Financial innovation: Credit default hybrid model for SME lending
Introduction
Most firms around the world are SMEs, and they are recognized worldwide as the engine of economic growth. It has also been suggested in the literature, that from a credit risk point of view, SMEs are different from large corporations for a number of reasons (Altman & Sabato, 2007). Jacobson, Lindé, and Roszbach (2005) and Dietsch and Petey (2004) show that bank loan portfolios of SME loans are usually riskier than corporate credit. Altman and Sabato (2007) further suggest that more accurate credit scoring models in the market for SME loans have many potential benefits; First of all, if banks are able to improve the accuracy of their credit scoring models, their capital requirements may be lower. Second of all, if banks are able to reduce their capital requirements in SME lending, this could result in lower interest rates for their SME customers. Tsai and Wu (2008) further suggest that even a slight improvement in credit scoring accuracy might reduce credit risk and translate into significant future savings. It may therefore be plausible to assume, that more accurate credit risk evaluation can strengthen the competitive advantage of financial institutions and reduce the financing difficulties for SMEs.
Financial innovation has been described as the life blood of efficient and responsive capital markets, and one of the most important innovations since the 1990 s is banks’ use of credit risk evaluation in SME business lending (Akhavein, Frame, & White, 2005). SME business lending based on credit risk evaluation is a relatively new technology that involves processing data about the firm and its owner by using statistical methods to predict applicants’ expected future loan performance (Hand & Henley, 1997). Previous studies indicate that many methods can be applied in credit risk evaluation. These methods can be classified into traditional statistical methods and recent artificial intelligence approaches.
Statistical methods are the first and most frequently used methods in credit scoring or credit risk evaluation. Many researchers used statistical methods to build a credit risk model (Altman, 1968, Altman and Sabato, 2007, Banasik et al., 2001, Boyes et al., 1989, Durand, 1941, Ewert, 1968, Makowski, 1985, Myers and Forgy, 1963, Orgler, 1970, Šarlija et al., 2004, Steenackers and Goovaerts, 1989, Wiginton, 1980). With the development of information and computational technologies, recently more accurate credit risk models have been developed based on sophisticated intelligence approaches which are more capable of modelling nonlinear or extremely complex functions. Credit risk modeling is one of the main areas in accounting and finance which artificial intelligent technologies have been applied into successfully (Angelini et al., 2008, Arminger et al., 1997, Chatterjee and Barcun, 1970, Desai et al., 1996, Piramuthu, 1999, Tsai and Wu, 2008, West, 2000). In recent years, hybrid prediction models, which combine traditional statistical methods and artificial intelligence technologies have been suggested to have better prediction ability than either of the two components. For instance, in environmental engineering, Schafer (2008) shows that the accuracy of prediction of a logistic regression model is lower than that of the ANN/logistic hybrid model. Lin (2009) obtained similar results when she investigate financially distressed banks. It has been suggested that neural network may not be as stable as standard statistical techniques, and combinations in certain applications might be more valuable (Paliwal & Kumar, 2009).
The bulk of previous studies in credit scoring models concentrate on either traditional statistical methods or artificial intelligence technologies and there is little evidence on hybrid credit risk models in the literature especially for SMEs business lending. The aim of this paper is to propose a novel hybrid credit risk model (ANN/logistic hybrid model) for SMEs lending integrating ANN with the logistic regression method. By using data Finnish SMEs from the fiscal years from 2004 to 2012, we are able to test whether the performance of credit risk models based on artificial intelligent technology can be improved by combining them with traditional statistical methods. To the best of our knowledge, our study is the first one to investigate the ANN/logistic hybrid model for data on SMEs.
Our main empirical results suggest that the ANN/logistic hybrid model is more accurate than either of the separate ANN or logistic regression approach. Additionally, we find that credit risky firms are less profitable than non-credit risky firms. Furthermore, they are more levered and have higher asset turnover than non-credit risky firms. Also, credit risky firms are more likely to be larger and younger than non-credit risky firms. This study extends the findings of previous studies on four aspects. First, our study is one of few that sheds light on the hybrid model for financial data, while most of the previous ones concentrate on statistical methods or computational (artificial) intelligence methods. Second, SMEs have unique accounting characteristics compared to large firms. In order to develop SME business lending, instead of using a credit risk model for corporations, a powerful and reliable credit risk estimation model just for SME is necessary. Third, we propose a credit risk model for Finnish firms, whereas most of existing studies have used data on the US and UK. Fourth, this study demonstrates that with the inclusion of traditional statistic methods (e.g. logistic regression), the accuracy of artificial intelligent technologies (e.g. ANN) in credit scoring models can be improved.
The remainder of this paper is structured as follows. Section two of the study discusses the relevant literature. Section three presents the methodology in this study. Section four presents the data and descriptive statistics on the variables. Section five presents the main empirical results. Section six is the discussions and conclusions.
Section snippets
Literature review
Previous studies have applied various methods to credit risk evaluation. These methods can be classified into traditional statistical methods and artificial intelligence approaches. Statistical methods are the first and the most frequently used methods in credit risk evaluation. These methods include linear regression, discriminate analysis, and logistic regression, etc. Artificial intelligence methods have been presented to credit risk evaluation much more recently. Even more recent and much
The principle of a feed-forward Neural Network
A BP neural network is one type of artificial intelligence algorithm in ANN. In this study, we chose a feed-forward neural network to build the model. The BP neural network is a two-tier or multi-layer feed forward neural network whose neurons transfer function is Sigmoid-function. The output of the network is continuous between 0 and 1 volume. The feed-forward neural network can realize the mapping from input to output of any non-linear function. A typical feed-forward neural network is an
Sample selection
The data used in this study is a combination of credit related data and financial data of Finnish SMEs from the fiscal years from 2004 to 2012. The credit related data were collected through a VOITTO database from Suomen Asiakastieto Ltd, which is a credit rating and financial information company. The firms’ financial data were collected from Amadeus Database of Bureau van Dijk, which covers firms’ financial and business information all over the Europe. Finally, the total number of observations
First step logistic regression in modeling samples
We put the selected six components as the inputs and CREDIT_RISK as the output to run a logistic regression model. In Table 5 we can see that the total number of the training samples for modeling is 1883 (70% of total sample 2681) and the pseudo R2 is 0.179. Based on the results, we define a new variable named Logistic_Results, which is the real prediction value from the first step logistic regression.
Second step basic neural network in modeling samples
We added 22 dependent variables as the inputs and CREDIT_RISK as the output to train the basic
Conclusions and future research
The purpose of this study is to investigate the accuracy of a hybrid credit risk model (ANN/logistic hybrid model) for SME lending using data of Finnish SMEs for the period 2004 to 2012. To the best of our knowledge, our study is the first one to investigate the ANN/logistic hybrid model with data on SMEs.
Our main empirical results indicate that the ANN/logistic hybrid model is more accurate in evaluating credit risk in SMEs lending. Also, we find that, at least in our data on Finnish SMEs,
Acknowledgements
We would also like to acknowledge the financial support from OP-Pohjola Group Research Foundation, Liikesivistysrahasto, CSC and Jenny ja Antti Wihurin Rahasto.
References (66)
- et al.
A neural network approach for credit risk evaluation
The Quarterly Review of Economics and Finance
(2008) - et al.
Using genetic algorithms to select architecture of a feedforward artificial neural network
Physica A: Statistical mechanics and its applications
(2001) - et al.
A real-coded genetic algorithm for training recurrent neural networks
Neural networks
(2001) - et al.
An econometric analysis of the bank credit scoring problem
Journal of Econometrics
(1989) - et al.
Combination of feature selection approaches with SVM in credit scoring
Expert Systems with Applications
(2010) The importance of relationships to the availability of credit
Journal of Banking and Finance
(1998)- et al.
Recent developments in consumer credit risk assessment
European Journal of Operational Research
(2007) - et al.
A comparison of neural networks and linear scoring models in the credit union environment
European Journal of Operational Research
(1996) - et al.
Should SME exposures be treated as retail or corporate exposures? A comparative analysis of default probabilities and asset correlations in French and German SMEs
Journal of Banking & Finance
(2004) - et al.
Logistic regression and artificial neural network classification models: A methodology review
Journal of biomedical informatics
(2002)