Credit is the cornerstone of modern economic market, with enterprises being the main participants. It is through fair exchange based on the credit foundation that all participants achieve their goals in market activities. Therefore, among various forms of credit, enterprise credit is the core of the economic bond (Lessmann, Baesens, Seow, & Thomas, 2015). However, along with the substantially increasing number of new industries, new formats, and new technologies, the issue of insufficient regulatory power has become more and more prominent (Zhou, Tam, & Fujita, 2016). Meanwhile, the regulators have also been put more demanding requirement of being “Accurate, Efficient and Reliable” in routine supervision. Based on this, regulatory authorities propose to utilize historical credit information, big data monitoring information, and other associated information to classify enterprises into multiple credit risk classes and take this as the rating result to implement differentiated supervision on enterprises, i.e., reducing the frequency of spot-checks on enterprises with better credit class and increase that on those with inferior credit class for the improvement the efficiency of supervision.
As for the management demands of regulatory authorities, regulators need to implement differentiated regulation oriented by objective and accurate results, so as to suit the remedy to the case for each enterprise. Therefore, the results of enterprise credit risk rating ought to meet the demands of regulators for classification accuracy (Lessmann et al., 2015). Also, regulators need to control the size of reference features by selecting some key risk features from the enterprise information database, because each risk feature has a different impact on the enterprise credit risk, and regulators are limited in capacity to focus on hundreds of risk features during routine supervision. Moreover, modern economic market is subject to many uncertainties, and the impact of certain factors on the result may be temporary or negligible (Ramírez et al., 2018). Given the huge cost of shifting regulatory focus and reformulating regulatory schemes, regulators ought to avoid significant disruptions in rating results due to minor deviations in some risk features (Liu, Bi, & Fan, 2017). The stability of results should also be taken into account in enterprise credit risk rating. Therefore, how to classify enterprises into multiple credit risk classes and make sure the rating results can meet the multiple management demands of regulators is the key to enterprise credit risk rating (Liu, Diao, Cao, & Zhang, 2017).
A lot of research has been conducted on enterprise credit risk rating, but limitations still exist. First, most studies construct a binary classification model from the perspective of financial approval instead of regulation (Kou et al., 2021). To be specific, based on the possibility that the borrower can repay the loan principal and interest on time or not, each borrower can be classified into Special Treatment (ST) object or non-ST one to approving its loan application or not (Kundu & Mallipeddi, 2022). This means such models cannot meet the basic demand of regulatory authorities for enterprise credit risk rating that classify enterprises into multiple classes. Second, there are many indicators such as classification accuracy and the size of reference features to evaluate when credit risk rating model is performed (Al-Tashi, Abdulkadir, Rais, Mirjalili, & Alhussian, 2020); however, most studies only evaluate the rating results based on a single indicator (Zhang et al., 2021). Such blind pursues on the performance based on a single indicator will lead to the biased interpretation of results. Third, many studies failed to take into account the stability of results in the field of enterprise credit risk rating; therefore, the results can hardly meet the multiple management demands, resulting in the decline of authority in regulators (Dhiman et al., 2021). In summary, current research cannot satisfy the actual demands of regulators for enterprise credit risk rating in terms of modeling objective design, risk feature selection, and evaluation indicators setting. Herein, constructing an enterprise credit risk rating model with the perspective of regulatory by classifying enterprises into multiple credit risk classes and combining the various demands of regulators, is in urgent need for its great practical significance and research value.
Based on that, this paper takes the various demands of regulators as evaluation indicators, proposes a novel multi-classification ensemble model with hybrid decomposition strategy and two-stage multi-objective feature selection for the enterprise credit risk rating. First, a hybrid decomposition strategy is developed to divide appropriate training subsets for subsequent feature selection and multi-classification operation, so that the purpose of multi-classification in differentiated supervision will be realized. Then, feature selection is treated as a two-stage multi-objective optimization problem, with classification accuracy, feature subset size, and stability of feature selection being evaluation indicators. In this case, the multiple demands of regulatory authorities will be met.
The main work and innovations of this paper are as follows:
This paper ensures the multiple credit risk classes of enterprises by integrating advantages of different decomposition strategies, such as the resampling technique as an adaptive operation and an O&R Hybrid Decomposition Strategy with Adaptive Resampling (O&RHDS-AR).
This paper meets the multiple demands of regulatory authorities by developing a two-stage multi-objective feature selection method that utilizes the output of Filter Integrated Feature Selection (FIFS) as priori information to the multi-objective Wrapper feature selection method, based on the Improved Ant Colony Optimization (MOFS-IACO).
This paper verifies the performance of the model in module design and the application in practical by analyzing the empirical results from two perspectives: the superiority of each module and the validity of the entire model.
The remainder of this paper is organized as follows: Section 2 reviews related work on multi-classification and feature selection methods; Section 3 elaborates the proposed model; Section 4 introduces the empirical datasets, evaluation indicators, and parameter settings; Section 5 analyzes the empirical results; Section 6 presents conclusions and discusses the future research direction.
Based on the purpose of enterprise credit risk rating, it can be found that multi-classification and feature selection are the main research field for the construct of enterprise credit risk rating model.
In this section, based on the multiple demands of regulators, a multi-classification ensemble model with hybrid decomposition strategy and two-stage feature selection for the enterprise credit risk rating is proposed. The ensemble model consists of three modules: O&RHDS-AR, FIFS, and MOFS-IACO, and the framework of the model is depicted in Fig. 1. The details of the three modules are elaborated in the following subsections.
In this section, data required for the empirical, the evaluation indicators, and parameter settings of the algorithm will be introduced in detail.
This section focuses on the analysis of empirical results from two perspectives: module superiority and model validity. In the section on module superiority analysis, the performance of each component algorithm in models is tested. It should be noted that the Variable-controllingapproach is used to compare the effects of the component algorithm on the model results. i.e., the parameter settings are the same in the compared models except for the difference in the test module. In the section on
Conclusions and future works
Enterprise credit risk rating has become an urgent and systematic decision problem in the field of market supervision. Based on the perspective of regulatory, this paper combines the purposes of enterprise credit risk rating and the multiple demands of regulatory authorities, proposes a multi-classification ensemble model with hybrid decomposition strategy and two-stage multi-objective feature selection. On the basis of the empirical comparison and analysis, this paper draws the following
Cao et al., 2013, Liu et al., 2017.
CRediT authorship contribution statement
Xiao Pei: Conceptualization, Methodology, Data curation, Formal analysis, Resources, Writing – original draft, Writing – review & editing. Hua Li: Conceptualization, Resources, Supervision, Funding acquisition. Aiping Wu: Conceptualization, Resources, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
This work was supported by the Technology Innovation Leading Program of Shaanxi (Program number 2022PT-49) and Shaanxi Province Innovation Capacity Support Program (Program number 2022KRW22).
© 2023 Elsevier Ltd. All rights reserved.