Cedit risk analysis of small and medium-sized enterprises based on Thai data

Taghizadeh-Hesary, Farhad; Yoshino, Naoyuki; Charoensivakorn, Phadet; Niraula, Baburam Working Paper Cedit risk analysis of small and medium-sized enterprises based on Thai data ADBI Working Paper Series, No. 905 Provided in Cooperation with: Asian Development Bank Institute (ADBI), Tokyo Suggested Citation: Taghizadeh-Hesary, Farhad; Yoshino, Naoyuki; Charoensivakorn, Phadet; Niraula, Baburam (2018) : Cedit risk analysis of small and medium-sized enterprises based on Thai data, ADBI Working Paper Series, No. 905, Asian Development Bank Institute (ADBI), Tokyo This Version is available at: https://hdl.handle.net/10419/222672 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/3.0/igo/ ADBI Working Paper Series CREDIT RISK ANALYSIS OF SMALL AND MEDIUM-SIZED ENTERPRISES BASED ON THAI DATA Farhad Taghizadeh-Hesary, Naoyuki Yoshino, Phadet Charoensivakorn, and Baburam Niraula No. 905 December 2018 Asian Development Bank Institute The Working Paper series is a continuation of the formerly named Discussion Paper series; the numbering of the papers continued without interruption or change. ADBI’s working papers reflect initial ideas on a topic and are posted online for discussion. Some working papers may develop into other forms of publication. Suggested citation: Taghizadeh-Hesary, F., N. Yoshino, P. Charoensivakorn, and B. Niraula. 2018. Credit Risk Analysis of Small and Medium-Sized Enterprises Based on Thai Data. ADBI Working Paper 905. Tokyo: Asian Development Bank Institute. Available: https://www.adb.org/publications/credit-risk-analysis-sme-based-thai-data Please contact the authors for information about this paper. Email: [email protected], [email protected] Farhad Taghizadeh-Hesary is Assistant Professor of Economics, Faculty of Political Science and Economics, Waseda University, Tokyo. Naoyuki Yoshino is Dean of the Asian Development Bank Institute and Professor Emeritus at Keio University, Tokyo, Japan. Phadet Charoensivakorn is Senior Executive Vice President, National Credit Bureau, Bangkok. Baburam Niraula is a Consultant at the World Bank, Nepal. The views expressed in this paper are the views of the author and do not necessarily reflect the views or policies of ADBI, ADB, its Board of Directors, or the governments they represent. ADBI does not guarantee the accuracy of the data included in this paper and accepts no responsibility for any consequences of their use. Terminology used may not necessarily be consistent with ADB official terms. Working papers are subject to formal revision and correction before they are finalized and considered published. Asian Development Bank Institute Kasumigaseki Building, 8th Floor 3-2-5 Kasumigaseki, Chiyoda-ku Tokyo 100-6008, Japan Tel: +81-3-3593-5500 Fax: +81-3-3593-5571 URL: www.adbi.org E-mail: [email protected] © 2018 Asian Development Bank Institute ADBI Working Paper 905 F. Taghizadeh-Hesary et al. Abstract SMEs often have severe difficulties raising money. Considering the bank-dominated characteristic of economies in Asia, banks are the main source of financing. In order to prevent the accumulation of non-performing loans in the small and medium-sized enterprise (SME) sector, it is crucial for banks to distinguish healthy SMEs from risky ones. This chapter examines how a credit rating scheme for SMEs can be developed when access to other financial and non-financial ratios is not possible by using data on lending by banks to SMEs. We employ statistical techniques on five variables from a sample of 3,272 Thai SMEs and classify them into subgroups based on their financial health. The source of data used for the credit risk analysis in this research is the National Credit Bureau of Thailand. By employing these techniques, banks could reduce information asymmetry and consequently set interest rates and lending ceilings for SMEs. This would ease financing to healthy SMEs and reduce the number of non-performing loans to this important sector. Keywords: small and medium-sized enterprises, SME, credit risk analysis, NCB JEL Classification: G21, G23, G24, G32 ADBI Working Paper 905 F. Taghizadeh-Hesary et al. Contents 1. INTRODUCTION ......................................................................................................... 1 2. IMPORTANCE OF SMALL AND MEDIUM-SIZED ENTERPRISES FOR THE THAI ECONOMY ........................................................................................ 2 3. SME FINANCIAL SUPPORT PROGRAM IN THAILAND ............................................ 5 3.1 Policy and Regulation ...................................................................................... 5 3.2 Bank-based Lending ....................................................................................... 6 3.3 Non-Bank-based Lending ................................................................................ 8 4. INTRODUCTION OF THE NATIONAL CREDIT BUREAU ......................................... 8 4.1 Thailand’s National Credit Bureau ................................................................... 8 5. ANALYSIS OF CREDIT RISK ..................................................................................... 9 5.1 Theoretical Background .................................................................................. 9 5.2 Data and Variables ........................................................................................ 11 5.3 Principal Component Analysis ....................................................................... 11 5.4 Cluster Analysis ............................................................................................. 13 5.5 Average Linkage Method .............................................................................. 14 5.6 Robustness Check of the Method ................................................................. 16 6. CONCLUDING REMARKS........................................................................................ 17 REFERENCES ..................................................................................................................... 20 ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 1 1. INTRODUCTION According to calculations by Thailand’s Office of Small and Medium Enterprises Promotion (OSMEP), the country is home to more than 2.7 million enterprises, 99.7% of which are small and medium-sized enterprises (SMEs) and 0.3% large enterprises (at the end of 2014) (OSPEM 2015). SMEs in Thailand make a structural contribution to the economy. They account for a significant share of employment (78.5% of total employment in 2016) and almost 42.2% of contribution to the GDP (OECD/ERIA 2018). Since SMEs are an important sector of the Thai economy, it is important to increase their resilience. One way to do this is to provide them with stable finance. SME credit, which amounted to 33.51% of total commercial bank loans in 2018 Q1, is still small in scale, whereas the ratio of non-performing loans (NPLs) remains high in SME lending, at 4.5% compared with the gross NPL rate of 2.9% in 2018 Q1 (Bank of Thailand 2018a). The lack of collateral is a critical barrier to raising business funds for Thai SMEs. Moreover, new start-up enterprises face increased difficulty in obtaining credit due to a lack of credit history, leading to them being perceived as high risk for finance by banks. Asymmetry of information makes the recognition of healthy SMEs among risky borrowers difficult. If the credit history of SMEs is assessed thoroughly, then it is possible for banks to ask for lower collateral from SMEs that will make the lending of banks to SMEs easier. A main requirement for achieving this goal could be establishing a comprehensive nationwide SME database. One existing SME database is Japan’s Credit Risk Database (CRD), which contains data from 14.4 million SMEs, including default data from 3.3 million corporations and sole proprietorships. The data are collected from credit guarantee corporations and financial institutions (Yoshino and Taghizadeh-Hesary 2015a). Another existing example is Thailand’s National Credit Bureau (NCB) database. In the absence of a nationwide comprehensive SME credit risk database, however, it is important for banks to start to accumulate SME data and carry out credit risk assessments on them by applying credit rating techniques. For the credit rating of SMEs, Yoshino and Taghizadeh-Hesary (2014a) proposed a statistical analysis of the quality of SMEs that can be helpful in facilitating bank financing. The motivation for this paper comes from the fact that, unlike for large firms, an extensive credit rating scheme/index for small and medium-sized firms is lacking. Developing a credit-rating index would not only shield banks from risky lending by reducing information asymmetry, but also lower borrowing costs for SMEs that have good financial health and prospects to grow. In this research, we seek to show that even if comprehensive data on SMEs is not available or is difficult to collect, it is possible to carry out the credit risk assessment and credit scoring of SMEs by relying on their borrowing history from banks. We believe that following a specific criterion like the one we are developing in this chapter for assessment of the creditworthiness of SMEs is more rational and fair than is the case if lending institutions do not consider any index but rely on the personal judgement of the bank clerk who might be subject to moral hazard, biased output and possible corruption. In the following Section 2, we illustrate the importance of SMEs in the context of Thailand. In Section 3, we highlight the SMEs’ financial support program in Thailand. Section 4 introduces the NCB. Section 5 explains the credit risk analysis using Thai SME data, followed by the methodologies we use. We show that credit ratings for SMEs can be based on variables that are easily obtained—such as total loans, the ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 2 amount of outstanding loans, initial loan amounts, and past due amounts—which is particularly useful when access to SMEs’ financial statements is not available. Section 6 provides the concluding remarks. 2. IMPORTANCE OF SMALL AND MEDIUM-SIZED ENTERPRISES FOR THE THAI ECONOMY SMEs are the key drivers of the Thai economy. In Table 1 and Figures 1, 2, and 3, we present the number of SMEs, employment, GDP growth rate and credit allocated to SMEs in Thailand. These figures show that the contributions of SMEs to the Thai economy in terms of number, employment, and GDP are all very large. However, bank loans to this sector, considering its contribution to the economy, are still small. The total number of enterprises in Thailand at the end of 2014 was 2,744,198, of which 2,736,744 were SMEs, or 99.73% of the total number of enterprises. The SMEs in Thailand experienced 0.76% growth compared to 2013 as a whole. The number of small enterprises totaled 2,723,932, accounting for 99.26% of the country’s total number of enterprises and 99.53% of the country’s total number of SMEs. SMEs are predominantly found in the wholesale and retail trade sectors and automobile repair (42.37%). The next largest share is for the service sector (37.87%). The third share is for the manufacturing sector (18%), while SMEs in agriculture account for about 1.1% of the total number of SMEs in Thailand (OSMEP 2015). There is a particularly high concentration around the capital: Bangkok and its environs account for 27.6% of SMEs, with Bangkok alone accounting for around 18.1%. The next highest concentrations can be found in Chonburi (3.4%) and Chiang Mai (3.2%) (OECD/ERIA 2018). Table 1: Number of Enterprises in Thailand Classified by Size Size of Enterprises 2013 2014 Number of Enterprises Ratio to Total Number of Enterprises Ratio to SMEs Number of Enterprises Ratio to Total Number of Enterprises Ratio to SMEs Small and Medium Enterprises (SMEs) 2,716,038 99.73 100 2,736,744 99.73 100 Small Enterprises (SEs) 2,716,038 99.27 99.53 2,723,932 99.26 99.53 Medium Enterprises (MEs) 12,645 0.46 0.47 12,812 0.47 0.47 Large Enterprises (LEs) 6,966 0.26 – 7,062 0.26 – Unknown 392 0.01 – 392 0.01 – Total 2,723,396 100.00 – 2,744,198 100.00 – Source: Office of Small and Medium Enterprises Promotion (OSMEP 2015). In common with much of Southeast Asia, there appears to be a “missing middle”1 in Thailand’s production structure (OECD/ERIA 2018). As is clear in Table 1, only 0.47% of enterprises, or fewer than 13,000, are observed to be medium-sized. Conversely, there are slightly more large firms than in other ASEAN countries. Micro-enterprises are not disaggregated in official SME statistics but are included in the small enterprise count. 1 A lack of medium size companies occurring in less developed countries is called a “missing middle”. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 3 Figure 1: Employment by Small and Medium-Sized Enterprises in Thailand Note: Left axis unit is millions of employees. Right axis is the share of SMEs in total employment (percentage). Source: Office of Small and Medium Enterprises Promotion (OSMEP 2015). Figure 1 shows contributions of SMEs to employment in Thailand. SMEs in Thailand account for a significant share of employment. More than 12 million employees are working in SMEs in Thailand, and the contribution of SMEs to employment in Thailand has risen from less than 70% in 2002 to more than 80% in 2017. Figure 2: Growth Rate of Total GDP and GDP of SMEs Note: The horizontal axis shows the year; the vertical axis is the GDP growth rate (percentage). Source: Bank of Thailand (2018b). While the contribution of SMEs to employment in Thailand is significant, SMEs account for a relatively low share of GDP (42.2%). Figure 3 compares the total GDP growth rate of Thailand with the GDP growth rate of the SME sector. The dashed line is the GDP growth rate of the whole economy and the constant line is for the SME sector. As is clear, since 2014 the SME sector GDP growth has overtaken the whole economy GDP ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 4 growth rate. This means that the supportive government plans were effective2 (Oxford Business Group, 2018) and the share of SMEs in Thailand’s GDP is gradually rising. Figure 3: Total Loans Outstanding and SME Loans in Thailand, 2010 Q1–2018 Q1 Note: Total NPL is total non-performing loans outstanding of all commercial banks registered in Thailand. Total Loan is total loans outstanding of all commercial banks registered in Thailand. SME NPL is non-performing loans outstanding for all commercial banks registered in Thailand to SMEs, and SME Loan is loans outstanding for all commercial banks registered in Thailand to SMEs. Source: Created by the authors using data from the Bank of Thailand (2018a). Figure 3 shows the commercial banks’ total loans outstanding and loans outstanding to SMEs and their non-performing loan ratios from 2010 Q1 until 2018 Q1. The share of SME loans in total commercial banks loans is almost constant. This share in 2010 Q1 was 32.28% and in 2018 Q1 slightly rose to 33.51%. On the other hand, as it is clear that SMEs’ non-performing loan ratio of SME loans is much higher compared to the ratio of total NPLs in total loans. In 2010 Q1, 6.45% of total SME loans were NPL; however, at the same time, only 4.97% of total commercial loans in Thailand were NPL. In 2014, SMEs’ NPL ratio reduced to less than 3.2%; however, in recent years it has been rising. As Figure 3 shows, this ratio is currently increasing again and the NPLs of SMEs is slightly diverging from total NPLs. In 2018 Q1, SME NPLs divided by total SME loans was almost 4.5%, while total NPLs divided by total loans was 2.91%. This shows the necessity of credit risk assessment for SMEs in order to direct the bank loans only to healthy SMEs and avoid lending to risky SMEs, thus reducing the NPL ratio of SME loans. 2 The Thai Government launched a series of support mechanisms aimed at boosting SME exports, supporting digital business development and e-commerce activities, and increasing access to finance at commercial banks and the Thai bourse. These efforts helped SMEs strengthen domestic operations and expand their presence regionally, boosting macroeconomic development and stability, and further supporting ongoing efforts to develop a digital economy. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 11 5.2 Data and Variables We used 2015 Commercial Credit Scoring Data from the NCB of Thailand. The dataset contains 1 million SMEs with their credit history: loan amounts, default status, past due amount, past due days, etc. For simplicity, we selected 3,272 observations for the principal component analysis.5 For the cluster analysis, we took 1,197 observations that were used in the principal component analysis.6 Table 2 describes the variables we used in the empirical part of this chapter. Table 2: Description of Examined Variables Series No. Variable Description Mean Minimum Maximum 1 Initial amount Principal amount 11,393,080 10,192 8,790,000,000 2 Past due days Overdue days code 6.924817 1 11.00 3 Past due amount Past due incurred 12,929,693 2 3,748,031,686 4 Total loans Total loans lent 39,638,340 4 49,995,000,000 5 Outstanding amount Outstanding amount 19,636,365 0 8,852,398,916 Note: Definitions of overdue days are as follows: 004: Overdue 91–120 days 005: Overdue 121–150 days 006: Overdue 151–180 days 007: Overdue 181–210 days 008: Overdue 211–240 days 009: Overdue 241–270 days 010: Overdue 271–300 days 011: Overdue more than 301 days Source: Yoshino et al. (2016). In the next stage, two statistical techniques are used: principal component analysis and cluster analysis. The underlying logic of both techniques is dimension reduction— summarizing information on multiple variables into just a few variables—but they achieve this in different ways. Principal component analysis reduces the number of variables into components (or factors). Cluster analysis reduces the number of SMEs by placing them in small clusters. In this survey, we use components (factors) that are the result of principal component analysis and then run the cluster analysis in order to group the SMEs. 5.3 Principal Component Analysis Principal component analysis is a standard data-reduction technique that extracts data, removes redundant information, highlights hidden features, and visualizes the main relationships that exist between observations. 7 Principal component analysis is a technique for simplifying a dataset by reducing multi-dimensional datasets to lower dimensions for analysis. Unlike other linear transformation methods, principal component analysis does not have a fixed set of basis vectors: its basis vectors 5 As there were too many zeros in the observations, we selected only SMEs with non-zero variables and randomly selected 3,272 observations from them. 6 Outlier observations were excluded. 7 Principal component analysis can also be called the Karhunen–Loève theorem (KLT), named after Kari Karhunen and Michel Loève. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 12 depend on the dataset. Principal component analysis has the additional advantage of indicating what is similar and different about the various models created (Bruce-Ho and Dash-Wu 2009; Jolliffe 2002). Through this method, we reduce the five variables listed in Table 2 to determine the minimum number of components that can account for the correlated variance among SMEs. We performed the Kaiser-Meyer-Olkin (KMO) test and Bartlett’s test of sphericity to examine the suitability of our data for factor analysis. KMO measures the sample adequacy, which indicates the proportion of common variance that might be caused by underlying factors. High KMO means that factor analysis may be useful, while a value of less than 0.5 indicates lower suitability for factor analysis. KMO in this study was 0.50; thus we proceeded with factor analysis. We then determined how many factors to use in our analysis. Table 3 reports the factors and their estimated eigenvalues. We used factors that explained more than 20% of the variance—i.e., an eigenvalue greater than or equal to 1. Thus, only three factors were retained, which together explain 69.33% of the variance. Accordingly, five variables in the dataset can be explained by the three main components. Table 3: Total Variance Explained Component Eigenvalue % of Variance Cumulative % Z1 1.39 27.85 27.85 Z2 1.07 21.42 49.27 Z3 1.00 20.06 69.33 Z4 0.92 18.50 87.83 Z5 0.61 12.17 100.00 Note: Extraction method: principal component analysis. When components are correlated, the sums of squared loadings cannot be added to obtain the total variance. Source: Yoshino et al. (2016). In running the principal component analysis, we used direct oblimin rotation. This is the standard method to obtain a non-orthogonal (oblique) solution—that is, one in which the factors are allowed to be correlated. In order to interpret the information revealed in the principal component analysis, the pattern matrix must then be studied. Table 4 presents the pattern matrix of factor loadings by use of the direct oblimin rotation method, where variables with large loadings—absolute value (>0.6) for a given factor— are highlighted in bold. In Table 4, we present three significant components. The first component, Z1, received main loadings from two variables, total loans and initial amount. This implies that the higher the total loans and initial amount, the larger Z1 will be. Z1 is called ‘flow amount of loan’. The second component that was significant in our credit analysis was Z2. It received the highest loadings from one variable, past due amount, with positive loading; it is called ‘expected NPL’. We can infer that the higher the past due amount, the higher Z2 will be. The third component was Z3, which received the highest loading from outstanding amount. This means that if the outstanding amount of loans of SMEs increases, Z3 will increase. Z3 is called ‘stock amount of loans’. Thus we have these three components based on the significant loadings they received. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 13 Table 4: Factor Loadings of Loan Variables after Direct Oblimin Rotation Variable Component Z1 Z2 Z3 Total loans 0.834 –0.009 0.004 Initial amount 0.833 0.018 0.017 Past due amount 0.029 0.834 –0.242 Outstanding amount 0.023 –0.083 0.896 Past due days 0.034 –0.590 –0.407 Notes: Extracted from principal component analysis using direct oblimin rotation with Kaiser normalization. Values greater than 0.5 in absolute terms are in bold. Components receive their main loadings from the bold variables. Source: Yoshino et al. (2016). Table 5 shows the correlation matrix for components Z1, Z2, and Z3. As can be seen, none of the correlations is significant, suggesting that a regular orthogonal rotation method to force an orthogonal rotation would have been applicable. Nevertheless, our application of the oblique rotation method provided an orthogonal rotation since there is no significant correlation between the components. Table 5: Component Correlation Matrix Z1 Flow Amount of Loan Z2 Expected NPL Z3 Stock Amount of Loans Z1 Flow amount of loan 1.000 –0.008 –0.025 Z2 Expected NPL –0.008 1.000 0.065 Z3 Stock amount of loans 0.025 0.065 1.000 Note: Extracted from principal component analysis with direct oblimin rotation with Kaiser normalization. Source: Yoshino et al. (2016). 5.4 Cluster Analysis In this section, we group SMEs with similar characteristics using the cluster analysis method. Clustering divides observations into groups with certain similar traits. In this survey, we use the three significant components that we obtained from the principal component analysis to group SMEs into different clusters. Clustering is useful to compare a group that has similar characteristics within the group to another group of observations with different characteristics from the former but similar within-group features. In this case, SMEs were organized into distinct groups according to the three significant components derived from the principal component analysis used in the previous section. Cluster analysis techniques can themselves be broadly grouped into three classes: hierarchical clustering, optimization clustering, and model-based clustering. 8 We used the method most prevalent in the literature, hierarchical 8 The main difference between the hierarchical and optimization techniques is that in hierarchical clustering the number of clusters is not known beforehand. The process consists of a sequence of steps where two groups are either merged (agglomerative) or divided (divisive) according to the level of ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 14 clustering. This produced a nested sequence of partitions by merging (or dividing) clusters. At each stage of the sequence, a new partition is optimally merged with (or divided from) the previous partition according to some adequacy criteria. The sequence of partitions ranges from a single cluster containing all the individuals to a number of clusters (n) containing a single individual. The series can be described by a tree display, or dendrogram (Figure 5). Agglomerative hierarchical clustering proceeds by a series of successive fusions of the n objects into groups. In contrast, divisive hierarchical methods divide the n individuals into progressively finer groups. Divisive methods are not commonly used because of the computational problems they pose (Everitt, Landau, and Leese 2001; Landau and Chis Ster 2010; Rousseeuw 2005). Below, we use the average linkage method, which is a hierarchical clustering technique. 5.5 Average Linkage Method The core idea of the average linkage method is that the distance between each cluster is regarded as the average distance from all observations in one cluster to all points in another cluster. Average linkage averages all distance values between pairs of cases from different clusters. At different distances, different clusters are formed, which can be represented by a dendrogram. The average linkage method is robust and takes cluster structure into account (Martinez and Martinez 2005). The basic algorithm for the average linkage method can be summarized as: 1. N observations start out as N separate groups. The distance matrix D=(dij) is searched to find the closest observations—for example, Y and Z. 2. The two closest observations are merged into one group to form a cluster (YZ), producing N–1 total groups. The process continues until all the observations are merged into one large group (Rousseeuw 2005). Figure 5 shows a dendrogram, where the y-axis marks the distance at which the clusters merge, while the objects (SMEs) are placed along the x-axis such that the clusters do not mix. The dendrogram shows the formation of three main groups. The resultant dendrogram does not tell us in which cluster SMEs of a particular financial category lie. This can be achieved by plotting the distributions for factors for each member of the major categories. The scatter plot in Figure 6 shows that there are three different groups of SMEs in three different dimensions.9 This is proof of our dendrogram, because our dendrogram also categorized the SMEs into three groups. Moreover, we picked random samples from each group the dendrogram gave us and found that Group A is the healthiest group. As we move on the horizontal line to the right, soundness declines, meaning that Group C has the lowest soundness, and Group B is in between. similarity. Eventually, each cluster can be subsumed as a member of a larger cluster at a higher level of similarity. The hierarchical merging process is repeated until all subgroups are fused into a single cluster (Martinez and Martinez 2005). Optimization methods, on the other hand, do not necessarily form hierarchical classifications of the data, as they produce a partition of the data into a specified or predetermined number of groups by either minimizing or maximizing some numerical criteria (Feger and Asafu-Adjaye 2014). 9 Scatter plots of other sets of components (Z1–Z2) and (Z1–Z3) show almost similar classifications. Here we keep the one that was the clearest. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 15 Figure 5: Dendrogram Note: Group A is the most creditworthy group of SMEs. Group C has the least credit healthiness. Group B is between the two. Source: Yoshino et al. (2016). Figure 6: Distribution of Factors Note: The x-axis is the Z2 component (expected NPL) and the y-axis is the Z3 component (stock amount of loans). Each dot shows one SME. Group A is the most creditworthy group of SMEs. Group C has the least credit healthiness. Group B is between the two. Source: Yoshino et al. (2016). Our dendrogram results are contrary to the dendrogram of Yoshino and Taghizadeh- Hesary (2014a). For the credit risk analysis, Yoshino and Taghizadeh-Hesary introduced 11 financial ratios of SMEs (equity [book value)/total liabilities, cash/total assets, working capital/total assets, cash/net sales, retained earnings/total assets, etc.), which represent the positive characteristics of the examined SMEs. This means that the larger these variables are, the healthier a certain SME is. Their cluster analysis shows the healthier SMEs on the right side of the dendrogram, and the SMEs with ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 16 reduced financial health toward the left of the x-axis. However, the variables in our analysis are the other way around. Our variables are lending variables (past due amount, past due days, outstanding amount, etc.) and mainly represent the negative characteristics of SMEs. Therefore the resultant components (Z1, Z2, and Z3) also show the negative features of SMEs. This means that the lower these variables/components are, the better the health of a certain SME is. This is the reason why in our results, Group A (the left group on x-axis of the dendrogram) is the healthiest group, and as we move on the horizontal line toward the right, soundness declines, meaning that Group C has the lowest financial healthiness. 5.6 Robustness Check of the Method To highlight the validity of the method we employed for predicting risk outcomes, there is one more step to go, because it might be questioned that the framework developed in this chapter derives risk factors based on factor analysis and clustering without demonstration of the usefulness of the factors for predicting actual risks or alternative outcomes. To show the validity of this model, additional work has been done on a sample consisting of 999 SMEs from an Asian country. The reason that we used another sample database for the robustness check is that, due to limitations of the current research’s sample database, the default data of SMEs were not available; however, we could obtain the default data of SMEs for this second sample. The definition of a default enterprise in this new sample is 90 days’ delay in interest payment. 11 22 33 Yc A A Au αα α =++ + + (5) From the financial ratios10 of the sample, and by applying principal component analysis, technique, three significant components were released ( 123 ,,AAA ). Using these three components, we ran a regression for an equation (Eq. 5) wherein the dependent variable is the default, being a binary variable (0 or 1) and the right hand side of the equation is a constant ( c ), three components ( 123 ,,AAA ) and u the error term. Eq. 5 is a maximum likelihood-binary probit (quadratic hill climbing) model. We ran the regression using the ordinary least squares method. Using probit models to assess the impact of macro-level variables and micro-level variables on the default has long been popular among scholars (Amaral, Abreu, and Mendes 2014; Mizen and Tsoukas 2012; Moulton, Haurin, and Shi 2015). However, the advantage of our method compared to a normal probit employing just financial ratios as the explanatory variable is that our prediction is based on factor analysis. Each factor contains information on several variables (financial ratios), and in preparing these factors, unnecessary information is eliminated by statistical techniques. Hence we believe this is a more complete method and the robustness check stated in this sub-section affirms it. The pseudo R-squared11 10 To see the financial variables used, see Yoshino and Taghizadeh-Hesary (2015b). 11 When analyzing data with a probit regression, an equivalent statistic to R-squared does not exist. The model estimates from a probit regression are maximum likelihood estimates arrived at through an iterative process. They are not calculated to minimize variance, so the ordinary least squares approach to goodness-of-fit does not apply. However, to evaluate the goodness-of-fit of logistic models, several pseudo R-squareds have been developed. These are pseudo R-squared because they look like R-squared in the sense that they are on a similar scale, ranging from 0 to 1 (though some pseudo R-squareds never achieve 0 or 1), with higher values indicating better model fit, but they cannot be interpreted as one would interpret an ordinary least squares R-squared, and different pseudo R-squareds can arrive at very different values. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 17 represents the percentage of variation in the dependent variable (Y=default) explained by variation in the independent variables (components). The results of probit model regression are demonstrated in Table 6. Table 6: Probit Regression Results Variable Definition Coefficient Std. Error Z-Statistic Prob. C Constant 1.14 0.09 13.06** 0 1 A Liabilities 1.00 0.16 6.31** 0 2 A Short-term assets –2.17 0.14 –15.40** 0 3 A Liquidity –1.02 0.21 –4.75** 0 McFadden R-squared:12 0.76. Note: The dependent variable in this regression is default. The regression method is ordinary least squares. ** shows a significant result in 0.01. Source: Yoshino et al. (2016). The first component in Table 6, which represents the liability13 of the SMEs, has a positive impact on default. It means default SMEs had higher liabilities, and it is statistically significant. The second component, representing short-term assets, had negative alteration of the default variable which was statistically significant. This indicates that the default SMEs in this survey primarily had lower short-term asset reserves. The last component (component 3), which represents liquidity, had a significant negative role in alteration of the default variable. This means the higher the liquidity of the SME, the lower the default risk will be. As Table 6 shows, the McFadden R-squared—which is a pseudo R-squared—is 0.76. This means that in 76% of cases the default variable (Y) is explained by independent components, which is quite large and acceptable. However, when instead of the component we used financial ratios as themselves, the R-squared was 0.65. This means that components, because they include information about different financial ratios, are able to forecast the default better. 6. CONCLUDING REMARKS SMEs play a significant role in the Thai economy. The country counts more than 2.7 million enterprises, of which 99.7% are SMEs and 0.3% large enterprises (as at the end of 2014) (OSPEM 2015). SMEs in Thailand demonstrate a structural contribution 12 11 22 33 ˆ Yc A A A αα α =++ + where ˆ Y is the estimated default by use of the components, and Y is the actual default ( Y ), where u is the error term. The R-squared can be calculated as below: ( ) ( ) ( ) ( ) 22 11 22 33 22 ˆ YY c A A AY R squared YY YY αα α −++ + − −= = −− ∑∑ ∑∑   where the Y  is the average of Y . 13 The first component, A1, receives significant loading from two variables, one of which is positive (total debt/total asset) and one negative (equity/total debt). For A1, the ratios with large loadings include total debt: hence A1 generally reflects the liabilities of an SME. As this factor explains the most variance in the data, it is the most informative indicator of the overall financial health of an SME. Z2 reflects shortterm assets. This component has two major loading variables: (a) retained income/total assets and (b) accounts receivable/total debt, both of which are positive. Z3 reflects the liquidity of SMEs. This factor has two variables with large loadings (liquidity/sales and cash/total assets), both with positive values, which shows an SME that is cash-rich. Hence, it mainly reflects the liquidity conditions of an SME. ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 18 to the economy. They account for a significant share of employment (78.5% of total employment in 2016) and almost 42.2% of contribution to the GDP. However, SMEs often have severe difficulties raising money. The under-supply of credit to SMEs is mainly because of asymmetrical information, high default risk and lack of collateral. SMEs have more difficulties accessing finance compared to large enterprises. Lending institutions mainly prefer to increase the flow of funds to the latter sector, since the aforementioned factors are lower in this group. The high cost of monitoring SMEs in the absence of credible credit rating companies results in many financial institutions being less interested in lending to SMEs. The setting up of a credit rating scheme for SMEs is possible by employing statistical techniques on various financial and non-financial variables of SMEs. These techniques could be applied by financial institutions, lending agencies, and credit risk analyzing bureaus. This chapter examined how a credit rating scheme for SMEs can be developed when we do not have access to all financial ratios and only have data on lending from banks to SMEs. By using a sample of 3,272 Thai SMEs and applying statistical analysis techniques, our results show that by using lending data available from lending institutions—i.e., data on total loans, initial amounts of loans, and past due days—we can describe the soundness of SMEs. The variables in this analysis are used to cluster SMEs into different subgroups and sort them based on their financial health. The robustness check stated in this chapter shows that risk factors achieved using statistical techniques can explain 76% of the default cases in a sample consisting of 999 SMEs. This means that that credit risk analysis can clarify the risk factors that have an impact on increasing the default risk ratio of loans, which is in accordance with the theoretical part of this paper. The main policy implication of this chapter is that banks and financial institutions can use credit risk analysis techniques like the one demonstrated in this chapter. Credit risk analysis will help banks to recognize healthy SMEs among non-healthy ones and expand their lending to sound SMEs. On the other hand, it will help to avoid the risk of default for firms with poor financial health. Use of this method will help banks to set borrowing ceilings and interest rates for different SMEs based on their creditworthiness. This will reduce borrowing costs—i.e., lower interest rates for financially healthy SMEs. Banks will also benefit, as the number of non-performing loans for SMEs would diminish. In addition, with the adoption of the Basel Capital Adequacy Requirements, banks are reluctant to fund riskier borrowers such as small enterprises and start-up businesses. The second policy recommendation is introducing hometown investment trust (HIT) funds in Thailand as a suitable financing schemes for risky business sectors (Yoshino 2013). HIT funds are a new form of financial intermediation that have become popular in Japan in a relatively short span of time. They are spreading in Japan and have already entered many other Asian countries, including Cambodia and Viet Nam, as well as outside the Asian region—for example, in Peru. HIT funds can be sold by regional banks and post offices. Such trust funds would not be guaranteed by credit associations or banks. The terms of a trust fund would have to be fully explained to investors—e.g., where their funds would be invested and what would be the risks associated with the investment—in order to strengthen trust fund investors’ confidence and help the trust fund market gro. (Yoshino and Taghizadeh-Hesary, 2014b). ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 19 I. HIT funds reduce information asymmetry. Often the lenders and borrowers know each other. II. HIT funds supply risk capital stably. As bank lending for risky capital has been tightened, SMEs and start-up ventures find it difficult to gather funds. HIT funds can be very helpful for these businesses. III. HIT funds are project-driven. Unlike mutual funds, where investors actually do not know the projects they are investing in, they can choose to invest in projects in their locality. IV. HIT funds are a form of crowdfunding, where the investors are mainly local people. The investors can receive the product and output of the project that they are investing in (for example, investing in a HIT fund for a farm and using the product of that farm instead of a cash return). There is a ‘warm’ feeling behind HIT funds, as investors are sympathetic to the company/project owners. V. HIT funds are now advertising through several internet companies in Japan so that even non-local investors can select the funds that they are interested in it and invest. The backbone of these funds is transparency, that allows trust for the investors. Investors monitor exactly where their money is invested, who the owner of the project is, what the products are, etc. VI. The internet companies that act as project intermediaries assess the projects and select those that are expected to have success; they do not accept and do not advertise very risky projects. These intermediaries need to receive a license from the government. In Japan, they need to receive a Type II Financial Instruments Business license from the Financial Services Agency (FSA), which is the regulator and supervisor of the financial system. VII. The intermediary companies are not asset management companies; they only introduce HIT projects for the development of various products. They simply act as intermediaries and therefore do not provide any guarantee (Yoshino and Taghizadeh-Hesary 2018b). ADBI Working Paper 905 F. Taghizadeh-Hesary et al. 20 REFERENCES Amaral, A., M. Abreu, and V. Mendes. 2014. “The Spatial Probit Model—An Application to the Study of Banking Crises at the End of the 1990s.” Physica A: Statistical Mechanics and its Applications 415: 251–260. Asian Development Bank (ADB). 2015. Asia SME Finance Monitor 2014. Manila: ADB. ADB and OECD. 2014. ADB–OECD Study on Enhancing Financial Accessibility for SMEs Lessons from Recent Crises. Manila: ADB. Bank of Thailand. 2018a. “Loan Outstanding of Commercial Banks Classified by Asset Classification Type Data.” Accessed October 25, 2018. http://www2.bot.or.th/ statistics/BOTWEBSTAT.aspx?reportID=829&language=ENG ———. 2018b. Thailand’s Key Macroeconomic Chart Pack. Bangkok: Statistics and Data Management Department, Bank of Thailand. Bruce-Ho, C.-T. and D. Dash-Wu. 2009. “Online Banking Performance Evaluation using Data Envelopment Analysis and Principal Component Analysis.” Computers & Operations Research 36 (6): 1835–1842. Everitt, B. S., S. Landau, and M. Leese. 2001. Cluster Analysis. 4th ed. London: Arnold. Feger, T., and J. Asafu-Adjaye. 2014. “Tax Effort Performance in Sub-Sahara Africa and the Role of Colonialism.” Economic Modelling 38: 163–174. Jolliffe, I. T. 2002. Principal Component Analysis. 2nd ed. New York: Springer. Klein, N. 2013. “Non-Performing Loans in CESEE: Determinants and Impact on Macroeconomic Performance.” IMF Working Paper 13/72, Washington: International Monetary Fund. Landau, S. and I. Chis Ster. 2010. “Cluster Analysis: Overview.” In International Encyclopedia of Education, 3rd ed., 72–83. Oxford: Elsevier. Martinez, W. L., and A. R. Martinez. 2005. Exploratory Data Analysis with Matlab. Chapman and Hall/CRC Press. Mizen, P., and S. Tsoukas. 2012. “Forecasting US Bond Default Ratings allowing for Previous and Initial State Dependence in an Ordered Probit Model.” International Journal of Forecasting 28 (1): 273–287. Moulton, S., D. R. Haurin, and W. Shi. 2015. “An Analysis of Default Risk in the Home Equity Conversion Mortgage (HECM) Program.” Journal of Urban Economics 90: 17–34. OECD/ERIA. 2018. SME Policy Index: ASEAN 2018: Boosting Competitiveness and Inclusive Growth. Paris/Economic Research Institute for ASEAN and East Asia, Jakarta: OECD Publishing. https://doi.org/10.1787/9789264305328-en OSMEP. 2015. SME White Paper. Bangkok: Office of Small and Medium Enterprises Promotion. Oxford Business group (2018). https://oxfordbusinessgroup.com/analysis/ground-small- business-heart-government-growth-strategy (Accessed November 3, 2018) Rousseeuw, L. K. 2005. Finding Groups in Data: An Introduction to Cluster Analysis. Hoboken, NJ: Wiley.