Models for performance benchmarking: measuring the eect of e-business activities on banking performance

Transcription

1 Available online at Omega 32 (24) Models for performance benchmarking: measuring the eect of e-business activities on banking performance Wade D. Cook a, Lawrence M. Seiford b, Joe Zhu c; a Schulich School of Business, York University, 47 Keele Street, Toronto, Ont., Canada M3J 1P3 b Department of Industrial and Operations Engineering, The University of Michigan, Ann Arbor, MI , USA c Department of Management, Worcester Polytechnic Institute, Worcester, MA 169, USA Received 12 December 22; accepted 15 January 24 Abstract The current paper presents mathematical programming models for use in benchmarking where multiple performance measures are needed to examine the performance and productivity changes. The standard data envelopment analysis method is extended to incorporate benchmarks through (i) a variable-benchmark model where a unit under benchmarking selects a portion of benchmarksuch that the performance is characterized in the most favorable light, and (ii) a xed-benchmarkmodel where a unit is benchmarked against a xed set of benchmarks. The models are applied to a large Canadian bank where some branches services are automated to reduce costs and increase the service speed, and ultimately to improve productivity. The empirical investigation indicates that although the performance appears to be improved at the beginning, productivity gain has not been discovered. Our nding can facilitate the bankin examining its business options and further point to weaknesses and strengths in branch operations.? 24 Elsevier Ltd. All rights reserved. Keywords: Benchmarking; Data envelopment analysis; E-business; Performance; Productivity 1. Introduction Performance evaluation and benchmarking has become an important continuous improvement tools for business units in the high-technology world of computers and telecommunications where competition is intense and grows more so each day. Benchmarking activities positively force any business unit to constantly evolve and improve in order to survive and prosper in a business environment facing global competition. In fact, as reported in a recent Wall Street Journal poll [1], benchmarking is one of the top three important and popular tools for continuous improvement. Gap analysis is often used as a fundamental method in Corresponding author. Tel.: ; fax: addresses: wcook@schulich.yorku.ca (W.D. Cook), seiford@umich.edu (L.M. Seiford), jzhu@wpi.edu (J. Zhu). performance evaluation and benchmarking. However, as pointed out by Camp [2], one of the dilemmas that we face is how to show benchmarks where multiple measurements exist. It is rare that one single measure can suce for the purpose of performance evaluation. As a result, some multi-factor based gap analysis methods have been developed, e.g., Spider charts, AHP maturity index, and Z charts. Although gaps can be identied within each performance measure, it remains a challenging taskto combine the multiple measures in the nal stage. Therefore, benchmarking models that can deal with multiple performance measures and provide an integrated benchmarking measure are needed. Note that a mathematical programming based method data envelopment analysis (DEA) [9] has been proven an eective tool for evaluating the relative eciency of peer decision making units (DMUs) when multiple performance measures are present. DEA evaluates eciency without the need to specify the relationships or tradeos among the performance /$ - see front matter? 24 Elsevier Ltd. All rights reserved. doi:1.116/j.omega

2 314 W.D. Cook et al. / Omega 32 (24) measures prior to the computation, and DEA identies an ecient frontier (tradeo curve) along with eciency scores for all DMUs. This type of eciency evaluation has also been regarded as benchmarking [3], since the ecient frontier identied can be regarded as an empirical benchmarkfrontier. However, when a new DMU outperforms the identied ecient frontier, a new frontier is generated by DEA. As a result, we do not have the same benchmarkfor other new DMUs, i.e., the original DEA method cannot be viewed as a multi-criteria performance benchmarking tool. Zhu [18] developed two DEA-based benchmarking models where the identied ecient frontier (benchmark) remains the same during the benchmarking process. One is called variable-benchmarkmodel where each DMU under benchmarking is allowed to choose a portion of benchmark frontier so that the benchmarking performance of the DMU is characterized in the most favorable light. The other is called xed-benchmarkmodel where each DMU is benchmarked against xed components from the benchmark frontier. The current study applies the two DEA-based benchmarking models to a large Canadian bank (thereafter, CBANK) in measuring the eectiveness service delivery conguration. There are many DEA studies on banking performance. For example, Sherman and Gold [4] published the rst signicant DEA bankanalysis and started what turned out to be a long list of DEA applications to banking from several dierent angles [5]. Sherman and Ladino [6] reported that a use of DEA in the restructuring process of the 33 branches of a US bankled to an annual savings of over $6 million. Oral and Yolalan [7] introduced a DEA model that forced each of 2 branches in a sample to compare itself with the global leader a Turkish bank. According to the Canadian Bankers Association, the Canadian banking industry includes 16 domestic banks, 31 foreign banksubsidiaries and 21 foreign bankbranches operating in Canada. In total, these institutions manage over $1.7 trillion in assets. Technology innovation is the most important factor contributing to the dramatic changes taking place in Canada s nancial services marketplace. Canada s banknancial groups have led the way in providing Canadians with many new ways to access nancial services. Canadians have embraced debit cards, ABMs, telephone banking, the internet and hand-held wireless devices. Advances in technology continue to revolutionize the industry, breaking down geographic barriers and permitting customers to access nancial services from virtually anywhere, at any time. In recent years, Canada s banks have demonstrated a consistent performance, with prots rising signicantly from 1995 to The rest of the paper is organized as follows. The next section introduces some basic DEA models. We then introduce our DEA-based benchmarking models described in Zhu [18]. The models are applied to benchmarka set of e-branches against the best-practice of traditional branches. Concluding remarks are given at last. 2. DEA models We assume that there are n DMUs. In the current paper, the DMUs represent bankbranches. Each DMU j(j = 1; 2;:::;n) consumes a vector of inputs, x j(=x 1j;x 2j;:::;x mj) to produce a vector of outputs, y j (=y 1j;y 2j;:::;y sj). The relative eciency of DMU j, among others, can be measured by the following DEA model: min j n jx ij 6 j x ij ; j=1 n jy rj y rj ; j=1 i =1; 2;:::;m; r =1; 2;:::;s; j : (1) The dual program to model (1) is max u ry rj m u ry rj v ix ij 6 ; m v ix ij =1; j =1;:::;n; v i ; u r : (2) To rene the results obtained from model (1) or(2), value judgment in forms of weight restrictions can be incorporated into model (2). A r 6 ur 6 B r; u r C i 6 vi 6 D i; (3) v i where A r, B r, C i, and D i are pre-specied constants. Model (3) is called assurance region in [8] and is closely related to the cone ratio concept of Charnes et al. [1]. When the optimal value to model (1) or(2) is equal to one, we say that a specic DMU under evaluation is a best-practice DMU. Model (1) or(2) assumes that the best-practice frontier exhibits constant returns to scale (CRS). i.e., a best-practice DMU is both technically and scale ecient. If scale ineciency is allowed in best-practice DMUs, we can assume variable returns to scale (VRS) and incorporate n j=1 j = 1 into model (1). (For a complete discussion on standard DEA models, please refer to [11].) Benchmarking is a process of dening valid measures of performance comparison among peer units, using them to determine the relative positions of the peer units, and ultimately, establishing a standard of excellence. Through the use of DEA, we can establish a standard of excellence via best-practice identication. Once we have the standard

3 W.D. Cook et al. / Omega 32 (24) of excellence, we compare the performance of a set of new DMUs to the best-practice frontier. If we use model (1), since the reference set or the benchmarkconsists of all DMUs under evaluation, including the best-practice frontier and the new DMUs, model (1) does not directly compare the new DMUs to the established standard. Thus, the original DEA method needs to be modied and extended if we use DEA as a benchmarking tool. 3. DEA-based benchmarking models 3.1. DEA-based benchmarking model I: variable-benchmark model Let E represent the benchmarks or the best-practice identied by the DEA. We develop the following linear programming problem: min CRS jx ij 6 CRS xi new ; r ; jy rj y new j ; j E ; (4) where a new observation is represented by DMU new with inputs xi new (i =1;:::;m) and outputs yr new (r =1;:::;s). The superscript of CRS indicates that the benchmarkfrontier composed by benchmarkdmus in set E exhibits CRS. Model (4) measures the performance of DMU new with respect to benchmarkdmus in set E when outputs are xed at their current levels. Model (4) is dierent from the super-eciency DEA model where a DMU under evaluation is excluded from the reference set [12,13]. Because in model (4), the reference set remains unchanged for each evaluation while in the super-eciency DEA model the reference set changes for each evaluation. Model (4) is also a direct generalization of the DEA/acceptance model developed by Seiford and Zhu [14] where only one output is present. Similarly, we can have a model that measures the performance of DMU new in terms of outputs when inputs are xed at their current levels. max CRS jx ij 6 xi new ; jy rj CRS yr new ; j ; j E : (5) Based upon models (4) and (5), we have Proposition 1. CRS =1= CRS, where CRS is the optimal value to model (4) and CRS is the optimal value to model (5). Model (4)or(5) yields a benchmarkfor DMU new. The ith input and the rth output for the benchmarkcan be expressed as j x ij (ith input); j y ij (rth output): (6) Note also that although the DMUs associated with set E are given, the resulting benchmarkmay vary for each new DMU under evaluation. Because for each new DMU under evaluation, (6) may represent a dierent combination of DMUs associated with set E. Thus, models (4) and (5) represent a variable-benchmarkscenario. Proposition 2. (i) CRS 1 or CRS 1 indicates that the performance of DMU new is dominated by the benchmark in (6); (ii) CRS =1 or CRS =1 indicates that DMU new achieve the same performance level of the benchmark in (6); (iii) CRS 1 or CRS 1 indicates that input savings or output surpluses exist in DMU new when compared to the benchmark in (6). Proof. (i) and (ii) are obvious results in terms of the DEA eciency concept. Now, CRS 1 indicates that DMU new can increase its input vector to reach the benchmark. This in turn indicates that the input saving achieved by DMU new is CRS 1. Similarly, CRS 1 indicates that DMU new can decrease its output vector to reach the benchmark. This in turn indicates that the output surplus achieved by DMU new is 1 CRS. Figs. 1a and b illustrate the three cases described in Proposition 2. ABC in Fig. 1a (A B C in Fig. 1b) represents the benchmarkfrontier constructed from A, B and C (or A,B, and C ). D, H and G (or D,H, and G ) represent the new DMUs to be benchmarked against ABC (or A B C ). We have CRS D 1 for DMU D ( CRS D 1 for DMU D ) indicating that DMU D can increase its input values by CRS D while producing the same amount of outputs generated by the benchmark(dmu D can decrease its output levels while using the same amount of input levels consumed by the benchmark). Thus, CRS D 1 is a measure of input savings achieved by DMU D and CRS D 1 is a measure of output surpluses achieved by DMU D. For DMU G and DMU G, CRS G = 1 and CRS G =1 indicate that they achieve the same performance level of the benchmarkand no input savings or output surpluses exist. For DMU H and DMU H, CRS H 1 and CRS H 1

4 316 W.D. Cook et al. / Omega 32 (24) D A X2 D B H G C Y2 A H B G C (a) O X1 (b) O Y1 Fig. 1. (a) Input-oriented variable-benchmarkmodel; (b) Output-oriented variable-benchmarkmodel. indicate that ineciency exists in the performance of these two DMUs. Note that for example, in Fig. 1a, a convex combination of DMU A and DMU B is used as the benchmarkfor DMU D while a convex combination of DMU B and DMU C is used as the benchmarkfor DMU G. From Proposition 2, we can dene CRS 1or1 CRS as the performance gap between DMU new and the benchmark. Based upon CRS or CRS, a ranking of the benchmarking performance can be obtained. It is likely that scale ineciency may be allowed in the benchmarking. We therefore modify models (4) and (5) to incorporate scale ineciency by assuming VRS. Similar to Proposition 2, we have 1 Proposition 3. (i) VRS 1 or VRS 1 indicates that the performance of DMU new is dominated by the benchmark in (6); (ii) VRS =1 or VRS =1 indicates that DMU new achieve the same performance level of the benchmark in (6); (iii) VRS 1 or VRS 1 indicates that input savings or output surpluses exist in DMU new when compared to the benchmark in (6). However, unlike models (4) and (5), models (7) and (8) may be infeasible. 2 min max VRS jx ij 6 VRS xi new ; r ; jy rj y new j =1; j ; j E ; (7) VRS jx ij 6 xi new ; jy rj VRS yr new ; j =1; j ; j E ; (8) Proposition 4. (i) If model (7) is infeasible, then the output vector of DMU new dominates the output vector of the benchmark in (6); (ii) If model (8) is infeasible, then the input vector of DMU new dominates the input vector of the benchmark in (6). Proof. The proof follows directly from the necessary and sucient conditions for infeasibility in the super-eciency DEA model provided in [13]. The implication of the infeasibility associated with models (7) and (8) needs to be carefully examined. Consider Fig. 2 where ABC represents the benchmarkfrontier. Models (7) and (8) yield nite optimal values for any DMU new located below EC and to the right of EA. Model (7) is infeasible for DMU new located above ray E C and model (8) is infeasible for DMU new located to the left of ray E E. Both models (7) 1 In this case, j in (6) are determined by either model (7) or (8). 2 Model (5) is always feasible. Model (4) is infeasible only if certain patterns of zero data are present (see [15]). Thus, if we assume that all the data are positive, (4) is always feasible.

5 W.D. Cook et al. / Omega 32 (24) Y (output) Case I F Input-oriented benchmarking model is infeasible The benchmarking performance is indicated by output surpluses E E Case II C Case IV Output-oriented benchmarking model is infeasible The benchmarking performance is indicated by input savings Case III A B Case V (underperforming) E X (input) Fig. 2. VRS benchmarking model. and (8) are infeasible for DMU new located above E E and to the left of ray EF. Note that if DMU new is located above E C, its output value is greater than the output value of any convex combinations of A, B and C. Note also that if DMU new is located to the left of E F, its input value is less than the input value of any convex combinations of A, B and C. Based upon Proposition 4 and Fig. 2, we have four cases Case I: When both models (7) and (8) are infeasible, this indicates that DMU new has the smallest input level and the largest output level compared to the benchmark. Thus, both input savings and output surpluses exist in DMU new. Case II: When model (7) is infeasible and model (8) is feasible, the infeasibility of model (7) is caused by the fact that DMU new has the largest output level compared to the benchmark. Thus, we use model (8) to characterize the output surpluses. Case III: When model (8) is infeasible and model (7) is feasible, the infeasibility of model (8) is caused by the fact that DMU new has the smallest input level compared to the benchmark. Thus, we use model (7) to characterize the input savings. Case IV: When both models (7) and (8) are feasible, we use both of them to determine whether input savings and output surpluses exist DEA-based benchmarking model II: xed-benchmark model In the previous section, although the benchmarkfrontier is given, a DMU new under benchmarking has the freedom to choose a set of best-practice DMUs so that the performance of DMU new can be characterized in the most favorable light. Situations when the same benchmarkshould be xed are likely to occur. For example, the management may indicate that DMUs A and B in Fig. 1 should be used as the xed benchmark, i.e., DMU C in Fig. 1 should not be used in constructing the benchmark. To cope with this situation, we turn to model (2). Model (2) determines a set of referent best-practice DMUs represented by a set of binding constraints in optimality. Let set B = {DMU j : j I B} be the selected subset of benchmark set E, i.e., I B E. We then modify model (2) to the following linear programming problem: CRS = max u ryr new u ry rj u ry rj m v ix ij =; j I B; v ix ij 6 ; j I B; ; v ix new i =1; u r;v i : (9) By applying equalities in the constraints associated with benchmarkdmus, model (9) measures DMU new s performance against the benchmarkconstructed by set B.

6 318 W.D. Cook et al. / Omega 32 (24) Note that model (9) may be infeasible because the DMUs in set B cannot be t into the same facet when they number greater than m + s 1, where m is the number of inputs and s is the number of outputs. In this case, we need to adjust the set B. Three possible cases are associated with model (9). CRS 1 indicates that DMU new outperforms the benchmark. CRS = 1 indicates that DMU new achieves the same performance level of the benchmark. CRS 1 indicates that the benchmarkoutperforms DMU new. Finally, if we assume VRS in the benchmark, we have VRS = max u ryr new v u ry rj 4. An application v ix ij v =; j I B; u ry rj v ix ij v 6 ; j I B;j E ; m v ix new i =1; u r;v i and v free in sign: (1) In the nancial services industry worldwide, the traditional face-to-face customer contacts are being replaced by electronic points of contact to reduce the time and cost of processing an application for various products. To best respond to this new marketplace, the CBANK identied a need to conduct research into the design and delivery of nancial services by the most ecient and eective means while meeting internal operational performance goals. CBANK created a set of 12 e-business branches (thereafter, e-branches 3 ) using a new banking concept intended to create customer convenience with more ecient platforms for performing transactions. The e-branches are aimed at increasing the speed of service delivery and decreasing costs in signicant proportions through branch operation automation via internet, ATMs, telephone banking and other electronic means. From a business perspective, these e-branches are a result of application of technology toward the automation of business transactions and workows [16]. Fig. 3 presents the impact of e-business on banking performance. Based upon Harvey [17], sales eectiveness/customer satisfaction leads to better relationships with a bankbranch s current clients who are inclined to bring more of their business to it. This increases market share, 3 We code these e-branches as e1, e2, :::, e12. Sales Effectiveness Customer satisfaction Happier bankers Higher earnings e-business activities Better relations (e-branches) Lower costs/improved profits Productivity Fig. 3. The eect of e-business activities on banking performance. as does the inux of new customers who hear about the branch s legendary levels of service. Increased share means that more transactions are being processed, presumably with the same amount of xed cost. That, in turn, lowers unit cost and increases revenues, both of which lead to improved productivity and higher earnings. More earnings result in increases in bonus pool, higher merit increases, and a higher stockprice, which benets all shareholders and the workforce. This leads to happier banks. We should not forget that the reverse of the cycle is also true. If a bankbranch does not provide the level of service that people want, it will lose customers to the competition. Since the e-branches are in a new form of business and banking is done in very dierent ways, uncertainty surrounds the development of this new delivery model. Performing a benchmarking study is extremely critical to the CBANK in undertaking the e-business activities and to examine whether the e-branches enhance the productivity and sales eectiveness while reducing expenditures, i.e., the CBANK wants to examine whether the e-business activities have a positive impact on the business cycle presented in Fig. 3. Clearly, productivity in this case is characterized by a number of measures including, labor, information technology investment, transactions, and others. We develop two DEA inputs: FTE (full time equivalent) counts which include sales, service, support and other sta, and operating expenses which include spending on stationary, communications, shortage & losses, business development, employee training, advertising & publications, computer costs, and others. There are two types of transactions: sales and service. We select seven transactions (DEA outputs) presented in Table 1 that account for over 9% of the volume of sales and service related workcarried out Identication of benchmark frontier This section identies the best practice frontier of traditional bankbranches in each quarter from 1995 to The current study uses the real data from 1995 to 1996, since the CBANK had just started the service redesigning for the 12 branches. After that period, the CBANK had stopped the re-designing of these branches.

7 W.D. Cook et al. / Omega 32 (24) Table 1 Transactions and processing time Processing time (h) DEA Transaction description Min. Max. outputs 1 (Tran1) Set up a new collateral or non collateral loan 2 (Tran2) Open a new account (menu account) 3 (Tran3) Process a branch deposit to.44.7 menu account 4 (Tran4) Process a withdrawal from.4.55 menu account 5 (Tran5) Update passbookfor menu account in branch 6 (Tran6) Process visa cash advance (Tran7) Process a business deposit The identied best practice in each quarter is later used as the benchmarkfrontier to evaluate the quarterly performance of e-branches. Table 1 reports the minimum and maximum process times for the seven transactions. The minimum and maximum process times are used as lower and upper bounds for output multipliers in model (2) to develop the following weight restrictions as presented in (3). 1:475 6 Tran1 6 5:929; 11:271 6 Tran1 Tran 2 Tran :568; 14:345 6 Tran1 6 52:325; Tran 4 3:346 6 Tran :278; 7:812 6 Tran1 Tran 5 Tran :49; 8:576 6 Tran1 6 53:667: Tran 7 There are about 12 traditional branches within the CBANK. We will identify and use the ecient ones in each quarter as a benchmarkdata set. Both CRS and VRS DEA models with restrictions of (3) are applied to identify the quarterly best practice frontier Benchmarking the e-branches against the traditional branches We benchmarkthe 12 branches (e-branches) against the identied best practice of traditional branches in each quarter from 1995 to The last quarter of 1996 is regarded as the turning point, since the e-branches were created during the last quarter of Note that the best practice of traditional branches was changing quarterly. Thus, we here capture a dynamic picture of the performance change of these 12 e-branches. The results from model (4) indicate a dramatic performance change from the third quarter to the fourth quarter in 1995 when the automation is implemented: all Benchmarking Score Benchmarking Score Benchmarking Score Benchmarking Score CRS Variable-benchmark CRS Variable-benchmark CRS Variable-benchmark CRS Variable-benchmark e5 e1 95Q1 95Q2 95Q3 95Q4 96Q1 96Q2 96Q3 96Q4 95Q1 95Q2 95Q3 95Q4 96Q1 96Q2 96Q3 96Q4 95Q1 95Q2 95Q3 95Q4 96Q1 96Q2 96Q3 96Q4 95Q1 95Q2 95Q3 95Q4 96Q1 96Q2 96Q3 96Q4 e8 e11 Fig. 4. Representative performance change patterns when the e-branches are benchmarked against the traditional branches. the e-branches outperform the best practice of traditional branches. However, the performance of the e-branches decline into Based upon the optimal value to model (4), we classify the e-branches into four categories with respect to the performance change during 1996 (see Fig. 4): (i) branches, e1, e3, and e6 where the performance improves over the rst three quarters of 1996 and then declines; (ii) branches e2, e5, e7, and e12 where the performance improves from 96Q1 to 96Q2, then declines from 96Q2 to 96Q3, and then improves from 96Q3 to 96Q4; Q1 stands for the rst quarter of 1996.

8 32 W.D. Cook et al. / Omega 32 (24) Table 2 Benchmarking within e-branch: variable-benchmark model e-branches 95Q4 96Q1 96Q2 96Q3 96Q4 e e e e e e e e e e e e The numbers are optimal values to model (5). (iii) branches e4, e8 e1 where performance improves from 96Q1 to 96Q2 and then declines; (iv) branches e11 where the performance declines from 96Q1 to 96Q2, and then improves. Only ve e-branches, in the second quarter of 1996, and three e-branches in the third quarter of 1996 outperformed the best practice, i.e., the optimal value to model (4) CRS is greater than one. The majority of the e-branches did not show performance improvement compared to the best practice of traditional branches. Recall that model (4) assumes CRS, i.e., scale ineciency is not allowed. We therefore turn to models (7) and (8). Similar results are obtained. This indicates that scale is not a factor related to the productivity in the e-branches. Under the case of VRS, no infeasibility occurs in models (7) and (8). Since most of the e-branches are under-performing units compared to the benchmark, this indicates that most of the e-branches are of Case V in Fig Benchmarking within e-branches The previous analysis indicates that there was no productivity gain as a result of branch automation when the performance of e-branches is compared to the best practice of traditional branches. Recall that the 12 e-branches were re-engineered from 12 existing traditional branches which were under-performing units compared to the best practice in the rst three quarters in Thus, it might be dicult for the newly established e-branches to close the performance gap between the best practice and their predecessors. Therefore, we next study the performance change within the 12 branches, i.e., we compare the e-branches to the best practice of these 12 branches before the automation. First, we compare the e-branches in each quarter of 1996 to the best practice of e-branches in each quarter of The results from models (4), (7) and (8) show that (i) when the best practice of the e-branch predecessors in the rst quarter of 1995 is used as the benchmark, most of the e-branches outperform the benchmark, although the performance of most e-branches declines into the last quarter of 1996, and (ii) under the assumption of VRS, two e-branch moved from Case II into Case IV described in Fig. 2, indicating a productivity decline and most of the e-branches can be categorized by Case III in Fig. 2. Only one e-branch in the second quarter of 1996 is of Case I, representing the best scenario with lowest costs and highest performance. Overall, the performance of e-branches declines as the benchmarkis changed from the rst quarter to the last quarter of Next, we assume each branch in each quarter of 1995 represents a branch. Since automation happened during the last quarter of 1995, we exclude the branches in that quarter from the identication of best practice. Thus, we have 12 (branches) 3 (quarters in 1995) = 36 branches. We then benchmarkthe e-branches in the last quarter of 1995 and in each quarter of 1996 against the best practice of these 36 branches. 6 The e-branches in 1995 are referred to the 12 branches before the automation. 7 However, as pointed by one reviewer, since the conversion to e-branches tookplace during the last quarter of 1995, this particular quarter may be tainted with the eects of the conversion. Also, not all the branches are converted in the same day. As a result, quarter totals include a mix of new and traditional branches. On the other hand, it is possible that these branches were already producing at a higher level during 1995, and therefore there were no noticeable changes once e-branch conversion tookover. The one quarter performance jump can be due to the novelty aspect as customers would come and checkthings out.

9 Table 3 Benchmarking within e-branch: xed-benchmark model W.D. Cook et al. / Omega 32 (24) e-branches 95Q4 96Q1 96Q2 96Q3 96Q4 e e e e e e e e e e e e The numbers are optimal values to model (9). In these 36 branches, eight branches, namely, e1-95q1, e1-95q2, e3-95q1, e5-95q2, e6-95q2, e5-95q3, e6-95q3, and e1-95q3, are best practice branches. These eight branches are used as benchmarks in model (4). In model (9), we select e5-95q3, e6-95q3 and e1-95q3 as three xed benchmarks, since they represent the best practice right before the establishment of e-branches. Tables 2 and 3 report the benchmarking scores from models (4) and (9), respectively. These are optimal values to models (4) and (9). A larger value indicates a better performance. For example, under 96Q1 of Table 2, e1 has a score of when e1 is compared to the best practice of the traditional branches in the rst quarter of This indicates that e1 in the rst quarter of 1996 outperformed the traditional branches. For e2, the corresponding benchmarking score is.3942, indicating that e2 was dominated by the traditional branches. Overall, the performance of these e-branches declines from the rst quarter to the last quarter of When the variable-benchmarkmodel is used, the benchmarking performance of e3, e6 and e9 constantly declines. The performance of e1 which outperforms the best practice improves during the rst three quarters and then declines. Note that only one branch (e11) s performance declines, outperform the best practice, and then declines. The remaining seven branches show a performance improvement from the rst to the second quarter of 1996 and then show a constant performance decline with respect to the best practice. When the xed-benchmarkmodel is used, as expected the benchmarking scores decrease, implying a worse performance with respect to the best practice. The performance change of e1 and e11 remains the same patterns as those under the variable-benchmarkmodel. The benchmarking performance of e3, e4, e6 and e8 constantly declines. The remaining six branches show a performance improvement from the rst to the second quarter of 1996 and then show a constant performance decline with respect to the best practice. Finally, we note that an additional factor may contribute to the dramatic performance improvement when the CBANK launched the e-branches in the last quarter of There were some additional sta employed by the e-branches that were not reported as they were under the balance sheet of the head oce and not paid for by the e-branches. The CBANK phased out these unreported sta over the rst two periods (the last quarter of 1995 and the rst quarter of 1996). Thus, the number of sta in the last quarter of 1995 and the rst quarter of 1996 was understated. As a result, the resulting benchmarking scores should be decreased in the last quarter of 1995 and the rst quarter of Such adjustment indicates that the performance change of e-branches does not move in a favorable direction. The above analysis indicates that the e-business activities (establishing the e-branches) do not lead to an increased productivity. This empirical nding helps the CBANK to further examine its current e-business options. 5. Conclusions To aid the CBANK in undertaking e-business activities, the current study is directed at evaluating and benchmarking branch bankperformance. Two DEA-based benchmarking models are developed to study the change in performance that branches undergo when moving from the old to the new structure where transactions are automated. The study reveals that e-branches (new structure) did not exhibit productivity gain when compared to both the best practice of traditional branches and e-branches predecessors. This nding allows the bankto examine its business options, and gain an understanding of what does not workwell in terms of the makeup of new branches. This further can point to weaknesses and strengths in e-branch operations. The current study provides tools needed to monitor the performance change and further facilitates the development of the best strategic option for the organization with regard to branch makeup.

10 322 W.D. Cook et al. / Omega 32 (24) Finally, as one reviewer points out, one general issue that has not been addressed by banks and consequently cannot realistically be solved by this paper relates to the transactions that occur in ATMs versus human tellers and customer service representatives versus e-transactions in a bank. For example, if a customer goes to a teller to cash a cheque and have to wait on line and ll out the cheque, the customer may want to economize on his/her time and take a larger amount of cash out than he/she would take from an ATM. The multitude of ATMs has resulted in many taking out small amounts of cash frequently. This means more transactions but these added transactions are not accompanied by proportionately higher resource usage or costs to the bank. The cost of servicing an ATM is not proportional to the transactions. In that case, the total number of transactions will be greater with ATM transactions, so the cost per transaction and similar ratios cannot be used to determine if there is an increase in real eciently. This is clearly an issue for future researches. Acknowledgements The authors want to thanktwo anonymous reviewers for their helpful comments and suggestions on an earlier version of the paper. References [1] Lancaster H. Managing your career. Wall Street Journal, Tuesday, April 28, B1, [2] Camp RC. Business process benchmarking, nding and implementing best practices. Milwaukee, WI: ASQC Quality Press; [3] Hawdon D, Hodson M. The use of data envelopment analysis in benchmarking. The Business Economist 1992;27(3): [4] Sherman HD, Gold F. Bankbranch operating eciency: evaluation with data envelopment analysis. Journal of Banking and Finance 1985;9(2): [5] Paradi JC, Vela S, Yang Z. Assessing bankand bankbranch performance: modeling considerations and approaches. In: Cooper WW, Seiford LM, Zhu J, editors. Handbookon data envelopment analysis. Boston: Kluwer Academic Publishers; 23. [6] Sherman HD, Ladino G. Managing bankproductivity using data envelopment analysis (DEA). Interfaces 1995;25(2): [7] Oral M, Yolalan R. An empirical study on measuring operating eciency and protability of bankbranches. European Journal of Operational Research 199;46: [8] Thompson RG, Langemeier LN, Lee C, Lee E, Thrall RM. The role of multiplier bounds in eciency analysis with application to Kansas farming. Journal of Econometrics 199;46: [9] Charnes A, Cooper WW, Rhodes E. Measuring the eciency of decision making units. European Journal of Operational Research 1978;2: [1] Charnes A, Cooper WW, Huang ZM, Wei QL. Cone ratio data envelopment analysis and multi-objective programming. International Journal of Systems Science 1989;2: [11] Cooper WW, Seiford LM, Tone K. Data envelopment analysis: a comprehensive reference text with models, applications, references, and DEA-Solver software. Boston: Kluwer Academic Publishers; 2. [12] Andersen P, Petersen NC. A procedure for ranking ecient units in data envelopment analysis. Management Science 1993;39: [13] Seiford LM, Zhu J. Infeasibility of super-eciency data envelopment analysis models. INFOR 1999;37: [14] Seiford LM, Zhu J. An acceptance system decision rule with data envelopment analysis. Computers and Operations Research 1998;25(4): [15] Zhu J. Robustness of the ecient DMUs in data envelopment analysis. European Journal of Operational Research 1996;9: [16] Kalakota R, Whinston AB. Electronic commerce: a manager s guide. Reading, MA: Addison-Wesley; [17] Harvey T. The banking revolution: positioning your bank in the new nancial service marketplace. IRWIN, A Bankline Publication; [18] Zhu J. Quantitative Models for Performance Evaluation and Benchmarking: Data Envelopment Analysis with Spreadsheets. Boston: Kluwer Academic Publishers; 22.