非均衡数据下企业碳减排信用风险预警模型的构建与检验

Credit risk of enterprise carbon emission reduction based on entropy weight TOPSIS-KNN-BPNN model under unbalanced data

碳中和作为应对气候变化的关键策略,对利益相关者和国家可持续发展具有重要影响.鉴于此,为提高企业碳减排信用风险的预测准确性,本文以2003—2020年2939家上市企业为研究对象,并构建了一种融合熵权TOPSIS-Kmeans-BPNN的新型企业碳减排信用风险预警模型.本文首先运用熵权TOPSIS(technique for order preference by similarity to ideal solution)对企业碳减排信用风险进行综合评分;然后对评估结果进行聚类,获得5种企业信用风险的等级,帮助BPNN(back propagation neural network)更好地进行监督学习;再是引入SMOTE算法(synthetic minority over-sampling technique),在少数等级企业样本中进行插值并生成新样本,以解决各等级企业样本不均衡问题;最后通过消融和多模型对比实验,验证本文所建模型的预测性能.结果表明:第一,各项碳减排指标对各信用风险等级企业的影响程度存在明显差异.其中,影响程度最高的是煤炭碳排放量指标,影响程度最低的是企业碳排放量指标;第二,利用XGBoost(extreme gradient boosting)算法筛选指标有效提高了模型的预测性能,平均提升了3.55%;第三,与其它模型相比,本文模型的预测准确率达99.05%,平均提升了17.38%,表明该模型是可行的,可为金融机构进行信用评级提供技术支撑.

Carbon neutralization,as a key strategy to cope with climate change,has an important impact on stakeholders and national sustainable development.In view of this,in order to improve the prediction accuracy of corporate carbon emission reduction credit risk,this paper takes 2,939 listed enterprises from 2003 to 2020 as the research subjects and proposes a corporate carbon emission reduction credit risk early-warning model integrated with entropy weight TOPSIS-KNN-BPNN.The paper applies firstly entropy weight TOPSIS to give a comprehensive score in regards of the credit risk of enterprise carbon emission reduction;then conducts a treatment of Kmeans clustering on the scoring results to effectively obtain the grade interval of credit risk,thus to lay a foundation for the supervised learning of BP neural network;furthermore SMOTE algorithm is introduced to value-interpolating operation for those enterprises with numerically-small samples to generate some new samples,so as to solve the problem of imbalance of samples of various grades of enterprises;finally,through the ablation and multi-model comparative experiments,the predictive performance of the model constructed in this paper is verified.The results show that:first,there is a significant difference in terms of the degree of influence of various carbon emission reduction indicators on enterprises with various credit risk levels,of which the highest degree of influence is found to be the coal carbon emission index,and the lowest degree of influence is from the enterprise carbon emission index;second,the use of the XGBoost algorithm to effectively screen the indexes improves the predicting performance of the model,with an average improvement of 3.55%;third,compared with other models,the predicting accuracy of the model reaches 99.05%,with an average improvement of 17.38%,indicating that the model is feasible and capable of providing technical support for financial institutions performing credit rating.