Expert Systems with Applications 39 (2012) 62386253

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications

journal homepage: www.elsevier.com/locate/eswa

Financial early warning system model anddata mining application for risk detection
Ali Serhan Koyuncugil a,, Nermin Ozgulbas b
a b

Capital Markets Board of Turkey, Research Department, Eskisehir Yolu 18.km., Ankara, Turkey Baskent University, School of Health Sciences, Department of Healthcare Management, Eskisehir Yolu 20.km., Ankara, Turkey

a r t i c l e

i n f o

a b s t r a c t
One of the biggest problems of SMEs is their tendencies to nancial distress because of insufcient nance background. In this study, an early warning system (EWS) model based on data mining for nancial risk detection is presented. CHAID algorithm has been used for development of the EWS. Developed EWS can be served like a tailor made nancial advisor in decision making process of the rms with its automated nature to the ones who have inadequate nancial background. Besides, an application of the model implemented which covered 7853 SMEs based on Turkish Central Bank (TCB) 2007 data. By using EWS model, 31 risk proles, 15 risk indicators, 2 early warning signals, and 4 nancial road maps has been determined for nancial risk mitigation. 2011 Elsevier Ltd. All rights reserved.

Keywords: CHAID Data mining Early warning systems Financial risk Financial distress SMEs

1. Introduction All enterprises especially SMEs need to think about global dimensions of their business earlier than ever. Especially in developing countries, in addition to the administrative insufciencies, competition, economical conditions, the permanent threat towards SMEs from globalization, and nancial crisis have caused distress and affect rms performance. SMEs are dened as enterprises in the non-nancial business economy (NACE, Nomenclature statistique des activits conomiques dans la Communaut europenne (Statistical classication of economic activities in the European Community)) that employ less than 250 persons. The complements of SMEs enterprises that employ 250 or more persons are large scale enterprises (LSEs). Within the SME sector, the following size-classes are distinguished:  Micro enterprises, employing less than 10 persons.  Small enterprises, employing at least 10 but less than 50 persons.  Medium-sized enterprises that employ between 50 and 250 persons. This denition is used for statistical reasons. In the European denition of SMEs two additional criteria are added: annual turnover should be less than 50 million , and balance sheet total should be less than 43 million (Commission Recommendation, 2003/361/EC). SMEs play a signicant role in all economies and are the key generators of employment and income, and drivers of innovation
Corresponding author. Tel.: +90 5326657084; fax: +90 3122466670.
E-mail address: askoyuncugil@gmail.com (A.S. Koyuncugil). 0957-4174/$ - see front matter 2011 Elsevier Ltd. All rights reserved. doi:10.1016/j.eswa.2011.12.021

and growth. Access to nancing is the most signicant challenges for the creation, survival and growth of SMEs, especially innovative ones. The problem is strongly exacerbated by the nancial and economic crisis as SMEs have suffered a double shock: a drastic drop in demand for goods and services and a tightening in credit terms, which are severely affecting their cash ows (OECD, 2009). As a result, all these factors throw SMEs in nancial distress. The failure of a business is an event which can produce substantial losses to all parties like creditors, investors, auditors, nancial institutions, stockholders, employees, and customers, and it undoubtedly reects the economics of the countries concerned. When a business with nancial problems is not able to pay its nancial obligations, the business may be driven into the situation of becoming a non-performing loan business and, nally, if the problems cannot be solved, the business may become bankrupt and forced to close down. Those business failures inevitably inuence all businesses as a whole. Direct and indirect bankruptcy costs are incurred which include the expenses of either liquidating or an attempting to reorganize businesses, accounting fees, legal fees and other professional service costs and the disaster broadens to other businesses and the economics of the countries involved (Ross, Westereld, & Jordan, 2008; Terdpaopong, 2008; Warner, 1977). The awareness of factors that contribute to making a business successful is important; it is also applicable for all the related parties to have an understanding of nancial performance and bankruptcy. It is also important for a nancial manager of successful rms to know their rms possible actions that should be taken when their customers, or suppliers, go into bankruptcy. Similarly, rms should be aware of their own status, of when and where they should take necessary actions in response to their nancial

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

6239

problems, as soon as possible rather than when the problems are beyond their control and reach a crisis. Therefore, to bring out the nancial distress risk factors into open as early warning signals have a vital importance for SMEs as all enterprises. There is no specic method for total prevention for a nancial crisis of enterprises. The important point is to set the factors that cause the condition with calmness, to take corrective precautions for a long term, to make a exible emergency plan towards the potential future crisis. The aim of this paper is to present an EWS model based on data mining. EWS model was developed for SMEs to detect risk proles, risk indicators and early warning signs. Chi-Square Automatic Interaction Detector (CHAID) Decision Tree Algorithm was in the study as a data mining method. Remaining of this paper is organized as follows: Section 2 presents denition of EWS. Section 3 contains data mining model for risk detection and early warning system. Implementation of data mining for risk detection and early warning signals is presented in Section 4. Concluding remarks and strategies were suggested in Section 5.

2. Financial early warning systems An early warning system (EWS) is a system which is using for predicting the success level, probable anomalies and is reducing crisis risk of cases, affairs transactions, systems, phenomena, rms and people. Furthermore, their current situations and probable risks can be identied quantitatively (Ozgulbas & Koyuncugil, 2010). Financial EWS is a monitoring and reporting system that alerts for the probability of problems, risks and opportunities before they affect the nancial statements of rms. EWSs are used for detecting nancial performance, nancial risk and potential bankruptcies. EWSs give a chance to management to take advantage of opportunities to avoid or mitigate potential problems. Nearly, all of the nancial EWSs are based on nancial statements. Balance sheets and income tables are the data sources that reect the nancial truth for early warning systems. In essence, the early warning system is a nancial analysis technique, and it identies the achievement analysis of enterprise due to its industry with the help of nancial ratios. The efforts towards the separation of distressed enterprises started with the z-score that are based on the usage of ratios by Beaver (1966) for single and multiple discriminant analysis of Altman in 1968. The examples of other important studies that used multi variable statistical models, are given by Deakin (1972), Altman, Haldeman, and Narayanan (1977), Tafer and Tisshaw (1977) with the usage of multiple discriminant model; are also given by Zmijewski (1984), Zavgren (1985), Jones (1987), Pantalone and Platt (1987), with the usage of logit and probit models; are at the same time given by Meyer and Pifer (1970) with the usage of multiple regression model. Beside the business distressed studies, researchers focused on monitoring ongoing situations to detect sudden changes or unexpected risk factors of enterprises. These attempts made important the early warning systems for research. Some of previous studies conducted in SMEs, banks, insurers, i.e., and their research methods are presented below. Brockett and Cooper (1990) developed an EWS by using neural network method. The model was developed with 24 variables rstly, and then the numbers of variables were decreased to 8. These variables were equities, capitalization ratio, return on assets, turnover of assets, account receivables to equities, changing of loses, and debt to current assets. Lee and Urrutia (1996) compared the models of logit, hazard, neural networks and discriminant for developing an early warning system. They found different indicators or signs for each model. Also they determined that forecast power of all models were same.

Barniv and Hathorn (1997) developed an early warning model based on logistic regression by evaluating the studies of Trieschmann and Pinches (1973), Ambrose and Seward (1998), and Barniv and McDonald (1992) in insurance rms. Laitinen and Chong (1999) presented a model for predicting crises in small businesses using early-warning signals. Study summarized the results of two separate studies carried out in Finland (with 72% response) and the UK (26%) on the decision process of corporate analysts (Finland) and bank managers (UK) in predicting the failure of small and medium-sized enterprises (SMEs). Both studies consisted of seven main headings and over 40 sub-headings of possible factors leading to failure. Weighted averages were used for both studies to show the importance of these factors. There were signicant similarities in the results of the two studies. Management incompetence was regarded as the most important factor, followed by deciencies in the accounting system and attitude towards customers. However, low accounting staff morale was considered a very important factor in Finland but not in the UK. Yang, Ling, Hai, and Jing (2001) used articial neural networks (ANN) for detecting nancial risk of banks as an early warning, and tested the method. Salas and Saurina (2002) compared the determinants of problem loans of Spanish commercial and savings banks in the period 19851997, taking into account both macroeconomic and individual bank level variables. The GDP growth rate, rms, and family indebtedness, rapid past credit or branch expansion, inefciency, portfolio composition, size, net interest margin, capital ratio, and market power are variables that explain credit risk. The ndings raised important bank supervisory policy issues: the use of bank level variables as early warning indicators, the advantages of bank mergers from different regions, and the role of banking competition and ownership in determining credit risk. Edison (2003) developed an operational early warning system (EWS) that can detect nancial crises. The system monitored several indicators that tend to exhibit an unusual behavior in the periods preceding a crisis. When an indicator exceeded (or falls below) a threshold, then it was said to issue a signal that a currency crisis may occur within a given period. The model was tested in 1997/ 1998 crises, but several weaknesses to the approach were identied. The paper also evaluated how this system can be applied to an individual country. The results suggested that an early warning system should be thought of as a useful diagnostic tool. El-Shazly (2003) investigated the predictive power of an empirical model for an early warning system of currency crises. EWS employed qualitative response models within a signals framework that monitors the behavior of key economic variables and issues a warning when their values exceed certain critical levels. Author conducted a case study in Egypt. Results showed that this model, and in particular the extreme value model, captured to a good extent the turbulence in the foreign exchange market and the onset of crises. Jacobs and Kuper (2004) presented an EWS for six countries in Asia. Financial crises were distinguished in three types; currency crises, banking crises, and debt crises. The signicance of the indicator groups was tested in a multivariate logit model on a panel of six Asian countries for the period 19702001. Author founded that some currency crises dating schemes outperform others by using EWS. Berg, Borensztein, and Pattillo (2004), developed early warning system models of currency crisis for Mexican and Asian crisis. Since the beginning of 1999, IMF staff has been systematically tracking, on an ongoing basis, various models developed in-house and by private institutions, as part of its broader forward-looking vulnerability assessment. This study examined in detail at the performance of these models in practice. The forecasts of the in-house model were statistically and economically signicant predictors of

6240

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

actual crises. On the whole, the short-horizon private sector models examined performed poorly out of sample, despite stellar in-sample performance. Brockett, Golden, Jang, and Yang (2006) examined the effect of statistical model and neural network methods to detect nancially troubled life insurers. They considered two neural network methods; back-propagation (BP), and learning vector quantization (LVQ), two statistical methods; multiple discriminant analysis, and logistic regression analysis. The results showed that BP and LQV outperform the traditional statistical approaches. Abumustafa (2006) detected early warning signs for predicting currency crises in Egypt, Jordan and Turkey. The study proposed real exchange rate, exports, imports, trade balance/gross domestic product (GDP), foreign liabilities/foreign assets, domestic real interest rate, world oil prices, and government consumption/GDP as indicators to predict currency risk. The results showed that all crises were predictable, and EWSs should use for detecting crises. Kyong, Tae, Chiho, and Suk (2006) presented the construction process of a daily nancial condition indicator (DFCI), which can be used as an early warning signal using neural networks and nonlinear programming. The procedure of DFCI construction was completed by integrating three sub-DFCIs, based on each nancial variable, into the nal DFCI. The study, then examined the predictability of alarm zone for the nancial crisis forecasting in Korea. Katz (2006) proposed to use EWS and early warning signs. Study listed often common warning signs and the best ways to solve the problems. These are: payroll taxes, sales tax, and other duciary obligations; communications with executive management and company leaders; accounts receivable; customers and product protability; accounts payable; inventory, management; for capital-intensive or manufacturing operations; and checks as an indicator of problems. Koyuncugil and Ozgulbas (2007a) aim to develop d a nancial early warning model for the SMEs listed in Istanbul Stock Exchange (ISE) in Turkey by using data mining. Authors conducted another study (2007b) and detected early warning signs for nancial risk. A data mining method, Chi-Square Automatic Interaction Detector (CHAID) Decision Tree Algorithm, was used in the study for nancial proling and detecting signs. The study covered 697 SMEs listed in ISE between 2000 and 2005. As a result of the study, the covered SMEs listed in ISE were categorized into 19 nancial proles and it was determined that 430 of them had poor nancial performance, in other words 61.69%. According to the proles of SMEs in nancial distress, return on equity (ROE) will be a nancial early warning signal for SMEs listed in ISE. Koyuncugil and Ozgulbas (2008a) emphasized the affect and importance of operational risk in nancial distressed of SMEs, beside the nancial risk. Authors developed an early warning model that qualitative (operational) and quantitative (nancial) data of SMEs taken into consideration. During the formation of system; an easy to understand, easy to interpret and easy to apply utilitarian model that is far from the requirement of theoretical background was targeted by the discovery of the implicit relationships between the data and the identication of effect level of every factor. This model was designed by data mining. Koyuncugil and Ozgulbas (2009a) developed a nancial early warning model that detected operational risk factors for hedging nancial risk. For this purpose study used CHAID (Chi-Square Automatic Interaction Detector) Decision Trees. The study covered 6.185 rms in Organized Industrial Region of Ankara in 2008. It was found that rms should emphasize the educational background of managers, status of managers, annual turnover, operating length of rms, makers of nancial strategies, expenditure of energy, knowledge about BASEL-II, quality standards, and usage of credit as operational risk factors for hedging operational risk and raising nancial performance.

Koyuncugil and Ozgulbas (2009b) to develop an intelligent nancial early warning system model based on operational and nancial risk factors by using data mining for SMEs in Turkey. This model was aimed to not remain in theoretical structure, be practicable for SME, and available for the utilization of SMEs managers. According to model, nancial data of Turkish SMEs was obtained by means of nancial analyses of balance sheets and income statements through Turkish Central Bank. Operational data couldnt be access by balance sheets and income statements was collected by a eld study from SMEs. Next step of model was analyzed the nancial and operational data by data mining and detecting early warning signs. Davis and Karim (2008a) successful predicted a majority of banking crises in emerging markets and advanced countries in 19702003. Karim also, suggested that logit was the most appropriate approach for global EWS and signal extraction for country specic EWS. Davis and Karim (2008b) searched to assess whether early warning systems based on the logit and binomial tree approaches on the UK and US economies could have helped to warn about the crisis. The study suggested that a broadening of approaches of macro prudential analysis was appropriate for early warning. It can be seen that risk detection oriented early warning systems have a very large implementation domain from the picture given above. Furthermore, last generation Business Intelligence approach data mining accelerated the accuracy of those systems. Operational logic of early warning systems is based on nding unexpected and extraordinary behaviors in subject area. On the other hand, data mining is the way of uncover previously unknown, useful and valuable knowledge, patterns, relations from big amount of data via sophisticated evolutionary algorithms of classical techniques such as statistics, pattern recognition, articial intelligence, machine learning. The denitions of EWS and data mining given lead an interesting similarity. An EWS developed for SMEs must design according to the needs of SMEs managers. Therefore, system must be easy to understand and easy to use, must design according to nancial and operational risk factors (as banks and BASEL II requirements), and must be intelligence for using update data. 3. Data mining model for risk detection and early warning system The identication of the risk factors by clarifying the relationship between the variables denes the discovery of knowledge. Automatic and estimation oriented information discovery process coincides the denition of data mining. Data mining is the process of sorting through large amounts of data and picking out relevant information. Frawley, Piatetsky-Shapiro, and Matheus (1992) has been described data mining as the nontrivial extraction of implicit, previously unknown, and potentially useful information from data. Also, Hand, Mannila, and Smyth (2001) described data mining as the science of extracting useful information from large data sets or databases. Data mining, the extraction of hidden predictive information from large databases, is a powerful new technology with great potential to help companies focus on the most important information in their data warehouses. Data mining tools predict future trends and behaviors, allowing businesses to make proactive, knowledge-driven decisions. The automated, prospective analyses offered by data mining move beyond the analyses of past events provided by retrospective tools typical of decision support systems. Data mining tools can answer business questions that traditionally were too time consuming to resolve. They scour databases for hidden patterns, nding predictive information that experts may miss because it lies outside their expectations (Thearling, 2004).

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

6241

Koyuncugil and Ozgulbas (2010) described the data mining as collection of evolved statistical analysis, machine learning and pattern recognition methods via intelligent algorithms which are using for automated uncovering and extraction process of hidden predictional information, patterns, relations, similarities or dissimilarities in (huge) data. Data mining is used by business intelligence organizations, and nancial analysts to get information from the large data sets. Data mining in relation to enterprise resource planning is the statistical and logical analysis of large sets of transaction data, looking for patterns that can aid decision making (Monk & Wagner, 2006). Today, data mining technology integrated measurement of different kinds of is moving into focus to measure and hedging risk. Data mining techniques have been successfully applied like fraud detection and bankruptcy prediction by Tam and Kiang (1992), Lee, Han, and Kwon (1996), Kumar, Krovi, and Rajagopalan (1997), strategic decision-making by Nazem and Shin (1999) and nancial performance by Eklund, Back, Vanharanta, and Visa (2003), Hoppszallern (2003), Derby (2003), Chang, Chang, Lin, and Kao (2003), Kloptchenko et al. (2004), Magnusson, Arppe, Eklund, and Back (2005). Also, some earlier studies of Koyuncugil and Ozgulbas (2006a, 2006b, 2006c, 2007a, 2007b, 2008a, 2008b, 2009a, 2009b), Ozgulbas and Koyuncugil (2006, 2009, 2010) conducted on nancial performance, nancial risk and operational risk of Small and Medium Enterprises (SMEs) and hospitals by data mining. Fayyad, Piatetsky-Shapiro, and Symth (1996), proposed main steps of DM:  Retrieving the data from a large database.  Selecting the relevant subset to work with.  Deciding on the appropriate sampling system, cleaning the data and dealing with missing elds and records.  Applying the appropriate transformations, dimensionality reduction, and projections.  Fitting models to the preprocessed data. Data mining techniques can yield the benets of automation on existing software and hardware platforms, and can be implemented on new systems as existing platforms are upgraded and new products developed. When data mining tools are implemented on high performance parallel processing systems, they can analyze massive databases in minutes. The most commonly used techniques in data mining are (Koyuncugil, 2006; Thearling, 2004):  Articial neural networks: Non-linear predictive models that learn through training and resemble biological neural networks in structure.  Decision trees: Tree-shaped structures that represent sets of decisions. These decisions generate rules for the classication of a dataset. Specic Decision Tree methods include Classication and Regression Trees (CART) and Chi Square Automatic Interaction Detection (CHAID).  Genetic algorithms: Optimization techniques that use process such as genetic combination, mutation, and natural selection in a design based on the concepts of evolution.  Nearest neighbor method: A technique that classies each record in a dataset based on a combination of the classes of the k record (s) most similar to it in a historical dataset. Sometimes called the k-nearest neighbor technique.  Rule induction: The extraction of useful if-then rules from data based on statistical signicance. Decision trees are tree-shaped structures that represent sets of decisions. The Decision Tree approach can generate rules for the classication of a data set. Specic Decision Tree methods include

Classication and Regression Trees (CART) and Chi Square Automatic Interaction Detection (CHAID). CART and CHAID are Decision Tree techniques used for classication of a data set. They provide a set of rules that can be applied to a new (unclassied) data set to predict which records will have a given outcome. CART typically requires less data preparation than CHAID (Lee & Siau, 2001). During the developing EWS; an easy to understand, easy to interpret and easy to apply utilitarian model that is far from the requirement of theoretical background is targeted by the discovery of the implicit relationships between the data and the identication of effect level of every factor. Because of this reason, data mining is the ideal method for nancial early warning system. Developing an EWS for SMEs focused segmentation methods. The main approach in analysis is discovering different risk levels and identifying the factors affected nancial performance. By means of Chi-Square metrics CHAID is able to separately segment the groups classied in terms of level of relations. Therefore, leaves of the tree have not binary branches but as much branches as the number of different variables in the data. So, it was deemed convenient to use CHAID algorithm method in the study. CHAID modeling is an exploratory data analysis method used to study the relationships between a dependent measure and a large series of possible predictor variables those themselves may interact. The dependent measure may be a qualitative (nominal or ordinal) one or a quantitative indicator. For qualitative variables, a series of chi-square analyses are conducted between the dependent and predictor variables. For quantitative variables, analysis of variance methods are used where intervals (splits) are determined optimally for the independent variables so as to maximize the ability to explain a dependent measure in terms of variance components (Thearling, 2004). Model of EWS The model of EWS based on data mining and data ow diagram of the EWS is shown in Fig. 1. The steps of the EWS are:  Step I: Preparation of data collection.  Step II: Implementation of DM method.

Fig. 1. Data ow diagram of the EWS.

6242

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

 Step III: Determination of risk proles.  Step IV: Identication for current situation of SMEs from risk proles and early warning signs.  Step V: Description of roadmaps for SMEs. The details of the EWS are given below: Step I: Preparation of data collection Financial data that are gained from balance sheets: Items of balance sheets will be entered as nancial data and will be used to calculate nancial indicators of system:  Calculation of nancial indicators like in Table 1.  Reduction of repeating variables in different indicators to solve the problem of collinearity/multicollinearity.  Imputation of missing data.  Solution of outlier and extreme value problem. Step II: Implementation of DM method In the scope of the methods of data mining  Logistic regression.  Discriminant analysis.  Cluster analysis.  Hierarchical cluster analysis.  Self Organizing Maps (SOM),  Classication and Regression Trees (C& RT).  Chi-Square Automatic Interaction Detector (CHAID). can be the principal methods, in addition to this several classication/segmentation methods can be mentioned. However, during the preparation of an early warning system for SMEs, one of the basic objectives is to help SME administrators and decision makers, who does not have nancial expertise, knowledge of data mining and analytic perspective, to reach easy to understand, easy to interpret, and easy to apply results about the risk condition of their enterprises. Therefore, Decision Tree algorithms that are one of the segmentation methods can be used because of their easy to understand and easy to apply visualization. Although, several Decision Tree algorithms have widespread usage today, CHAID is separated from other Decision Tree algorithms because of the number of the branches that are produced by CHAID. Other Decision Tree algorithms are branched in binary, but CHAID manifests all the different structures in data with its multi-branched characteristic. Hence, the method of Chi-Square Automatic Interaction Detector (CHAID) is used in the scope of this study. Assume that X1, X2, . . . , XN1, XN denote discrete or continuous independent (predictor) variables and Y denotes dependent vari-

able as target variable where X1 2 [a1, b1], X2 2 [a2, b2], . . . , XN 2 [aN, bN] and Y 2 {Poor, Good}. While Poor shows poor nancial performance in red1 bar and Good shows good nancial performance in green bar. CHAID Decision Tree is given in Fig. 2. In Fig. 2 we can see that,  Only 3 variables of N have a statistically signicant relationship with the target Y,  X1 has most statistically signicant relation with target Y,  X2 has statistically signicant relation with X1 where, X1 6 b11.  X3 has statistically signicant relation with X1 where b11 < X1 6 b12.

Table 1 Financial indicators. Financial variables Current ratio Quick ratio (liquidity ratio) Absolute liquidity Inventories to current assets Current liabilities to total assets Debt ratio Current liabilities to total liabilities Long term liabilities to total liabilities Equity to assets ratio Current assets turnover rate Fixed assets turnover rate Days in accounts receivables Inventories turnover rate Assets turnover rate Equity turnover rate Prot margin Return on equity Return on assets

Step III: Determination of risk proles CHAID algorithm organizes Chi-square independency test among the target variable and predictor variables, starts from branching the variable which has the strongest relationship and arranges statistically signicant variables on the branches of the tree due to the strength of the relationship. An example of a CHAID Decision Tree is seen in Fig. 2. As it is observed from Fig. 2, CHAID has multibranches, while other Decision Trees are branched in binary. Thus, all of the important relationships in data can be investigated until the subtle details. In essence, the study identies all the different risk proles. Here the term risk means the risk that is caused because of the nancial failures of enterprises. Fig. 2 shows that there are six risk proles: Prole B1 shows that There are n11 samples where X1 6 b11 and X2 6 b21% m111 has poor nancial performance,% m211 has good nancial performance. Prole B2 shows that There are n12 samples where X1 6 b11 and X2 > b21% m112 has poor nancial performance,% m212 has good nancial performance Prole C1 shows that There are n21 samples where b11 < X1 6 b12 and X3 6 b31% m121 has poor nancial performance,% m221 has good nancial performance. Prole C2 shows that There are n21 samples where b11 < X1 6 b12 and X3 > b31% m122 has poor nancial performance,% m222 has good nancial performance. Prole D shows that There are n3 samples where b12 < X1 6 b13% m13 has poor nancial performance,% m23 has good nancial performance. Prole E shows that There are n4 samples where X1 > b13% m14 has poor nancial performance,% m24 has good nancial performance. If all of the proles are investigated separately, Prole B1 shows that if any rms variables X1 and X2 have values where X1 6 b11 and X2 6 b21, poor nancial performance rate or in another words risk rate of the rm will be RB1 = m111. Prole B2 shows that if any rms variables X1 and X2 have values where X1 6 b11 and X2 > b21, poor nancial performance rate or in another words risk rate of the rm will be RB2 = m112. Prole C1 shows that if any rms variables X1 and X3 have values where b11 < X1 6 b12 and X3 6 b31 poor nancial performance rate or in another words risk rate of the rm will be RC1 = m121. Prole C2 shows that if any rms variables X1 and X3 have values where b11 < X1 6 b12 and X3 > b31 poor nancial performance rate or in another words risk rate of the rm will be RC2 = m122. Prole D shows that if any rms variable X1 have values where b12 < X1 6 b13 poor nancial performance rate or in another words risk rate of the rm will be RD = m13. Prole E shows that if any rms variable X1 have values where

1 For interpretation of color in Fig. 2, the reader is referred to the web version of this article.

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

6243

Fig. 2. CHAID Decision Tree.

X1 > b13 poor nancial performance rate or in another words risk rate of the rm will be RE = m14. Step IV: Identication for current situation of SME according to risk proles and early warning signs The part of study until this point is based on the identication of risk proles from all of the data. In the scope of the data that is about the past of SMEs, the part of the study until this point denes the relationships between nancial risk and variables, and also the risk proles. At this step, risk proles that all of the rms belong to are identied in the study. This identication is realized with taking the group of variables in the risk proles into consideration. All of the rm will look at the values of their own enterprises, in the light of the statistically signicant variables in the Decision Tree. According to Fig. 2 these variables are X1, X2 and X3. The rm compares the values of X1, X2 and X3 between decision tree and rms. Then, they can identify their risk prole. For example if any rm has X1 > b13. Therefore, the risk prole of the rm must be Prole E. According to the risk proles of SMEs, it is possible to detect the early warning signs that show highest nancial risk. Step V: Description of roadmap for SMEs According to Fig. 2 we can easily determine the risk grades of the rms. Assume that, the risk rates of the rms in the order of E > D > C2 > C1 > B2 > B1. Therefore, the best risk prole will be B1. Then, every rm tries to be in Prole B1. There are two variables X1 and X2 related with prole B1. If any rm want to be in Prole B1, the rm must make arrangements to make values X1 6 b11 and X2 6 b21. Enterprise will identify the suitable road map after dening its risk prole. The enterprise can identify the path to reach upper level risk prole and the indicators that require privileged improvement in the light of the priorities of the variables in the roadmap. Furthermore, enterprise can pass to upper level risk proles step by step at the same time can reach to a targeted risk prole in the upper levels for improving indicators due to this target. For example, any rm in Prole E has the biggest risk rate. The rm must be rehabilitating rst the variable X1 to decrease it between (b12, b13). Therefore, the rm will be in prole D and so on.

4. Application of model for risk detection and early warning signs Application of our model, early waning signs, nancial road maps and other results are presented below. Step I: Data preparation Application of our model covered SMEs in Turkey in 2007. Data of rms was obtained from Turkish Central Bank (TCB) after permission. Total number of rms had nancial data were 8.979 in TCB in 2007. Since scope of our study only covered micro, medium, and small-scaled enterprises, which are often referred to as SMEs, those 7.853 rms were classied to identify the rms, which can be categorized as a SME. We based on SME denition of the EU in an attempt to participate to Turkeys efforts to align with the EU acquits and to ensure comparability of the analysis provided herein. The thresholds used to classify SME on basis of the EUs SME denitions are 50 million. Financial data that are gained from balance sheets and income statements was used to calculate nancial indicators of system. Steps of preparation of data:  Calculation of nancial indicators like in Table 2.  Reduction of repeating variables in different indicators to solve the problem of collinearity/multicollinearity.  Imputation of missing data.  Solution of outlier and extreme value problem. Financial ratios as nancial risk indicators were calculated with variables collected from balance sheets of SMEs. These indicators and their denitions are presented in Table 2.

Step II: Implementation of data mining method (CHAID) During the preparation of an early warning system for SMEs, one of the basic objectives is to help SME managers and decision makers, who does not have nancial expertise, knowledge of data mining and analytic perspective, to reach easy to understand, easy to interpret, and easy to apply results about the risk condition of their enterprises. Therefore, Decision Tree algorithms that are one of the segmentation methods can be used because of their easy to understand and easy to apply visualization. CHAID algorithms used in this study are developed on basis of two groups of variables, namely target variable and predictor

6244 Table 2 Financial variables and denitions. Code A A1 A2 A3 A4 A5 A6 A7 A8 B B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 B11 B12 B13 B14 B15 B16 B17 C C1 C2 C3 C4 C5 C6 C7 C8 D D1 D1a D1b D1c D1d D1e D1f D2 D2a D2b D2c D2d D2e D2f D3 D3a D3b Variables

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

Denition Current assets/current liabilities (Cash, banks, marketable securities, account receivables)/current liabilities (Cash, banks, marketable securities)/current liabilities Total inventories/current assets Total inventories/total assets (Short-term liabilities (liquid assets + marketable securities))/inventories Current account receivables/total assets Short-term receivables/total assets total assets (Short-term liabilities + long-term liabilities)/total assets (leverage ratio) Own funds/total assets Own funds/(short-term liabilities + long-term liabilities) Short-term liabilities/ total liabilities Long-term liabilities/ total liabilities long-term liabilities/(long-term liabilities and own funds) Tangible xed assets/own funds Tangible xed assets/long-term liabilities Fixed assets/(short-term liabilities + long term liabilities) Fixed assets/own funds xed assets/(long term loans + own funds) Short-term liabilities/ total loans Bank loans/total assets (Short-term bank loans + principal installments and interest payments of long-term bank loans)/shortterm liabilities (Short-term bank loans + principal installments and interest T payments of long-term bank loans + longterm bank loans)/(short-term liabilities + long term liabilities) Current assets/total assets Tangible xed assets/total assets Cost of goods sold (current year)/(previous years inventory + current years inventory)/2 Net sales/(short-term trade receivables + long-term) Net sales/current asset Net sales/(current assetsshort-term liabilities) Net sales/tangible xed assets (net) Net sales/xed assets Net sales/own funds Net sales/total assets

Liquidity ratios Current ratio Quick ratio (liquidity ratio) Absolute liquidity Inventories to current assets Inventories to total assets Inventory dependency ratio Current account receivables to Total assets Short-term receivables to total assets total assets Ratios of nancial position Total loans to total assets Own funds to total assets Own funds to total loans Short-term liabilities to total liabilities Long-term liabilities to total liabilities Long-term liabilities to long-term liabilities and own funds Tangible xed assets to own funds Tangible xed assets to long-term liabilities Fixed assets to total loans Fixed assets to own funds Fixed assets to long term loans + own funds Short-term liabilities to total loans Bank loans to total assets Bank loans to short-term liabilities Bank loans to total loans Current assets to total assets Tangible xed assets to total assets Turnover ratios Inventory turnover Receivables turnover Working capital turnover Net working capital turnover Tangible xed assets turnover Fixed assets turnover Own funds turnover Total assets turnover Protability ratios Ratios relating prot to capital Net prot to own funds Prot before tax to own funds Prot before interest and tax to prot before tax + nancing expenses Net prot to total assets Operating prot to assets used in carrying out of the operations Cumulative protability ratio Ratios relating prot to sales Operating prot to net sales Gross prot to net sales Net prot to net sales Cost of goods sold to net sales Operating expenses to net sales Interest expenses to net sales Ratios relating prot to nancial obligations Prot before interest and tax to prot before tax + nancing expenses Net prot and interest expenses net prot + nancing expenses to interest expenses

Net prot (prot after tax)/own funds Prot before tax/own funds Prot before interest and tax/(prot before tax + nancing expenses) Net prot/total assets Operating prot/total assets-nancial xed assets Reserves from retained earnings/ total assets Operating prot/net sales Gross prot/net sales Net prot/net sales Cost of goods sold/net sales Operating expenses/net sales Interest expenses/net sales Prot before interest and tax/(prot before tax + nancing expenses) (Net prot and interest expenses net prot + nancing expenses)/interest expenses

variables that will explain the target variable. In this study nancial performance is explained by means of all nancial variables of a SME. Therefore, the nancial performance indicator is considered as the target variable and all nancial variables (see Table 2) are considered as the predictor variables. Fig. 3 shows the CHAID Decision Tree and all proles obtained from CHAID. Tables 3 and 4 are explained and summarized the proles in Fig. 3.

Step III: Determination of risk proles CHAID has multi-branches, and all of the important relationships in data can be investigated until the subtle details. As can be seen from Fig. 3 and its partitions Figs. 47 which explain SMEs proling and nancial performance statuses based on CHAID method, although it was possible to supercially categorize the covered SMEs into two groups as SMEs with good nancial performance and with poor nancial performance with CHAID method it was

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

6245

Fig. 3. CHAID Decision Tree and nancial proles of SMEs.

possible to categorize the covered SMEs in 31 different proles in terms of level of nancial performance. These proles show us what nancial indicators should focus on for good nancial performance as well as those proles those SMEs should take example to improve their nancial performances. As required under CHAID method SMES proling is based on prot before tax to own funds ratio (D1B), which has the strongest relation with the nancial performance (p < 0.001). SMEs with this ratio lower than and equal to 0 are grouped in the 1st prole. According to this prole, all of 1718 SMEs in this prole, or 21.88% of all covered SMEs in the study have poor nancial performance. Those SMEs with prot before tax to own funds ratio between 0 and 0.20 are grouped in 2nd to 5th proles given in Fig. 4. Also, it was determined that return on equity ratio (D1A, p < 0.001) and cumulative protability ratio (D1F, p = 0.001) affected nancial risk of SMEs in these proles. 2nd prole comprises 27 SMEs, or 0.34% of total covered SMEs with return on equity ratio lower than and equal to 0, and all those SMEs have poor nancial performance. 3rd prole with return on equity ratio between 0 and 0.02 and cumulative protability ratio lower than and equal to 0.0000002 covered total 101 SMEs. In this prole 78.22% (79 SMEs) have good nancial performance and remaining 21.78% (22 SMEs) have poor nancial performance. On the other hand, 4th prole comprises 13 SMEs with return on equity ratio higher than 0.02 and cumulative protability ratio lower than and equal to 0.0000002, and 46.15% of which have good nancial performance and 53.85% of which have poor nancial performance. Last prole with prot before tax to own funds ratio between 0 and 0.20 is 5th prole. In this prole, return on equity ratio is higher than 0 and cumulative protability ratio is higher than 0.0000002. 90.12% (447 SMEs) of SMEs have good nancial performance and remaining 9.88% (49 SMEs) have poor nancial performance in this prole.

SMEs with prot before tax to own funds ratio between 0.20 and 0.36 are grouped in 6th to 21st proles. Proles 6th to 16th given in Fig. 5 and proles 17th to 21st given in Fig. 6. Beside this ratio, return on equity ratio (D1A, p < 0.001), cumulative protability ratio (D1F, p = 0.001), short-term liabilities to total loans (B12, p = 0.0001), and total loans to total assets (B1, p = 0.0230) affected nancial risk of SMEs in 6th prole to 9th prole. 6th prole comprises 8 SMEs, or 0.1% of total covered SMEs with return on equity ratio lower than and equal to 0, and all those SMEs have poor nancial performance. 7th prole includes SMEs with return on equity ratio higher than 0, cumulative protability ratio lower than and equal to 0.0000002, short-term liabilities to total loans lower than and equal to 0.86, and total loans to total assets lower than and equal to 0.20. In this prole, 20% (1 SMEs) of SMEs have good nancial performance, and 80% (4 SMEs) of SMEs have poor nancial performance. 8th prole contains SMEs with return on equity ratio higher than 0, cumulative protability ratio lower than and equal to 0.0000002, short-term liabilities to total loans lower than and equal to 0.86, and total loans to total assets higher than 0.20. In this prole, 77.45% (285 SMEs) of SMEs have good nancial performance, and 22.55% (83 SMEs) of SMEs have poor nancial performance. 9th prole includes SMEs with return on equity ratio higher than 0, cumulative protability ratio lower than and equal to 0.0000002, and short-term liabilities to total loans higher than 0.86. In this prole, 88.57% (341 SMEs) of SMEs have good nancial performance, and 11.43% (44 SMEs) of SMEs have poor nancial performance. In 10th to 16th proles, beside prot before tax to own funds ratio (D1B, p < 0.001), return on equity ratio (D1A, p < 0.001), cumulative protability ratio (D1F, p = 0.001), interest expenses to net sales (D2F, p = 0.0011), xed assets to long term loans + own funds (B9, p = 0.0027), and long-term liabilities to total liabilities (B5, p < 0.0001) affected nancial risk of SMEs. 10th prole includes SMEs with return on equity ratio higher than 0,

6246

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

Table 3 Proles of SMEs according to CHAID Decision Trees. Proles Nodes Financial indicators D1B 1 2 3 4 5 6 7 0, 1 0, 2, 5 0, 2, 6, 11, 20 0, 2, 6, 21 0, 2, 6, 12 0, 3, 7 0, 3, 8, 13, 22, 36 0, 3, 8, 13, 22, 37 0, 3, 8, 13, 23 0, 3, 8, 14, 24 0, 3, 8, 14, 25, 38 0, 3, 8, 14, 25, 39 0, 3, 8, 14, 25, 40 0, 3, 8, 14, 26 0, 3, 8, 14, 27, 41 0, 3, 8, 14, 27, 42 0, 3, 8, 15, 28, 43 0, 3, 8, 15, 28, 44 0, 3, 8, 15, 28, 45 0, 3, 8, 15, 29, 46 0, 3, 8, 15, 29, 47 0, 4, 9, 16, 30, 48 0, 4, 9, 16, 30, 49 0, 4, 9, 16, 30, 50 0, 4, 9, 16, 31 0, 4, 9, 17, 32 0, 4, 9, 17, 33, 51 0, 4, 9, 17, 33, 52 0, 4, 10, 60 0 0.20 0 0.20 0 0.20 0 0.20 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 0.20 0.36 >0.36 D1A D1F D2F B12 B1 B9 B5 D2B B6 B13 C7 A8 D2E C2

60 0 0.02 >0.02 >0 60 >0 60.0000002 60.86 60.20 60.0000002 60.0000002 >0.0000002

>0

60.0000002

60.86

>0.20

9 10 11

>0 >0 >0

60.0000002 0.0000002 0.04 0.0000002 0.04 0.0000002 0.04 0.0000002 0.04 0.0000002 0.04 0.0000002 0.04 0.0000002 0.04 >0.04 60 0 0.000005 0 0.000005 0 0.000005 0.000005 0.06 >0.06

>0.86

60.74

12

>0

0.74 0.95 >0.95

13

>0

14 15

>0 >0

60.22

16

>0

>0.06

>0.22

17

>0

60.14

60.52

18

>0

>0.04

0.14 0.38 >0.38

60.52

19

>0

>0.04

60.52

20

>0

>0.04

60.13

>0.52

21

>0

>0.04

>0.13

>0.52

22

60.75

60.26

60.015

23

>0.36

60.75

60.26

60.015

24

>0.36

60.75

60.26

60.015

>0.03

25 26 27

>0.36 >0.36 >0.36

60.75 >0.75 >0.75

60.26 60.26 60.26

>0.015 60.03 >0.03 60.02

28

>0.36

>0.75

60.26

>0.02

29

>0.36

>0.26

60.05

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253 18 0, 4, 10, 19, 34 0, 4, 10, 19, 35

6247

30 31

>0.36 >0.36

>0.26 >0.26

60.0000006 >0.0000006

>0.05

cumulative protability ratio between 0.0000002 and 0.04, and interest expenses to net sales lower than and equal to 0. In this prole, 89.82% (247 SMEs) of SMEs have good nancial performance, and 10.18% (28 SMEs) of SMEs have poor nancial performance. 11th prole covers SMEs with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, interest expenses to net sales between 0 and 0.000005, and xed assets to long term loans + own funds lower than and equal to 0.74. In this prole, 97.12% (202 SMEs) of SMEs have good nancial performance, and 2.88% (6 SMEs) of SMEs have poor nancial performance. 12th prole contains SMEs with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, interest expenses to net sales between 0 and 0.000005, and xed assets to long term loans + own funds ratio between 0.74 and 0.95. In this prole, 76.92% (20 SMEs) of SMEs have good nancial performance, and 23.08% (6 SMEs) of SMEs have poor nancial performance. 13th prole comprises SMEs with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, interest expenses to net sales between 0 and 0.000005, and xed assets to long term loans + own funds ratio higher than 0.95. In this prole, 94.51% (86 SMEs) of SMEs have good nancial performance, and 5.49% (5 SMEs) of SMEs have poor nancial performance. 14th prole with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, and interest expenses to net sales between 0.000005 and 0.06 covers total 1441 SMEs. In this prole, 88.41% (1274 SMEs) of SMEs have good nancial performance, and 11.59% (167 SMEs) of SMEs have poor nancial performance. 15th prole includes SMEs with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, interest expenses to net sales higher than 0.06, and long-term liabilities to total liabilities lower than and equal to 0.22. In this prole, 88.49% (269 SMEs) have good nancial performance and remaining 11.51% (35 SMEs) have poor nancial performance. 16th prole contains SMEs with return on equity ratio higher than 0, cumulative protability ratio between 0.0000002 and 0.04, interest expenses to net sales higher than 0.06, and long-term liabilities to total liabilities higher than. In this prole, 70.31% (90 SMEs) have good nancial performance and remaining 29.69% (38 SMEs) have poor nancial performance. In 17th to 21st proles, prot before tax to own funds ratio (D1B, p < 0.001), return on equity ratio (D1A, p < 0.001), cumulative protability ratio (D1F, p = 0.001), long-term liabilities to total liabilities (B5, p < 0.0001), gross prot to net sales (D2B, p = 0.0332), and bank loans to total assets (B13, p < 0.0012) affected nancial risk of SMEs. 17th prole with return on equity ratio higher than 0, cumulative protability ratio higher than 0.04, long-term liabilities to total liabilities lower than and equal to 0.14, and bank loans to total assets lower than and equal to 0.52 contains total 1131 SMEs. In this prole, 94.16% (1065 SMEs) of SMEs have good nancial performance, and 5.84% (66 SMEs) of SMEs have poor nancial performance. 18th prole comprises SMEs with return on equity ratio higher than 0, cumulative protability ratio higher than 0.04, long-term liabilities to total liabilities between 0.14 and 0.38, and bank loans to total assets lower than and equal

to 0.52. In this prole, 88.26% (218 SMEs) of SMEs have good nancial performance, and 11.74% (29 SMEs) of SMEs have poor nancial performance. All of 35 SMEs in prole 19th have good nancial performance. This prole covers SMEs with return on equity ratio higher than 0, cumulative protability ratio higher than 0.04, long-term liabilities to total liabilities higher than 0.38, and bank loans to total assets lower than and equal to 0.52. 20th prole contains SMEs with return on equity ratio higher than 0, cumulative protability ratio higher than 0.04, gross prot to net sales lower than and equal to 0.13, and bank loans to total assets higher than 0.52. In this prole, 93.94% (31 SMEs) have good nancial performance and remaining 6.06% (38 SMEs) have poor nancial performance. 21st prole contains SMEs with return on equity ratio higher than 0, cumulative profitability ratio higher than 0.04, gross prot to net sales higher than 0.13, and bank loans to total assets higher than 0.52. In this prole, 64.29% (18 SMEs) have good nancial performance and remaining 35.71% (10 SMEs) have poor nancial performance. SMEs prot before tax to own funds ratio higher than 0.36 are grouped in 22nd to 31st proles given in Fig. 7. Beside this ratio, total loans to total assets (B1, p = 0.0230), inventory dependency

Table 4 Financial situation of SMEs. Proles Financial performance Good n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 Total 0 0 79 6 447 0 1 285 341 247 202 20 86 1274 269 90 1065 218 35 31 18 15 3 101 236 27 3 93 66 1 132 5391 % 0.00 0.00 78.22 46.15 90.12 0.00 20.00 77.45 88.57 89.82 97.12 76.92 94.51 88.41 88.49 70.31 94.16 88.26 100.00 93.94 64.29 100.00 75.00 100.00 91.83 100.00 33.33 80.17 67.35 25.00 85.71 68.65 Poor n 1718 27 22 7 49 8 4 83 44 28 6 6 5 167 35 38 66 29 0 2 10 0 1 0 21 0 6 23 32 3 22 2462 % 100.00 100.00 21.78 53.85 9.88 100.00 80.00 22.55 11.43 10.18 2.88 23.08 5.49 11.59 11.51 29.69 5.84 11.74 0.00 6.06 35.71 0.00 25.00 0.00 8.17 0.00 66.67 19.83 32.65 75.00 14.29 31.35 Total n 1718 27 101 13 496 8 5 368 385 275 208 26 91 1441 304 128 1131 247 35 33 28 15 4 101 257 27 9 116 98 4 154 7853 % 21.88 0.34 1.29 0.17 6.32 0.10 0.06 4.69 4.90 3.50 2.65 0.33 1.16 18.35 3.87 1.63 14.40 3.15 0.45 0.42 0.36 0.19 0.05 1.29 3.27 0.34 0.11 1.48 1.25 0.05 1.96 100.00

6248

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

ratio (B6, p < 0.0001), bank loans to total assets (B13, p < 0.0012), own funds turnover (C7, p = 0.0432), short-term receivables to total assets total assets (A8, p = 0.0121), operating expenses to net sales assets (D2E, p = 0.0149), receivables turnover (C2, p < 0.0001) affected nancial risk of SMEs in these proles. In 22nd to 25th proles, total loans to total assets ratio is lower than and equal to 0.75, and inventory dependency ratio is lower than and equal to 0.26. Beside these ratios, 22nd prole contains SMEs with bank loans to total assets ratio lower than and equal to 0.015. All of 15 SMEs have good nancial performance in this prole. 23rd prole covers SMEs with bank loans to total assets ratio lower than and equal to 0.015, and receivables turnover higher lower and equal to 0.03. In this prole, 75 % (3 SMEs) have good nancial performance and 25% (1 SMEs) have poor nancial performance. All of 101 SMEs have good nancial performance in 24th prole. This prole contains SMEs with bank loans to total assets ratio lower than and equal to 0.015, and receivables turnover higher 0.03. 25th prole contains SMEs with bank loans to total assets ratio higher than 0.015. In this prole, 91.83% (236 SMEs) have good nancial performance and 8.17% (21 SMEs) have poor nancial performance. In 26th to 28th proles, total loans to total assets ratio is higher than 0.75 and inventory dependency ratio is lower than and equal to 0.26. Beside these ratios, 26th prole contains SMEs with own funds turnover lower than and equal to 0.03. All of 27 SMEs have good nancial performance in 26th prole. 27th prole covers SMEs with own funds turnover higher than 0.03, and short-term receivables to total assets ratio lower than and equal to 0.02. In this prole, 33.33% (3 SMEs) have good nancial performance and 66.67% (6 SMEs) have poor nancial performance. Last prole in this group covers SMEs with short-term receivables to total

assets ratio higher than 0.02. In 28th prole, 80.17% (93 SMEs) have good nancial performance and 19.83% (23 SMEs) have poor nancial performance. Step IV:Identication forcurrent situation of SME from risk proles and early warning signs According to the CHAID, it is possible to indicate the nancial position or situation with risk proles that all of the rms belong to, determine the proles with highest nancial risk, identify nancial indicators that affect nancial distress of SMEs, and detect early warning signs. As you can see in Fig. 3 and Table 4, SMEs are classied in 31 different proles, according to indicators that affected their nancial performance and nancial situation. It was determined that 5391 SMEs (68.6%) out of 7853 covered SMEs had good nancial performance while 2462 of them had poor nancial performance. Results showed that 31.4% of the covered SMEs nancially distress. These distress rm are in 27 different proles except 19th, 22nd, 24th, and 26th proles depend on different nancial indicators. All of SMEs in proles 1st, 2nd, and 6th have poor nancial performance and these SMEs are exactly distressed rms. These proles contain SMEs with highest nancial risk. All of SMEs in proles 19th, 22nd, 24th, and 26th have good nancial performance and these SMEs are exactly non distressed rms. Results of the study revealed that there are 15 indicators (in total 41 ratios), which had effects on nancial performance or in other words distress of the covered SMEs. As seen in Table 5, these are prot before tax to own funds ratio (D1B, p < 0.001), return on equity ratio (D1A,

Fig. 4. CHAID Decision Tree and 2nd and 5th nancial proles of SMEs.

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

6249

Fig. 5. CHAID Decision Tree and 6th and 16th nancial proles of SMEs.

Fig. 6. CHAID Decision Tree and 17th and 21st nancial proles of SMEs.

6250

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253

Fig. 7. CHAID Decision Tree and 22nd and 31st nancial proles of SMEs.

p < 0.001), cumulative protability ratio (D1F, p = 0.001), short-term liabilities to total loans (B12, p = 0.0001), total loans to total assets (B1, p = 0.0230), interest expenses to net sales (D2F, p = 0.0011), xed assets to long term loans + own funds (B9, p = 0.0027), long-term liabilities to total liabilities (B5, p < 0.0001), gross prot to net sales (D2B, p = 0.0332), bank loans to total assets (B13, p < 0.0012), inventory dependency ratio (B6, p < 0.0001), own funds turnover (C7, p = 0.0432), short-term receivables to total assets total assets (A8, p = 0.0121), operating expenses to net sales assets (D2E, p = 0.0149), receivables turnover (C2, p < 0.0001). We determined that 15 indicators affected nancial risk and distress position of SMEs. When we consider risk proles and risk indicators together, only 2 indicators can be identied as early warning signals. These are prot before tax to own funds and return on equity (ROE). (i) If prot before tax to own funds was lower than and equal to 0. (ii) If ROE was lower than and equal to 0. (iii) If prot before tax to own funds was between 0 and 0.20, and ROE was lower than and equal to 0. (iv) If prot before tax to own funds was between 0.20 and 0.36, and ROE was lower than and equal to 0 nancial distress were indispensable for SMEs. Step V:Description of roadmap for SMEs When we determine the best risk prole, we can suggest this prole for benchmarking as a road map. Financial road maps are tools for decision making and give information as inputs in decision process. SMEs can identify the path to reach upper level risk prole and the indicators that require privileged improvement in the light of the priorities of the variables in the roadmap. Furthermore, enterprise can pass to upper level risk proles step by step at

the same time can reach to a targeted risk prole in the upper levels for improving indicators due to this target. According to our model, the best risk Proles are 19th, 22nd, 24th, and 26th that contained SMEs without risk. Then, every rm tries to be in these proles. The best proles and road maps to reach best performance are presented in Table 6.1st road map contains 19th prole and their indicators. If any SME wants to be in Prole 19, the SME must make arrangements to make values of prot before tax to own funds between 0.20 and 0.36, return on equity higher than 0, cumulative protability ratio higher than 0.04, long term liabilities to total liabilities higher than 0.38, and bank loans to total assets lower than and equal to 0.52. 2nd road map contains 22nd prole and their indicators. If any SME wants to be in Prole 22 the SME must make arrangements to make values of prot before tax to own funds higher than 0.36, total loan to total assets lower than and equal to 0.75, inventory dependency ratio lower than and equal to 0.26, and bank loan to total assets lower than and equal to 0.02.
Table 5 Financial indicators affected nancial distress. Code D1B D1A D1F B12 B1 D2F B9 B5 D2B B13 B6 C7 A8 D2E C2 Financial indicators Prot before tax to own funds Return on equity Cumulative protability ratio Short-term liabilities to total loans Total loans to total assets Interest expenses to net sales Fixed assets to long term loans + Own funds Long-term liabilities to total liabilities Gross prot to net sales Bank loans to total assets Inventory dependency ratio Own funds turnover Short-term receivables to total assets total assets Operating expenses to net sales Receivables turnover p <0.0001 <0.0001 =0.0001 =0.0001 =0.0230 =0.0011 =0.0027 <0.0001 =0.0332 =0.0012 <0.0001 =0.0432 =0.0121 =0.0149 <0.0001

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253 Table 6 Road maps. Rod maps Proles Probability of no risk (%) Financial indicators D1B Prot before tax to own funds 1 2 3 4 19 22 24 26 100 100 100 100 0.200.36 >0.36 >0.36 >0.36 D1A Return on equity >0 D1F Cumulative protability ratio >0.04 60.75 60.75 >0.75 B1 Total loans to Total assets B5 Long term liabilities to Total liabilities >0.38 60.26 60.26 60.26 B6 Inventory dependency ratio B13 Bank loans to total assets 60.52 60.02 60.02 C2 Receivables turnover C7

6251

Own funds turnover

>0.03 >0.03

3rd road map contains 24th prole and their indicators. If any SME wants to be in Prole 24 the SME must make arrangements to make values of prot before tax to own funds higher than 0.36, total loan to total assets lower than and equal to 0.75, inventory dependency ratio lower than and equal to 0.26, bank loan to total assets lower than and equal to 0.02, and receivables turnover higher than 0.03. 4th road map contains 26th prole and their indicators. If any SME wants to be in Prole 26 the SME must make arrangements to make values of prot before tax to own funds higher than 0.36, total loan to total assets higher than 0.75, inventory dependency ratio lower than and equal to 0.26, own funds turnover higher than 0.03. 5. Conclusion Financial early warning system is a technique of analysis that is used to predict the achievement condition of enterprises and to decrease the risk of nancial distress. By the application of this technique of analysis, the condition and possible risks of an enterprise can be identied with quantity. Risk management has become a vital topic for all institutions, especially for SMEs, banks, credit rating rms, and insurance companies. The nancial crisis has pushed all rms to active risk management and control nancial risks. All enterprises need EWS to warn against risks and prevent from nancial distress. But, when we consider the issues of poor business performance, insufcient information and insufciencies of managers in nance education, it is clear that EWS is vital for SMEs. Benets of an EWS can summarize as early warning before nancial distress, road maps for good credit rating, better business decision making, and greater likelihood of achieving business plan and objectives. Developing practical solutions will not only help to SMEs but also to the economies of countries. Having information about their nancial risk, monitoring this nancial risk and knowing the required roadmap for the improvement of nancial risk are very important for SMEs to take the required precautions. Data mining, that is the reection of information technologies in the area of strategically decision support, develops a system for nding solutions to the nancial administration as one of the most suitable application area for SMEs as the vital point of economy. In this study, we developed a nancial EWS based on nancial risk by using data mining. As results of the study we classied 7853 SMEs in 31 different risk proles via CHAID. Results showed that 31.4% of the covered SMEs nancially distress. All of SMEs in proles 1st, 2nd, and 6th have poor nancial performance and these SMEs are exactly distressed rms. These proles contain SMEs with highest nancial risk. All of SMEs in proles 19th, 22nd, 24th, and 26th have good nancial performance and these SMEs are exactly non distressed rms. According to these proles, we identied that prot before tax to own funds ratio, return on equity ratio, cumulative protability ratio, short-term liabilities to total loans, total loans to total assets, interest expenses to net sales, xed assets to long term loans + own

funds, long term liabilities to total liabilities, gross prot to net sales, bank loans to total assets, inventory dependency ratio, own funds turnover, short-term receivables to total assets total assets, operating expenses to net sales assets, receivables turnover affect nancial performance or in other words distress of the covered SMEs. When we consider risk proles and these 15 risk indicators together, only 2 indicators can be identied as early warning signals. Financial early warning signs for covered SMEs are prot before tax to own funds and return on equity (ROE). If prots before tax to own funds and ROE ratios are lower than and equal to 0, nancial distress is indispensable for SMEs. Beside these ndings we determine nancial road maps for risk mitigation and improve nancial performance. Financial road maps can use for decision making process as inputs. According to our study ndings, we developed 4 nancial road maps. All of 4 road maps provide risk indicators and their values for successful management and risk hedging. EWSs should develop and implement in every business, to provide information relating to the actions of individual ofcers, supervisors, and specic units or divisions. In deciding what information to include in their early warning system, business should balance the need for sufcient information for the system to be comprehensive with the need for a system that is not too cumbersome to be utilized effectively. The system should provide supervisors and managers with both statistical information and descriptive information about the function of business. In case of using our EWS model by SMEs, some of expected contributions can be summarized as:  Determine nancial performance and position of rms.  Determine nancial strategies by minimum level of nance education and information.  Financial and operational risk detection.  Roadmaps for risk mitigation.  Prevent for nancial distress.  Decrease the possibility of bankruptcy.  Decrease risk rate.  Efcient usage of nancial resources.  By efciency in resources;  Increase the competition capacity.  New potential for export.  Decrease the unemployment rate.  More taxes for government.  Adaptation to BASEL II capital accord Developing a nancial EWS based on nancial risk is not enough for to understand and manage the nancial risks that can cause insolvency and distress. Managers need also to manage operational risks that can arise from execution of a companys business functions, and strategically risks that can undermine the viability of their business models and strategies or reduce their growth prospects and damage their market value. For this reason we suggest to develop EWS that contain all kind of risk factors.

6252

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253 Koyuncugil, A. S., & Ozgulbas, N. (2007a). Developing nancial early warning system via data mining. In Proceedings Book of 4th Congress of SMEs and Productivity, Istanbul, pp. 153166. Koyuncugil, A. S., & Ozgulbas, N. (2007b). Detecting nancial early warning signs in Istanbul Stock Exchange by data mining. International Journal of Business Research VII (3), 188193. Koyuncugil, A. S., & Ozgulbas, N. (2008a). Early warning system for SMEs as a nancial risk detector. In Hakikur Rahman (Ed.), Data mining applications for empowering knowledge societies (pp. 221240), New York: Idea Group Inc. Koyuncugil, A. S., & Ozgulbas, N. (2008b). Strengths and weaknesses of SMEs listed in ISE: A CHAID Decision Tree application. Journal of Dokuz Eylul University, Faculty of Economics and Administrative Sciences, 23(1), 122. Koyuncugil, A. S., & Ozgulbas, N. (2009a). Measuring and hedging operational risk by data mining. In Proceedings Book of World Summit on Economic-Financial Crisis and International Business, Washington, pp. 16. Koyuncugil, A. S., & Ozgulbas, N. (2009b). An intelligent nancial early warning system model based on data mining for SMEs. In Proceedings of the International Conference on Future Computer and Communication, Kuala Lumpur, Malaysia. doi:10.1109/ICFCC.2009.118, pp. 662666. Koyuncugil, A. S., & Ozgulbas, N. (2010). Social Aid fraud detection system and poverty map model suggestion based on data mining for social risk mitigation. In A. S. Koyuncugil, & N. Ozgulbas (Eds.). Surveillance Technologies and Early Warning Systems: Data Mining Applications for Risk Detection (pp. 173193), New York: Idea Group Inc. Kumar, N., Krovi, R., & Rajagopalan, B. (1997). Financial decision support with hybrid genetic and neural based modeling tools. European Journal of Operational Research, 103, 339349. Kyong, J. O, Tae, Y. K, Chiho, K., & Suk, J. L. (2006). Using neural networks to tune the uctuation of daily nancial condition indicator for nancial crisis forecasting. Advances in Articial Intelligence, DOI:10.1007/11941439_65 Volume 4304/ 2006. Laitinen, K., & Chong, H. G. (1999). Early warning system for crisis in SMEs: Preliminary evidence from Finland and the UK. Journal of Small Business and Enterprise Development, 6(1), 89102. Lee, K. C., Han, I., & Kwon, Y. (1996). Hybrid neural network models for bankruptcy predictions. Decision Support Systems, 18, 6373. Lee, S. J., & Siau, K. (2001). A review of data mining techniques. Industrial Management & Data Systems, 101(1), 4146. Lee, S. H., & Urrutia, J. L. (1996). Analysis of insolvency prediction in the propertyliability insurance industry: A comparison of logit and hazard models. Journal of Risk and Insurance, 63, 121130. Magnusson, C., Arppe, A., Eklund, T., & Back, B. (2005). The language of quarterly reports as an indicator of change in the companys nancial status. Information & Management, 42, 561570. Meyer, P. A., & Pifer, H. W. (1970). Prediction of bank failures. Journal of Finance, XXV(4), 853886. Monk, E., & Wagner, B. (2006). Concepts in Enterprise Resource Planning.Second Edition. Boston: Thomson Course Technologies. Nomenclature statistique des activits conomiques dans la Communaut europenne (NACE). (2009). Retrieved August 2010, from <http:// epp.eurostat.ec.europa.eu/statistics_explained/index.php/SMEs>. Nazem, S., & Shin, B. (1999). Data mining: New arsenal for strategic decision making. Journal of Database Management, 10(1), 3942. OECD (2009). Policy responses to the economic crisis: Investing in innovation for long-term growth. Retrieved August 2010, from <http://www.oecd.org/ dataoecd/59/45/42983414.pdf>. Ozgulbas, N., & Koyuncugil, A. S. (2006). Proling and determining the strengths and weaknesses of SMEs listed in ISE by the Data Mining Decision Trees Algorithm CHAID. In 10th National Finance Symposium, Izmir. Ozgulbas N., & Koyuncugil A. S. (2009). Developing road maps for nancial decision making by CHAID Decision Tree. In Proceeding of International Conference on Information Management and Engineering (pp. 723727), IEEE Computer Society Press. Ozgulbas, N., & Koyuncugil, A. S. (2010) Financial early warning system for risk detection and prevention from nancial crisis. In Surveillance Technologies and Early Warning Systems: Data Mining Applications for Risk Detection. In A. S. Koyuncugil, & N. Ozgulbas (Eds.). Surveillance Technologies and Early Warning Systems: Data Mining Applications for Risk Detection (pp. 76108). New York: Idea Group Inc. Pantalone, C., & Platt, M. (1987). Predicting failures of savings and loan associations. AREUEA Journal, 15, 4664. Ross, S. A., Westereld, R., & Jordan, B. D. (2008). Fundamentals of corporate nance. New York: McGraw-Hill Publishing Company. Salas, V., & Saurina, J. (2002). Credit risk in two institutional regimes: Spanish commercial and savings banks. Journal of Financial Services Research, 22(3), 203224. Tafer, R. J., & Tisshaw, H. (1977). Going, going, gone-four factors which factors which predict. Accountancy, March, 5054. Tam, K. Y., & Kiang, M. Y. (1992). Managerial applications of neural networks: The case of bank failure predictions. Decision Sciences, 38, 926948. Terdpaopong, K. (2008). Financially distressed, small and medium-sized enterprises characteristics and discriminant analysis model: Evidence from the Thai market. In: Proceedings of the Small Enterprise Association of Australia & New Zealand (SEAANZ) Conference 2008, Sydney, Australia, ISBN: 978-0-64650668-5 p.125-165. Retrieved August 2010, from <http://www.seaanz.org/ documents/SEAANZ2008ConferenceProceeding_000.pdf>.

Acknowledgment This research was funded by The Scientic and Technological Research Council of Turkey (TUBITAK).

References
Abumustafa, N. I. (2006). Development of an early warning model for currency crises in emerging economies: An empirical study among Middle Eastern countries. International Journal of Management, 23(3), 403. Altman, E. (1968). Financial ratios, discriminant analysis and the prediction of corporate bankruptcy. Journal of Finance, September, 589609. Altman, E. I., Haldeman, G., & Narayanan, P. (1977). Zeta analysis: A new model to identify bankruptcy risk of corporations. Journal of Banking and Finance, June, 2954. Ambrose, J. M., & Seward, J. A. (1998). Bests ratings nancial ratios and prior probabilities in insolvency prediction. Journal of Risk and Insurance, 55, 229244. Barniv, R., & Hathorn, J. (1997). The merger or insolvency alternative in the insurance industry. Journal of Risk and Insurance, 64(1), 89113. Barniv, R., & McDonald, J. B. (1992). Identifying nancial distress in the insurance industry: A synthesis of methodological and empirical issues. Journal of Risk and Insurance, 59, 543573. Beaver, W. (1966). Financial ratios as predictors of failure. Journal of Accounting Research, 4, 71111. Berg, A., Borensztein, E., & Pattillo, C. (2004). Assessing early warning systems: How have they worked in practice? IMF Working Paper, March 2004. Retrieved April, 2009, from <http://www.ksri.org/bbs/les/research02/wp0452.pdf>. Brockett, P. L., & Cooper, W. W. (1990). Report to the State Auditor and the State Board of Insurance on Early Warning Systems to Monitor the Performance of Insurance Companies in Texas. Ofce of the State Auditor, Austin, TX. Brockett, P. L., Golden, L. L., Jang, J., & Yang, C. (2006). A comparison of neural network, statistical methods and variable. Journal of Risk and Insurance, 73(3), 397419. Chang, S., Chang, H., Lin, C., & Kao, S. (2003). The effect of organizational attributes on the adoption of data mining techniques in the nancial service industry: An empirical study in Taiwan. International Journal of Management, 20(1), 497503. Commission Recommendation of 6 May 2003 Concerning the Denition of Micro, Small and Medium-sized Enterprises (2003/361/EC), L 124/36 2003. Ofcial Journal of the European Union,Retrieved August 2010, from <http://eur-lex. europa.eu/LexUriServ/LexUriServ.do?uri=OJ:L:2003:124:0036:0041:en:PDF>. Davis, E. P., & Karim, D. (2008a). Comparing early warning systems for banking crises. Journal of Financial Stability, 4(2), 89120. Davis, E. P., & Karim, D. (2008b). Could early warning systems have helped to predict the sub-prime crisis? National Institute Economic Review, 206(1), 3547. Deakin, E. B. (1972). A discriminant analysis of predictors of business failure. Journal of Accounting Research, 10(1), 167179. Derby, B. L. (2003). Data mining for improper payments. The Journal of Government Financial Management, 52(1), 1013. Edison, H. J. (2003). Do indicators of nancial crises work? An evaluation of an early earning system. International Journal of Finance and Economics, 8(1), 1153. Eklund, T., Back, B., Vanharanta, H., & Visa, A. (2003). Using the self-organizing map as a visualization tool in nancial benchmarking. Information Visualization, 2(3), 171181. El-Shazly, A. (2003). Early warning of currency crises: An econometric analysis for Egypt. The Middle East Business and Economic Review, 18(1), 3448. Fayyad, G., Piatetsky-Shapiro, P., & Symth, P. (1996). From data mining to knowledge discovery in databases. AI Magazine, 17(3), 3754. Frawley, W., Piatetsky-Shapiro, G., & Matheus, C. (1992). Knowledge discovery in databases: An overview. AI Magazine: Fall, 213228. Hand, D., Mannila, H., & Smyth, P. (2001). Principles of data mining. Cambridge: MIT Press. Hoppszallern, S. (2003). Healthcare benchmarking. Hospitals & Health Networks, 77, 3744. Jacobs, L. J. & Kuper, G. H. (2004) Indicators of nancial crises do work! An earlywarning system for six Asian countries. CCSO Working Paper 13. Department of Economics, University of Groningen, the Netherlands. Jones, F. (1987). Current techniques in bankruptcy prediction. Journal of Accounting Literature, 6, 131164. Katz, M. (2006). Multivariable analysis: A practical guide for clinicians. New York: Churchill-Livingstone. Kloptchenko, A., Eklund, T., Karlsson, J., Back, B., Vanhatanta, H., & Visa, A. (2004). Combining data and text mining techniques for analyzing nancial reports. Intelligent Systems in Accounting Finance and Management, 12(1), 2941. Koyuncugil, A. S. (2006). Fuzzy Data Mining and its application to capital markets. Unpublished doctoral dissertation, Ankara University, Ankara. Koyuncugil, A. S., & Ozgulbas, N. (2006a). Financial proling of SMEs: An application by data mining. In Proceedings Book of the European Applied Business Research (EABR) Conference, Clute Institute for Academic Research, pp. 122. Koyuncugil, A. S., & Ozgulbas, N. (2006b). Is there a specic measure for nancial performance of SMEs? The Business Review, Cambridge, 5(2), 314319. Koyuncugil, A. S., & Ozgulbas, N. (2006c). Determination of factors affected nancial distress of SMEs listed in ISE by data mining. In Proceedings Book of 3rd Congress of SMEs and Productivity, KOSGEB and Istanbul Kultur University, Istanbul.

A.S. Koyuncugil, N. Ozgulbas / Expert Systems with Applications 39 (2012) 62386253 Thearling, K. (2004). Information about data mining and analytic technologies. Retrieved August 2010, from <http://www.thearling.com/>. Trieschmann, J. S., & Pinches, G. E. (1973). A multivariate model for predicting nancially distressed property-liability insurers. Journal of Risk and Insurance, 40, 327338. Warner, J. (1977). Bankruptcy costs: some evidence. Journal of Finance, 32, 337 347.

6253

Yang, B., Ling, X. L., Hai, J., & Jing, X. (2001). An early warning system for loan risk assessment using articial neural networks. Knowledge Based Systems, 14(56), 303306. Zavgren, C. (1985). Assessing the vulnerability to failure of American industrial rms: A logistics analysis. Journal of Accounting Research, 22, 5982. Zmijewski, M. E. (1984). Methodological issues related to the estimation of nancial distress prediction models. Journal of Accounting Research, Supplement, 5982.