Coventry University Repository for the Virtual Environment (CURVE) Author names: Gaganis, C., Pasiouras, F., Tanna, S. and Zopounidis, C.

Transcription

1 Coventry Unversty Coventry Unversty Repostory for the Vrtual Envronment (CURVE) Author names: Gagans, C., Pasouras, F., Tanna, S. and Zopounds, C. Ttle: Bnary choce models for external audtors decsons n Asan banks. Artcle & verson: Post-prnt verson Orgnal ctaton: Gagans, C., Pasouras, F., Tanna, S. and Zopounds, C. (2008) Bnary choce models for external audtors decsons n Asan banks. Operatonal Research, volume 8 (2): Publsher statement: The fnal publcaton s avalable at Copyrght and Moral Rghts are retaned by the author(s) and/ or other copyrght owners. A copy can be downloaded for personal non-commercal research or study, wthout pror permsson or charge. Ths tem cannot be reproduced or quoted extensvely from wthout frst obtanng permsson n wrtng from the copyrght holder(s). The content must not be changed n any way or sold commercally n any format or medum wthout the formal permsson of the copyrght holders. Ths document s the author s fnal manuscrpt verson of the journal artcle, ncorporatng any revsons agreed durng the peer-revew process. Some dfferences between the publshed verson and ths verson may reman and you are advsed to consult the publshed verson f you wsh to cte from t. Avalable n the CURVE Research Collecton: August

2 Bnary choce models for external audtors decsons n Asan banks Chrysovalants Gagans 1, Fotos Pasouras 2, Salesh Tanna 3, Constantn Zopounds 1 1 Fnancal Engneerng Laboratory, Department of Producton Engneerng and Management, Techncal Unversty of Crete, Unversty Campus, Chana, P.O 73100, Greece 2 School of Management, Unversty of Bath, BA2 7AY, Bath, UK 3 Department of Economcs, Fnance and Accountng, Faculty of Busness, Envronment and Socety, Coventry Unversty, Prory Street, CV1 5FB, Coventry UK Abstract The present study nvestgates the effcency of four classfcaton technques, namely dscrmnant analyss, logt analyss, UTADIS multcrtera decson ad, and nearest neghbours, n the development of classfcaton models that could assst audtors durng the examnaton of Asan commercal banks. To develop the audtng models and examne ther classfcaton ablty, the dataset s splt nto two dstnct samples. The tranng sample conssts of 1,701 unqualfed fnancal statements and 146 ones that receved a qualfed opnon over the perod The models are tested n a holdout sample of 527 unqualfed fnancal statements and 52 ones that receved a qualfed opnon over the perod The results show that the developed audtng models can dscrmnate between fnancal statements that should receve qualfed opnons from the ones that should receve unqualfed opnons wth a satsfactory accuracy. Both fnancal varables and the envronment n whch banks operate appear to be mportant. The hghest classfcaton accuracy s acheved by UTADIS, followed by logt analyss, nearest neghbours and dscrmnant analyss. Keywords: Audtng, Bank, Classfcaton Author for correspondence: Prof. Constantn Zopounds, Tel: , E-mal: kostas@dpem.tuc.gr 1

3 1. Introducton The recent crses n Asa and Latn Amerca have shown that wthout effectve regulaton and supervson by the central banks and fnancal authortes the bankng system mght experence serous problems, wth adverse consequences for the economy as a whole. As a result, the regulatory and supervsory framework n many countres s currently experencng sgnfcant changes. In a recent report, focusng on the mplcatons of Basel II on Asan banks, Ftch (2005) ponts out that a number of Asan supervsors have expressed a clear ntent to adopt Basel II as a key part of ther bank supervsory regme. However, mplementng Basel II and supervsng Internal Ratngs Based (IRB) banks wll present challenges for Asan supervsors as more techncal sklls and resources wll be requred (Ftch, 2005). Another challenge wll be to mprove the qualty of accountng and audtng standards. Obvously, accurate fnancal statements are necessary to all the stakeholders that want to assess the fnancal condton of banks. Hence, audtng standards are crucal, whatever reforms n the regulatory framework may be ntroduced, snce wth poor accountng, and audtng requrements, the qualty and dsclosure of fnancal statements can be out-of date or unrelable. However, Asa has been noted for an nadequate reportng, accountng and audtng framework that can partly explan why there was a lack of awareness among market partcpants and regulators [Shra (2001)]. Therefore, much work needs to be done to ensure the effectve and meanngful dsclosure of fnancal nformaton through mproved accountng and audtng standards [Parrenas (2002)]. As Gunter and Moore (2003) pont out, the ncreasng emphass on the relablty of fnancal reportng has underlned the need for better understandng of the factors contrbutng to accountng naccuraces and the methods by whch such msstatements mght be detected and corrected. However, despte ts mportance, research focusng on banks audtng s lmted compared to non-fnancal enterprses, wth a few studes manly focusng on the U.S and examnng ssues such as the prcng of audt servces for fnancal nsttutons [Sten et al. (1994), Felds et al. (2004)], the loss underreportng and the audtng role of bank exams [Berger et al. (1991), Gunther and Moore (2003a)], and the effectveness of bank audt [Sddqu and Podder (2002)]. The objectve of the present study s to contrbute towards the detecton of banks fnancal statements that should receve qualfed audt opnons. To accomplsh ths task, we employ four classfcaton technques, namely dscrmnant analyss, logt analyss, Utltes Addtves Dscrmnantes (UTADIS), and nearest neghbours, for the development of audtng models that could assst audtors durng the examnaton of Asan banks. Most of the prevous studes on the feld have developed models for non-fnancal frms. Examples of such studes are: Mutchler (1985), Levtan and 2

4 Knoblett (1985), Dopouch et al. (1987), Keasey et al. (1988), Spaths et al. (2002, 2003), Fannng et al. (1995). However, these studes generally exclude banks and other fnancal frms due to ther specfc characterstcs, dfferences n the envronment n whch banks operate, as well as dfferences n the fnancal statements of banks whch make many of the emprcal proxes used n these studes napproprate for the bankng sector. Furthermore, to the best of our knowledge no study nvestgates the development of such models for Asan banks. Snce the detaled audt of all transactons of a bank would not only be tmeconsumng and expensve but also mpractcable [Bank for Internatonal Settlements- BIS (2002)], through the employment of such models audtors can save tme and money. For example, classfcaton models can provde the bass for a decson tool for audtors when predctng what opnon other audtors would ssue n smlar crcumstances, when evaluatng potental clents, n determnng the scope of an audt for exstng clents, n peer revews, to control qualty wthn frms and as a defense n law suts [Latnen and Latnen (1998)] as well as to avod dffcultes n analyzng large quanttes of data. The analyss s also mportant to partes other than audtors, such as the managers of the bank, and the fnancal regulators. The later, rely heavly on the work of external audtors as they make ther evaluatons of banks fnancal condton [Felds et al. (2004)]. As Fernandez and Gonzalez (2005) pont out better accountng and audtng systems that provde the regulator wth more nformaton about the real rsk of bank assets can not only ncrease the effectveness of mnmum captal requrements but also serve to gude dscplnary acton mposed by supervsors on bank management n order to reduce nstablty. Ther emprcal results ndcate that accountng and audtng systems can be effectve devces to counteract tendences for frm rsk-takng assocated wth bank safety nets. In addton, such systems appear as complements for mnmum captal requrements, and substtutes for restrctons on bank actvtes and offcal dscplne. Credt agences mght also have an nterest n such models as they usually take nto account audtor s opnon durng the ratng of banks. The remander of the paper s as follows: Secton 2 descrbes the methodology (sample, varables, classfcaton technques) of the study. Secton 3 presents the emprcal results of the analyss, whle the last secton dscusses the concludng remarks. 2. Methodology 2.1 Sample The sample of ths study conssts of an unbalanced panel dataset of 2,426 fnancal statements from 258 commercal banks operatng n the man South and Southeastern 3

5 Asan countres of Chna, Hong Kong, Inda, Korea, Malaysa, Sngapore, Tawan, Thaland, over the perod These banks were ncluded n the sample after fulfllng the requrement of data avalablty n terms of fnancal statements and audtor s opnon (.e. qualfed or unqualfed) n Bankscope database of Bureau van Djk s company. We should menton at ths pont that the audtor consders how the fnancal statements mght be materally msstated and consders whether fraud rsk factors are present that ndcate the possblty of fraudulent fnancal reportng or msappropraton of assets [BIS (2002)]. At the end of the examnaton, the audtor must prepare a report that contans a clear expresson of opnon on the fnancal statements. An unqualfed opnon means that the audtor does not dsagree wth the fnancal statements presented by the management mplyng that they meet at least the mnmum acceptable standards of presentaton. When the audtor dscovers a msstatement materal to the fnancal statements taken as a whole he/she asks management to adjust the fnancal statements. If the management refuses to make the adjustment, the audtor ssues a qualfed opnon on the fnancal statements. Alternatvely, a qualfed opnon can be expressed n cases that management has not provded the audtor wth all the nformaton or explanatons he/she requres (BIS, 2002). In our case, some of the banks receved qualfed opnons for more than one year, whle others were not ncluded n the sample for all years, due to mssng nformaton (n terms of fnancal data and/or audtor s opnon). Furthermore, for some banks both consoldated and unconsoldated fnancal statements were avalable and consdered. Consequently, the fnal sample conssts of an unbalanced dataset of 198 qualfed fnancal statements and 2,228 unqualfed ones. The geographcal coverage s as follows: Chna (289), Hong Kong (365), Inda (488), Korea (252), Malaysa (397), Sngapore (152), Tawan (339), Thaland (144). To ensure the proper evaluaton and comparson of the four audtng models we splt the sample nto two dstnct sub-samples, one used for tranng and one used for testng. The former one conssts of 1,701 unqualfed fnancal statements and 146 ones that receved a qualfed opnon over the perod , and s used to develop the models. The later one, conssts of 527 unqualfed fnancal statements and 52 ones that receved a qualfed audt opnon between 2002 and 2004, and s used to valdate the models Varables selecton Felds et al. (2004) pont out that snce the managers of banks are eventually answerable to ther regulators t seems reasonable to assume that the audt functon should be drven by fnancal varables and ratos that these regulators consder mportant. Although there are many ratos that can be employed to assess the fnancal condton of banks, these are n general classfed under the man categores 4

6 of Captal strength, Asset qualty, Management, Earnngs and Lqudty, known as the CAMEL model that s beng used by U.S. regulators snce the 1970s. In a smlar manner, the Internatonal Audtng Practces Commttee (2000) mentons that: The audtor consders the ratos obtaned by one bank n the context of smlar ratos acheved by other banks for whch the audtor has, or may obtan suffcent nformaton. These ratos generally fall nto the followng categores: Asset qualty, Lqudty, Earnngs, and Captal Adequacy. Potental reasons for whch varables from these categores can have an mpact on the audt decson are dscussed below. Startng wth asset qualty, a low qualty portfolo may have a negatve mpact on bank proftablty, by reducng nterest ncome and by ncreasng the provsonng costs, thus decreasng net profts. As a result, banks may set provsons outsde the range commensurate wth ther credt qualty. Thus, they can reduce the varablty of reported ncome by makng hgher provsons than necessary when credt qualty and net ncome are hgh and keepng provsons low once credt qualty deterorates. Obvously, ths approach makes the fnancal condton of a bank less transparent to shareholders, nvestors and authortes [Gunther and Moore (2000)] and t s expected that nadequate provsons would ncrease the lkelhood of dsagreements and the ssuance of a qualfed audt opnon. The nfluence of lqudty on audtor s decson s not so clear. For nstance, Ireland (2003) mentons that hgh lqudty may ncrease dsagreement type modfcatons because assets may have been overstated. In contrast, Spaths (2003) ponts out that the possblty of a qualfed audt report s hgher when the fnancal health of a company deterorates (.e. low lqudty). Numerous studes that examne non-fnancal sectors ndcate that the frms whch receve qualfed opnons are the less proftable ones [Loebbecke et al. (1989), Summers and Sweeney (1998), Beasley et al. (1999), Spaths (2002, 2003), Spaths et al. (2003)]. As Spaths (2002) ponts out the proftablty orentaton s tempered by manager s own utlty maxmzaton defned (partally) by job securty (p. 185). Wth respect to captal, the captal adequacy requrements mposed by the 1988 Accord (Basel I) as well as the new captal framework (Basel II) requre from banks to hold captal on the bass of ther assets rsk. The reason s that captal serves as the last lne of defense aganst the rsk of bank falure snce any losses a bank suffers could be fnally wrtten off aganst captal. Therefore, an adequate supply would seem to obvate the need for more specfc controls over rsk [Goln (2001)]. However, banks are hghly geared enterprses that do not usually mantan much captal relatve to ther labltes, unless constraned by regulatons. Consequently, bank management may manpulate fnancal statements, gven the need to meet certan requrements. Obvously, the later could ncrease the lkelhood of dsagreements and the ssuance of a qualfed audt opnon. Furthermore, poor captal strength could ncrease 5

7 qualfed opnons as fnancally dstressed banks are more lkely to attempt to overstate ther fnancal poston. In the present study, on the bass of data avalablty we ntally consder a set of 18 ratos, from the ones that are pre-calculated n Bankscope, and cover the abovementoned categores. We also consder the natural logarthm of total assets to examne the relatonshp between sze and audtors opnon. At one hand, large companes are more lkely to have good accountng systems and nternal controls, thus reducng dsagreements and lmtatons on scope [Ireland (2003)]. On the other hand, assets overstatng or msappropraton s among the typcal fnancal statement fraud technques [Zegenfuss (1996), Beasley et al. (1999)]. Table 1 presents the 19 varables. Table 1 Lst of avalable fnancal varables Formula Defnton Log of Total Assets (LOGAS) Equty / Total Assets (EQAS) Equty / Net Loans (EQLOAN) Equty / Customer & ST Fundng (EQCUST) Equty/Labltes (EQLIAB) Cap Funds / Labltes (CAPLIAB) Net Interest Margn (NIM) The natural logarthm of bank s total assets expressed n mllon US dollars. Ths rato measures the amount of protecton afforded to the bank by the Equty they nvested n t. The hgher ths fgure the more protecton there s. Smlarly to equty/total assets ths rato measures the Equty cushon avalable to absorb losses on the loan book. Ths rato measures the amount of permanent fundng relatve to short term potentally volatle fundng. The hgher ths fgure the better. Ths leverage rato s smply another way of lookng at the Equty fundng of the balance sheet and s another of lookng at captal adequacy. Ths rato s smlar to equty/labltes but adds hybrd captal and subordnated debt to shareholders equty n the numerator. Ths rato s the net nterest ncome expressed as a percentage of earnng assets. The hgher ths fgure the cheaper the fundng or the hgher the margn the bank s commandng. Hgher margns and proftablty are desrable 6

8 Net Interest Revenue / Average total Assets (NIRAS) Other Operatng Income / Average Assets (OPIAS) Non nterest expenses/ Average Assets (EXPAS) Return On Average Assets (ROAA) Return on Avg. Equty (ROAE) Cost To Income Rato (COST) Recurrng Earnng Power (RECUR) Net Loans / Total Assets (LOANAS) Net Loans / Customer & ST Fundng (LOANCUST) Net Loans / Tot Deposts & Borrowngs (LOANDEP) as long as the asset qualty s beng mantaned. Ths rato s smlar to net nterest margn but expressed as a percentage of the total balance sheet. Ths rato ndcates to what extent fees and other ncome represent a greater percentage of earnngs of the bank. As long as ths s not volatle tradng ncome t can be seen as a lower rsk form of ncome. In general, the hgher ths fgure s the better. Non nterest expenses or overheads plus provsons gve a measure of the cost sde of the banks performance relatve to the assets nvested. Ths s perhaps the most mportant sngle rato n comparng the effcency and operatonal performance of banks as t looks at the returns generated from the assets fnanced by the bank. The return on average equty s smlar to ROAA but ndcates the return on shareholder funds. Ths rato measures the overheads or costs of runnng the bank, the major element of whch s normally salares, as percentage of ncome generated before provsons. Ths rato s a measure of before tax profts addng back provsons for bad debts as a percentage of Total Assets. Effectvely ths s a return on assets performance measurement wthout deductng provsons. Ths lqudty rato ndcates what percentage of the assets of the bank are ted up n loans. The hgher ths rato the less lqud the bank wll be. Ths loans to depost rato s a measure of lqudty n as much as hgh fgures denotes lower lqudty. Ths rato s smlar to net loans/customer & 7

9 Lqud Assets / Customer & ST Fundng (LIQCUST) Lqud Assets / Total Deposts & Borrowngs (LIQDEP) shot term fundng has as ts denomnator deposts and borrowngs wth the excepton of captal nstruments. Ths s a depost run off rato and looks at what percentage of customer and short term funds could be met f they were wthdrawn suddenly, the hgher ths percentage the more lqud the bank s and less vulnerable to a classc run on the bank. Ths rato s smlar to Lqud Assets / Customer & ST Fundng but looks at the amount of lqud assets avalable to borrower as well as depostors From a practcal pont of vew, developng a model that consders such a large number of varables ntroduce problems to the applcablty of the model on a daly bass snce t wll requre collecton of varous data leadng to ncreased tme and cost for data preparaton and management [Spaths et al. (2003)]. In addton, n a multvarate analyss, multcollnearty among the varables s another ssue that should be kept n mnd. Therefore, reducng the set of varables to an easly manageable number s essental. In the present study, ths s acheved through a combnaton of a unvarate test of sgnfcance, correlaton analyss and human judgment as n Doumpos and Zopounds (2002a), Spaths et al. (2003), Doumpos et al. (2004), Gagans et al. (2005), Pasouras et al. (2005), among others. Obvously, n order to classfy the qualfed and unqualfed fnancal statements effectvely, the varables should be able to dscrmnate between the two groups. In ths case, the rule of thumb s to keep the number of varables small and exclude a varable unless ts dscrmnatng power s statstcally sgnfcant [Kocagl et al. (2002]. Therefore, n selectng the approprate varables to be ncluded n the audtng models, we focus on ther statstcal sgnfcance at the unvarate level usng a Kruskal Walls test of means dfferences. Table 2 presents descrptve statstcs (mean and standard devaton) and the results of the Kruskal Walls test. Table 2 Descrptve statstcs and Kruskal-Walls test Qualfed Unqualfed Kruskal Walls Mean St. Dev. Mean St. Dev. p-value LOGAS EQAS

10 EQLOAN EQCUST EQLIAB CAPLIAB NIM NIRAS OPIAS EXPAS ROAA ROAE COST RECUR LOANAS LOANCUST LOANDEP LIQCUST LIQDEP The unvarate test suggests that the mean values of the ndependent varables for the qualfed versus the unqualfed fnancal statements are sgnfcantly dfferent n a number of cases. The varables that measure captal strength are all sgnfcant ndcatng that banks wth qualfed fnancal statements are not as well captalzed as the ones wth unqualfed fnancal statements. Bank wth qualfed fnancal statements are n general less proft (ROAA, ROAE, OPIAS, RECUR) and less cost effcent (EXPAS, COST) than the ones wth unqualfed fnancal statements. Fnally, the varables measurng lqudty ndcate that banks wth qualfed fnancal statements appear to be more lqud on average than the ones wth unqualfed statements. The next step n the analyss was to examne the correlatons among the aforementoned sgnfcant varables and exclude varables that were hghly correlated (among 0.75 n absolute terms). 9

11 Table 3-Correlaton analyss (1) (2) (3) (4) (5) (6) (7) (8) (1) EQAS (2) EQLOAN (3) EQCUST (4) EQLIAB (5) CAPLIAB (6) OPIAS (7) EXPAS (8) ROAA (9) ROAE (10) COST (11) RECUR (12) LOANAS (13)LOANCUST (14) LOANDEP (15) LIQCUST (16) LIQDEP Note: Correlatons above 0.75 (n absolute values) are denoted wth bold 10

12 Table 3-Correlaton analyss (contnue) (9) (10) (11) (12) (13) (14) (15) (16) (9) ROAE (10) COST (11) RECUR (12) LOANAS (13) LOANCUST (14) LOANDEP (15) LIQCUST (16) LIQDEP Note: Correlatons above 0.75 (n absolute values) are denoted wth bold 11

13 The results n Table 3 ndcate that both EQAS and EQCUST are hghly correlated wth EQLIAB and CAPLIAB whch are also correlated to each other. From these correlated varables, we select EQAS that s consdered one of the basc ratos whose use dates back to the early 1900s [Goln (2001)] and has been employed n numerous studes n bankng [e.g. Kocagl et al. (2002), Gunther and Moore (2003a,b)]. We also nclude EQLOAN and EQCUST that are not hghly correlated wth EQAS. Between the two proftablty ratos that are correlated (0.764), we select ROAE rather than ROAA because t s more lkely that managers wll manpulate the former to keep shareholders pleased. Fnally, we observe that among the lqudty ratos, LOANAS and LOANCUST are correlated wth LOANDEP, as are the ratos LIQDEP and LIQCUST. From these ratos we select LOANCUST and LIQDEP that are both consdered basc measures of lqudty. In addton to the 10 fnancal varables selected above, we consder a non-fnancal varable to control for the bankng envronment n whch banks operate. Obvously, bankng supervson wll have an mpact on almost every scheme of bank s governance, through prudental regulaton (e.g. captal requrements), dsclosure requrements and constrants on ther busness actvtes. We therefore employ the Hertage Bankng and Fnance Factor that measures the relatve openness of a country s bankng and fnancal system. The score for ths factor, s estmated by determnng: (1) whether foregn banks and fnancal servces frms are able to operate freely, (2) how dffcult t s to open domestc banks and other fnancal servces frms, (3) how heavly regulated the fnancal system s, (4) the presence of state-owned banks, (5) whether the government nfluences the allocaton of credt, and (6) whether banks are free to provde customers wth nsurance and nvest n securtes. In general, the factor may take the values of 1 (very low restrctons on banks), 2 (low restrctons on banks), 3 (moderate restrctons on banks), 4 (hgh restrctons on banks) or 5 (very hgh restrctons on banks). Table 4 presents the number of qualfed and unqualfed fnancal statements and the relatve bankng envronment for the banks n sample. Two conclusons can be drawn. Frst, all banks operate n bankng envronments that receved values between 1 and 4, whle none has receved a value of 5. Second, most of the qualfed opnons were assgned n an envronment wth hgh restrctons on banks (.e. bankng envronment 4) whle none operates n an envronment wth very low restrctons on banks (.e. bankng envronment 1). On the other hand, most unqualfed opnons (44%) were assgned n an envronment wth moderate restrctons (bankng envronment 3), followed by hgh restrctons (bankng envronment 4) and very low restrctons (.e. bankng envronment 1). 12

14 Table 4 - Audtors opnon and bankng envronment Audtor s opnon Bankng envronment Qualfed Unqualfed Total Very low restrctons (=1) Low restrctons (=2) Moderate restrctons (=3) ,034 Hgh restrctons (=4) Very hgh restrctons (=5) Total 198 2,228 2,426 To ntroduce ths varable n the analyss three dummy 0-1 varables are used. The frst dummy varable (BANKING 1) ndcates whether a bank operates n a market wth very low restrctons (very low restrctons=1) or not (very low restrctons=0). Smlarly, the second varable (BANKING 2) ndcates whether the bank operates n a market wth low restrctons (low restrctons = 1) or not (low restrctons = 0). Fnally, the thrd varable (BANKING 3) ndcates whether the bank operates n a market wth moderate restrctons (moderate restrctons = 1) or not (moderate restrctons = 0). Banks operatng n markets wth hgh restrctons (BANKING 4) are represented wth zero values n all these dummy varables. 2.3 Classfcaton technques The problem consdered n ths study s a classfcaton one that n general nvolves the assgnment of a fnte set A = {a 1,a 2,,a m ) of m alternatves, along a set g = {g 1, g 2,, g n ) of n crtera nto a set of q groups {C 1,C 2,,C 3 ). In ths study the alternatves nvolve the fnancal statements n the sample, the varables correspond to the 11 ndependent varables and there are two classes, the qualfed fnancal statements and the unqualfed ones. Ths paper employs four classfcaton technques namely Dscrmnant Analyss (DA), Logt Analyss (LA), k-nearest Neghbours (k-nn) and UTltés Addtves DIScrmnantes (UTADIS) that are brefly dscussed below Dscrmnant Analyss DA seeks to obtan a lnear combnaton of the ndependent varables whose objectve s to classfy observatons nto mutually exclusve groups as accurately as possble by maxmzng the rato of among-groups to wthn-groups varance. The DA method therefore estmates a dscrmnant functon of the followng form: Da w0 w1 g1 w2 g2... w m g m (1) 13

15 where D a s the score (for a fnancal statement ), w 0 s the ntercept term and (j=1,,m) represent the slope coeffcents assocated wth the ndependent varables g (j=1,, m) for each frm. j A cut-off pont s calculated accordng to the a-pror probabltes of group membershp and the costs of msclassfcaton. In the fnal step, each fnancal statement s classfed nto the qualfed or the unqualfed group, dependng on ts score and the cut-off pont. Fnancal statements wth dscrmnant scores greater than the cut-off pont are classfed nto the one group, whle fnancal statements wth dscrmnant scores less than the cut-off pont are classfed nto the other group. Alternatvely, frms can be classfed on the bass of the probablty of belongng to one of the groups and a cut-off probablty pont. w j Logt Analyss In logt analyss the probablty of a fnancal statement to be qualfed s based on a set of ndependent varables s gven by the followng functon: where P 1 1 e Z P ln (3) Z w0 w1 g1 w2 g 2... wm g m, 1 P (2) s the probablty that frm wll receve a qualfed audtor s opnon, w 0 s the ntercept term and w j (j = 1,,m) represents the coeffcents assocated wth the correspondng ndependent varables g (j= 1,,m) for each fnancal statement. The j coeffcent estmates are obtaned by regresson whch nvolves maxmsng a loglkelhood functon. The model s then used to estmate the group-membershp probabltes for all fnancal statements under consderaton. The fnancal statement s classfed as qualfed or unqualfed usng an optmal cut-off pont, attemptng to mnmse type I and type II errors Nearest Neghbours Nearest Neghbours s a non-parametrc densty estmaton method that classfes an object (.e. fnancal statement) to the class of ts nearest neghbour n the measurement space usng some knd of dstance measure lke the local metrcs [Short and Fucunaga (1980)], the global metrcs [Fukunaga and Flck (1984)], the 14

16 Mahalanobs or the Eucldean dstance. The later s the most commonly used one and s also employed n the present study. The modfcaton of the nearest neghbour rule, the k-nearest neghbour (k-nn) method that s employed n the present study, classfes an object (.e. fnancal statement) to the class (.e. qualfed or unqualfed) more heavly represented among ts k nearest neghbours. Assumng a fnancal statement x descrbed by the feature vector th g x g x,..., g m ) where g r x s used to denote the values of the r 1, 2 x characterstc of frm x, the dstance between two nstances follows: j m r 1 r r j x and x j s estmated as d x, x g x g x (4) Then, the algorthm for approxmatng a dscrete-valued functon of the form n f : C, where C s a fnte set of classes C 1, C2,..., Cq proceeds as follows: Step 1: For each tranng example (.e. fnancal statement) x, f x, add the frm to the lst of tranng examples. Step 2: Gven a query frm x to be classfed, let x 1, x2,..., xk denote the k nstances from the tranng examples that are nearest to x. Step 3: Return f where a,b) 0 otherwse. ^ x k arg max c, f x, where a,b) 1 f a b and c C 1 2 Thus, the algorthm returns the value x as an estmate of f x common value of f among the k tranng examples nearest to x. f^, whch s the most UTADIS UTADIS leads to the development of an addtve utlty functon that s used to score the fnancal statements and decde upon ther classfcaton. The developed addtve utlty functon has the followng general form: n U a w u ( g ) [0,1] (5) 1 15

17 where g g, g,..., g } s the set of the evaluaton crtera, whch n ths case { 1 2 n correspond to the 11 varables, w s the weght of crteron g (the crtera weghts sum up to 1) and u ( g ) s the correspondng margnal utlty functon normalzed between 0 and 1. The margnal utlty functons provde a mechansm for decomposng the aggregate result (global utlty) n terms of ndvdual assessment to the crteron level. To avod the estmaton of both the crtera weghts and the margnal utlty functons, t s possble to use the transformaton u g ) w u ( g ). ( ( g Snce u g ) s normalzed between 0 and 1, t becomes obvous that u ) ranges ( n the nterval [0, w ]. In ths way, the addtve utlty functon s smplfed to the followng form: n U ( a) u ( g ) [0,1] (6) 1 The developed utlty functon provdes an aggregate score U a for each fnancal statement along all crtera. In the case of audtng decsons, ths score provdes the bass for determnng whether the fnancal statement could be classfed n ether the group of qualfed or unqualfed fnancal statements. The classfcaton rule n ths case s the followng (C 1 and C 2 denote the group of unqualfed and qualfed fnancal statements respectvely, whle u 1 s a cut-off utlty pont defned on the global utlty scale,.e. between 0 and 1): U(a) U(a) u u 1 1 The estmaton of the addtve value functon and the cut-off threshold s performed through lnear programmng technques so that the sum of all volatons of the classfcaton rule (7) for all the fnancal statements n the tranng sample s mnmzed. Detaled descrpton of the mathematcal programmng formulaton used n the UTADIS method can be found n the works of Zopounds and Doumpos (1999) and Doumpos and Zopounds (2002b). 3. Emprcal Results After selectng an approprate set of varables the classfcaton models are developed usng the tranng data and tested on the future holdout sample. Summary statstcs, the sgnfcance of the varables n dscrmnatng between qualfed and unqualfed fnancal statements and classfcaton accuraces are presented n Tables 5 and 6. a a C C 1 2 (7) 16

18 Table 5 Varables sgnfcance and models summary Varables LA DA UTADIS EQAS (3.223) (-0.017) 4.42% EQLOAN (30.003)** (-0.233) 9.62% EQCUST (0.632) (-0.06) 1.50% OPIAS (10.225)** (0.119) 62.67% EXPAS (9.132)** (0.151) 0.34% ROAE (3.25) (-0.259) 8.42% COST (17.418)** (0.372) 0.10% RECUR (30.215)** (-0.373) 7.37% LOANDEP (5.179)* (-0.039) 3.02% LIQCUST (30.229)** (0.326) 2.53% BANKING (2.245) (-0.667) 0.01% BANKING (95.523)** (-0.645) BANKING (81.086)** (-0.802) Constant (1.022) Models Summary Statstcs Ch-square ** - Nagelkerke R Square Wlks' Lambda Ch-square Notes: * Statstcally sgnfcant at the 5% level ** Statstcally sgnfcant at the 1% level In the case of UTADIS, bankng enters the analyss as a sngle varable that takes the values of 1, 2, 3 and 4 for very low restrctons, low restrctons, moderate restrctons and hgh restrctons respectvely. 17

19 Among the captal strength ratos only EQLOAN appears to be statstcally sgnfcant at the LA model, whle smlar results are observed both n the case of DA model (as ndcated by the standardzed coeffcents) and the UTADIS model (as ndcated by the weghts of the crtera). OPIAS s also mportant n all the models and s actually the most mportant one n the case of UTADIS. From the remanng proft and cost effcency ratos that are statstcally sgnfcant n the LA model, EXPAS and COST are postvely related to the probablty of recevng a qualfed audtor s opnon, whle RECUR s negatvely related. The results n the DA model for these ratos are qute smlar however n the case of UTADIS only ROAE and RECUR appear to have some mportance. The lqudty ratos are both statstcally sgnfcant n the LA model. LOANDEP s negatvely related to the probablty of a qualfed audtor s opnon whle, as expected, the opposte occurs n the case of LIQCUST, ndcatng that, consstent wth the unvarate results, the more lqud a bank appears to be the hgher the probablty of recevng a qualfed opnon. The mportance of these two ratos s smlar and moderate n the UTADIS model, whle only LIQCUST s mportant n the DA model. As t concerns the coeffcents of the three dummy varables that correspond to the envronmental condtons, they all carry a negatve sgn. The hghest coeffcent (n absolute terms) s observed n the case of the bankng envronment wth very low restrctons ( ), followed by low restrctons (-4.077) and moderate restrctons (-1.933). Hence, the probablty of a qualfed report s lower n for banks operatng n an envronment wth very low restrctons, followed by banks operatng n envronments wth low and moderate restrctons. Nevertheless, t should be mentoned that ths varable s the less mportant one n the case of the UTADIS model. Table 6 Classfcaton accuraces (n %) LA DA UTADIS k-nn Tranng Unqualfed Qualfed Average Testng Unqualfed Qualfed Average The classfcaton results obtaned from the applcaton of the four methods n dscrmnatng between qualfed and unqualfed fnancal statements are presented n Table 6. Wth regard to the tranng sample, the average classfcaton accuracy of k- 18

20 NN s 95%, hence hgher than that of the other three methods whch all obtan average classfcaton accuraces around 85%. Of course, hgher model ft does not ensure hgher generalzng ablty and the results n the future holdout sample become of partcular nterest towards a more approprate evaluaton of the four methods. As expected the classfcaton accuraces n the holdout sample are lower than the ones acheved n the tranng sample. The model developed through UTADIS s now the one that acheves the hghest classfcaton accuracy (64.9%), followed by LA (63.2%), k-nn (62.9%) and DA (54.8%). Further nspecton of the results ndcates that the superorty of UTADIS and LA s due to ther ablty to classfy relatvely well both qualfed and unqualfed fnancal statements. On the other hand, DA acheves the hghest classfcaton n terms of qualfed fnancal statements, but ts performance s nferor n terms of unqualfed fnancal statements, whle the opposte occurs n the case of k-nn. These dfferences ndcate that the combnaton of all methods nto an ntegrated model could possbly lead to hgher classfcaton accuraces. 4. Conclusons The regulatory and supervsory framework n many Asan countres s currently experencng sgnfcant changes. However, audtng standards are crucal, whatever reforms n the regulatory framework may be ntroduced, snce wth poor accountng and audtng requrements, the qualty and dsclosure of fnancal statements can be out-of date or unrelable. Nevertheless, Asa has been noted for an nadequate reportng, accountng and audtng framework. The objectve of the present study was to contrbute towards the detecton of banks fnancal statements that should receve qualfed audt opnons. To accomplsh ths task we nvestgated the effcency of four classfcaton technques, namely dscrmnant analyss, logt analyss, UTADIS multcrtera decson ad, and nearest neghbours, n the development of audtng models. Snce the detaled audt of all transactons of a bank would not only be tme-consumng and expensve but also mpractcable, the employment of such models could provde the bass for a decson tool for audtors durng the examnaton of Asan commercal banks. The analyss s also mportant to partes other than audtors, such as the managers of the bank, and the fnancal regulators. To develop the audtng models and examne ther classfcaton ablty, the dataset was splt nto two dstnct samples. The tranng sample conssted of 1,701 unqualfed fnancal statements and 146 ones that receved a qualfed opnon over the perod The models were then tested n a holdout sample of 527 unqualfed fnancal statements and 52 ones that receved a qualfed opnon over the perod The results showed that the developed audtng models can dscrmnate between fnancal statements that should receve qualfed opnons from 19

21 the ones that should receve unqualfed opnons wth a satsfactory accuracy. Both fnancal varables and the envronment n whch banks operate appeared to be mportant. The hghest classfcaton accuracy was acheved by UTADIS, followed by logt analyss, nearest neghbours and dscrmnant analyss. Future research could be drected towards the analyss of other types of banks (e.g. nvestment, co-operatve), the employment of alternatve classfcaton technques (e.g. neural networks, support vector machnes), and the use of addtonal varables (e.g. audt fess, audtor s ndependence) that were not ncluded n the present study due to data avalablty. Acknowledgments We would lke to thank two anonymous revewers for valuable comments. Earler versons of the manuscrpt were wrtten whle Dr. Pasouras was Research Assocate n Techncal Unversty of Crete and Honorary Research Fellow of Coventry Unversty. References Bank for Internatonal Settlements. (2002). The relatonshp between bankng supervsors and banks external audtors. Basel Commttee on Bankng Supervson, January, avalable at: Basel Commttee on Bankng Supervson. (1999). A New Captal Adequacy Framework: Consultatve paper. June, Basel. Basel Commttee on Bankng Supervson. (2001). The New Basel Captal Accord: an explanatory note. Bank for Internatonal Settlements, January. Basel Commttee on Bankng Supervson. (2003). Overvew of The New Basel Captal Accord: Consultatve Document. Bank for Internatonal Settlements, Aprl. Basel Commttee on Bankng Supervson. (2004). Internatonal Convergence of Captal Measurement and Captal Standards: A Revsed Framework. Bank for Internatonal Settlements, June. Beasley S.M, Carcello J.V. and Hermanson D.R. (1999). Fraudulent Fnancal Reportng: : An Analyss of US Publc Companes. Research Report, COSO. Berger A.N., Kng K.K. and O'Bren J.M. (1991). The lmtatons of market value accountng and a more realstc alternatve. Journal of Bankng and Fnance vol. 15, Dopouch N., Holthausen R. and Leftwch R. (1987). Predctng audt qualfcatons wth fnancal and market varables. Accountng Revew vol. 62,

22 Doumpos M., Kosmdou K. and Pasouras F. (2004). Predcton of acquston targets n the UK: A multcrtera approach. Operatonal Research An Internatonal Journal vol. 4, Doumpos M. and Zopounds C. (2000a). Busness Falure Predcton: A Comparson of Classfcaton Methods. Operatonal Research An Internatonal Journal vol. 2, Doumpos M. and Zopounds C. (2002b). Multcrtera Decson Ad Classfcaton Methods. Kluwer Academc Publshers: Dordrecht. Fannng K, Cogger K. and Srvastava R. (1995). Detecton of management fraud: a neural network approach. Internatonal Journal of Intellgent Systems n Accountng, Fnance and Management vol. 4, Fernandez A.I. and Gonzalez F. (2005). How accountng and audtng systems can counteract rsk-shftng of safety-nets n bankng: Some nternatonal evdence. Journal of Fnancal Stablty vol. 1, Felds L.P., Fraser D.R. and Wlkns M.S. (2004). An nvestgaton of the prcng of audt servces for fnancal nsttutons. Journal of Accountng and Publc Polcy vol. 23, Ftch. (2005). Asan Banks and Basel II. Specal Report. 19 January, Gagans Ch., Pasouras F. and Tzanetoulakos A. (2005). A comparson and ntegraton of classfcaton technques for the predcton of small UK frms falure. The Journal of Fnancal Decson Makng vol. 1, Goln J. (2001). The Bank Credt Analyss Handbook: A Gude for Analysts, Bankers and Investors. John Wley & Sons Pte Ltd: Asa. Gunther J.W. and Moore R.R. (2000). Fnancal Statements and Realty: Do Troubled Banks Tell All?, Federal Reserve Bank of Dallas Economc and Fnancal Revew, Thrd Quarter, Gunther J.W. and Moore R.R. (2003a). Loss underreportng and the audtng role of bank exams. Journal of Fnancal Intermedaton vol. 12, Gunther J.W. and Moore R.R. (2003b). Early warnng models n real tme. Journal of Bankng & Fnance vol. 27, Internatonal Federaton of Accountants. (2000). The Audt of Internatonal Commercal Banks: Proposed Internatonal Audtng Practce Statement. Internatonal Audtng Practces Commttee, September. Ireland J. (2003). An emprcal nvestgaton of determnants of audt reports n the UK. Journal of Busness Fnance & Accountng vol. 30, Keasey K., Watson R. and Wynarzcyk P. (1988). The small company audt qualfcaton: A prelmnary nvestgaton. Accountng and Busness Research vol. 18, Kocagl A.E., Reyngold A., Sten R.M. and Ibarra E. (2002). Moody s RskCalc TM Model for Prvately-Held U.S, Banks. Moody s Investor Servce Global Credt Research, July. 21

23 Latnen E.K. and Latnen T. (1998). Qualfed audt reports n Fnland: evdence from large companes. European Accountng Revew vol. 7, Levtan A.S. and Knoblett J.A. (1985). Indcators of exceptons to the gong concern opnon decson. Audtng: A Journal of Practce and Theory vol. 5, Loebbecke J., Enng M. and Wllngham J. (1989). Audtor s experence wth materal rregulartes: frequency, nature, and detectablty. Audtng: A Journal of Practce and Theory vol. 9, Mutchler J.F. (1985). A multvarate analyss of the audtor s gong concern opnon decson. Journal of Accountng Research vol. 23, Parrenas J.C. (2002). The New Basle Accord: Challenges for Asa s Banks and Regulators, PECC Fnance Forum Conference, Issues and Prospects for Regonal Cooperaton for Fnancal Stablty and Development, August 11-13, Honolulu. Pasouras F., Tanna S. and Zopounds C. (2005). Applcaton of Quanttatve Technques for the Predcton of Bank Acquston Targets. World Scentfc: Sngapore. Shra S. (2001). Searchng for New Regulatory Frameworks for the Intermedate Fnancal Structure n Post-Crss Asa. Wharton School, Workng Paper Seres 01-28, Unversty of Pennsylvana. Sddqu J. and Podder J. (2002). Effectveness of bank audt n Bangladesh, Manageral Audtng Journal vol. 17, Spaths Ch. (2003). Audt Qualfcaton, Frm Ltgaton, and Fnancal Informaton: an Emprcal Analyss n Greece. Internatonal Journal of Audtng vol. 7, Spaths Ch. (2002). Detectng false fnancal statements usng publshed data: some evdence from Greece. Manageral Audtng Journal vol. 17, Spaths Ch., Doumpos M. and Zopounds C. (2003). Usng clent performance measures to dentfy pre-engagement factors assocated wth qualfed audt reports n Greece. The Internatonal Journal of Accountng vol. 38, Spaths Ch., Doumpos M. and Zopounds C. (2002). Detectng falsfed fnancal statements: a comparatve study usng multcrtera analyss and multvarate statstcal technques. The European Accountng Revew vol. 11, Sten M., Smunc D. and O Keefe T. (1994). Industry dfferences n the producton of audt servces. Audtng: A Journal of Practce and Theory 13 (Supplement), Summers S.L. and Sweeney J.T. (1998). Fraudulently msstated fnancal statements and nsder tradng: an emprcal analyss. The Accountng Revew vol. 73, Zegenfuss D.E. (1996). State and local government fraud survey for Manageral Audtng Journal vol. 9, Zopounds C. and Doumpos M. (1999). A multcrtera decson ad methodology for sortng decson problems: The case of fnancal dstress. Computatonal Economcs vol. 14,

24 23