Generating priority series via AHP for conducting statistical tests on CAMELS dimension priorities in evaluating bank failure risk 1

1 Introduction

The prediction of bank or business failure is considerably prominent for financial analysts, managers, investors, and other decision makers. The financial ratios are accepted as the most significant components of bankruptcy prediction. Financial health measurement of a bank can be accepted as a compulsive need. Failure/bankruptcy risk is an important problem that is interested not only by bankers but also by all stakeholders in the business world since the risk may critically endanger the affairs of the business. Accordingly, appropriate evaluation of failure risk is required for investments, market stability and fortune [37]. The financial sufficiency for enterprises, and restructuring and improving competitiveness in developing economies critically depend on the efficiency of the banking system. For the last decades, “lots of countries have experienced severe banking crisis such as Sweden, 1990s; Thailand, Malaysia, Korea, Philippines, and Indonesia, 1997; Paraguay, 1995–98; Russia, 1998; Turkey, 1994, 2000, and 2001; Argentina, 2001” [28]. The mortgage crisis, formerly has been experienced in the USA than spread out to the world, can be count as a bank crisis in 2008. In 2009, 199 banks failures were detected in the US [2 : 206]. These experiences show bank failures are crucial for state economies, and states should take precautions in preventing them.

We may mention about bank crisis in Turkey. The banking system rapidly transformed technology in the early 1990 s and came to hand international banking services. In the 1990 s, the Turkish banking sector’s return on asset ratio has been as high as five times the OECD average. IMF-backed disinflation program, which preloads with several structural changes, was launched in December 1999 [16]. After the program, Turkey has got a regulation system of banks as Turkey experienced an extensive banking crisis in 2000–2001. In Turkey, between August 2000 and Mart 2005, BDDK (Banking regulation and supervision agency of Turkey), seized or canceled licenses of 22 banks. The result of operations was accepted as a huge loss of more than 50 billion USD. The reason behind this loss is not only banking crises but also misconduct. There were 48 banks continuing their banking activity on March 24, 2005 [4]. These lose urge managers and researchers to take into account the bank failure risk for also goodwill of firms and the economy of the country.

Altman conducts first original empirical academic study on the bankruptcy of firms. Altman et al. [10] for evaluating the bankruptcy risk of firms. Many studies exist in the academic literature about firm bankruptcy prediction but not many of them focuses on banks. In the related literature about failure/bankruptcy prediction or/and assessment, discriminant analysis (DA), logit regression (LR), least square regression (LSR), neural networks (NN) and multi-criteria decision making (MCDM) techniques can be used.

Bank failure prediction cannot be determined via one criterion/factor, and there are many criteria, which have to be considered simultaneously. In the literature of bank failure/bankruptcy prediction, lots of units/alternatives and criteria exist, and then MCDM techniques can be used. So many techniques that have different advantages are promoted for MCDM problems [7]. Some of them are Analytic Hierarchy Process (AHP), ANP, ELECTRE, PROMETHEE, TOPSIS, UTADIS, grey relational analysis, etc. The literature of MCDM techniques is growing in especially about product selection, investment decisions, facility location, and facility layout planning, assessing firm performances, achievement order, etc. Gaganis et al. [6] used UTADIS, LR, and DA; Ginevičius & Podviezko [40] used COPRAS, TOPSIS and PROMETHEE for evaluating the soundness of the banks. Akhisar & Karpak [16] used AHP for evaluating bank performances. AHP is an eigenvalue approach to the pair-wise comparisons (p-w-c) that is promoted by Satty [44] for various MCDM problems, generally used for determining the criteria priorities (priority weights) via the p-w-c matrix if it is consistent [7]. In bank industry evaluation, MCDM techniques are usually used for bank rating [26] and/or bank performance evaluations[15, 41].

The motivation of this paper is showing the ability to conduct statistical tests on generated priority series which is calculated via AHP, in evaluating the CAMELS dimensions priorities of experts’ view in bank failure risk assessments. By this way, we tried to show that generated priority series via AHP method combined with consistency boundary of Dodd et al. [13] approach in order to overcome the rigid consistency boundary of Saaty, can be more commonly usable than present usage of AHP. Along with this main motivation, determining subcriteria of CAMELS dimensions, and noticing the advantages of AHP method in determining the significant criteria in comparison with some commonly used alternative statistical methods like DA, LR, LSR, are also following motivations of this paper.

The aim of our research is along with determining the priorities of subcriteria in the evaluation of bank failure/bankruptcy risk, to show generating priority series from experts’ views for each criterion for carrying out statistical tests with respect to expert subgroups, and so produce information for researchers/decision makers. To our knowledge, this study along with other studies of Pekkaya [28–34] contributes to academic literature of especially AHP, since (1) an approach of Dodd et al. [13] to tolerated consistency boundary for the p-w-c is adapted to AHP methodology for the sake of less information lose obtained from respondents, that is one of the highlights of this study, accordingly the priorities and priority series are generated from the each experts’ views of consistency controlled p-w-c, (2) the study uses relatively quite a big sample in volume, and experts generally consist of top managers who study finance and banks in particular, (3) variety of scenarios are conducted to obtain the priorities via AHP, and the results have been compared, and (4) the study shows that lots of statistical hypothesis testing on priorities can be enabled by using generated series for each criterion. Then, these advantages can take AHP method as more preferable in academic empirical researches. Moreover, with comparing to the bank failure researches, some advantages of this study can be listed as follows: (5) This study takes into account 24 subcriteria of the CAMELS’ dimension where subcriteria are pre-selected by a group of experts. The CAMELS approach, which is generally used for bank performances [12, 20], can be accepted as new for bankruptcy evaluations; as usually relatively less financial ratios are used for such kind of evaluation in related literature. (6) An improved early warning system (EWS) for bank failure risk can be promoted with using determined subcriteria priorities in this study along with the main criteria priorities of Pekkaya and Demir [28]. (7) MCDM or AHP are rarely used in determining the priorities of the dimensions/ratios since the other studies generally use DA, LR, and LSR. (8) With relation to AHP method, our data consists of experts views, ordinary studies in related literature use market data which has some advantages as representing real market conditions and some disadvantages as having difficulties in satisfying the assumptions of regression analysis. Accordingly, AHP usage is one of the originalities of this study. This study is an advanced phase of the presentation of Pekkaya and Demir [28] which reports only obtained the main dimension priorities of CAMELS from the geometric mean of p-w-cs.

To carry out the research, we used the main criteria data of Pekkaya and Demir [28] and p-w-c survey to experts to get data for the calculations of subcriteria that affect the bank failure risk. The experts consist staffs of BDDK, TCMB (Central Bank of Turkey), and academics, etc. After getting the data, AHP calculation is conducted to determine priorities/priority series for each criteria, and evaluation at differentiations in priorities with respect to experts’ characteristics are carried out via statistical hypothesis tests.

The rest of the study is organized as follows. The second part consists of a report of bankruptcy literature research, especially for banks. The third part of the paper emphasizes CAMELS approach and items of CAMELS dimensions for the bankruptcy risk factors of the banks. In the fourth part, AHP method calculation methodology, highlights of the AHP method, and the approach to AHP method with a tolerated consistency in this study is presented. The fifth part consists of an application of the criteria /subcriteria priority determination in evaluating the bank failure risk via AHP by using a p-w-c survey data of experts. In the conclusion part, the results of the study are interpreted, evaluated, and compared with the related academic literature.

2 Literature summary on bank bankruptcy/financial failure risk

Failure prediction of firms is studied for years. Back et al. [5] states that in the 1930 s, Ramser and Foster; Fitzpatrick; Winakor and Smith; and in 1942 Merwin studied the fundamentals for firm failure prediction. In the 1960 s, Beaver presented the univariate analysis approach and later on Altman pioneered the use of a multivariate approach which is accepted as a dominant study for exploring firm bankruptcy. In the 1970 s, Deakin; Edminster; Blum and Altman et al. studied multiple DA technique in related literature. In the bankruptcy literature, Altman’s usage of DA have some problems in the assumptions which must be satisfied, in order to overcome the limitations, Ohlson [21] used LR in the bankruptcy problem that can be accepted as the entrance of the new era until 1990 the times of usage NN and operation research methods.

Altman et al. [10] developed the ZETA model for evaluating the bankruptcy risk of firms using their financial ratios via DA. The financial ratios are working capital/total assets; retained earnings/total assets; earnings before interest and taxes/total assets; market value equity/total debt; sales/total assets. They compared their model performance to other studies and especially a prior study of Altman in 1968 and developed more accurate model in bankruptcy classification. Many studies exist in the academic literature about firm bankruptcy prediction but not much about banks. Literature summary for research purposes, data, approach and methods of some of the researchers are listed in Table 1 for bank failure.

Olmeda & Fernandez [19] used nine financial ratios of 66 banks to evaluate the bank bankruptcy risk, and they found NN produces best results compared to others such as DA, LR, etc. Vilen [36], used 25 ratios selected among 32 financial ratios for 124 banks to evaluate the bank failure risk and they found LR has better results compared to probit regression (PR).

Lanine and Vennet [14] studied 582 failed banks over the period 1988–2004, which is about 17% of all registered Russian banks (3393 banks) over that period. They used profitability, liquidity risk, government debt securities, capital adequacy, credit risk and loans to total assets financial ratios, and size variable to test LR and trait recognition (TR) in order to predict failures among Russian commercial banks. Dataset of this study is quite big with related to bank failure literature. They found modified nonparametric TR approach outperformed LR and the non-modified TR. Moreover, mainly liquidity and also asset quality – capital adequacy are observed as having an important role in bank failure prediction. Jin et al. [25] also with using such a big sample, predicted failed banks out of 4877 banks in financial crisis via LSR with using 13 accounting /auditing variables. The results declared that banks audited by credited auditors have less risk in failure, the recent banking crisis in the US was primarily driven by credit problems, and the loan mix, problematic loans, and growth in real estate loans increase the probability of bank failure. The last big sampled study observed on banks by us is conducted by Cinca and Nieto [8] by using 17 financial ratios of 8293 banks. They tested the partial LSR-DA method for predicting the 2008 USA baking crisis by comparing the performance of 8 algorithms commonly used in bankruptcy prediction. The suggested method has similar performance with linear DA and support vector machine (SVM), and it performs well in the existence of multicollinearity.

Kumar & Kavita [38] and Khaddafi et al. [27] used Z score model of Altman for evaluating the bankruptcy risk of 10 Indian banks and 29 Indonesian banks, respectively. These studies applied a prediction of the banks Z score and ranking/grouping them according to bankruptcy risk.

A different and quite simple approach is conducted by Santoni and Arbia [2] which shows rating agencies and LACE rating variables are quite important for predicting bank failure. They searched the rating agencies’ ability to predict the bank failure by evaluating the descriptive statistics and frequencies of 2005–2009 LACE rating of 116 out of a total 199 US bank failures of 2009. According to the analysis the rating agency has identified and informed, the banking market about the risk and default probability of US banks via downgraded to E, 94% of such banks, two quarters prior to failure (72% in the fourth quarter prior to failure).

Bellovary et al. [22] compared 14 academic studies subjected bankruptcy prediction and reported that NN models are the best, and PR is the worst satisfiers in accuracy power. Demyanyk and Hasan [47] states operation research and MCDM methods can be preferred and applicable to explain, predict, or suggest remedies for financial crises or banking defaults. Gaganis et al. [6] used six financial and four non-financial ratios of 894 banks from 79 countries to evaluate the soundness of the banks using MCDM techniques which name is UTADIS along with LR and DA. Akhisar and Karpak [16] evaluated the performance of commercial banks in Turkey via AHP as an EWS.

In the related literature about firm bankruptcy prediction or/and assessment, DA, LR, PR, NN and MCDM techniques with some selected financial ratios can be used commonly.

According to the bank failure and AHP researches in related literature, there is not an analogous methodological study has been observed yet with a variety of scenarios used in this study. Our study differs from the related literature in several ways. Meanwhile, the CAMELS approach is not used for bank failure evaluations, as this approach is used for performance evaluation of banks. MCDM/AHP are rarely used in determining the priorities of the dimensions/ratios, since the other studies mostly use DA, LR, LSR for searching the effective dimensions/ratios in a bank failure. With relation to AHP method, our data is consist of relatively a large volume expert views, ordinary studies in related literature uses market data which has some advantages (they especially represents real market conditions, not attitudes) and some disadvantages (especially lots of problems of regression analysis, especially assumptions, such as unit root existence for time series/panel data; multicollinearity problem which do not let together usage of correlated ratios in a model; however, not forgetting that many ratio may have high correlations among the same clustered firms). In some cases, some variables cannot be evaluated without any violation in case of the unit root problem, and some effective ratios may not be taken into consideration together in the analysis because of multicollinearity problem. However, no such criteria/variable evaluation problem is observed in priority calculation of AHP technique except consistency of replier’s views, and consistency powered priority calculation as AHP method lets variables which may have high correlations between each other. This opportunity is important if these criteria would be in the model.

Banks, similar to any other commercial firms, study to maximize shareholders value by generating earnings that exceed expenses. The crucial elements of bank problems are mainly dependent on loan characters, liquidity, the quality of borrowers and investments, and exogenous shocks [36]. Loans comprise an important proportion of bank assets, and particularly loan loss provisioning is crucial in determining the health of a bank. Higher loan loss provision could signal future trouble since managers have special information on loan quality and default risk. A higher capital ratio provides a bigger safety for banks to write-off bad loans in the future and banks with higher capital may bear less from debt problems and have more flexibility to respond to adverse shocks. Therefore, it can be accepted that capital ratio is related to the probability of bank failure [25]. More generally, lots of financial ratios may indicate the robustness of the banks to financial failures. For financial performance, operating soundness, rating or/and regulatory compliance assessments of the banks, CAMELS that is formed in 1997, is commonly used not only by academicians but also by banking supervisors and regulators. The CAMELS rating is based on assessing both qualitative and quantitative information of a bank. In this study, we thought CAMELS dimensions as the determinants of bank failure criteria.

CAMELS refers to the six components of a bank’s condition that are assessed: Capital adequacy (CAP), Asset quality (ASS), Management (MAN), Earnings (EAR), Liquidity (LIQ), and bank’s Sensitivity to market risk (SMR). CAMELS rating is usually accepted as a supervisory rating of the bank’s overall condition. Three federal banking supervisors (the Federal Reserve, the FDIC, and the OCC) and other financial supervisory agencies to provide a convenient summary of bank conditions use this rating system. In the public monitoring of banks, private supervisory information in CAMELS rating also appears to be useful [11 : 622, 3].

The CAP contains the level and quality of capital and the overall financial condition of the institution, the ability of management to address emerging needs for additional capital, and balance sheet composition. ASS may base upon the adequacy of underwriting standards, soundness of credit administration practices and appropriateness of risk identification practices, the level and severity of the problem, the adequacy of the allowance for loan and lease losses and other asset valuation reserves. The MAN contains level and quality of oversight and support of all institution activities by the board of management, the ability of the board of management, in their respective roles, to plan for, and respond to, risks that may arise from changing business conditions. The EAR contains the level of earnings, including trends/stability, the earning quality/sources, and the expenditure level in operations. The LIQ contains the liquidity adequacy for present/future needs and the institution ability in liquidity needs for its operations, the readily convertible assets availability to cash. The SMR contains the sensitivity of the financial institution’s earnings or the economic value of its capital to adverse changes in interest rates, adequate foreign exchange, the management ability to identify, measure, monitor [11 : 622–623, 3 : 10–12, 49 : 2–5].

In this study, subcriteria are selected based on both academic literature and pre-survey analysis which is applied to 19 experts who are academics /bank managers in the related area. Main criteria of CAMELS and their subcriteria selected for this study are presented in Table 2.

In this section, AHP method calculation methodology, and the reason of preferring the approach to AHP method with a tolerated consistency are explained. The usage of tolerated consistent AHP methodology is one of the highlights of this study.

The decision making for a firm is a selection process of one of the alternatives, which is in the context of management functions in terms of the achievement of the goal/objectives, by mental and mathematical methods. Researches have shown that it is convenient to make intuitive daily decisions, but that modern decision-making techniques must be used in complex and vital decisions. In a decision-making problem, selecting one of the most advantageous alternatives in the presence of probable contradictory criteria is being investigated. This type of decision-making process is called as the MCDM problem, and its solution can be made by the MCDM methods. MCDM methods are advantageous over intuitive or non-numerical methods because they offer the opportunity to sort, classify, and select the alternative with the ideal scorecard among the options [32, 34]. The number of MCDM methods is very and increasing, and some of their names are AHP, ANP, DEMATEL, TOPSIS, ELECTRE, entropy, PROMETHEE, GRA, etc. There is no so much commonly used alternative method used for priority calculation of criteria except AHP, entropy, but there are lots of alternative method for selecting /sorting the alternatives, in the academic literature of MCDM [33 : 297].

AHP is widely used MCDM method for determining criteria priorities and/or choosing alternatives in solving almost every MCDM problems of life which contain lots of criteria and alternatives, from site selection problem to job-spouse selection [33 : 297]. In solving /investigating the problems of stock selection [7], career choice [31], selecting organized industrial zones [33], evaluation of service quality [32, 34], bank bankruptcy risk [28], laptop selection [30], and bank performance evaluation [16, 48], etc., AHP method is commonly used in academic literature.

AHP structure of solving the problem is consisting of the main criteria and their subcriteria structure as reported in Table 2. The p-w-c and AHP calculation process are conducted for main criteria priority calculations and then separately for the subcriteria of six main dimensions of CAMELS. Totally, seven separate groups of AHP calculation process is conducted.

The calculation process of AHP, developed by Saaty [43, 44] in the 1970 s, can be summarized in 5 steps for n criteria as follows, for determining the criteria priorities [7, 30].

Step 1: Constructing a p-w-c matrix ( A  = [aij]), by obtained p-w-c scores of expert/respondent views. The scale of p-w-c is double-sided 9-point interval scale. The scale has 17 points, numerically 1/9, 1/8,  ...  , 1/2, 1, 2,..., 8, 9 values. Matrix A is an nxn square matrix, diagonal values are 1, and each cell of the matrix show dominance severity of column criteria on row criteria. The scale values of p-w-c for AHP indicate how many times more important/dominant one criterion is over another criterion.

Step 2: Constructing matrix B, by normalizing matrix A. The normalizing process is carried out by dividing each value of matrix A to its column sum. b ij = a ij ∑ i = 1 n a ij(1)

Step 3: The priority weight matrix (W) contains criteria priorities, and its elements are calculated by getting arithmetic mean of each row values of matrix B which is nxn matrix. W matrix is an nx1 column matrix; each priority is settled in the rows, respectively. w i = ∑ j = 1 n b ij n(2)

Step 4: The λ value is calculated by dividing the sum of ci/wi values to the number of criteria n. At this stage, vector multiplication of A and W calculate ci values of column matrix C. λ = ∑ c i w i n where C = A * W(3)

Step 5: The consistency ratio (CR) is calculated by dividing the consistency index (CI) to random index (RI) as in the equation (4). CR is a consistency value of all p-w-c scores of matrix A, attributed by each expert/respondent. CR = CI RI(4) where CI values can be calculated by (λ-n)/(n-1). And, RI values in the equation (4) can be obtained according to counterparts’ values of criteria number n in RI table (Table 3).

AHP is an eigenvalue approach to the p-w-c, and it uses the numeric scale in order to measure the quantitative as well as qualitative performances. This priority calculation process via AHP is a kind of obtaining the weight /priority vector W by solving the equation of “A*W=λ max*W”, where λ max is the largest eigenvector of matrix A which contains p-w-c scores [7].

One of the important advantages of AHP is presenting p-w-c cross consistencies for each p-w-c as a sole number. In this way, consistency of the data can be obtained an easy/objectively and commonly accepted numerical method. Secondly, the AHP method can be thought as a converter of the p-w-c judgments to priorities for each criteria. In other words, it converts ordinal/interval scaled expert views to ratio scaled weights which can be individually interpreted or can be used as a device for calculation process of sorting /ranking the alternatives. Meanwhile, p-w-c scale of AHP has a higher sensitivity in comparing measurements than Likert scale, and AHP lets an objectively easy consistency calculation of all p-w-c of each person by generating an index [30, 32]. These properties put forward the AHP method as an essential technique for priority calculations.

AHP and DEMATEL methods are used in so many studies among the priority calculation techniques with getting views from experts. Entropy is also one of the common MCDM for calculating priorities, but it uses market values and its variances, not expert p-w-c views. Since this study uses expert views, entropy is not useful for the purpose of this study. According to the results of Pekkaya & Aslan [33 : 308], AHP method is accepted as appropriate in determining the criteria priorities, as DEMATEL method is for investigating the interaction and interrelations between the criteria. Accordingly, AHP method along with its advantages is decided as an appropriate method in priority calculations of this study. In comparing the AHP method with the other ordinary statistical methods such as DA, LR, LSR; these methods are used for investigating a model which variable/criteria are effective on bankruptcy risk of a firm in the related academic literature by using the real market data. Meanwhile, AHP method is used for determining the priorities of the criteria, which are accepted as effective on bank failure risk evaluations according to related literature and expert views.

In this research, an approach of Dodd et al. [13] to consistency boundary for the p-w-c is adapted to AHP methodology. The approach of Dodd et al. is accepted for the sake of the possibility of using much more views in the priority calculations, less loss of information, and in series much more expert views’ priorities which let us to conduct more respectable statistical hypothesis testing especially with respect to characteristic properties of experts [28–34]. Accordingly, consistency tolerated priorities are taken into account in the analyses, and statistical hypothesis tests are also conducted.

The priority series are acquired from consistency tolerated experts’ views using Dodd et al. approach with consistency boundary of.4113 (=0.50996/1.24) for six criteria (Table 4). The value of 1.24 stands for the RI, declared by Saaty. The value of 0.50996 stands for a tolerance value of CI at the 95% confidence level for six criteria. Dodd et al. explained it as “we could then be 95% confident that a given matrix which achieves this has done so by virtue of consistent judgment on the part of the decision maker rather than by chance”. Therefore, p-w-c of experts are scored consciously /not by chance, and those p-w-c can be accepted as containing information. As the Saaty’s consistency boundary (SCB) for CR is 0.100, the Dodd et al.’s consistency boundary (DCB) for CR is not rigid, their values with respect to the number of criteria (n) are reported at Table 4. The adaptation of tolerated consistent AHP methodology is one of the contributions of this study.

The aim of this analysis is to determine the priorities of CAMELS dimension subcriteria in assessing the bank failure risk via AHP. Moreover, the analysis of this study shows statistical tests can be conducted with respect to expert groups in terms of gender, experience, etc. by generating the priority series from each experts’ views. We used the main criteria data of Pekkaya and Demir [28] and p-w-c survey to experts to get data for the calculations of subcriteria that affect the bank failure risk. The experts consist of staffs of BDDK, TCMB, and academics, etc. The sample is selected among volunteer experts; then the results can be accepted as the views of our expert sample.

The selection of subcriteria of CAMELS is carried out in three steps. First, Likert scaled pre-survey questions are prepared by taking into account commonly used subcriteria items of CAMELS dimensions in academic literature. Second, the pre-survey is conducted to assess the subcriteria of CAMELS dimensions by obtaining the views of 19 experts who are academicians/bank managers on this topic. Third, along with the pre-survey results, co-authors and two other academics selected the subcriteria which are presented in Table 2. In order to obtain better consistencies of p-w-c, and pure assessment of determining the priorities, using fewer subcriteria is desired. On the other side, we avoid to drop out any important subcriteria during selection for the analyses. 24 (=4 + 5 + 4 + 3 + 4 + 4) subcriteria are selected from 32 (=7 + 8 + 3 + 4 + 5 + 5) nominee plus possible suggested subcriteria which are asked by open-ended questions for each dimension.

After selecting the subcriteria, p-w-c survey for main/subcriteria is conducted on 108 experts working at BDDK, TCMB, and top managers of some banks, academics who study finance /banks in Turkey. Priorities and consistencies of criteria of each expert are calculated via AHP. In academic literature of AHP, ordinarily very little sample in volume (usually less than ten experts’ views) can be used for such calculations, the volume of expert sample may not report or/and there may be almost no consistency calculation conducted/reported for each unit, but only one consistency calculation conducted for joined views. The joined views are obtained by calculating the geometric mean of experts’ views for each p-w-c. We think that, since inconsistent p-w-c violate the whole data; inconsistent p-w-c must not take into account in the joined views of experts.

However, the SCB which is 0.100, is strict in lack of possibility of getting re-p-w-c for surveys of such a big sample. For this study, the experts have very restricted time for a survey and having almost no possibility to re-survey them in the conditions of the inconsistency of p-w-c. Numerically for main criteria, p-w-c of 16 experts among 108 can be accepted as consistent which consistency boundary is 0.100 according to SCB [44]. However, according to DCB [13], p-w-c of 81 experts are not randomly scored ones which consistency boundary is 0.4113 for six criteria, and then their p-w-c can be accepted as having information. In this study, the approach of DCB is accepted, so the priorities calculated from the views of 81 experts (some characteristic properties are reported in Appendix 1) are accepted as consistency tolerated priorities. Accordingly, consistency tolerated priorities are taken into account in the analyses, and statistical hypothesis tests are also conducted on this sample.

In Table 5, calculated main priorities with respect to experts’ views are presented. Priorities-DM, which stands for priorities picked up by calculating the arithmetic mean of generated priority series for each criteria. The priority series are created from consistencytolerated priorities (DCB) of expert views. So, Priorities-SM is calculated with respect to SCB [44] by using 16 experts’ views, the methodology is the same as to Priorities-DM, but the sample.

Prior(ities)-DG is calculated via only one p-w-c matrix which is computed by getting a geometric mean of p-w-c scores of experts. In related literature of AHP, using whole sample views to incorporate a group consensus (or common sense) for priorities are preferred, and some of the studies do not reveal whether each experts’ consistency is taken into account. However, priorities of whole sample values do not take into account in this study, since containing inconsistent p-w-c may violate the data. In order to represent the higher sample volume, DCB is accepted. So, Priorities-DM and Prior-DG results are explained/analyzed. Priorities-DM results are used especially for statistical hypothesis testing since they have priority series for each criteria/individual, and Prior-DG represents the common sense of experts’ views. Since consistencies are become weaker in subcriteria, for representing bigger sample view, p-w-c’s scores which have a sample of 81 are used for calculating the priorities of subcriteria.

The differences among priorities of main criteria are statistically significant at the 0.05 level, with respect to one way repeated measures ANOVA test. In other words, all the main criteria can be rejected as equally important in evaluating the bank failure risk (at least two of them are different). Accordingly, LIQ (24.76%) can be evaluated as the most important main criteria, ASS and CAP are also significantly important. These three main criteria have a total priority of 66.54%. In the assessment of homogeneities of main criteria priorities, MAN series has the biggest heterogeneity, followed by the SMR series, but ASS and LIQ series can be accepted as more homogenous than others. In other words, we can state that views about the priorities of MAN and SMR changing in a wider range than others from person to person, but views about the priorities of ASS and LIQ are more rigid.

CAMELS dimension priorities, mentioned only for bank performance evaluation in related literature, are reported in Table 6 since the similar study is not encountered for bank failure risk except Pekkaya and Demir [28]. These studies subject to evaluate bank performances, but not evaluating bank failure risk, moreover they use these priorities as a tool for their analyses calculations. As it is observed at Table 6, CAP, ASS, MAN and SMR priorities of bank performance studies can be accepted quite close to the Prior-DG [28], and according to t-test or non-parametric test of Wilcoxon signed ranks except EAR and LIQ. Since the observed studies which have got priorities for CAMELS dimensions are so few, in case of non-normality nonparametric test is conducted for testing the existing differences are statistically significant at 0,05 or not. At 0.05 level, EAR and LIQ priorities (and at 0.10 level MAN priorities) of bank performance studies can be accepted significantly different from the reference Prior-DG. Meanwhile, we accept that the priority differentiates in bank failure risk and performance assessments strongly for EAR-LIQ, weakly for MAN dimensions. Although the majority of differences in priorities have a big gap, small sample/heterogeneity in literature not let us get evidence in significant differences. To sum up, it can be expressed that, priorities of EAR, MAN and LIQ dimensions of CAMELS can be differently evaluated for bank failure risk and performance assessments. Then it can be stated that as LIQ priority is more important in the evaluation of bank failure risk than bank performance assessments, but EAR priority is more important in the bank performance evaluation than the bank failure risk.

The difference in priorities of CAP, MAN and SMR are statistically significant at the 0.05 level, with respect to the gender of the experts, according to independent samples t-test /Mann-Whitney U Test. As male experts think MAN priority about 95.40% times more important than females, female experts think CAP priority about 32.68% and SMR about 60.14% more important than males (Appendix 2). Other differences with respect to gender are not statistically significant at the 0.05 level, by this way Priorities-DM of Table 5 is valid for the other main criteria priorities.

Other results according to characteristic properties of experts can be listed as follows. Differences in the priorities with respect to sub-categories of the experts’ experience are not statistically significant at the 0.05 level, but LIQ and SMR are statistically significant at the weak level of 0.10. As SMR has more importance, according to relatively less experienced experts, LIQ has more importance, relatively more experienced experts (Appendix 3). MS/PhD experts pay more attention to LIQ than those with a graduate/bachelor degree; it is statistically significant at the weak level of 0.10 (Appendix 4). There are no statistically significant differences in priorities with respect to sub-categories of the experts’ vocational status even in weak level (Appendix 5). So, education level and vocation situation can be accepted as not producing a significant difference in priority determinations. Based on age and experience of experts, some scenarios up to alternative sub-categories are conducted for searching differences in views, but no statistically significant difference has been detected, so such results are not reported in this research.

CAMELS’ subcriteria priorities are calculated according to 3 scenarios (Appendix 6). DG81 line shows the subcriteria priorities, calculated via geometric mean of p-w-c of tolerated consistent experts’ views (DCB) where Priorities DA of p-w-c in main criteria, for the sake of more views of experts. The consistencies of subcriteria p-w-c are worse than the main criteria p-w-c, since evaluated views are quite less according to DCB, maybe because of smaller DCB value for fewer criteria. Therefore, subpriority values of DG81 scenario is accepted since (1) it contains bigger sample common views with consistency at most 0.027 which is accepted also for SCB, as conducted some of the studies in AHP literature, (2) the priority results are not changing so much according to other scenarios of DGs and DAs, and (3) the p-w-c of the experts of this sample for main criteria is acceptable for DCB, that they proved their conscious p-w-c evaluation. The subcriteria priorities of other scenarios are reported for inter-comparison in Appendix 6.

Table 7 presents the results of the priorities of main and subcriteria of DG81 scenario. Subcriteria of lA/sL (12.89%), E/CMO (10.97%) and dC/tC (9.97%  + 6.49%) have the highest private priorities among 24 criteria in assessing the failure risk of the banks. These three subcriteria totally have 40.33% importance, which is more than one-third of this evaluation. Thus, decision-makers should take into considerate especially the criteria of LIQ with its subcriteria of lA/sL, ASS (and also MAN) with its subcriteria of dC/tC and CAP with its subcriteria of E/CMO for assessing the bank failure risk.

Evaluating the banks’ failure risk is very prominent for banks, firms, and the economy of the country. Thus, along with determining the priorities of subcriteria in the evaluation of bank failure/bankruptcy risk, generating priority series from experts’ views for each criterion for carrying out statistical tests with respect to expert subgroups are conducted. AHP technique is used for determining criteria priorities by using the tolerated consistent views of 81 experts. This study is original for AHP literature, especially because of the DCB approach adaption to AHP methodology for the p-w-c, and resulting with less information lose obtained from respondents, accordingly the priorities and priority series are generated from each experts’ views of consistency controlled p-w-c. And then, this study gives an opportunity of considering each consistent experts’ views, using such a big sample mostly comprising top manager-experts in Turkey, and also by showing that lots of statistical hypothesis testing on the priorities can be carried out via generated series for each criteria from separately computed experts’ consistent priorities. To sum up, this study may pave the way of DCB adapted usage of AHP technique, as popular as Likert scale, since (1) each respondent tolerated consistencies can be calculated by taking into account each scoring, (2) via AHP calculation for each criteria, we can obtain ratio scaled scores which can also be used for much more analysis by using interval scaled survey data, and (3) statistical tests can be conducted (not observed in AHP literature except Pekkaya [28–34]) on bigger generated priority series with respect to sample sub-groupcharacteristics. Obtaining/comparing the bank failure risk priorities using CAMELS dimensions may give an opportunity to decision makers/researchers in the bank assessment procedure since no similar study has encountered which concentrates on such subject. Accordingly, an improved “Early Warning System” for bank failure risk can be promoted by taking into consideration of determined criteria priorities in this study. Since all the CAMELS dimensions are not used in the evaluation of bank failure risk, our results are not compatibly compared with the related literature. However, the used dimensions of the LACE approach contains four CAMELS dimensions (LIQ, ASS, CAP, EAR), are used by rating agencies, and the supportive findings of LACE bank failure prediction is presented by the study of Santoni and Arbia [2], are overlapped with the three main CAMELS dimensions with determined priority of 66.54% (LIQ, ASS, and CAP) totally, except the EAR dimension having the priority of only 7.89% . The priority findings may have a new proposal for “Early Warning System” and/or “bank rating agencies” with calculating the banks’ indexes for bank failure prediction with the obtained priorities which may produce better predictions than LACE approach.

Homogeneities in priorities show that the views about the priorities of MAN and SMR may change more from person to person, but views on priorities of ASS and LIQ are more rigid and like-minded which means experts accept these priorities score in general than other dimensions. As male experts pay more attention to MAN than females, female experts pay more attention to CAP and SMR than males. Among 24 subcriteria, lA/sL, E/CMO, and dC/tC individually have the 40.33% importance which is more than one-third of overall priorities.

Academic literature for the CAMELS approach concentrate on only bank performance evaluation, so the priority results of this study may get differentiating priorities which can be discussed. Those studies aimed to evaluate bank performances, not evaluating bank failure risk, and mainly use the discussed priorities as a tool for their research calculations. LIQ priority is observed more important in the evaluation of bank failure risk than bank performance assessments, but EAR priority is more important in the evaluation of bank performance than the bank failure risk.

Overall, in evaluating the financial failure risk of the banks, decision makers, financial managers, analysts, investors and other users of financial statements, must pay most regard to LIQ with its subcriteria of “Liquid assets/Short term liabilities”. Then, ASS with its subcriteria of “Delinquent credits/Total credits +accounts receivable” and CAP with its subcriteria of “Equity/(Credits+Market+Operational risk based amount)” are also very important criteria (totally % 40.33) for evaluating the financial health of a bank, for the investments along with the business world and market stability.

Since the data is not obtained via random sampling methods; the determined priorities are representing the views of experts in the sample. The calculated priorities are the experts’ views, not market data and comparison cannot be conducted since no similar study is observed which uses the CAMELS priorities for bank failure risk evaluation. For the following studies, other priority calculation approaches can be traced such as hesitant fuzzy preference relations for especially improving consistencies [50], randomly selected bigger expert sample than this one can be used for obtaining views in more representative priority calculations, and bank failure risks evaluations can be conducted by using the determined priorities to test the appropriateness of the priority values and getting information about the banks status.