Elsevier

Applied Soft Computing

Volume 73, December 2018, Pages 914-920
Applied Soft Computing

Application of eXtreme gradient boosting trees in the construction of credit risk assessment models for financial institutions

https://doi.org/10.1016/j.asoc.2018.09.029Get rights and content

Highlights

  • This paper construct a credit risk assessment model for financial institutions.

  • This paper constructs a credit risk assessment model with the XGBoost method.

  • The research results can improvement of the loan business efficiency.

Abstract

The majority of the studies on credit risk assessment models for financial institutions during recent years focus on the improvement of imbalanced data or on the enhancement of classification accuracy with multistage modeling. Whilst multistage modeling and data pre-processing can boost accuracy somewhat, the heterogeneous nature of data may affects the classification accuracy of classifiers. This paper intends to use the classifier, eXtreme gradient boosting tree (XGBoost), to construct a credit risk assessment model for financial institutions. Cluster-based under-sampling is deployed to process imbalanced data. Finally, the area under the receiver operative curve and the accuracy of classifications are the assessment indicators, in the comparison with other frequently used single-stage classifiers such as logistic regression, self-organizing algorithms and support vector machine. The results indicate that the XGBoost classifier used by this paper achieve better results than the other three and can serve as a superior tool for the development of credit risk models for financial institutions.

Introduction

The larger the loan book, the greater the risks of non-performing loans. To stabilize financial markets and control risks, credit risk evaluation has been on the top of the agenda for financial institutions. The Basel Committee on Banking Supervision in 2004 released the New Basel Capital Accord (Basel II), allowing banks to create their own internal credit ratings to assess credit risks and enhance credit sensitivity.

The studies on credit scoring models have attracted significant attention. For example, Baesens et al. [1] compare the frequently used credit scoring methods, such as logistic regression, discriminant analysis, K-nearest neighbor (KNN), neural networks, decision trees and support vector machine. The study samples the credit scoring data from eight databases for analysis. The research findings indicate that neutral networks and support vector machine yield better classification results. Lee et al. [2] use hybrid models to construct risk assessment tools. The first step is to conduct a discriminant analysis on the sample and extract the significant variables as the inputs in the back-propagation network for the hybrid models. An analysis of the discriminant analysis, logistic regressions and back indicates that hybrid models produce better classification accuracy. Min and Lee [3] apply support vector machine and multiple discriminant analysis, logistic regression, back-propagation network methods to predict corporate crisis for the small-and-medium enterprises in the heavy industry of South Korea. The study posits that support vector machine combined with radial basis function yield better predictive results.

Most of the credit data used for credit scoring is imbalanced. In general, the bulk of the data are normal loans, and only a small portion of data is about defaults. A direct input of the imbalanced data may cause higher accuracy for major classifications and lower accuracy for minor classifications. In reality, the costs for defaults are often higher so financial institutions tend to emphasize the accuracy of default data [4]. In the past, single-stage classifiers produced less-than-ideal accuracy for minority classification data. This is why over recent years many scholars have proposed the development of hybrid models so as to better the accuracy of minority classifications. For instance, Huang et al. [5] employ two-stage genetic programming for the construction of credit risk assessment models. Genetic programming is used in the first stage to identify the algorithmic rules. These rules are then applied to separate the data into two types, easy to discriminate and difficult to discriminate. In the second stage, the “difficult-to-discriminate” data is processed with generic programming in order to identify a group of discriminant function to determine on defaults or not. The results indicate that two-stage models have greater accuracy than one-stage ones. Huang et al. [6] combine support vector machine and genetic programming in the construction of a two-stage credit risk assessment model. The study finds that two-stage models are not much different from genetic algorithms and back-propagation networks in terms of accuracy. However, the selection of variables can lead to better accuracy with fewer variables. Lin [7] uses logistic regression along with neural networks to build a two-stage credit risk assessment model. The cut-off point is adjusted so that the accuracy rate for defaults and the accuracy rate for non-defaults are equal. The purpose is to avoid overly low accuracy for defaults. Chuang and Lin [8] integrate case-based reasoning and neutral networks for the construction of a two-stage credit risk assessment model. In the first stage, the neural network approach is deployed to separate the data into two groups, good credit ratings and poor credit ratings. During the second stage, the good credit ratings subsequently determined to be poor credit ratings are re-distributed with case-based reasoning. The results suggest that two-stage models report better accuracy than single-stage ones.

Although previous studies indicated that two-stage models could delivery more accurate classification results than one-stage ones. However, they are far more complicated than one-stage model in terms of modeling building as well as when applying the model to practice. This paper first applied eXtreme gradient boosting tree (XGBoost) method proposed recently by Chen and Guestrin [9], a method has drawn attention in some world-wide major big-data competitions like Kaggle and DataCastle due to its speed and accuracy. Instead of using complex method for build models for financial institutions to apply in practice, we do think a powerful yet effective solution is a promising improvement in this area. Aiming to improve the predictive power of credit risk scoring, this method is based on decision trees, supported with gradient boosting (an improvement of the traditional gradient boosting trees). This paper is arranged as follows. Section 2 is a literature review on boosting and gradient boosting, eXtreme Gradient Boosting Trees and Receiver Operative Curve. Section 3 introduces the research methodology, research flows and structure. Section examines the empirical findings. Section 5 presents the conclusions and suggestions.

Section snippets

Boosting and gradient boosting

Breiman [10] proposes the concept of bagging, i.e. random sampling to train classifiers. These classifiers are then assembled to synthesize into a higher-accuracy one. Freund and Schapire [11] come up with another boosting technique. Different from bagging in terms of sampling, boosting gives a weight to each observation and changes the weight after the training of a classifier. The weight to the wrongly classified observation is increased and the weight to the correctly classified observation

Research method

This research comprises of four steps. Step 1 is the collection and consolidation of data. Step 2 involves the selection of most frequently used financial and non-financial variables in Taiwan and overseas. Step 3 process imbalanced data with cluster-based under-sampling. Finally, XGBoost is used for model construction to calculate accuracy and relevant values. Fig. 1 is a flowchart of the research.

Step 1: Data collection and consolidation

This paper gathers the historical data of the loan books

An empirical study

To verify the feasibility of the credit risk assessment model built by this paper, the credit data from the loan book of a financial institution in Taiwan is sourced for empirical research. Finally, this paper tests the validity of the model constructed with the chosen classifier with the results produced with logistic regression, self-organizing algorithms and support vector machine.

Conclusion and suggestions

This paper constructs a credit risk assessment model with the XGBoost method and conducts an empirical test by using the loan book data of a financial institution in Taiwan. The validation results suggest that the credit risk assessment model built by this paper exhibits superior discriminating accuracy measured with AUC values (compared to the single-stage classified used frequently in the past such as logistic regression, GMDH and SVM). The research findings can serve as a reference to the

Acknowledgment

The authors would like to thank the Ministry of Science and Technology, Taiwan for supporting this research financially under Contract No. MOST 105-2410-H-009-027 and MOST 107-2410-H-145-001.

References (25)

  • ChangY.C. et al.

    Establish decision tree-based short-term default credit risk assessment models

    Commun. Stat. Theory Methods

    (2016)
  • T. Chen, C. Guestrin, Xgboost: A scalable tree boosting system, in :Proceeding of the 22Nd ACM SIGKDD International...
  • Cited by (187)

    View all citing articles on Scopus
    View full text