Analysing the Determinants of Bank Efficiency: The Case of Italian Banks

Transcription

1 Analysng the Determnants of Bank Effcency: The Case of Italan Banks Clauda Grardone, Phlp Molyneux *, Edward P.M. Gardener School of Accountng, Bankng and Economcs, Unversty of Wales, Bangor, Gwynedd, LL57 DG, UK. Abstract Ths paper nvestgates the man determnants of Italan banks cost effcency over the perod , by employng a Fourer-flexble stochastc cost fronter n order to measure X-effcences and economes of scale. Qualty and rskness of bank outputs are explctly accounted for n the cost functon and ther mpact on cost effcency levels are evaluated. The results show that mean X-neffcences range between 3 and 5 per cent of total costs and they tend to decrease over tme for all bank szes. Economes of scale appear present and sgnfcant, beng especally hgh for popular and credt co-operatve banks. Moreover, the ncluson of rsk and output qualty varables n the cost functon seems to reduce the sgnfcance of the scale economy estmates. Followng Spong et al. (995) we undertake a proftablty test that allows for the dentfcaton of banks that are both cost and proft effcent. The results suggest that the most effcent and proftable nsttutons are more able to control all aspects of costs, especally labour costs. Fnally, we pool the data to carry out a logstc regresson model n order to examne bank- and market-specfc factors that nfluence Italan banks neffcency. Confrmng Mester (993 and 996), neffcences appear to be nversely correlated wth captal strength and postvely related to the level of non-performng loans n the balance sheet. The analyss also shows that there s no clear relatonshp between assets sze and bank effcency. Fnally, t s possble to nfer that quoted banks seem to be on average more effcent than ther non-quoted counterparts. JEL classfcaton: G, D. Keywords: Italan banks; Cost functon; Ineffcences; Economes of scale. * Correspondng author: School of Accountng, Bankng and Economcs, Unversty of Wales Bangor, Gwynedd, Bangor LL57 DG (UK). E-mal: p.molyneux@bangor.ac.uk, Fax , Tel Acknowledgements Thanks to Professor G.B. Pttaluga from the Unverstà degl Stud d Genova for hs precous comments and for provdng the Blbank database. We also wsh to thank M. Brown, B. Casu and J. Wllams from the Unversty of Wales, Bangor, for ther valuable suggestons.

2 . Introducton Bankng has experenced dramatc changes over the last decade or so. Deregulaton, fnancal nnovaton and automaton have been maor forces mpactng on the performance of the bankng sector. In such a context, banks have become ncreasngly concerned about controllng and analysng ther costs and revenues, as well as measurng the rsks taken to produce acceptable returns. In lne wth these developments, an extensve lterature has evolved examnng fnancal frm effcency ssues (see, for a comprehensve survey, Berger and Humphrey, 997), and dfferent methodologcal approaches have been employed to nvestgate fnancal frm effcency (.e. parametrc and non-parametrc technques). However, only a handful of studes have so far nvestgated how rsk and output qualty factors nfluence bank effcency levels (for example, Mester, 996; Berger and Mester, 997; Altunbas, Lu, Molyneux and Seth, 999). Ths paper has two man obectves. Frst t ams to extend the establshed lterature by examnng the determnants of Italan banks cost effcency over the perod , by employng a Fourer-flexble stochastc cost fronter to evaluate X-effcency and scale economes. Qualty and rskness of bank output are explctly accounted for n the cost functon and ther mpact on effcency levels are evaluated and dscussed. Secondly, the paper attempts to dentfy the man characterstcs of effcent banks. Followng the approach suggested by Spong et al. (995) our sample of banks s subect to a proftablty test that allows us to dentfy nsttutons that are both cost and proft effcent. Followng Mester (993 and 996) we also use a

3 logstc regresson model to examne bank- and market- specfc factors whch nfluence banks neffcency levels. The paper s set out as follows. Secton provdes a bref lterature revew on recent developments n fnancal frm effcency analyss placng partcular emphass on varous Italan studes. Secton 3 outlnes the methodology and secton 4 reports the results. Secton 5 s the concluson.. New vews on bank effcency Over the last decade parametrc studes of Italan bank cost effcency (for example, Baldn and Land, 990; Conglan et al., 99; Parg et al., 99; European Commsson, 997; Rest, 997; Inzerllo et al., 999) have found evdence of economes of scale across a wde range of bank szes, and average X-neffcency levels are usually n lne wth those found n the nternatonal lterature typcally rangng between 5 and 0 percent. However, none of these studes have ncluded the level of captal and/or loan losses as arguments n the cost functon to control for rsk of default and/or output qualty, respectvely. A relatvely new approach (for example, Hughes and Mester, 993; McAllster and McManus, 993; Clark, 996; Mester, 996; Berger and Mester, 997; Altunbas, Lu, Molyneux and Seth, 999) ponts to the mportance of ncludng measures of output qualty and default rsk on the grounds that unless they are accounted for n the cost functon, bank levels of X- effcences and economes of scale may be mscalculated. In the case of Italan bankng the ncluson of these two varables could be crucal because recently credt nsttutons have suffered from a dramatc ncrease n ther level of non-performng 3

4 loans (NPLs). For example, Italan banks based n the south of the country recorded NPLs (as a percentage of total loans) of 4. per cent n 996 and.8 percent n 997, ratos that were more than twce the natonal average and hgh by any standard (Bank of Italy, 998). The achevement of more compettve condtons n the Italan bankng market durng the 990s has also brought about stuatons of crss for varous bankng frms. Durng the perod the man banks n the south of Italy experenced substantal reductons n ther net nterest ncome and were no longer able to cover ther relatvely hgh operatng expenses and loan losses. As a consequence, these banks ether reduced ther sze or were acqured by larger (healther) banks. Another mportant recent ssue n bank effcency analyss concerns the choce of the functonal form for the cost functon. Studes by leadng researchers n the feld (see, for example, Mtchell and Onvural, 996; Berger and DeYoung, 997; Berger et al., 997; Berger and Mester, 997; Altunbas, Lu, Molyneux and Seth, 999) have abandoned the typcal U-shaped translog for the snusodal Fourer-flexble, whch combnes the stablty of the translog specfcaton near the average of the sample data wth the flexblty of the Fourer specfcaton for observatons far from the averages. The Fourer functonal form s preferred to the translog because t better approxmates the underlyng cost functon across a broad range of outputs. Ths paper advances the lterature on bank effcency by nvestgatng the effects of the ncluson of rsk and output qualty factors n the cost functon. Moreover, we adopt a stochastc Fourer-flexble cost functon to calculate X-effcency levels and economes of scale for a sample of Italan banks between 993 and 996. The paper 4

5 also uses a proftablty test as suggested by Spong et al. (995) to nvestgate the characterstcs of the most and least effcent banks n the Italan system. 3. Methodology 3.. Input and Output Defnton and Data Sample Choosng the approprate defnton of bank output s an mportant ssue for research nto banks cost effcency. Whle the multproduct nature of the bankng frm s wdely recognsed, there s stll no agreement as to the explct defnton and measurement of banks nputs and outputs. Generally, each defnton of nput and output carres wth t a partcular set of bankng concepts, whch nfluence and lmt the analyss of the producton characterstcs of the ndustry. The approach to output defnton used n ths study s a varaton of the ntermedaton approach, whch was orgnally developed by Sealey and Lndley (977) and posts that total loans and securtes are outputs, whereas deposts along wth labour and captal are nputs to the producton process of bankng frms. (Table A. n the Appendx provdes descrptve statstcs on the outputs and nput prces ncluded n our model). The data used to construct the estmates for the cost functon parameters are derved from Blbank, an Italan database of the Assocazone Banche Prvate Italane. Ths database provdes annual ncome and balance sheet data for credt nsttutons belongng to dfferent bank categores. The sample comprses an unbalanced panel of,958 bank observatons dstrbuted n the followng way: 545 banks n 993, 53 n 994, 466 n 995 and 44 n 996. The sample excludes: ) 5

6 banks that are subsdares of foregn banks; ) the central nsttutons for each category of banks; and ) specal credt nsttutons (medum- and long- term banks). Our sample consttutes nearly 50 per cent of the Italan bankng market n terms of number of banks. (As dsplayed n Table A., the dfference n banks sze across the fve classes s relatvely hgh. The groups of very bg banks have approxmately 400 tmes the average assets of the very small banks and 40 tmes the average assets of small banks). We also nvestgate effcency of banks across geographcal regons and therefore dentfy those nsttutons located n four maor regons: north-west, north-east, centre, south and slands Stochastc Cost Fronter Model, X-Effcences and Economes of Scale Researchers nvestgatng bank cost effcency postulate a relatonshp between costs, nput prces and output quantty. Ths relatonshp s based on the dualty concept between producton and cost functons. The producton functon Q = Q( X ) summarses the technology of a frm, that s the exstng relatonshp between nputs, It should also be noted that n the emprcal analyss most of the estmaton s carred out usng the FRONTIER 4. and TSP 4.0 packages. Snce FRONTIER 4. does not tolerate mssng values, banks wth ncomplete accountng data could not be ncluded n the data sample (Coell, 996 and Coell et al., 998). The Bank of Italy categorses banks accordng to fve sze groups: very bg; bg; medum; small and very small. 3 These dfferent regons are defned as follows: ) North-West: Lgura, Lombarda, Pemonte, Valle d Aosta; ) North-East: Emla-Romagna, Frul-Veneza Gula, Trentno Alto Adge, Veneto; 3) Centre: Lazo, Marche, Toscana, Umbra; 4) South and Islands: Abruzzo, Baslcata, Calabra, Campana, Molse, Pugla, Sardegna, Scla. It should be noted that a bank was assgned to a gven regon f, over the perod, t had ts head offce n that area. 6

7 X, and outputs, Q. The cost functon TC TC( Q, P) = shows the relatonshp between total producton costs, TC, and the prces of varable nputs. The dualty condton between the producton and the cost functon ensures that they contan the same nformaton about producton possbltes and that there s a unque correspondence between both functons. Moreover, observable producton plans and cost levels usually do not follow from perfectly ratonal and effcent decsons. On the contrary, such factors as errors, lags between the choce of the plan and ts mplementaton, nerta n human behavour and dstorted communcatons and uncertanty are amongst the factors that mght cause X-neffcences to drve real data away from the optmum (Rest, 997). Ths study employs the stochastc cost fronter approach to generate estmates of X-effcences for each bank over the years along the lnes frst suggested by Agner et al. (977) and Meeusen and van den Broeck (977). For the -th frm, the ln TC = lntc Q, P ; B ε, sngle equaton cost functon model s represented by ( ) where TC s the observed total cost of producton for bank, Q s the vector of bankng output for bank, P s the vector of nput prces for bank and B s a vector of parameters. ln ( Q, P ; B) bank operatng at ( Q P, B) th frm can be wrtten as follows: TC s the predcted log cost functon of a cost mnmsng,. Fnally ε s a two-components error term that for the - ε = v u () where v s a two-sded error term representng statstcal nose whch s assumed to be ndependently and dentcally dstrbuted; and u s a non-negatve (or one-sded) random varable representng neffcency and assumed to be dstrbuted 7

8 ndependently of the v. It s also assumed that the v are normally dstrbuted wth mean zero and varance σ v, and the u are the absolute values of a varable that s normally dstrbuted wth mean µ and varance σ u. In the present study banks data over the perod are organsed n a panel. Specfcally, we employ the Battese and Coell (99) model of a stochastc fronter functon for panel data wth frm effects whch are assumed to be dstrbuted as halfnormal random varables (that s, wth µ=0) 4 and are also permtted to vary systematcally wth tme. Therefore, t s possble to express ths model as TC = x β v u, [wth = (,,..., N) and t = (,,..., T)], where TC t s (the t t t t logarthm) of the total costs for the -th frm n the t-th tme perod; x t s a k vector of (transformatons of the) nput and output quanttes of the -th frm n the t-th tme perod; β s the vector of unknown parameters; and the v t and u t are defned as above, wth u t = {exp[η(t-t)]}, where η s an unknown scalar parameter to be estmated that represents the hypothess about the evoluton or steadness of ndvdual neffcences over the perod under study. Moreover, the parametersaton of Battese and Corra (977) s employed, who replaced σ v and σ u wth σ = σ σ and v u γ = σ u ( σ v σ u). 5 As recently emphassed by Coell et al. (998) the γ-parametersaton has an advantage n 4 There are many varatons on ths assumpton n the lterature (for detals, see Greene, 993 and Coell et al., 998). 5 In the lterature, the lkelhood functon has often been expressed n terms of the two varance parameters = and σ σ v σ u λ σ u σ v (Agner et al., 977; Jondrow et al., 98; Coell, 996). See also Battese and Coell (993) and Coell et al. (998) for the log-lkelhood functon of the model used here gven these dstrbutonal assumptons. 8

9 seekng to obtan the maxmum lkelhood estmates because the parameter space for γ can be searched for a sutable startng value for the teratve maxmsaton algorthm nvolved. In partcular, a value of γ of zero ndcates that the devatons from the fronter are due entrely to nose, whle a value of one would ndcate that all devatons are due to neffcency. As concerns the choce of the functonal form for the cost functon, ths study employs a Fourer-flexble form because t s a global approxmaton that has proved to domnate the commonly specfed translog form (see, for example, McAllster and McManus, 993; Mtchell and Onvural, 996; Berger et al., 997; Altunbas, Lu, Molyneux and Seth, 999). The resultng mxed cost functon can be wrtten as: lntc = = = = = = = = 4 [ λ cosz θ snz ] [ λ cos( z z ) θ sn( z z )] = s [ λ cos( z z zk ) θ sn( z z zk )] ε = k k α = δ lnq lnq 3 0 ρ lnq lnp α lnq lns α lnq 3 3 = = s 3 k = γ lnp lnp α lnq lnk β lnp lns τ 3 = = β lnp τ lnk τ lns ks k τ lnk lnk τ 3 kk = k lnk lns s ss β lnp lnk lns lns () where TC s a measure of the normalsed costs of producton, comprsng operatng costs and fnancal costs (nterest pad on deposts); Q and Q are output quanttes, that s total loans and total securtes, respectvely; P s the normalsed prce of labour; P s the normalsed prce of deposts; K s the level of fnancal captal; S s 9

10 the rato of NPLs to total loans; α, β, δ, γ, ρ, λ, θ,τ are parameters to be estmated; and ε s the two-component error term as defned n (). Moreover, the z are adusted values of the natural log of output ln Q so that they span the nterval [. π,.9 π ] where [ b]. In partcular, the formula for z used here s (. π µ a µ ln Q ), a, s the range of Q ln and µ ( 9 π. π ) ( b a).. The Fourer-flexble form s a global approxmaton because the terms such as z cos, sn z, cos z, z sn are mutually orthogonal over the [,π ] 0 nterval, so that each addtonal term can make the approxmatng functon closer to the true path wherever t s most needed. As observed by Gallant (98), by restrctng the z to span the nterval [. π,.9 π ] endponts are reduced., the approxmaton problems arsng near the In addton, standard symmetry has to be mposed on the translog porton of the functon: δ = δ and γ = γ, where ( =, ) and ( =,,3), and the followng lnear restrctons on () are necessary and suffcent for lnear homogenety n factor 3 prces: β = 3 = ; γ = 3 = 0 ; ρ = 3 = 0; β = k 3 = 0 ; β = s = 0. In accordance wth the assumed constrant of lnear homogenety n prces, TC, P and P are normalsed by the prce of captal, P 3. It s also mportant to menton that consderaton of nput share equatons embodyng Shephard s Lemma restrctons s excluded n order to allow for the possblty of allocatve neffcency (see, for example, Berger and Mester, 997). The Fourer terms are ncluded only for the outputs, leavng the nput prce effects to be descrbed solely by the translog term (see, for example, Berger et al., 997 and Altunbas, Lu, Molyneux and Seth, 999). The Fourer terms for the 0

11 nput prces are also excluded n order to conserve the lmted number of Fourer terms for the output quanttes used to measure economes of scale. The nput prces also show very lttle varaton, thereby provdng greater ustfcaton for our methodologcal approach. In ths research the parameters of the stochastc fronter cost functon, defned by eq. (), are estmated usng the Maxmum-Lkelhood (ML) approach. For nstance, the ML estmates of β and γ are obtaned by fndng the mnmum of the loglkelhood functon as specfed n Coell et. al (998). The nature of ths loglkelhood functon gven the dstrbutonal assumptons on v and u can also be found n Battese and Coell (99). Once the model s estmated, bank level measures of X-effcences are calculated usng the resduals and are usually gven by the mean of the condtonal dstrbuton of u gven ε. For the half-normal stochastc model the E(u ε ) s consdered as a consstent estmator for ndvdual X-effcences (Coell et al., 998). The resultng cost effcency rato may be thought of as the proporton of costs or resource that are used effcently. For example, a bank wth a cost effcency of 0.80 s 80 per cent effcent or equvalently wastes 0 per cent of ts costs relatve to a best practce frm facng the same condtons. A natural way to express the extent of scale economes s the proportonal ncrease n cost resultng from a small proportonal ncrease n the level of output, that s the elastcty of total cost wth respect to output. The degree of economes of

12 scale (SCALE) used here s gven by SCALE = m = lntc ln Q whch represents the sum of ndvdual cost elastctes. 6 Ths can be rewrtten as: SCALE= = α lnk = = k = k k α δ lnq = = = = = λ sn λ k = α lns = ( z z ) θ cos( z z ) sn ( z ) θ cos( z ) ( z z z ) θ cos( z z z ) s k λ sn k 3 ρ lnp k (3) where there are economes of scale f SCALE<, constant returns to scale f SCALE=, and dseconomes of scale f SCALE>. The degree of scale economes s computed by usng the mean values of output, nput prces and rsk and output qualty varables Proftablty Test and Correlates wth Ineffcency Followng Spong et al. (995), we subect our cost effcency measures derved from the Fourer model to a proftablty test. Ths s undertaken n order to dentfy banks that are both cost and proft effcent. Ths approach s taken because the cost sde may provde naccurate rankngs of effcency because a seemngly cost neffcent bank mght be offsettng hgher expenses wth hgher revenues. Moreover, ths test 6 Evanoff and Isralevch (985) dstngush between scale elastcty and scale effcency where the former s measured as n equaton (3) and the latter s measured as the change n output requred to produce at mnmum effcent scale. Throughout ths paper reference to economes and dseconomes of scale relates to scale elastctes.

13 provdes an alternatve to some of the consstency condtons recently suggested by Bauer et al. (997). It follows that banks whch do well on both cost effcency and proftablty tests wll comprse the most effcent bank category; banks that fare poorly on the two tests are n the least effcent category. The two groups, comprsng the most effcent and least effcent banks were parttoned n the followng way: ) most effcent group: banks that rank n the upper quartle of Italan banks on the cost effcency estmates and rank n the upper half n terms of ROA (return on assets), and ) least effcent group: banks that rank n the bottom quartle on the cost effcency estmates and rank n the bottom half n term of ROA. Subsequently, we analyse the fnancal characterstcs of effcent and neffcent banks by carryng out a comparson between the maor sources of ncome and expenses as well as several other fnancal ratos for the year 996. Lastly, n order to nvestgate possble determnants of bank effcency, frmspecfc measures of neffcency derved exclusvely from the model excludng rsk and output qualty varables that s, equaton () wthout the varables K and S, are regressed on a set of ndependent varables relevant to the bankng busness. Ths set of potental correlates wth bank neffcency s chosen n such a way that many aspects of bankng actvtes are consdered: for nstance, bank sze, market characterstcs, geographc poston, captal, performance and retal actvtes (see also Mester, 996; Berger and Mester, 997; and Altunbas, Lu, Molyneux and Seth, 999). A logstc functonal form rather than a lnear regresson model s used because the values of the neffcency estmates, Eˆ ( ε ) applcatons of ths technque to the bankng system, see Mester, 993 and 996). 3 u, range between 0 and (for

14 4. Results 4.. Structural Tests Structural tests are undertaken to see f the cost functon that ncluded the rsk and output qualty varables dffers sgnfcantly from the standard cost fronter specfcaton, the translog form, and the models ncludng ndvdual rsk and qualty varables. Table A.3 n the Appendx reports the lkelhood rato test statstcs and shows that the cost fronter ncludng rsk and qualty varables provdes the best ft to the data. More specfcally the cost functon that ncludes the rsk and qualty varables (RQCF) dffers sgnfcantly from the standard model (CF) snce the lkelhood rato test reects the hypothess that the two models are not sgnfcantly dfferent. Wth partcular regard to the choce of the functonal form for the cost functon, the translog model s reected at the 0.0 per cent level, thus supportng the choce of the Fourer-flexble functon. Smlarly, the models excludng ndvdual rsk and qualty varables are all reected aganst our model defned n () at the 0.0 per cent level. 4.. X-Effcences and Economes of Scale Tables and report the average X-effcency levels for banks grouped by sze classes, geographcal areas (a bank s assgned to a gven regon f t has ts head offce n that area) and bank types (the category commercal banks ncludes those banks that are not savngs, popular or credt co-operatve banks). We dstngush between dfferent types of banks because mutual banks (savngs and co-op banks) 4

15 may have dfferent manageral obectves compared wth commercal banks and ths could be reflected n ther cost effcency. For nstance, t could be the case that mutual banks have obectves other than cost mnmsaton such as servng the local communty and maxmsng returns to members. Tables & here Average neffcency levels range between approxmately 3 and 5 per cent. Smlar fgures can be found n varous recent studes of Italan banks (see, for example, Allen and Ra, 996; European Commsson, 997) and, although dfferences are not large, the most effcent banks seem to be the bg, medum and very small banks. Moreover, X-neffcency levels seem to decrease over tme and for all szes of banks. When the X-effcency results are grouped accordng to banks operatng n dfferent geographcal areas, t appears that the lowest effcency levels are generally found for banks havng ther head offce n the centre and south of the country, thus confrmng the apparent exstence of sgnfcant dspartes among geographcal regons. Fnally, accordng to the dfferent bank types, t s possble to observe that on average the better performng banks seem to be the credt co-operatves together wth popular banks, possbly reflectng a greater homogenety of the co-operatve bankng sector. Altunbas, Evans and Molyneux (999) found smlar results for German banks and they argue that mutual banks tend to have a lower cost of funds than other bank types due, for example, to ther (possble) local monopoles. 5

16 Table 3 provdes the pooled effcency scores from the estmaton of our model n relaton to each asset-based quartle n order to check whether sze can effectvely be consdered as a determnant of potental cost reductons. From the table t appears that both the mean and the medan of frm-specfc X-effcency estmates tend to decrease from Quartle to Quartle 4 [Quartle (4) contans the smallest (largest) frms]. Ths suggests that, on average, larger banks devate more from ther respectve cost-effcent fronter than do smaller banks. Relatvely speakng, the smallest banks appear to be less neffcent than ther larger counterparts Table 3 here The same data (assets n logarthms) are reported n Fgure. The scatter dagram suggests that the dsperson of the effcency scores s qute hgh, thus mplyng that n many cases smlar-szed banks have dfferent effcency levels and, supposedly, dfferent costs. From a general vewpont, a slght nverse trend between total assets and X-effcences seems to preval. Fgure : Bank Sze and Effcency Scores X-EFF TOTAL ASSETS (values n logs) 6

17 Table 4 reports economes of scale estmates for each year together wth ther sgnfcance levels. The results seem to confrm that bg and medum banks show hgh and sgnfcant scale economes n the maorty of cases, and very small banks also appear to enoy substantal economes. Very bg banks do not appear to experence potentally realsable scale economes n the years under study and manly exhbt constant returns to scale. Table 4 here Table 5 shows scale economy estmates accordng to dfferent bank types and geographc areas. The table reveals several nterestng patterns: ) commercal banks show hgh and sgnfcant economes of scale especally n the northern part of the country; ) savngs banks appear on average the less lkely to gan from potental scale economes; and ) popular and credt co-operatve banks appear best able to explot cost reductons n terms of economes of scale. Table 5 here Fnally, f we compare the results on the X-effcences (Table ) and economes of scale (Table 4) derved from the cost functon ncludng rsk and output qualty factors (RQCF) wth those obtaned from the standard cost functon (CF) we fnd that X-effcency results are smlar. However, t does appear to be the case that the extent of scale economes seems to be overstated f rsk and qualty factors are not ncluded 7

18 n the cost functon specfcaton. Ths result s broadly n accordance wth the recent fndngs of Berger and Mester (997) and Altunbas, Lu, Molyneux and Seth (999) Proftablty Test and Correlates wth Ineffcency So far our analyss has only focused on cost effcency. Spong et al. (995), however, note that t s mportant to combne both cost effcency estmates wth a proftablty test so as to evaluate fnancal frm effcency. Ths s because one needs to evaluate banks ablty to use resources effectvely n producng products and servces (cost effcency), and ther skll at generatng ncome from these servces (proftablty). Followng Spong et al. (995) we partton our sample of banks nto two categores the most and least effcent n terms of ther cost effcency and profts performance. The most (least) effcent banks are those that rank n the upper (lower) quartle accordng to the cost effcency estmates and n the upper (lower) half n terms of return on assets. Table 6 shows the number of banks and ther effcency and proftablty characterstcs accordng to the aforementoned parttonng. Table 6 here For the four years under study, an average of 75 banks satsfy the selecton crtera for the most effcent group and 84 banks are classfed n the least effcent group. The mean bank n the least effcent group has a cost effcency of only 0.77, whch ndcates that the bank wth the hghest effcency n the sample could have produced the same amount of output as the least effcent bank at only 77 per cent of ther cost. 8

19 In contrast, the average cost effcency level for the most effcent banks s approxmately 0.95, thus ndcatng less dsparty wth the best bank n the sample. Moreover, as an average for the four years, the ROA for the most effcent banks s equal to.66% compared wth 0.09% for the poorest performers. In order to analyse the fnancal characterstcs of effcent and neffcent banks Table 7 shows a comparson between the maor sources of ncome and expenses as well as several other fnancal ratos for these banks. The table reveals how effcent and neffcent banks dffer. The table also shows fnancal characterstcs for the sample all banks and also for a sub-sample of large banks only. Ths latter group ncludes the followng bank szes: very bg, bg, medum and small (see Table A. for ther assets sze range). The choce of creatng such a sub-sample s motvated by the fact that the relatvely large number of very small banks found n the least and most effcent bank categores could bas our nterpretaton of the results. Table 7 here On the earnngs sde, the advantages held by the most effcent banks seems to relate to ther ncome generatng capacty and, as expected, expenses control. Focusng on the sample all banks the most effcent group has, for nstance, an advantage over neffcent banks n terms of hgher nterest receved on assets. On the other hand, the most neffcent banks have relatvely hgh non-nterest revenues compared wth the most effcent banks, thus suggestng that there mght be some dfferences n the way the two groups generate ncome. However, these results 9

20 change f the very small banks are excluded from the sample, whch shows a greater mportance of non-nterest ncome for the most effcent group. As concerns the expense sde, the most effcent and least effcent banks show smlar nterest expenses. Ths means that the most effcent banks do not have mportant advantages n fundng costs, and therefore they are achevng ther performance by other means, other factors beng equal. Furthermore, t seems apparent that effcent banks are more effectve n controllng operatng costs, and partcularly staff expenses. Wth regard to ther balance sheet structures, n 996 the most effcent banks hold more securtes and have hgher levels of equty than ther neffcent counterparts, thus provdng a hgh level of protecton to ther customers. Moreover, the most effcent banks appear to have better asset qualty, thereby mplyng that they are assgnng more attenton and resources to loan orgnaton, montorng and other credt udgement actvtes. Fnally, n order to nvestgate the possble determnants of bank effcency, frmspecfc X-neffcences are regressed on a set of ndependent varables relevant to the bankng busness. As stressed by Mester (996), the fndngs are ntended manly to ndcate where banks mght look for clues toward ncreasng ther effcency. As shown n Table 8 the estmates ndcate that dfferent varables sgnfcantly correlate wth neffcences n the Italan bankng sector. Agan our sample banks are dvded nto large banks (very bg, bg, medum and small banks), and very small banks. Table 8 here 0

21 In accordance wth Mester s fndngs (993 and 996), neffcences are always nversely correlated wth fnancal captal. On the one hand, ths s qute predctable snce banks wth low neffcency wll tend to have more profts as they wll be able (holdng dvdends constant) to retan more earnngs as captal. However, ths result should not be nterpreted as sayng that f a bank ncreases ts captal/assets rato then ts neffcency wll decrease. As Mester (996) ponts out, ths could also be explaned as an ndcaton that hgher captal ratos may prevent moral hazard both for the bank and ts managers. Moreover, neffcences are usually nversely correlated wth bank performance varables, although ths relatonshp s nsgnfcantly dfferent from zero. Wth regard to the coeffcent for the level of NPLs, t s always postvely related to bank neffcency. In fact, hgher effcency s expected to be correlated wth better credt rsk evaluaton (see also Mester, 993; Berger and DeYoung, 997; Altunbas, Lu, Molyneux and Seth, 999). Over the perod , neffcent banks also tended to have a hgher ntensty of retal bankng busness, a hgher number of branches than effcent banks, and a hgher nterest margn to assets rato. These factors appear to be mportant n determnng varatons n effcency among sze classes. The results concernng the relatonshp between total assets sze and bank effcency are mxed. The coeffcent s not sgnfcantly dfferent from zero for both the all and large bank categores. These fndngs show that there s no statstcal evdence that larger banks are more or less X-effcent than ther smaller counterparts. In fact, from the results t s possble to see that an nverse relatonshp between assets sze and neffcency appears only to hold wthn a specfc bank sze

22 group (.e. very small banks). The results also ndcate that quoted banks are, on average, more effcent than ther non-quoted counterparts. Fnally, the sgnfcance and negatve sgn of the dummy varable for prvate banks suggests that, at least large prvate banks tend to have lower levels of neffcency. 5. Conclusons Durng the 990s the Italan bankng system has experenced a declne n proftablty brought about by the fall n nterest margns and persstently hgh levels of staffng costs. In addton, adverse macroeconomc condtons have lead to a substantal ncrease n non-performng loans especally for banks located n specfc geographcal areas of the country. The response of the bankng system to these pressures has been to undertake a substantal consoldaton movement resultng n an ncrease n the number of bankng groups and promotng the prvatsaton process. Evdence on effcency and proftablty gans followng M&A actvtes n Italy are mxed (see for example Rest, 998; Focarell et al., 999). As to the prvatsaton process, by the end of 993 over 90 per cent of those banks prevously actng as publc foundatons had become ont-stock companes. Despte ths only some 30 banks are publcly lsted and more than 50 per cent of the bankng system s stll n publc hands (Rest, 998). The am of ths paper was to provde an emprcal analyss of the cost effcency of the Italan bankng sector over the perod takng nto account the rsks assocated wth banks operatons. We found that mean X-neffcency levels range between 3 and 5 per cent of ther total costs and they tend to decrease over tme

23 and for all szes of banks. Smlarly, economes of scale appear present and sgnfcant n the Italan bankng system when consdered as a whole. These are qute mportant results f we consder that durng the process of consoldaton and restructurng of the system has amed at gradually ncreasng banks sze. The most cost effcent banks, both n terms of X-effcency and economes of scale, are bg and medum szed banks generally located n the northern part of the country. Furthermore, the results for credt co-operatve banks confrm other studes that economes of scale can be partcularly hgh at a local level and for very small banks because of possble local monopoles. We also compare the results on X-effcences and economes of scale derved from the cost functon ncludng rsk and output qualty factor and the standard cost functon. In lne wth recent fndngs by Berger and Mester (997) and Altunbas, Lu, Molyneux and Seth (999), the X-effcency estmates are smlar across the two dfferent cost functon specfcatons. In contrast, the level and sgnfcance of scale economy estmates appears lower as a result of the ncluson of rsk and output qualty factors n the cost functon. Followng the proftablty test as suggested by Spong et al. (995), the man dfferences between the most effcent and least effcent bank seem to be manly related to staff expenses. In the context of mportant technologcal mprovements n banks productve processes, ths suggests an urgent need for greater labour market flexblty and the consequent substtuton of labour for captal. Moreover, neffcent banks always appear to have lower levels of equty/assets and hgher levels of nonperformng loans. 3

24 Fnally, we pooled the data to carry out a logstc regresson model n order to examne bank- and market-specfc factors that nfluence banks effcency. Confrmng Mester (993 and 996), neffcences appear to be nversely correlated wth captal and postvely related to the level of non-performng loans. Ths latter fndng suggests that effcent banks are assgnng more attenton and resources to loan orgnaton, montorng and other credt udgement actvtes. Interestngly, over the perod neffcent banks also tended to have (on average) a greater retal bankng orentaton, hgher nterest margns and more branches compared wth ther effcent counterparts. Fnally, the analyss also shows that there s no clear relatonshp between assets sze and bank effcency. However, from the results t s possble to nfer that quoted banks, on average, appear to be more effcent than ther non-quoted counterparts. 4

25 Appendx Table A.: Varable Descrpton Varables Symbol Descrpton Total Costs TC Staff expenses other non-nterest expenses nterest pad Output Q Total customer loans Output Q Other earnng assets Input Prce P Staff expenses / average number of personnel Input Prce P Interest expenses / total customer deposts Input Prce 3 P 3 Other non-nterest expenses / total fxed assets Fnancal Captal K Total equty Asset Qualty S Non-performng loans / total loans Table A.: Dstrbuton of Sample Banks by Average Assets a,b,c Bank sze Mean Very Bg 4,86. (8) 40,645. (8) 4,3. (8) 36,3. (8) 40,06. (8) Bg 35,360.8 () 34,875.7 () 37,597. () 37,755.9 () 36,397.4 () Medum,77.8 (5),000.9 (5),770.4 () 3,844.6 (),583.4 (3) Small 3,849. (67) 3,58.9 (6) 3,708.9 (60) 3,853.0 (58) 3,735.0 (6) Very Small 364. (433) 34.4 (46) (364) 39. (35) (385) a Bllons of Italan Lre. b In brackets the number of banks. c All monetary aggregates are expressed at 996 prces. Table A.3: Structural Tests a,b Test Performed Degrees Crtcal [versus RQCF] Test Statstcs of Freedom Value χ.0 Outcome CF = Standard cost fronter 98.4 k = Reected Translog form c 0 k = Reected Npls/Total loans only (S only) 3 k = Reected Equty only (K only) 83.4 k = Reected a The lkelhood rato statstcs s calculated as { ln[ L( H )] [ L( H )]} were L ( ) and ( ) lkelhood functon under the null and the alternatve hypotheses, 0 ln H 0 H and 0 H, respectvely. L are the values of the b RQCF = cost functon estmates wth rsk and output qualty varables. The CF (standard cost fronter) s lke the model specfed H n eq. () excludng K and S (rsk and output qualty varables, respectvely). c The translog form tested here ncludes K and S. 5

26 References Agner, D.J., Lovell C.A.K. and P. Schmdt, 977, Formulaton and Estmaton of Stochastc Fronter Producton Functon Models, Journal of Econometrcs 6, -37. Allen, L. and A. Ra, 996, Operatonal Effcency n Bankng: An Internatonal Comparson, Journal of Bankng and Fnance 0, Altunbas, Y., Evans L. and P. Molyneux, 999, Bank Ownershp and Effcency, Journal of Money, Credt and Bankng, forthcomng. Altunbas, Y., Lu M.H., Molyneux P. and R. Seth, 999, Effcency and Rsk n Japanese Bankng, Journal of Bankng and Fnance, forthcomng. Baldn, D. and A. Land, 990, Econome d scala e complementaretà d costo nell ndustra bancara talana, L ndustra, Bank of Italy, 998, Summary Reports on Economc Developments n the Italan Regons n 997, Rome. Battese, G.E. and T.J. Coell, 99, Fronter Producton Functons, Techncal Effcency and Panel Data: Wth Applcaton to Paddy Farmers n Inda, Journal of Productvty Analyss 3, Battese, G.E. and T.J. Coell, 993, A Stochastc Fronter Producton Functon Incorporatng a Model for Techncal Ineffcency Effects, Workng Paper n Econometrcs and Appled Statstcs, Department of Econometrcs, Unversty of New England, Armdale Battese, G.E. and G.S. Corra, 977, Estmaton of a Producton Fronter Model: Wth Applcaton to the Pastoral Zone of Eastern Australa, Australan Journal of Agrcultural Economcs, Bauer, P.W., Berger A.N., Ferrer G.D. and D.B. Humphrey, 997, Consstency Condtons for Regulatory Analyss of Fnancal Insttutons: A Comparson of Fronter Effcency Methods, Fnance and Economcs Dscusson Seres, Federal Reserve Board Berger, A.N. and R. DeYoung, 997, Problem Loans and Cost Effcency n Commercal Banks, Journal of Bankng and Fnance, Berger, A.N. and D.B. Humphrey, 997, Effcency of Fnancal Insttutons: Internatonal Survey and Drectons for Future Research, Fnance and Economcs Dscusson Seres, Federal Reserve Board 97-. Berger, A.N., Leusner J.H. and J. Mngo, 997, The Effcency of Bank Branches, Journal of Monetary Economcs 40,

27 Berger, A.N. and L.J. Mester, 997, Insde the Black Box: What Explans Dfferences n the Effcences of Fnancal Insttutons?, Journal of Bankng and Fnance, Clark, J.A., 996, Economc Cost, Scale Effcency and Compettve Vablty n Bankng, Journal of Money, Credt and Bankng 8, Coell, T.J., 996, A Gude to FRONTIER Verson 4.: A Computer Program for Stochastc Fronter Producton and Cost Functon Estmaton, Workng Paper, Centre for Effcency and Productvty Analyss Coell, T.J., Prasada Rao D.S. and G. Battese, 998, An Introducton to Effcency and Productvty Analyss, Norwell, Kluwer. Conglan, C., De Bons R., Motta G. and G. Parg, 99, Econome d scala e d dversfcazone nel sstema bancaro talano, Tem d Dscussone, Banca d Itala 50. European Commsson, 997, Credt Insttutons and Bankng, The Sngle Market Revew, Subseres II: Impact on Servces, Vol. 3, Kogan Page-Earthscan, London. Evanoff, D.D. and P.R. Isralevch, 995, Scale Elastcty versus Scale Effcency Southern Economc Journal 6, Focarell, D., Panetta F. and C. Saleo, 999, Why Do Banks Merge: Some Emprcal Evdence from Italy, Banca d Itala, forthcomng. Gallant, A. R., 98, On the Bas n Flexble Functonal Forms and an Essentally Unbased Form: The Fourer Flexble Form, Journal of Econometrcs 5, -45. Greene, W.H., 993, The Econometrc Approach to Effcency Analyss, n: H.O. Fred, Lovell C.A.K. and S.S. Schmdt, eds., The Measurement of Productve Effcency: Technques and Applcatons (Oxford, Oxford Unversty Press) Hughes, J.P. and L.J. Mester, 993, A Qualty and Rsk-Adusted Cost Functon for Banks: Evdence on the Too-Bg-To-Fal Doctrne, Journal of Productvty Analyss 4, Inzerllo, U., Morell P. and G.B. Pttaluga, 999, Deregulaton and Changes n European Bankng Industry, n: CEPS-CSC, Regulatory Reform for the Better Functonng of Markets, forthcomng. Jondrow, J., Lovell C.A.K., Materov I.S. and P. Schmdt, 98, On the Estmaton of Techncal Ineffcency n the Stochastc Fronter Producton Functon Model, Journal of Econometrcs 9,

28 McAllster, P.H. and D. McManus, 993, Resolvng the Scale Effcency Puzzle n Bankng, Journal of Bankng and Fnance 7, Meeusen, W. and J. van den Broeck, 977, Effcency Estmaton from Cobb-Douglas Producton Functons wth Composed Error Term, Internatonal Economc Revew 8, Mester, L.J., 993, Effcency n the Savngs and Loan Industry, Journal of Bankng and Fnance 7, Mester, L.J., 996, A Study of Bank Effcency Takng nto Account Rsk Preferences, Journal of Bankng and Fnance 0, Mtchell, K. and N.M. Onvural, 996, Economes of Scale and Scope at Large Commercal Banks: Evdence from the Fourer Flexble Functonal Form, Journal of Money, Credt, and Bankng 8, Parg, G., Sestto P. and U. Vvan, 99, Econome d scala e dversfcazone nell ndustra bancara: ruolo dell eterogenetà tra mprese, Tem d Dscussone, Banca d Itala 74. Rest, A., 997, Evaluatng the Cost Effcency of the Italan Bankng Market: What Can Be Learned from the Jont Applcaton of Parametrc and Non-parametrc Technques, Journal of Bankng and Fnance, -50. Rest, A., 998, Can Mergers Foster Effcency? The Italan Case, Journal of Economcs and Busness 50, 57-69, March/Aprl. Sealey, C.W. and J.T. Lndley, 977, Inputs, Outputs and a Theory of Producton and Cost at Depostory Fnancal Insttutons, Journal of Fnance 3, Spong, K., Sullvan R. and R. DeYoung, 995, What Makes a Bank Effcent? A Look at Fnancal Characterstcs and Bank Management and Ownershp Structure, Fnancal and Industry Perspectves, Federal Reserve Bank of Kansas Cty, -9, December. 8

29 Table : Average Values of X-Effcency Grouped by Sze Classes a RQCF RQCF RQCF RQCF CF CF CF CF Very bg 83.% 84.0% 84.7% 85.4% 8.% 8.8% 8.3% 8.8% Bg 85.% 85.8% 86.4% 87.% 85.% 85.6% 86.% 86.5% Medum 85.6% 86.% 87.% 87.5% 85.9% 86.3% 87.0% 87.0% Small 8.% 8.0% 8.8% 83.5% 8.6% 8.3% 8.8% 83.3% Very small 85.% 86.% 87.0% 87.6% 85.4% 86.% 86.7% 87.% a RQCF= cost functon estmates wth rsk and output qualty varables; CF = standard cost functon estmates. 9

30 Table : Average Values of X-effcency Grouped by Bank Types and Geographcal Areas a By regon North-west 86.4% 87.6% 88.0% 88.4% North-east 84.9% 85.8% 87.0% 87.4% Centre 83.6% 84.5% 85.0% 85.3% South and slands 83.5% 84.4% 85.% 86.% By bank type Commercal 80.5% 8.8% 8.6% 83.% Savngs 8.3% 83.% 84.% 85.0% Popular 84.% 85.6% 86.6% 87.% Credt co-op. 86.4% 87.% 87.7% 88.% a Unless otherwse stated, results refer to RQCF, that s the cost functon estmates wth rsk and output qualty varables. 30

31 Table 3: Bank Sze and Effcency Levels a Assets range b,c Mean Medan Stdev Mn Max Quartle (58.8) 87.7% 87.7% % 99.% Quartle (88.9) 85.7% 85.6% % 99.5% Quartle ,40. (653.9) 86.% 86.5% % 99.6% Quartle ,043.9 (7,6.5) 83.7% 83.0% % 99.6% a Unless otherwse stated, results refer to RQCF, that s the cost functon estmates wth rsk and output qualty varables. b Mean values n brackets. c Values n bllons of Italan Lre. 3

32 Table 4: Economes of Scale (by Sze of Banks) a,b,c RQCF RQCF RQCF RQCF CF CF CF CF Very bg Bg 0.83* 0.849* 0.805** ** 0.785** 0.745** 0.84* Medum 0.8** 0.806** 0.88** 0.800** 0.76** 0.75** 0.767** 0.744** Small ** Very small 0.767*** 0.77*** 0.784*** 0.847*** 0.754*** 0.76*** 0.779*** 0.844*** All banks 0.80*** 0.80*** 0.89*** 0.859*** 0.779*** 0.784*** 0.803*** 0.844*** *, **, *** means statstcally sgnfcant at the 0%, 5% and.% respectvely. a If SCALE >, < or = then there are dseconomes, economes of scale or constant returns to scale respectvely. b In ths case the standard error refers to the hypothess H 0=. c RQCF= cost functon estmates wth rsk and output qualty varables; CF = standard cost functon estmates. 3

33 Table 5: Economes of Scale (by Bank Types and Geographcal Areas) a,b,c Commercal North-west 0.859*** 0.864*** 0.87** 0.878** North-east 0.89** 0.889** 0.9* 0.906** Centre * 0.96 South and slands Savngs North-west North-east Centre South and slands Popular North-west 0.890* 0.88* 0.897* 0.856** North-east 0.98* 0.97* 0.96* Centre 0.98* 0.870** 0.865** 0.88** South and slands 0.94** 0.90** 0.905** 0.90* Credt co-op. North-west 0.88*** 0.765*** 0.783*** 0.88** North-east 0.7*** 0.736*** 0.75*** 0.844** Centre 0.676*** 0.677*** 0.706*** 0.775*** South and slands 0.700*** 0.686*** 0.76*** 0.740*** *, **, *** means statstcally sgnfcant at the 0%, 5% and.% respectvely. a If SCALE >, < or = then there are dseconomes, economes of scale or constant returns to scale respectvely. b In ths case the standard error refers to the hypothess H 0=. c Unless otherwse stated, results refer to RQCF, that s the cost functon estmates wth rsk and output qualty varables. 33

34 Table 6: Proftablty Test (993-96) a Banks Number of banks Cost eff. (averages) % Roa (averages) % 993 Most effcent Least effcent Most effcent Least effcent Most effcent Least effcent Most effcent Least effcent a The two groups, comprsng the most effcent and least effcent banks were parttoned n the followng way: ) most effcent group: banks that rank n the upper quartle of Italan banks on the cost effcency estmates and rank n the upper half n term of ROA (Return On Assets), and ) least effcent group: banks that rank n the bottom quartle on the cost effcency estmates and rank n the bottom half n term of ROA. 34

35 Table 7: Sample Bank Informaton (Group Averages 996 data) a All banks Large banks only b 996 Most Least Most Least effcent effcent effcent effcent Number of banks Cost effcency ndex % ROA Interest receved Non-nterest ncome Interest pad Operatng costs Staff expenses/operatng costs Staff expenses Other non-nterest expenses Loans Deposts Securtes Equty Fxed assets Interest margn Npls/total loans a Unless otherwse stated, values are expressed as percentage of total assets. b The group of large banks ncludes the prevously defned very bg, bg, medum, and small banks. 35

36 Table 8: Logstc regressons (all banks, large banks and very small banks) a,b Parameter All banks Large banks Very small banks Intercept -3.00*** -.996*** *** Assets *** Margn 7.665***.6647** 5.43** Branches * *** Retal ***.5393*** 0.794*** Owners ** Non-perf *** *** Perform Captal ** * -.7** Quoted ** -0.69*** Northwe -0.58*** -0.07* *** Northea *** *** * Cent Com 0.433*** *** Sav 0.83*** ** 0.78* Popul 0.8*** ** 0.43** a Assets = total assets; Margn = nterest margn / total assets; Branches = number of branches; Retal = (customer loans customer deposts) / total assets; Owners = for prvate bank and 0 for publc; Non-perf = non-performng loans / total loans; Perform = net ncome / equty; Captal = equty / total assets; Quoted = for quoted banks and 0 for not quoted; Northwe = dummy for northwestern banks; Northea = dummy for north-eastern banks; Cent = dummy for banks located n the centre; Com = commercal banks; Sav = savng banks; Popul = popular banks. ndcates dummy varable. b All banks: number of obs.,958 log-lkelhood functon Large banks: number of obs. 40 log-lkelhood functon 67.6.Very small banks: number of obs.,538 log-lkelhood functon