128280200140218422l2_leaf_reg311060116567number of trials/10001000100010001000Parameter search range/learning_rate∈[0.01,1]; depth∈[1,10]; iterations∈[100,1000]; border_count∈[32,512]; l2_leaf_reg∈[1,200] Set parameters of the different optimization algorithms. Table 6Set parameters of the different models.ModelParametersRangeValueSVMkernel-“rbf”gamma-“auto”C[0.05, 10]9.90RFn_estimators[1, 200]131max_depth[1, 20]4min_samples_split[2, 12]6max_features[5, 30]29XGBoostn_estimators[1, 300]75max_depth[1, 10]9learning_rate[0.01, 1]0.15subsample[0.5, 1]0.78colsample_bytree[0.5, 1]0.98LightGBMn_estimators[1, 600]332max_depth[1, 10]4num_leaves[5, 200]181learning_rate[0.01, 1]0.68feature_fraction[0.1, 1]0.95bagging_fraction[0.5, 1]0.82min_data_in_leaf[1, 10]3 Table 6Set parameters of the different models.ModelParametersRangeValueSVMkernel-“rbf”gamma-“auto”C[0.05, 10]9.90RFn_estimators[1, 200]131max_depth[1, 20]4min_samples_split[2, 12]6max_features[5, 30]29XGBoostn_estimators[1, 300]75max_depth[1, 10]9learning_rate[0.01, 1]0.15subsample[0.5, 1]0.78colsample_bytree[0.5, 1]0.98LightGBMn_estimators[1, 600]332max_depth[1, 10]4num_leaves[5, 200]181learning_rate[0.01, 1]0.68feature_fraction[0.1, 1]0.95bagging_fraction[0.5, 1]0.82min_data_in_leaf[1, 10]3 Set parameters of the different models. Robustness test of the model Given the limited absolute number of ST samples in this study, to verify the robustness of the OptCAT model under varying sample intervals, a sensitivity analysis was performed within the 2009–2020 training period using rolling window cross-validation and expanding window validation:Rolling window cross-validation: The window length was set to 5 years with a step size of 1 year. The initial training set covered 2009–2013, with 2014 as the validation set. The window was then rolled forward iteratively: training on 2010–2014 and validating on 2015, and so on, until training on 2015–2019 and validating on 2020. In total, 7 pairs of training-validation results were obtained.Expanding window validation: The initial training set spanned 2009–2015, with 2016 as the validation set. The training set was then gradually expanded to 2009–2016, 2009–2017, 2009–2018, and 2009–2019, with corresponding validation on 2017–2020, respectively. In total, 4 training-validation splits were generated. Rolling window cross-validation: The window length was set to 5 years with a step size of 1 year. The initial training set covered 2009–2013, with 2014 as the validation set. The window was then rolled forward iteratively: training on 2010–2014 and validating on 2015, and so on, until training on 2015–2019 and validating on 2020. In total, 7 pairs of training-validation results were obtained. Rolling window cross-validation: The window length was set to 5 years with a step size of 1 year. The initial training set covered 2009–2013, with 2014 as the validation set. The window was then rolled forward iteratively: training on 2010–2014 and validating on 2015, and so on, until training on 2015–2019 and validating on 2020. In total, 7 pairs of training-validation results were obtained. Expanding window validation: The initial training set spanned 2009–2015, with 2016 as the validation set. The training set was then gradually expanded to 2009–2016, 2009–2017, 2009–2018, and 2009–2019, with corresponding validation on 2017–2020, respectively. In total, 4 training-validation splits were generated. Expanding window validation: The initial training set spanned 2009–2015, with 2016 as the validation set. The training set was then gradually expanded to 2009–2016, 2009–2017, 2009–2018, and 2009–2019, with corresponding validation on 2017–2020, respectively. In total, 4 training-validation splits were generated. In this study, MSE, RMSE, and MAE were used as evaluation metrics. The mean and standard deviation of the multiple validation results were calculated. The standard deviation was employed to measure the fluctuation of model performance: a smaller standard deviation indicates stronger robustness. Results and discussion Green credit risk measurement based on modified KMV model The mean results for DD and EDF are shown in the Table 7. The − 0.9065 for the ST group is less than the 0.6551 for the non-ST group, indicating that the ST group has a higher default risk compared to the non-ST group. Additionally, the 0.1902 for the ST group is greater than the 0.1740 for the non-ST group, further suggesting a higher default risk for the ST group compared to the non-ST group. Mann-Whitney U tests were conducted to assess the discriminatory power of DD and EDF in distinguishing between “ST” and “non-ST” based on their differences. From Table 6, it can be observed that DD is significant at the 10% level, and EDF is significant at the 5% level. This indicates that the modified KMV model possesses the capability to identify green credit risks in China’s construction industry to some extent. In this paper, the Mann-Whitney U test was adopted to verify the discriminative power of DD and EDF (calculated by the modified KMV model) between ST and non-ST firms. The results demonstrate that DD and EDF can effectively measure the credit default risk of listed construction companies. In the subsequent OptCAT model, EDF serves as the core target output for the regression task, while DD is incorporated into the input feature set as an auxiliary feature. Together with environmental indicators and financial indicators, they form a feature system for green credit risk prediction, realizing the integration of credit risk indicators and environmental risk indicators. This constitutes the core design of the green credit risk measurement and prediction framework proposed in this study. Table 7Mann-Whitney U test results of samples.DDEDFST groupNon-ST groupFull sampleST groupNon-ST groupFull sampleObs.102847949102847949Mean−0.90650.65510.48730.19020.17400.1757Standardized Test Statistic0.893*0.202**Mann-Whitney U1686091.151679292.86Asymptotic Significance (Two-Tailed Test)0.07170.0398 Table 7Mann-Whitney U test results of samples.DDEDFST groupNon-ST groupFull sampleST groupNon-ST groupFull sampleObs.102847949102847949Mean−0.90650.65510.48730.19020.17400.1757Standardized Test Statistic0.893*0.202**Mann-Whitney U1686091.151679292.86Asymptotic Significance (Two-Tailed Test)0.07170.0398 Mann-Whitney U test results of samples. Analysis of OptCAT model prediction performance The core task of the proposed OptCAT model in this study is the continuous regression prediction of the EDF value. Therefore, three classic evaluation metrics for regression tasks: MSE, RMSE, and MAE, are adopted to evaluate model performance. Smaller values of these metrics indicate higher regression prediction accuracy of the model. Meanwhile, to verify the practical economic significance of the predicted EDF values, we further correlate the predicted EDF values with the ST status of firms in period t. We analyze the differences in the mean EDF values among firms with different actual risk statuses, so as to validate whether the model predictions can effectively distinguish the actual green credit risk of construction enterprises. Comparing performance of Optuna with other optimization algorithms Five algorithms, namely DP, RS, GS, BO, Hyperopt, were selected for comparison with Optuna. The algorithm performance is shown in Table 8, with the highlighted values representing the optimal ones. Compared to the other five optimization algorithms, Optuna exhibits the best performance and higher efficiency. Table 8Comparison the performance of different optimization algorithms.AlgorithmsDPRSGSBOHyperoptOptunaMSE0.02890.02300.02280.02100.0217 0.0199 RMSE0.16990.15170.15110.14490.1472 0.1412 MAE0.14190.13010.13010.12990.1286 0.1275 Table 8Comparison the performance of different optimization algorithms.AlgorithmsDPRSGSBOHyperoptOptunaMSE0.02890.02300.02280.02100.0217 0.0199 RMSE0.16990.15170.15110.14490.1472 0.1412 MAE0.14190.13010.13010.12990.1286 0.1275 Comparison the performance of different optimization algorithms. To verify the statistical significance of performance differences between Optuna and other optimization algorithms, this study conducted 20 independent repeated experiments for each algorithm. Using Optuna as the control group, paired t-tests were performed with the other five algorithms at a significance level of α = 0.05. As shown in Table 9, the differences between Optuna and DP, RS, GS, BO, and Hyperopt were statistically significant for all three metrics (MSE, RMSE, and MAE), with all P < 0.01. This indicates that the superior optimization performance of Optuna over other algorithms is not caused by specific parameter settings, but reflects substantial statistical robustness. Meanwhile, Optuna achieved the smallest standard deviation across all metrics, demonstrating more stable optimization performance across multiple repeated experiments. Table 9Evaluation indicators for 20 repeated experiments of various hyperparameter optimization algorithms.AlgorithmsMSE (Mean ± Std.)RMSE (Mean ± Std.)MAE (Mean ± Std.)DP0.0289 ± 0.00320.1699 ± 0.00950.1419 ± 0.0082RS0.0230 ± 0.00250.1517 ± 0.00780.1301 ± 0.0069GS0.0228 ± 0.00230.1511 ± 0.00750.1301 ± 0.0068BO0.0210 ± 0.00180.1449 ± 0.00620.1299 ± 0.0061Hyperopt0.0217 ± 0.00200.1472 ± 0.00650.1286 ± 0.0059Optuna0.0199 ± 0.00120.1412 ± 0.00450.1275 ± 0.0042 Table 9Evaluation indicators for 20 repeated experiments of various hyperparameter optimization algorithms.AlgorithmsMSE (Mean ± Std.)RMSE (Mean ± Std.)MAE (Mean ± Std.)DP0.0289 ± 0.00320.1699 ± 0.00950.1419 ± 0.0082RS0.0230 ± 0.00250.1517 ± 0.00780.1301 ± 0.0069GS0.0228 ± 0.00230.1511 ± 0.00750.1301 ± 0.0068BO0.0210 ± 0.00180.1449 ± 0.00620.1299 ± 0.0061Hyperopt0.0217 ± 0.00200.1472 ± 0.00650.1286 ± 0.0059Optuna0.0199 ± 0.00120.1412 ± 0.00450.1275 ± 0.0042 Evaluation indicators for 20 repeated experiments of various hyperparameter optimization algorithms. Comparing the OptCAT with the other models Twenty experiments were conducted using single-year data and the entire sample data to verify whether OptCAT is the most suitable model for predicting green credit risk in China’s construction industry. The results are shown in Table 10. The highlighted values indicate the optimal model performance, while the numbers in parentheses represent the ranking values of evaluation metrics, with lower values indicating better predictive performance. The last row represents the average ranking values of each model across all experiments. From the results in Table 8, it can be observed that OptCAT outperforms other models. Table 10Comparison the performance of different models in 2009–2020.YearAlgorithmsSVMRFXGBoostLightGBMOptCAT2009MSE0.0226(5)0.0178(4)0.0165(3)0.0142(2) 0.0131(1) RMSE0.1505(5)0.1333(4)0.1286(3)0.1190(2) 0.1143(1) MAE0.1318(5)0.1106(4)0.1165(3)0.1024(2) 0.0905(1) 2010MSE0.0224(5)0.0213(4)0.0171(3)0.0123(2) 0.0115(1) RMSE0.1497(5)0.1459(4)0.1308(3)0.1108(2) 0.1072(1) MAE0.1409(5)0.1220(4)0.1041(3) 0.0905(1) 0.0905(1) 2011MSE0.0223(4)0.0295(5)0.0218(3)0.0182(2) 0.0163(1) RMSE0.1493(4)0.1719(5)0.1476(3)0.1351(2) 0.1277(1) MAE0.1252(3)0.1348(5)0.1322(4)0.1076(2) 0.1058(1) 2012MSE0.0204(5)0.0199(4)0.0197(3)0.0196(2) 0.0193(1) RMSE0.1430(5)0.1410(4)0.1403(3)0.1401(2) 0.1388(1) MAE0.1237(4)0.1277(5)0.1232(3)0.1227(2) 0.1160(1) 2013MSE0.0242(5)0.0233(4)0.0208(3)0.0187(2) 0.0185(1) RMSE0.1554(5)0.1527(4)0.1443(3)0.1368(2) 0.1359(1) MAE0.1398(4)0.1439(5)0.1338(3)0.1296(2) 0.1140(1) 2014MSE0.0237(4)0.0251(5)0.0227(3)0.0223(2) 0.0199(1) RMSE0.1538(4)0.1584(5)0.1506(3)0.1494(2) 0.1410(1) MAE0.1392(4)0.1441(5)0.1348(3)0.1293(2) 0.1267(1) 2015MSE0.0199(5)0.0159(4)0.0129(3)0.0124(2) 0.0102(1) RMSE0.1409(5)0.1261(4)0.1135(3)0.1114(2) 0.1010(1) MAE0.1247(5)0.1016(4)0.0997(3)0.0854(2) 0.0851(1) 2016MSE0.0259(3)0.0287(5)0.0268(4)0.0209(2) 0.0180(1) RMSE0.1611(3)0.1695(5)0.1636(4)0.1445(2) 0.1340(1) MAE0.1393(4)0.1388(3)0.1450(5)0.1286(2) 0.1166(1) 2017MSE0.0295(5)0.0276(4)0.0234(3)0.0151(2) 0.0135(1) RMSE0.1718(5)0.1660(4)0.1531(3)0.1231(2) 0.1163(1) MAE0.1590(5)0.1466(4)0.1384(3)0.1059(2) 0.0905(1) 2018MSE0.0249(5)0.0220(3)0.0225(4)0.0215(2) 0.0142(1) RMSE0.1578(5)0.1482(3)0.1499(4)0.1466(2) 0.1193(1) MAE0.1432(5)0.1361(3)0.1361(3)0.1297(2) 0.1033(1) 2019MSE0.0280(5)0.0264(4)0.0180(3)0.0178(2) 0.0176(1) RMSE0.1672(5)0.1626(4)0.1343(3)0.1334(2) 0.1328(1) MAE0.1200(5)0.1179(4)0.1178(3)0.1100(2) 0.1075(1) 2020MSE0.0239(5)0.0238(4)0.0217(3)0.0187(2) 0.0152(1) RMSE0.1545(5)0.1544(4)0.1472(3)0.1369(2) 0.1233(1) MAE0.1159(5)0.1117(3)0.1129(4)0.1112(2) 0.1072(1) Full sampleMSE0.0263(5)0.0255(4)0.0245(3)0.0215(2) 0.0199(1) RMSE0.1621(5)0.1598(4)0.1566(3)0.1467(2) 0.1412(1) MAE0.1412(5)0.1379(4)0.1323(3)0.1299(2) 0.1275(1) Average rank4.644.133.211.97 1.00 Table 10Comparison the performance of different models in 2009–2020.YearAlgorithmsSVMRFXGBoostLightGBMOptCAT2009MSE0.0226(5)0.0178(4)0.0165(3)0.0142(2) 0.0131(1) RMSE0.1505(5)0.1333(4)0.1286(3)0.1190(2) 0.1143(1) MAE0.1318(5)0.1106(4)0.1165(3)0.1024(2) 0.0905(1) 2010MSE0.0224(5)0.0213(4)0.0171(3)0.0123(2) 0.0115(1) RMSE0.1497(5)0.1459(4)0.1308(3)0.1108(2) 0.1072(1) MAE0.1409(5)0.1220(4)0.1041(3) 0.0905(1) 0.0905(1) 2011MSE0.0223(4)0.0295(5)0.0218(3)0.0182(2) 0.0163(1) RMSE0.1493(4)0.1719(5)0.1476(3)0.1351(2) 0.1277(1) MAE0.1252(3)0.1348(5)0.1322(4)0.1076(2) 0.1058(1) 2012MSE0.0204(5)0.0199(4)0.0197(3)0.0196(2) 0.0193(1) RMSE0.1430(5)0.1410(4)0.1403(3)0.1401(2) 0.1388(1) MAE0.1237(4)0.1277(5)0.1232(3)0.1227(2) 0.1160(1) 2013MSE0.0242(5)0.0233(4)0.0208(3)0.0187(2) 0.0185(1) RMSE0.1554(5)0.1527(4)0.1443(3)0.1368(2) 0.1359(1) MAE0.1398(4)0.1439(5)0.1338(3)0.1296(2) 0.1140(1) 2014MSE0.0237(4)0.0251(5)0.0227(3)0.0223(2) 0.0199(1) RMSE0.1538(4)0.1584(5)0.1506(3)0.1494(2) 0.1410(1) MAE0.1392(4)0.1441(5)0.1348(3)0.1293(2) 0.1267(1) 2015MSE0.0199(5)0.0159(4)0.0129(3)0.0124(2) 0.0102(1) RMSE0.1409(5)0.1261(4)0.1135(3)0.1114(2) 0.1010(1) MAE0.1247(5)0.1016(4)0.0997(3)0.0854(2) 0.0851(1) 2016MSE0.0259(3)0.0287(5)0.0268(4)0.0209(2) 0.0180(1) RMSE0.1611(3)0.1695(5)0.1636(4)0.1445(2) 0.1340(1) MAE0.1393(4)0.1388(3)0.1450(5)0.1286(2) 0.1166(1) 2017MSE0.0295(5)0.0276(4)0.0234(3)0.0151(2) 0.0135(1) RMSE0.1718(5)0.1660(4)0.1531(3)0.1231(2) 0.1163(1) MAE0.1590(5)0.1466(4)0.1384(3)0.1059(2) 0.0905(1) 2018MSE0.0249(5)0.0220(3)0.0225(4)0.0215(2) 0.0142(1) RMSE0.1578(5)0.1482(3)0.1499(4)0.1466(2) 0.1193(1) MAE0.1432(5)0.1361(3)0.1361(3)0.1297(2) 0.1033(1) 2019MSE0.0280(5)0.0264(4)0.0180(3)0.0178(2) 0.0176(1) RMSE0.1672(5)0.1626(4)0.1343(3)0.1334(2) 0.1328(1) MAE0.1200(5)0.1179(4)0.1178(3)0.1100(2) 0.1075(1) 2020MSE0.0239(5)0.0238(4)0.0217(3)0.0187(2) 0.0152(1) RMSE0.1545(5)0.1544(4)0.1472(3)0.1369(2) 0.1233(1) MAE0.1159(5)0.1117(3)0.1129(4)0.1112(2) 0.1072(1) Full sampleMSE0.0263(5)0.0255(4)0.0245(3)0.0215(2) 0.0199(1) RMSE0.1621(5)0.1598(4)0.1566(3)0.1467(2) 0.1412(1) MAE0.1412(5)0.1379(4)0.1323(3)0.1299(2) 0.1275(1) Average rank4.644.133.211.97 1.00 Comparison the performance of different models in 2009–2020. To quantify the incremental contribution of environmental indicators to green credit risk prediction, we designed an ablation study to compare the prediction performance differences between the complete model with environmental indicators (All features) and the benchmark model without environmental indicators (No ESG features). In addition, we also tested the performance of the model using only environmental indicators (Only ESG features) to evaluate the independent predictive ability of environmental factors. The experimental results are shown in Table 11. Table 11Ablation study on the contribution of environmental indicators to prediction performance (Full sample).Model configurationMSERMSEMAERankAll features (with ESG)0.01990.14120.12751No ESG features0.02350.15330.13682Only ESG features0.03870.19670.16233 Table 11Ablation study on the contribution of environmental indicators to prediction performance (Full sample).Model configurationMSERMSEMAERankAll features (with ESG)0.01990.14120.12751No ESG features0.02350.15330.13682Only ESG features0.03870.19670.16233 Ablation study on the contribution of environmental indicators to prediction performance (Full sample). The results of the ablation study indicate that the complete model with environmental indicators outperforms the benchmark model without environmental indicators in all evaluation metrics. Specifically, the introduction of environmental indicators reduces the MSE by 15.3% (from 0.0235 to 0.0199), the RMSE by 7.9% (from 0.1533 to 0.1412), and the MAE by 6.8% (from 0.1368 to 0.1275). This result confirms that environmental factors have significant incremental predictive value for green credit risk prediction of construction enterprises. Meanwhile, the model using only ESG indicators performs the worst, indicating that environmental factors need to be combined with financial indicators to achieve the optimal prediction effect. Analysis of model prediction performance To validate the practicality of the model, we utilized the latest data from 119 companies for prediction. We used the OptCAT model to predict the EDF for these companies in 2021 and 2022. Subsequently, we compared the results with the actual outcomes to verify the accuracy and effectiveness of the model. We examined the predicted results for 2021 and 2022, evaluating the predictive ability of the OptCAT model. We also compared the predicted results with the ST status of the companies’ stocks in 2021 and 2022, assessing the consistency of our model with the real-world status of the enterprises. By employing these two evaluation methods, we thoroughly assessed the predictive performance of the model in addressing real-world challenges. Table 12 presents the performance evaluation of the OptCAT model on the datasets for 2021 and 2022, showing comparable results with other datasets across three error-based metrics. Additionally, the model demonstrates good efficiency. These findings provide strong evidence for the accuracy of the OptCAT model in predicting green credit risk in the construction industry. As shown in Table 13, companies recently marked as ST had the highest mean EDF in 2021 and 2022, followed by companies whose stocks were labeled as ST in previous years, while the smallest EDF values were observed for normal samples in 2021 and 2022. These results indicate that the predicted outcomes of the model align with the actual situations, highlighting the model’s high accuracy and its effectiveness in addressing the challenges of predicting green credit risk in the real-world context of China’s construction industry. As shown in Table 13, the mean EDF value of newly identified ST enterprises during 2021–2022 is significantly higher than those of historical ST enterprises and normal enterprises. All analyses in this section are based on the original imbalanced distribution of the test set covering 2021–2022, without any resampling preprocessing. The model can still effectively distinguish enterprises with different risk statuses, further demonstrating the applicability of the OptCAT model in real-world imbalanced data scenarios. Table 12Analysis of the model prediction effects in 2021 and 2022.MSE202120220.02030.0210RMSE0.14260.1450MAE0.12920.1254 Table 12Analysis of the model prediction effects in 2021 and 2022.MSE202120220.02030.0210RMSE0.14260.1450MAE0.12920.1254 Analysis of the model prediction effects in 2021 and 2022. Table 13The value of EDF in different sample enterprises.Sample enterprisesEDFSample enterprises of ST in previous years0.2017Sample enterprises of ST in 20210.2412Sample enterprises of ST in 20220.2382Normal sample enterprises in 20210.1579Normal sample enterprises in 20220.1660 Table 13The value of EDF in different sample enterprises.Sample enterprisesEDFSample enterprises of ST in previous years0.2017Sample enterprises of ST in 20210.2412Sample enterprises of ST in 20220.2382Normal sample enterprises in 20210.1579Normal sample enterprises in 20220.1660 The value of EDF in different sample enterprises. Classification Metric Evaluation Based on EDF To further verify the practical identification ability of EDF values for green credit risk in the construction industry, this study converts continuous EDF values into classification prediction results. ST enterprises are defined as the positive class (default), and non-ST enterprises as the negative class (non-default). The optimal EDF classification threshold is determined using the Youden index method, yielding an optimal threshold of 0.185 (based on the 2009–2020 sample period), which is uniformly applied to the 2021–2022 out-of-sample period. In this section, classification metrics including AUC, accuracy, precision, recall, and F1-score are calculated for both the in-sample and out-of-sample periods. Confusion matrices under the optimal threshold are plotted to compare the classification performance of the OptCAT model with SVM, RF, XGBoost, and LightGBM. Table 14Comparison of EDF classification metrics among different models.In-sample periodModelAUCAccuracyPrecisionRecallF1-scoreSVM0.72510.70110.68940.62590.6700RF0.78300.75670.74240.70500.7234XGBoost0.82620.79850.78510.75330.7698LightGBM0.85110.82220.81010.78690.7987 OptCAT 0.9020 0.8761 0.8655 0.8430 0.8544 Out-of-sample periodModelAUCAccuracyPrecisionRecallF1-scoreSVM0.70220.68550.67120.62390.6464RF0.75810.73160.71560.67840.6967XGBoost0.80540.77660.76230.72530.7431LightGBM0.83330.80100.78980.75620.7727 OptCAT 0.8869 0.8522 0.8416 0.8187 0.8290 Table 14Comparison of EDF classification metrics among different models.In-sample periodModelAUCAccuracyPrecisionRecallF1-scoreSVM0.72510.70110.68940.62590.6700RF0.78300.75670.74240.70500.7234XGBoost0.82620.79850.78510.75330.7698LightGBM0.85110.82220.81010.78690.7987 OptCAT 0.9020 0.8761 0.8655 0.8430 0.8544 Out-of-sample periodModelAUCAccuracyPrecisionRecallF1-scoreSVM0.70220.68550.67120.62390.6464RF0.75810.73160.71560.67840.6967XGBoost0.80540.77660.76230.72530.7431LightGBM0.83330.80100.78980.75620.7727 OptCAT 0.8869 0.8522 0.8416 0.8187 0.8290 Comparison of EDF classification metrics among different models. Table 15Confusion matrix of OptCAT model at optimal EDF threshold.In-sample periodOut-of-sample periodPredicting STPredicting non STPredicting STPredicting non STActual ST76(TP)4(FN)20(TP)2(FN)Actual non ST89(FP)641(TN)29(FP)168(TN) Table 15Confusion matrix of OptCAT model at optimal EDF threshold.In-sample periodOut-of-sample periodPredicting STPredicting non STPredicting STPredicting non STActual ST76(TP)4(FN)20(TP)2(FN)Actual non ST89(FP)641(TN)29(FP)168(TN) Confusion matrix of OptCAT model at optimal EDF threshold. According to Table 14, the OptCAT model achieves significantly superior classification metrics in the in-sample period compared with other benchmark models. The AUC value reaches 0.9020, indicating its optimal discriminative ability between ST and non‑ST enterprises. The F1‑score is 0.8544, reflecting a favorable balance between precision and recall. The out‑of‑sample classification results are consistent with those in the in‑sample period. The OptCAT model still maintains the best classification performance, with an AUC of 0.8869 and an F1‑score of 0.8290, demonstrating excellent generalization ability on unseen new data. As observed from the confusion matrix in Table 15, the model achieves a 95% recognition rate for actual ST enterprises in the in‑sample period, with a misdetection rate of only 5%, which can effectively avoid credit losses for banks when lending to high‑risk enterprises.For the out‑of‑sample period, the misdetection rate for ST enterprises is only 9.1% and the false positive rate is 14.8%. While effectively identifying high‑risk enterprises, the model keeps the misclassification rate for normal enterprises within a reasonable range, which meets the practical requirements of green credit decision‑making in commercial banks. Robustness test To verify the robustness of the OptCAT model when the sample interval changes, this study conducted rolling window cross-validation and expanding window validation within the 2009–2020 training period. Taking SVM, RF, XGBoost, and LightGBM as benchmark models, the mean value (reflecting overall performance) and standard deviation (reflecting robustness) of the evaluation metrics for each model among multiple sets of validation results were calculated. The results are shown in Table 16. Table 16Robustness test results of each model during the training period.ModelRolling windowExpanding windowMSERMSEMAEMSERMSEMAEMeanStd.MeanStd.MeanStd.MeanStd.MeanStd.MeanStd.SVM0.02580.00210.16060.00650.11240.00430.02650.00180.16280.00570.11450.0039RF0.02490.00190.15780.00610.10870.00380.02560.00170.16000.00530.11020.0035XGBoost0.02310.00150.15200.00490.10360.00310.02380.00140.15430.00450.10510.0028LightGBM0.02080.00120.14420.00390.09740.00250.02150.00110.14660.00360.09890.0022 OptCAT 0.0192 0.0008 0.1386 0.0028 0.0921 0.0019 0.0199 0.0007 0.1412 0.0025 0.0937 0.0016 Table 16Robustness test results of each model during the training period.ModelRolling windowExpanding windowMSERMSEMAEMSERMSEMAEMeanStd.MeanStd.MeanStd.MeanStd.MeanStd.MeanStd.SVM0.02580.00210.16060.00650.11240.00430.02650.00180.16280.00570.11450.0039RF0.02490.00190.15780.00610.10870.00380.02560.00170.16000.00530.11020.0035XGBoost0.02310.00150.15200.00490.10360.00310.02380.00140.15430.00450.10510.0028LightGBM0.02080.00120.14420.00390.09740.00250.02150.00110.14660.00360.09890.0022 OptCAT 0.0192 0.0008 0.1386 0.0028 0.0921 0.0019 0.0199 0.0007 0.1412 0.0025 0.0937 0.0016 Robustness test results of each model during the training period. As can be seen from the results in Table 16, in both the rolling window and expanding window validation, the OptCAT model not only achieves the lowest mean values of MSE, RMSE, and MAE among all models, maintaining the optimal prediction performance, but also has the smallest standard deviation for all metrics. This indicates that when the sample interval changes continuously, the OptCAT model has the smallest fluctuation in prediction performance, and its robustness is significantly superior to other benchmark models. This result further verifies the reliability of the OptCAT model in green credit risk prediction of the construction industry, even when the number of ST samples is small, the model can still maintain a stable prediction effect. Conclusions The modified KMV model can accurately measure the credit dimension of green credit risk in China’s construction industry. The EDF output by this model serves as the regression target of the OptCAT model, and when combined with environmental indicators and financial indicators, it realizes a two-stage prediction of green credit risk: quantitative regression and status verification. Firstly, this paper utilizes the modified KMV model to assess green credit risk in the construction industry. Secondly, the Mann-Whitney U test and a feature ranking method based on machine learning are employed for indicator selection, making the final set of indicators more objective and scientific. Subsequently, a green credit risk assessment system suitable for China’s construction industry is proposed. Finally, the OptCAT algorithm is used to predict green credit risk in the construction industry. The following conclusions are drawn: the modified KMV model provides an accurate measurement of green credit risk in China’s construction industry. Compared with DP, RS, GS, BO, and Hyperopt, Optuna demonstrates superior optimization performance. When compared with SVM, RF, XGBoost, and LightGBM models, OptCAT exhibits optimal predictive performance. In this study, by setting a unified parameter search range and differentiated reasonable search budgets, and conducting 20 repeated experiments and paired t-tests for each algorithm, it is confirmed that the performance differences between Optuna and other optimization algorithms are statistically significant. Moreover, Optuna achieves the smallest standard deviation among all metrics, indicating that its optimization advantage is not accidental but possesses robustness and substantiality. In addition, further verification based on EDF-based classification metrics shows that the OptCAT model achieves significantly higher AUC, accuracy, F1-score and other metrics than the benchmark models in both the in-sample and out-of-sample periods. Under the optimal EDF threshold, the misdetection rates are only 5% and 9.1%, respectively, demonstrating stronger ability to distinguish the actual risks between ST and non-ST enterprises in the construction industry. The results of the confusion matrix indicate that the OptCAT model can effectively balance the dual needs of avoiding misdetection of high-risk enterprises and misclassification of normal enterprises, which is more in line with the actual scenarios of green credit decision-making in commercial banks. Through the empirical research conducted in this study, the following managerial insights can be derived: (1) Commercial banks can accurately assess green credit risk in the construction industry to make informed lending decisions. (2) Investors can gain an understanding of green credit risk in the construction industry, enabling them to prudently mitigate risks and maximize returns. (3) The OptCAT model can accurately predict green credit risk in the construction industry, providing decision support for risk managers. (4) Business decision-makers can comprehend the situation of green credit risk in the construction industry, making appropriate adjustments to achieve the organic integration of environmental practices, social responsibility, and corporate performance. In conclusion, this study has limitations, and there are areas that warrant improvement: (1) As the focus of this research is on establishing the OptCAT model for predicting green credit risk in the construction industry, the direct choice of using Borderline-SMOTE oversampling to balance the dataset may lead to issues such as fuzzy boundaries between majority and minority class samples. Future considerations could involve using undersampling, ADASYN, and cluster-based oversampling methods to balance the dataset. (2) Over time, as the number of default events of listed construction companies in China increases, incorporating more default samples into the model for training will further enhance the accuracy of model predictions.