The Economic Value of Network Externalities in an Electronic Inter-Bank Payment Network: An Empirical Evaluation

Eduardo S. Jallath-Coria Banco de Mexico 5 de Mayo 1, cuarto piso Mexico 06059, D.F. [email protected]

Tridas Mukhopadhyay Sandra Slaughter* Graduate School of Industrial Administration Carnegie Mellon University Pittsburgh, PA 15213 [email protected] [email protected]

Amir Yaron The Wharton School University of Pennsylvania and NBER [email protected]

August 1, 2001 * Contact Author for this paper Acknowledgements: We thank Alberto Espinosa, Mark Fichman, Margarita Moleres, Jose Luis Negrin and participants in seminars in Carnegie Mellon University, Banco de Mexico and the Workshop for Information Systems and Economics in Brisbane, Australia for helpful comments and discussions. Any remaining errors are our responsibility.

*** Do not cite, quote, or distribute without permission of the authors ***

i

The Economic Value of Network Externalities in an Electronic Inter-Bank Payment Network: An Empirical Evaluation

Abstract Theoretical work ascribes a positive value to the externalities arising from the adoption of network technologies. However, few studies have attempted to quantify the business value of these externalities. As a result, relatively little is known about the economic impact of externalities on firms that adopt network technologies. In this study, we investigate the value of an electronic inter-bank payment network in terms of its effect on the reserve management performance of commercial banks. Performance is measured by the opportunity and penalty costs generated by balances on the reserve account that commercial banks hold at the central bank. We propose that the electronic inter-bank payment network enhances banks’ reserve management performance (i.e., reducing opportunity and penalty costs) by providing more timely information on deposits and withdrawals affecting the banks’ reserve accounts. Technology impact is characterized as an initial stand-alone effect and a network externalities effect as more banks join the network. To evaluate the economic value of the technology impact, we analyze data from the implementation of an electronic inter-bank payment network adopted by all Mexican commercial banks. We find that early adopters of the electronic network, with a low ratio of electronic to overall operations, experience increasing opportunity and penalty costs. However, as additional banks join the network, the ratio of electronic operations increases, and costs decrease. After all banks have adopted the electronic network, we observe a reduction equivalent to 9.9% of each bank’s opportunity and penalty costs. The aggregated savings for all banks equal $5.3 million dollars for the 6 months following the technology adoption. Overall, the electronic inter-bank payment network project provides a significant positive net present value. Our findings provide support for the network externalities hypothesis. They also suggest an initial negative performance impact when there is concurrent use of old and new network technolo gies.

Keywords : Network Externalities; Business Value of Information Technology; Electronic Banking; Information Technology Investment; Economic Analysis.

ii

The Economic Value of Network Externalities in an Electronic Inter-Bank Payment Network: An Empirical Evaluation 1. Introduction Telecommunication technologies and networked applications have become increasingly prevalent. For instance, firms develop Electronic Data Interchange (EDI) links to share information with their network of customers and suppliers. In the financial services sector, banks offer direct connections to their corporate customers through PC banking applications and, more recently, via Internet banking services. Similarly, the World Wide Web is encouraging the creation of applications that support information sharing among different firms. These technologies have implications beyond the traditional transaction-processing gains. More timely and accurate information can enable a firm to improve its decision making process. Moreover, the benefits of the system are often positively related to the number of firms using the system, because an increase in the number of users can increase the available information, generating network externalities. Externalities are an important issue in the adoption of standards and network technologies (Katz and Shapiro 1986). For instance, word processor applications, such as Microsoft Word, derive a great part of their value from the number of compatible users of the product (Brynjolfsson and Kemerer 1996). Similarly, network applications such as electronic mail are more valuable when the number of users increases. Some studies have theoretically analyzed network externalities (Farrell and Saloner 1986, Saloner and Shepard 1995); others have empirically exa mined the likelihood of network adoption (Kauffman et al. 2000). However, few studies have attempted to empirically quantify the economic value of network externalities. Thus, it is not possible to address important questions concerning the strength of the network effect and whether the externalities effect is uniform across all users in the network. Our objectives in this study are to measure the performance impact on a firm when it adopts a network technology, to determine whether this impact can be attributed to network externalities, and to assess whether the externalities effect is uniform across all firms in the network. We hypothesize that, ceteris paribus, the adoption of a network technology and the ensuing network externalities increase the information set and thereby improve the performance of each firm in the network. To test this hypothesis we analyze detailed data on operations in an

1

inter-bank payment network operated by the Central Bank of Mexico (Banco de Mexico) and used by all commercial banks in the country. The Central Bank provides services to commercial banks similar to those provided by commercial banks to their corporate customers. Commercial banks have a reserve account that is used not only to fulfill reserve requirements but also to perform inter-bank transactions. Positive balances in this account pay no interest, generating an opportunity cost for the bank. Negative balances are charged a high penalty rate, generating an overdraft cost. Thus, excess balances may be necessary to avoid expensive overdrafts generated by uncertain operations. During the period of analysis, the banks adopted an electronic interbank payment network that allowed them to perform on- line fund transfers and queries on the reserve account balance, thereby improving the information used to forecast uncertain transactions affecting reserve balances. A unique feature of this setting is that the adoption of the network technology by the commercial banks occurred sequentially permitting us to directly analyze the costs, benefits, and externalities derived from network adoption for each bank. We find that early adopters of the electronic network, with a low ratio of electronic to overall operations, experience an initial decrease in performance, i.e., their opportunity and penalty costs rise. However, as additional banks join the network, the ratio of electronic operations increases, and costs decrease. Once all banks have adopted the electronic network, we observe a reduction equivalent to 9.9% of each bank’s opportunity and penalty costs. The aggregated savings for all banks are equivalent to $5.3 million dollars for the 6 months following system adoption. Thus, our results provide support for the network externalities hypothesis. We also find an initial negative performance impact when there is concurrent use of old and new network technologies. This paper is organized as follows. The second section reviews the related literature on network adoption and externalities. Section three develops the research model for the study. The fourth section describes the general characteristics of the Mexican financial system, the interaction of financial institutions with the central bank, the implementation of the inter-bank payment network, and the dataset. Section five describes the estimation of the model, and presents and discusses the results. The last section summarizes the contributions of this work and suggests opportunities for further research.

2

2. Related Literature The literature on externalities has characterized the role of network externalities in standards (Farrell and Saloner 1985, 1986) and in network adoption (Katz and Shapiro 1985, 1986). In the theoretical work on standards, Farrell and Saloner analyzed whether standardization benefits could lead to the adoptio n of an inferior standard when a better alternative was available. They found that the adoption of an inferior standard could not occur in the presence of complete information and also derived the conditions under which users would switch from an existing technology to a new one. The work of Farrell and Saloner is relevant to our study because it shows that users can find it optimal to adopt a new network technology, despite inertia in the use of an old technology. In the empirical work on standards, Gandal (1994) found support for the externalities hypothesis by estimating a hedonic price equation for spreadsheet programs. He showed that customers are willing to pay a significant premium for spreadsheets that are compatible with the Lotus 1-2-3 platform. Similar findings were reported by Brynjolfsson and Kemerer (1996) who measured externalities in terms of a product’s installed base. These studies attempt to characterize the economic value of externalities; however, the focus is on standards, not on value. In the research on network adoption, the general finding is that network externalities serve to promote the rate and likelihood of technology adoption. For example, using data on the estimated fax installed base, Economides and Himmelberg (1995) showed tha t the speed of adoption increases in the presence of network externalities. Similarly, Saloner and Shepard (1995) examined data on banks’ adoption of ATMs and found that the likelihood of adoption increases with the number of branches served (network effect) and the number of users (scale economies effect). Using a hazard model to evaluate the likelihood of adoption of a shared ATM network, Kauffman et al. (2000) found that banks that can generate a larger network size and a higher level of externalities tend to adopt ATM networks earlier than banks that own a large branch network. Finally, Gowrisankaran and Stavins (1999), examining data on Automated Clearing House (ACH) payments, found that network externalities increased the rate of ACH adoption. To summarize this work, there are but a few studies that measure the economic value of externalities, and these studies have focused on standards. Research on externalities for physical

3

networks has concentrated on the theoretical aspects of the problem, leaving aside the empirical measurement of the externalities. The few empirical studies in this area have examined the influence of network externalities on network adoption without attempting to assess the economic value of the externalities. As a result, relatively little is known about the economic impact of externalities on firms that adopt network technologies. A major contribution of our work to the literature in this area is to develop a general approach to modeling and evaluating the value of externalities ensuing from network technology adoption. 3. Theoretical Basis and Model In this section we define a model to evaluate the impact of network technology adoption on firm performance. Specifically, we model the impact of adoption of an inter-bank payment network on the opportunity and penalty costs (OPC) that banks incur in managing balances on their reserve account. The model proposes that, controlling for the main factors affecting performance, a bank can improve its performance (i.e., reduce its OPC) conditional on improvements in the information generated by adopting the network technology. We start this section by describing a general model to account for technological impact. The model is then detailed to address the specific technology examined in our study – an inter-bank payment network - and the implications of network adoption. The model is further developed to characterize the effect of both the adoption process and the network externalities. Finally, we enhance the model to control for changes not attributed to technological impact. 3.1. General Model of Technological Impact To account for the impact generated by the introduction of technology, we propose an augmented version of the classical growth-accounting framework developed by Solow (1957). The central equation of the Solow model is as follows: Y=A(•) f(K,L)

(1)

where Y is output, A(•) measures the impact of technology, and K and L are capital and labor. In this paper, we tailor the Solow model in two ways. First, we make explicit the functional form of the technological impact A(•). That is, we characterize the impact of technology adoption not by portraying a time trend but by depicting an initial stand-alone effect and a subsequent network

4

externalities effect as more users adopt the network. Second, we describe the input factors that affect the cost minimization process f(•) performed by a firm. The central hypothesis of our study is that, ceteris paribus, improvements in the information set generated by the adoption of network technology improve firm performance (reducing costs). This implies that, for a given set of inputs, the adoption of an electronic network allows a firm to improve its cost minimization process. Considering the structure suggested by the Solow model (1) and assuming a neutral technological change, 1 we can represent the effect of network technology adoption by a function A(•). This function depends on an observable vector of variables Z and a time invariant vector of parameters θ 1 . Similarly, we can characterize the initial OPC (i.e., before the network technology adoption) as a function f(•) that depends on an observable vector of time varying variables X, and a time invariant vector of parameters θ 2 . Thus, we can propose that, assuming no effects for technological change, the opportunity and penalty costs follow a normal distribution with conditional mean β'X and conditional variance σ2 , where {β ,σ2 }∈ θ2 . Consequently, if we put together the factors that account for technological impact A(•) and the functional form of the cost minimization process f(•) we have: OPC = A(Z; θ 1 ) f(X; θ 2 )

(2)

A graphical description of the role of the technological impact A(•), generated by the adoption of an electronic payment network, and the cost minimization process f(•) can be seen in Figure 1. The functional details of each factor are described in the following two sections. Figure 1 Factors Characterizing the Opportunity and Penalty Cost

Cost Minimization Process f( X ,θ 2 )

Opportunity and Penalty Costs OPC

Payment Network Adoption Effect A( Z , θ 1)

3.2. Technology Adoption In this section we depict the details of the technological impact A(•). We start by describing the general characteristics of a two-way payment network and how its users move from paperbased payments to electronic payments. Then we describe how adding new banks to the network creates network externalities for the banks that have already subscribed. Finally, we adapt the

5

functional form of the network externalities model proposed by Farrell and Saloner (1986) to incorporate it into the functional form of equation (2). 3.2.1. Payment Networks. Payment networks facilitate payments by concentrating the payment process in a switch. The switch has one account for every subscriber so that payments can take place by transferring funds from one account to the other. When a user wants to perform a payment, he or she refers the payment to the switch. The switch debits the user’s account and credits the account of the beneficiary. Then, the switch notifies the beneficiary about the amount of the payment. The above interaction can be framed in the context of a two-way network in which there is an interchange of information flows sent and received by users using a switch (Economides 1996). From a user stand point, there are two information flows: 1) the flow of the user sending the payment request to the switch (outgoing payment flow), and 2) the flow of the payment notification coming from the switch (incoming payment flow). In an inter-bank payment network, commercial banks are the users of the network, and the role of the switch is played by the central bank. Figure 2 provides an illustration of a two-way inter-bank payment network in which bank 1 sends a payment to bank 2. Figure 2 Two-way Network

bank 1 Outgoing payment flow

bank 2

bank 3

bank 4

Incoming payment flow

switch (Central Bank) The flows between the banks and the central bank can take place by sending paper-based documents or by sending electronic messages. In the paper-based mode, payments are performed by banks completing forms (similar to checks) that are sent by courier to the central bank (outgoing payment flow). The central bank processes each form by debiting the account of the payer and crediting the account of the payee. Once all forms have been processed, the central

6

bank produces a printed statement with the details of the operations of the day. The beneficiary bank receives by courier the printed statement with the payment information (incoming payment flow). In the electronic mode, banks have computer clients connected to the computer server of the central bank. The computer client allows every commercial bank to key in the payment requests into the system (outgoing payment flow). Once a payment has been keyed into the system, the computer server at the central bank credits the account of the payer, debits the account of the payee and updates the respective balances in real time. The beneficiary bank, using the computer client, can inquire on- line about the updated details of payment information (incoming payment flow). We can describe the adoption of an electronic inter-bank payment network in the context of a set of banks switching from a network of paper-based flows to a network of electronic flows. We expect that banks find it optimal to switch to the electronic network because of improvements in the level of aggregation and timeliness of their information flows (Ahituv 1989; Barua et al. 1989). The electronic network improves the level of aggregation of information by providing an on- line facility with a summary of incoming and outgoing payment flows. That is, the system provides a list of the operations generated by all of the bank’s departments as well as a list of the operations generated by other banks. Similarly, it improves the timeliness of information by speeding up information about incoming flows generated by the banks adopting the electronic network. Thus, a bank switching to an electronic network of size N experiences two effects. The first occurs when all the departments of a bank adopt the electronic network, and as a consequence all outgoing payment flows become electronic and are summarized into the computer system (improved level of aggregation). The second is a cumulative effect of speeding up incoming flows that depends on other banks adopting the electronic network (improved timeliness). Since we have two separate effects and the second effect depends on other banks joining the network, the effects for banks adopting the network will vary through time. The initial sequence of adoption is as follows: •

The first bank that switches to the electronic network will have all of its outgoing

payment flows electronically sent to the central bank. The system at the central bank will

7

summarize these operations (improved level of aggregation). However, the first bank will receive the information about incoming flows using the next period printed statement because the other N-1 banks have not yet joined the electronic network. Similarly, the N-1 banks experience no changes in their paper-based payments flows with the central bank. •

The second bank that joins the electronic network will send all of its outgoing flows to

the central bank electronically (improved level of aggregation) and it will receive electronic payment flows generated by the first bank that adopted the network (improved timeliness). The first bank that adopted the network will receive electronic incoming payment flows from the second bank (improved timeliness). However, the first and the second banks will receive the remaining incoming flows using the paper-based procedures because the other N-2 banks have not joined the network. The N-2 banks keep operating their paper-based flows with the central bank without changes. The adoption of the electronic network continues until all N banks adopt the electronic network and all incoming and outgoing flows with the central bank are performed using the new technology. An important point to highlight is that the summary of outgoing payments is complete from the first day of network adoption but the incoming transactions are not electronic until all banks have adopted the network. This gradual adoption of the electronic network forces banks to use two technologies. That is, during the transition period, banks must deal with incoming electronic transactions coming from the banks that have joined the network and with incoming paper-based transactions coming from the other banks. The concurrent existence of the two network technologies contrasts with the model presented by Farrell and Saloner (1986) in which the adopters of the new technology generate negative externalities to the old network installed base. In this context, the adopters of the new technology do not affect the utility of the users of the old technology but instead must deal with a transition period in which they concurrently use both old and new technologies. The problem of the dual use of technologies is eliminated once all banks have subscribed to the electronic network. 3.2.2. Network Externalities. Theoretical literature on network externalities suggests that the value of network adoption derives from two sources: the initial adoption effect and the

8

externalities effect based on the number of subscribers in the network (e.g. Farrell and Saloner 1986; Kauffman et al. 2000). The basic representation is given by: sa+ne(n)

(3)

where sa is the stand-alone effect that a firm derives from adopting the technology and ne(n) represent the network externalities effect derived when n firms have joined the network. 2 In the context of an electronic inter-bank payment network, the stand-alone effect sa is characterized by the switch to electronic mode of the outgoing payment flows, whereas the network externalities effect ne(n) signifies the automation of the incoming payment flows. If the focus of the analysis is on the information set of a bank, the benefits of the stand-alone effect are characterized by the improvements in the level of aggregation of outgoing payment flows, whereas the benefits of the network externalities derive from improvements in the level of aggregation and timeliness of the incoming payment flows. In an inter-bank payment network, the improved level of aggregation and timeliness directly relate to the amount of information generated in the system (i.e., the number of payment flows). Expression (3) assumes that all n banks generate a homogeneous number of payments. When we relax this assumption, and allow for heterogeneity in the number of payments per bank, the adoption benefits will depend on the relative number of payment transactions rather than on the number of banks. Thus, the adoption effect for a given bank can be expressed as: sa+ne(tr/TR)

(4)

where tr is the number of incoming transactions (incoming payment flows) generated by banks that have switched to the new network and TR is the total number of incoming transactions for a given bank. Therefore, we hypothesize that the improvement in the information set available to each bank is a function of the transactions generated by the initial stand-alone adoption effect and the network externalities generated by other banks in the network. We assume that a bank will use the new information to improve the forecast used in its cost minimization process. In this context, we can express the stand-alone and network externalities effects as the factors that constitute the technological impact on the OPC process, or A(•). The symbolic representation is expressed in (5) as follows:

9

OPC=A(sa+ne(tr/TR);θ 1 ) f(X;θ 2 )

(5)

3.3. Opportunity and Penalty Costs (OPC) In this section we depict the details of the cost minimization process f(•). We start by describing the general characteristics that generate a bank’s OPC. Then, we define the set of input factors that affect the cost minimization process. Reserve requirements and the provision of inter-bank payments require that commercial banks keep a reserve account at the central bank. As with any account, the reserve balance can be positive or negative. A positive balance may pay a low or zero interest rate for the bank, thus generating an opportunity cost. A negative balance results in a high interest rate charged on the overdraft amount creating a penalty cost. It would be optimal for a bank to keep the exact amount of money needed to pay for the transactions that take place during the day; that is, to keep a zero end-of-day balance. However, to avoid expensive overdrafts, a bank must keep some excess reserves to allow for uncertain operations that occur during the day and in particular, overnight. 3 Uncertain operations generate informational constraints under which banks try to forecast a balance to minimize the opportunity and penalty costs for the reserve account. In other words, banks try to minimize costs conditional on the available information. We can characterize the conditional mean of the cost minimization process (β'X) by the exogenous random transactions affecting the reserve balance and some exogenous and endogenous factors affecting the banks. In this context, the opportunity and penalty costs depend upon the value and volume of the uncertain transactions affecting the balance. For instance, a bank performing payments of high value might end up with a higher outstanding balance than a bank performing payments of low value. In addition, some endogenous factors depict the characteristics of the firm. For example, a bank with a network of nation-wide branches might experience more logistical problems aggregating information than a regional bank. The lack of aggregation might be reflected in higher costs. Exogenous factors can also reflect the financial and economic environment. For instance, if the prevailing rates in the market in a given day are higher than the rates on a subsequent day, banks will end up with different costs. The above descriptions suggest different levels of factors affecting the OPC of each bank: Transaction Factors (TF), Firm Factors (FF) and Economic Factors (EF). Transaction Factors

10

include the magnitude of incoming and outgoing operations that affect the bank's reserve account. We can split them into daylight and overnight operations. Daylight operations affect OPC because they are subject to operational errors, omissions and delays. Overnight operations are generated by third parties and create random amounts unknown to the bank. Firm Factors reflect intrinsic characteristics of a bank and therefore have an indirect impact on the costs. For instance, the size of a bank is an indication of its operational complexity. A large number of branches can be an indication of economies of scale or the level of geographical dispersion. Thus, firm-specific characteristics will either facilitate or complicate the processing of information used in the cost minimization process. Economic Factors have a direct impact on the costs. Higher interest rates will generate higher opportunity and penalty costs, ceteris paribus. Similarly, changes in the monetary base will generate changes in the balances of banks affecting cost. Hence, the conditional mean of the OPC process (β 'X) depends on a time variant vectors of variables {TF, FF, EF} ∈ X and a time invariant vector of parameters β ∈ θ 2 . The details of the elements included in the set of factors {TF, FF, EF} are described in the next section of this paper. To summarize, we have proposed a general model to evaluate whether the adoption of a network technology can improve firm performance, i.e., help banks reduce the opportunity and penalty costs generated by excess balances in their reserve accounts. We hypothesized that conditional on improvements in the information set derived from the electronic network, firms can improve their cost minimization process. We modeled the changes in the information set as an initial stand-alone network effect and a subsequent network externalities effect generated when more users adopt the network. Finally, to characterize the cost minimization process, we identified three levels of factors (transaction, firm, and economic factors) affecting costs. We now elaborate upon this general model by incorporating elements relevant to our empirical context. 4. Empirical Context, Measures and Data In this section we begin with a general description of the Mexican financial system, followed by a description of inter-bank operations. Then, we describe the implementation process of the electronic inter-bank payment network, and finally we depict the characteristics of the data set.

11

4.1. The Mexican Financial System In 1990, Mexico’s financial system included a Central Bank and 19 commercial banks. The Central Bank of Mexico (Banco de Mexico) offers banking services to commercial banks similar to those offered by commercial banks to their corporate customers. The main vehicle for settling inter-bank operations is the reserve account that banks have at the Cent ral Bank. Commercial banks pay and receive funds in the reserve account from 9 AM to 5 PM. The balance of the account changes as different departments of the bank perform operations that require electronic (or paper-based) payments. However, just before 5 PM, the bank has to decide how much money to leave in the account to support the overnight withdrawals generated by the settlement of the clearinghouse. If the end-of-day balance is positive, there is an opportunity cost. If the end-ofday balance is negative, a penalty is charged on the account. 4 During the first six months of our study, the aggregated outstanding balance of the 19 commercial banks was equivalent to $676 million dollars. This balance generated a combined holding and penalty cost equivalent to $87.7 million dollars. Consequently, an improvement in the management of the reserve account has the potential to generate significant benefits. Figure 3 provides an illustration of the intra-day behavior of the reserve account of a bank. Figure 3 Illustration of intra-day behavior of the balance of the reserve account of a bank Balance

Daylight Operations

9 AM

Overnight Operations

5 PM

12 PM

time

4.2. Inter-bank Operations The operation of inter-bank payments is relatively uncomplicated. For example, if Bank A buys a bond from Bank B, Bank A generates a payment from its reserve account to the reserve account of Bank B. The settlement of the payment is performed by the debit of the reserve account of Bank A and the credit of the reserve account of Bank B. Prior to adoption of the

12

electronic inter-bank payment network, fund transfers were performed by banks completing forms (similar to checks) that were later sent by courier to the Central Bank. When the forms reached the Central Bank, they were keyed into the computer system that controlled the reserve accounts. The system processed the operations and automatically credited and debited the reserve account of each bank in the Central Bank’s general ledger. At the end of the day, the computer system printed a balance statement showing the operations performed during the day. This statement was sent to each financial institution by courier the following day. Despite the existence of a computer system in the Central Bank, inter-bank operations were paper-based, because commercial banks had no direct access to the Central Bank’s computer system. Inter-bank operations affecting the reserve account can be divided into daylight and overnight operations. Daylight operations include fund transfers (used to settle foreign exchange and money market operations), federal tax related operations, and cash deposits and withdrawals. 5 Overnight operations were performed after all banks had closed and included the settlement of the check clearing houses. 6 Once daylight and overnight operations have been included, the final balance on the reserve account generates the respective opportunity or penalty cost. Although the optimal strategy for a bank is to leave a zero balance at the end of the day, possible errors, delays or omissions in the processing of daylight operations and uncertainty about the value of overnight operations make this goal infeasible. 4.3. Electronic Payments In 1990 the Mexican Central Bank started the implementation of an electronic inter-bank payment system to provide an on- line connection to every institution bearing an account at the central bank. The system enables financial institutions to switch from paper-based to electronic mode for the following operations: •

Fund transfers from reserve accounts (pesos checking account) to the accounts of other

banks, •

Fund transfers from U.S. dollar accounts to the accounts of other banks,

•

Government-bond transfers to the accounts of other banks, and

•

REPOs and direct purchasing and selling of government bonds to the central bank. 7

13

In addition, the electronic inter-bank payment system provides on- line queries to display the operations affecting the reserve account during the day. For instance, a screen could show outgoing payments keyed into the system by different divisions of the bank as well as incoming payments keyed into the system by other institutions. Hence, the system improves the level of aggregation by providing an automatic list of all operations affecting the bank's account (standalone effect). Similarly, the electronic inter-bank payment network improves the timeliness by providing on- line information of the incoming operations generated by the other banks (network externalities effect). In our analysis, we focus on examining the network technology impact on the reserve accounts of the banks in the network. We concentrate on the reserve account because it is the most significant in terms of value and number of operations. Moreover, the reserve account is the only account that is subject to the uncertainty of overnight operations and therefore makes the proposed evaluation more meaningful. 4.4. Implementation Process The implementation of the electronic inter-bank payment system took place from November 1990 to April 1991. We can distinguish three phases that characterize the process: the first, when the banks were only operating with paper-based payments; the second, when the banks were gradually switching to electronic payments; and the third, when all banks were operating using the electronic inter-bank payment network. A graphical representation of the system implementation process can be seen in Figure 4. Figure 4 Implementation Process 6 months

6 months

6 months

Bank 1 Bank 2 Bank 3 Bank 4 Bank 5 Bank 6

Bank 19 paper - based

Implementation

electronic - based

14

This implementation process has several characteristics of a quasi-experimental design (i.e., an experiment in which the researcher cannot control subject assignment to the stimuli): •

All banks adopted the same information system. This is important because, in general,

firms adopt different technologies making it difficult or impossible to conduct technology adoption evaluations across firms. •

For every bank, we have the equivalent of a within-subject interrupted time series design

in which we measure performance both before and after the implementation of the system. •

During the six- month implementation phase, the banks adopted the system sequentially.

About once a week, a new bank was added to the network while the remaining banks used paperbased operations. This strategy provides a between-subjects design for a given time t during which some banks have adopted the system while controlling for other banks that have no access to the electronic inter-bank payment network. In addition to the controls provided by our quasi-experimental design, we can control for other relevant aspects of the environment. First, the reserve account was the only account that banks used to settle inter-bank operations. Hence, there was no alternative account that we could not monitor. Second, the implementation of the system changed neither the rules for inter-bank transactions nor the deposit reserve requirements. For instance, trading practices and operational rules remained the same. Third, paper-based and electronic inter-bank payments were restricted to banks. Third parties such as companies and individuals did not have access to either paperbased or electronic inter-bank payments. Fourth, we have access to a wide set of variables to control for changes in the environment. For instance, we can characterize the magnitude of daily operations affecting the reserve account as well as exogenous factors such as inflation and interest rates. The disadvantage of this and every quasi-experiment is that the adoption sequence (i.e., assignment to the experiment) is not random. The lack of randomness can make it difficult to protect against some systematic bias. For instance, less efficient banks could adopt the system before more efficient banks. However, in our study, banks freely decided when they would adopt the system. Thus, despite the absence of random assignment, every subject had an equal chance to be assigned during the implementation process. Further, as we describe in our subsequent econometric analysis, we control for the possible impact of adoption sequence on performance.

15

4.5. Data The data we used in this study were collected from the electronic archives of the Central Bank of Mexico and from reports in the Bank´s supervisory office (Comision Nacional Bancaria). 8 The data include all daily reserve account operations from May 1, 1990 to October 31, 1991. We examine 6 months of daily operations before the adoption of the system, 6 months of daily operations during imp lementation, and 6 months of daily operations after all banks were operating electronically. Observations include both longitudinal (i.e., time-varying) and crosssectional (i.e., time-invariant) data. After eliminating holidays and weekends, there are 377 working days. This provides us with 377 time series observations for a cross section of 19 banks to yield 7,163 data points. In the following sections, we explain the details of each variable in the data set. 4.5.1. Opportunity and Penalty Costs (OPC). The observed OPC value is based on the ex-post observation of the end-of-day balance of the reserve account. The end-of-day balance is computed using the initial balance (Y) and the net of daylight (DL) and overnight (ON) operations. The opportunity and penalty costs incurred in the management of the reserve account are the result of multiplying the end-of-day balance by the given market rate of deposit for overnight funds (RD) if the balance is positive; or by the given penalty rate (RP) in the event of an overdraft. Thus, the accounting relationship of a bank’s OPC is: OPC = RD (Y + DL + ON )1( Y + DL +ON )≥0 − RP (Y + DL + ON )1(Y + DL+ ON ) <0

(6)

where 1 is an indicator function that equals “1” if the expression is true or equals “0” otherwise. The OPC and all other variables representing amounts were deflated to pesos as of May 1990 using the consumer price index. Then, using the peso to U.S. dollar exchange rate, we converted the amounts to equivalent U.S. dollars as of May 1990. A logarithmic transformation was used to correct for the skewed distribution of all transaction variables affecting the OPC. 4.5.2. Technology Adoption. We have presented a general model to characterize the impact of network technology adoption on a firm’s cost minimization process. The technological impact A(•) is characterized by the initial stand-alone effect of adoption sa and the network externalities generated by the transactions of other users in the system ne(tr/TR). To operationalize these

16

effects we use two variables. The first represents the stand-alone (SA) effect. The second represents the cumulative effect of the network externalities (NE). SA (Stand Alone) – represents the stand-alone benefit sa obtained by a bank when it joins the electronic network. We use a dummy variable to identify the adoption of the network. The variable has the value of 1 when the electronic network has been adopted and zero otherwise. The coefficient on this variable indicates the percentage change in OPC attributed to the standalone effect. We have hypothesized that improvements in the level of aggregation of the information set generated by the initial adoption of the network will reduce OPC implying a negative sign on the coefficient of this variable. NE (Network Externalities) - represents the ratio of the number of incoming electronic payments to the total number of incoming payments for a specific bank. 9 This variable has a value of zero when no electronic payments exist. During the implementation period, the ratio has a value that ranges from 0 to 1. When all banks have adopted the network the ratio equals 1. The coefficient on this variable shows the percentage change from having all incoming transactions in the system. We have hypothesized that the improvements in the timeliness and level of aggregation of the information set generated by the adoption of the network by other banks reduces OPC implying a negative sign on the coefficient of this variable. A hypothetical example of how the value of the NE variable evolves for the third bank that adopts the new network technology in a network of size N=6 banks can be seen in the row “Ratio of Incoming Electronic Payments to Total Incoming Payments” of Table 1. In this table, we can see that Bank 3 adopts the system in the third period (t=3). Before this time it can only operate with paper-based transactions. Thus, at t=0, t=1, and t=2, Bank 3 has no incoming electronic transactions, and the proportion of its electronic payments is 0/5 (NE=0.00). At t=3, Bank 3 joins the network after Banks 1 and 2 have already adopted the network, implying that Bank 3 has a proportion of electronic transactions of 2/5 (NE=0.40) in this period. The proportion of Bank 3’s incoming electronic transactions keeps increasing until it reaches 5/5 (NE=1.00) when all six banks have adopted the network at t=6. It is important to note that the banks in this example generate the same number of transactions. At our research site, banks generate a different number of transactions, and therefore the ratio of incoming electronic payments will vary according to the volume of operations that every bank performs.

17

Table 1 Ratio of incoming electronic payments for Bank 3 adopting a network of size N=6 Time Incoming Payment Flows Network Externalities Ratio of Incoming Effect Electronic Payments to Total Incoming Payments Banks Subscribed to the Electronic Network

t=0

t=1

t=2

t=3

t=4

t=5

t=6

0.00

0.00

0.00

0.40

0.60

0.80

1.00

0/5

0/5

0/5

2/5

3/5

4/5

5/5

0

1

2

3

4

5

6

bank 1 bank 2 bank 4 bank 5 bank 6

Incoming Paper-Based Payment Flows Incoming Electronic Payment Flows

4.5.3. Transaction Factors, Firm Factors, and Economic Factors. This subsection describes the variables in the three level model we use to characterize the cost minimization process f(X,θ 2 ). It includes the Transaction Factors, the Firm Factors and the Economic Factors {TF, FF, EF} ∈ X. Transaction Factors . Transaction Factors include operations affecting the balance of the reserve account. These variables can be subject to operational errors, delays or omissions, and therefore have a direct impact on the forecast of the end-of-day balance. This vector of variables includes fund transfers (FT), cash deposits and withdrawals (CASH), federal tax related operations (TAX), and the settlement of the check clearinghouse (CLH). Each variable has different operational characteristics and therefore influences OPC differently. It is important to point out that the end-of-day balance in equation (6) is computed using the net of withdrawals and deposits. The variables that we describe in this section do not represent the net value but rather the overall sum of withdrawals and deposits. We use the sum because we consider that it is the magnitude of the overall operations that has an effect on OPC. FT (Fund Transfers) – these operations are used for the settlement of money market and foreign exchange operations. Incoming payments are performed by other banks in the network. A large number of transfers, in particular incoming transfers, can be an operational burden for the bank. Thus, we expect a positive correlation between the amount of fund transfers and OPC. CASH (Cash Deposits and Withdrawals) - A bank performs cash deposits and withdrawals at the Central Bank to fulfill the vault cash requirements of its branches. Most withdrawals are scheduled one day in advance but correspondent banks can generate same-day withdrawals. 10

18

Deposits at the Central Bank come from the excess of vault cash that bank branches have at the end of the day. Hence, banks with a large number of branches or that are geographically dispersed will find it difficult to gather information about all nation-wide deposits and withdrawals performed in their network of branches. We expect that the collection of this information is an operational problem for the bank, and therefore it will be positively related to OPC. TAX (Federal Tax Collection and Disbursement) – these operations are done throughout the set of branches of the 6 largest banks. That is, when a company pays taxes it goes to a branch of the banking system to perform the payment. In like manner, banks distribute federal payments nation-wide. The bank’s treasury knows these operations two or three days in advance but the recip ient can claim the payments days later. Thus, the uncertainty of the claim should generate a positive impact on OPC. CLH (Clearinghouse Settlement) - is the total amount of checks cleared in a day. It can be divided in two parts. The first is the checks deposited in the bank’s network of branches. These deposits take place nation-wide, and the geographical dispersion of this information makes it a candidate for delays and operational errors. The second is the bank’s checks deposited in branches of other banks (withdrawals from the bank’s standpoint). These checks represent the main source of uncertainty because the bank does not know the exact value of the checks written by its customers. Therefore, we expect the coefficient of this variable to be positive since the value of these operations has a direct effect on the deviations of the reserve account balance and therefore on OPC. Firm Factors. Firm factors allow us to control for cross sectional differences in our sample. We include two measures: the number of branches (BRAN) as a measure of size and the financial margin (FM) as a measure of financial efficiency. BRAN (Branches) - the number of branches represents the size of a bank’s infrastructure. Size is relevant because it is directly related to the complexity of management, and it is an indication of geographical dispersion. We expect this variable to be positively correlated with OPC since the number of branches is an operational burden for a bank. FM (Financial Margin) – this amount is the net of interest paid and interest received. This measure of financial performance is relevant because the strategy for the management of the

19

reserve account is a subset of the overall financial strategy of a bank. We expect the financial margin to be negatively correlated with OPC since a good financial strategy will help reduce costs. We converted this variable to millions of dollars to make the regression coefficient more interpretable. Economic Factors. Economic Factors are the exogenous inputs that control for general characteristics of the financial system and the economy. We include the prevailing rates in the market that affect OPC. We assume that the rates are exogenous to the bank's decision since there is nothing the bank can do to influence them. The role of the rates in the accounting relationship given by equation (6) is similar to the role of the prices given in the total cost functions used in production economics (TC = wL + vK). In the cost minimization process, a firm can chose the level of inputs (L and K) but the prices (w, v) are given. Similarly, the rates (RD, RP) in equation (6) are given parameters in our cost minimization process, and consequently we include them in the characterization of f( X;θ 2 ). The problem of an accounting identity between equation (6) and f( X;θ 2 ) does not exist because the factors used in the computation of the end-of-day balance (Y+DL+ON) are not used in the characterization of f(X;θ 2 ). In addition to the market rates, we consider the growth in the monetary aggregates since they can nominally increase the aggregate reserve balance, even after inflation has been considered. Finally, we include a calendar dummy for Thursdays to control for the excess of operations that take place on the day of the settlement of the primary auction of government bonds. RD (Rate of Deposit) – identifies the real overnight interest rate that represents the opportunity cost of investing the funds. It is important to point out that this rate changes continuously when the money markets are operating. Therefore, every transaction is performed with a different rate. The value that we use is a weighted average of the rates used on the operations of the day. Since the interest rate has a direct impact on the opportunity cost, we expect a positive coefficient on this variable. 11 RP (Rate of Penalty) – represents the real discount rate that the banks are charged for an overdraft on the reserve account. The rate is defined as three times the rate of the last auction of Mexican T-bills (Cetes). The auction takes place every week and therefore this value changes accordingly. We expect a positive coefficient on this variable. 12

20

MB (Monetary Base) - is the sum of the currency that the public owns and the deposits of commercial banks at the central bank. Since we are analyzing the behavior of the account at the Central Bank, the introduction of this variable is relevant. To facilitate the interpretation of the coefficient of this variable we normalized it using the value of the first day of the sample. We expect a positive correlation between this variable and OPC since any change in the monetary base should be reflected in the overall reserve balance of the reserve account and eventually in OPC. THU (Thursday) – controls for the above average financial activity that takes place the day of the settlement of the weekly auction of government bonds. We expect a positive sign on the coefficient of this variable since the additional volume of operations should complicate the overnight forecast and therefore increase OPC. 5. Estimation and Results Descriptive statistics and a correlation matrix for the variables in our model are presented in Tables 2 and 3, respectively. We can see from Table 2 that the daily average opportunity and penalty cost per bank is equivalent to $30,532 U.S. dollars. This amount implies a daily average for the whole system of $580,108 U.S. dollars, and a monthly average of $17,403,240 U.S. dollars. Table 2 also shows that the standard deviations of the amount variables are higher than the means. This difference suggests a skewed distribution that we correct with a logarithmic transformation. 5.1 Econometric Estimation Equation (5) describes the behavior of OPC with and without the adoption of the system. If there is no effect attributed to the adoption of the network technology, the non- negative term A(•) must be equal to one. If the term is greater than zero and less than one, it indicates a reduction in OPC. If the term is greater than one, this indicates that technology adoption is increasing OPC. Given this behavior of the function accounting for technological impact, it can be represented with the exponential function that has been used in the Solow model. 13

21

Table 2 Descriptive Statistics Name

Variable

Opportunity and Penalty Costs Stand Alone Technology Adoption Factors Network Externalities Fund Transfers Cash Deposits and Withdrawals Transaction Tax Related Operations Factors Clearing House Settlements Total Transactions Number of Branches Firm Factors Financial Margin Economic Factors

Daily Mean per Bank

OPC

$30,532

$82,517

0.502

0.500

0.485 $150,239,783 $8,515,944 $18,451,214 $506,581,715 $683,788,656 237

0.488 $190,170,148 $11,633,070 $78,752,571 $953,391,106 $953,391,106 225

SA NE FT CASH TAX CLH BRAN

Standard Deviation

FM

$4,313,011

$6,727,549

Rate of Overnight Deposits

RD

86.0%

34.3%

Penalty Rate

RP

55.0%

15.4%

Monetary Base Index

MB

1.270

0.190

Thursdays

THU

0.198

0.399

Note: the figures of Costs, Transation Factors and Financial Margin are in equivalent US dollars.

Table 3 Correlation Matrix Stand Alone Network Externalities Fund Transfers Cash Deposits and Withdrawals Tax Related Operations Clearing House Settlements Number of Branches Financial Margin Rate of Overnight Deposits Penalty Rate Monetary Base Thursday

SA NE FT CASH TAX CLH BRAN FM RD RP MB THU

SA NE FT CASH TAX CLH BRAN FM RD RP 1.00 0.98 1.00 0.20 0.19 1.00 0.07 0.04 0.59 1.00 0.12 0.10 0.45 0.63 1.00 0.18 0.15 0.69 0.70 0.59 1.00 0.10 0.06 0.55 0.75 0.58 0.73 1.00 0.02 0.00 0.40 0.59 0.43 0.59 0.83 1.00 -0.50 -0.51 -0.03 0.02 -0.02 -0.05 0.00 -0.01 1.00 -0.82 -0.84 -0.12 0.01 -0.08 -0.10 0.00 -0.04 0.55 1.00 0.78 0.77 0.13 0.02 0.09 0.10 0.00 0.01 -0.47 -0.87 0.00 0.00 0.05 0.12 0.45 0.08 0.00 0.00 -0.03 -0.02

MB

1.00 0.03

THU

1.00

Thus, the model of OPC can be represented by: OPC=exp(γ1 sa+γ2 ne(tr/TR)) f(X;θ 2 )

(7)

where {γ1 , γ2 } ∈ θ2 and {sa, ne} ∈ Z. To estimate equation (7) we take natural logarithms on both sides of the equation to yield: Ln(OPC)=(γ1 sa+γ2 ne(tr/TR)) + Ln(f(X;θ 2 ))

(8)

Since we have a panel data set we need to consider the function for every bank i and for the time series dimension t. Thus, the general specification of equation (8) is:

22

Ln(OPCit ) = α it +γ1SAit +γ2NEit +β'Xit +ε it

(9)

where α it is a constant term, γ1 and γ2 are the coefficients indicating the percentage change of OPC, SAit is the dummy variable representing the stand-alone adoption effect sa, NEit is the cumulative variable representing the network externalities effect ne(tr/TR), β is a (k x 1) vector of coefficients, Xit is a (1xk) vector of control variables that includes {Ln(FTit ), Ln(CASHit ), Ln(TAXit ), Ln(CLHit ), BRAN it , FMit , RDit , RPit , MBit , THUit }, and ε it is a white noise error term distributed N(0,σ2 ). We initially performed a Pooled Least Squares (PLS) estimation of equation (9) assuming the same intercept α for all banks and the same coefficients for the cross sectional variables. The results of the PLS estimation can be seen in column labeled “PLS” in Table 4. With the exception of the monetary base and the technology variables, the remaining variables are significant. However, the PLS specification does not control for cross sectional differences among banks that are fixed over time. To control for these differences, we relax the restriction of a single intercept and allow for the estimation of fixed effects. Fixed effects models assume that differences across banks can be captured by differences in the constant term. This allows a different intercept α i for each bank. The specification of the fixed effects model is as follows. Ln(OPCit ) = α i+γ1SAit +γ2NEit +β'Xit +ε it

(10)

We performed a Least Squares estimation of the fixed effects model (10) and evaluated the Likelihood Ratio (LR) test of this model versus the restricted PLS specification. The LR test favors the fixed effects model (χ2 =875, p<0.01), suggesting that time invariant differences across banks can be captured by differences in the constant term. With the reduction of cross sectional variability, the coefficients of SA and NE are statistically significant. The heterogeneity of the Mexican banks is substantial: the three largest banks have 70% of the total liabilities of the market. Hence, we could expect cross sectional differences in the variance. We performed a Lagrange Multiplier (LM) test for group-wise heteroskedasticity (Greene 1997, p. 594). The result of this test yielded a value of 59.67 that is well above the 3.88 critical value, rejecting the null hypothesis of homoskedasticity. To identify if the variance is related to one of the variables in the model we performed White’s Test for heteroskedasticity (White 1980) regressing the residuals of the fixed effects estimation on the variables of the

23

model and their squared values. The LM value for the White Test is 56.29 which is above the critical value of 3.88. We found the coefficients of the variables FT, BRAN and FM significantly different from zero. Since we do not have a single variable to weight the regression or a known form of heteroskedasticity, we use a general weighting procedure. That is, we estimated model (10) using a Feasible Generalized Least Squares (FGLS) estimation of the fixed effects specification. The results of this estimation are shown in Table 4 in the column labeled “FGLS Fixed Effects Full Model”. Using the residuals of the FGLS specification, we performed an analysis to detect whether there was autocorrelation (the tables for the Durbin-Watson test do not provide values for a sample sizes greater than 200). We performed this analysis generating a correlogram for the first 20 lags. We found no evidence of autocorrelated disturbances (p>0.10). The existence of outliers was not relevant to the results of the estimation. We identified five standardized residuals that were more than 2.5 standard deviations away from the mean. When we removed these outliers, the results of the estimation did not exhibit any significant change. As an additional test of the significance of the coefficients of the variables accounting for the technology impact (SA and NE), we performed a LR test of the full model versus the restricted version with the coefficients of SA and NE equal to zero (γ1 =γ2 =0). The LR test indicates that the restricted model is significantly different from the unrestricted one (χ2 =8.80, p<0.01) giving additional support to the model including the technological impact. The results of the FGLS fixed effects estimations are in Table 4 in the columns "Restricted Model" and "Full Model". An alternative specification to the fixed effects model is the random effects model. Random effects assume that differences among different banks can be viewed as parametric shifts of the regression function (Greene 1997, pp. 464-479). In this model (11), we allow for a single intercept and a different specification of the error term for each cross sectional unit. Ln(OPCit ) = α+γ1SAit +γ2NEit +β'Xit +ui+ε it

(11)

The estimation results of the random effects model (shown in Table 4 in the column labeled “Random Effects”) are consistent with the estimation of fixed effects. However, the random effects model is used to account for the variability of units not included in the sample. In the case

24

of our data set, we have the entire population of Mexican banks, and therefore this specification is not needed. Table 4 Dependent Variable Ln(OPC)

Name

Technology Adoption

Transaction Factors (TF)

Firm Factors (FF)

Economic Factors (EF)

Variable

Expected Sign

Stand Alone

SA

(+)

Network Externalities

NE

(+)

Fund Transfers

Ln(FT)

(+)

Cash Deposits and Withdrawals

LN(CASH)

(+)

Tax Related Operations

Ln(TAX)

(+)

Clearing House Settlement

Ln(CLH)

(+)

Number of Branches

BRAN

(+)

Financial Margin

FM

(-)

Rate of Overnigh Deposits

RD

(+)

Penalty Rate

RP

(+)

Monetary Base

MB

(+)

Thursdays

THU

(+)

Adjusted R-squared Log Likelihood

Pooled Least Squares

Random Effects

Full Model

Full Model

FGLS Fixed Effects Restricted Model

Full Model

0.061 (0.155) -0.205 (0.163) -0.017 ** (0.008) 0.199 *** (0.015) 0.039 *** (0.006) 0.352 *** (0.013)

0.245 * (0.146) -0.372 ** (0.153) 0.035 *** (0.012) 0.014 (0.021) 0.011 * (0.007) 0.097 *** (0.019)

0.025 ** (0.010) 0.020 (0.019) 0.006 (0.006) 0.049 *** (0.018)

0.304 (0.140) -0.403 (0.146) 0.027 (0.010) 0.015 (0.019) 0.006 (0.006) 0.051 (0.018)

0.002 *** (0.000) -1.250 *** (0.094) 3.642 *** (0.383) 0.908 *** (0.179) -0.117 (0.102) -0.273 *** (0.033)

0.004 *** (0.000) -0.268 *** (0.098) 3.711 *** (0.359) 1.085 *** (0.169) 0.079 (0.097) -0.065 * (0.034)

0.001 (0.002) -0.213 ** (0.098) 3.489 *** (0.325) 1.371 *** (0.131) 0.071 (0.084) -0.031 (0.032)

0.002 (0.002) -0.265 *** (0.101) 3.374 *** (0.328) 1.122 *** (0.154) 0.061 (0.088) -0.029 (0.032)

0.559 -9138.12

0.6143

0.8886 -8704.8

0.8891 -8700.4

Standard Errors are in parenthesis *** Significant at 1% level. ** Significant at 5% level. * Significant at 10% level.

The impact of the sequence of adoption and any other time invariant effects are controlled for by the fixed effects specification. However, to determine if the sequence of adoption was significant before the inclusion of the fixed effects we investigated the sequence effect in two ways. First, we included a variable that indicates the seque nce of adoption (SEQ) and ranges from 1 to 19. Second, we used dummy variables to separate the banks into four groups (G1, G2, G3, and G4). The composition of the groups is: early adopters (16%), early majority (32%), late majority (32%), and laggards (20%) (Rogers 1995, p. 262). The result of the regression including the variable indicating the adoption sequence shows a statistically significant coefficient (p<0.01) with a negative sign. This result indicates that, before controlling for fixed effects, the opportunity and penalty costs of the late adopters were smaller than those of the early adopters.

25

** *** ***

***

The results for the four dummies show that all coefficients are not significantly different from zero (p>0.10). We also explored the possibility that the availability of the electronic system might have generated a structural change in the number of fund transfers and thus affect OPC. To test this, we include in the original model (10) a variable that accounts for the interaction of the Network Externalities effects (NE) and the fund transfers (Ln(FT)). The coefficient of the interaction variable (NE*FT) is not significant (p>0.10). Finally, we explored if commercial banks could have developed internal information systems to improve the cost minimization process implying the existence of an embodied technological change. To test for this effect, we included a variable that equals an exponential time trend (TREND) in the first six months of the data set and zero when the system had not been implemented. The estimated model is given by: Ln(OPCit ) = α i+γ4 TRENDit +γ1 SAit +γ2 NEit +β'Xit +ε it

(12)

The coefficient of the time trend was not significant (p>0.10). Thus, we find no evidence of embodied technological change. 5.2 Results In this section, we examine the results from our final model (in equation 10) shown in Table 4 in the column labeled “Full Model” of the FGLS Fixed Effects specification. We start by examining the results for the stand-alone and network externalities effects of the network technology adoption. We then analyze the results for the variables that characterize the cost minimization process. The coefficient of the variable (SA) representing the stand-alone adoption of technology is significant (p<0.05) but it has an unexpected positive sign. We explore reasons for this result in the discussion section. The coefficient of the network externalities (NE) is significant (p<0.01) and negative as expected. Among the Transaction Factors, the coefficients of the clearing house process (CLH) and the fund transfers (FT) are significant with the expected sign (p<0.05). The coefficients of the tax related operations (TAX) and the cash deposits and withdrawals (CASH) are not significant (p>0.10). The explanation might come from two sources. First, the amounts represented by these

26

variables do not vary much over time. Second, only six banks perform tax related operations, and seven operate cash facilities on behalf of the central bank. Thus, the fixed effects specification captures this firm- level difference. Among Firm Factors, the coefficient of the number of bank branches (BRAN) is not significant. This is not surprising since the number of branches does not vary much over time, and therefore the differences are captured by the different intercepts. In the case of financial margin (FM) the coefficient is significant (p<0.01) and with the expected negative sign. This is an indication that good financial performance in a bank is related to good performance in the management of the reserve account. Economic and Financ ial Factors have an important impact. The interest rates (RD, RP) are significant and with the expected sign (p<0.01). This is expected since they influence the determination of cost. The monetary base variable (MB) that was included to control for changes in aggregate reserves is not significant. This indicates that the changes in aggregate values are only attributed to inflation. The coefficient for Thursday (THU) was not significant probably because the increase in operations that takes place on these days could already be captured in the Transaction Factors. 5.3 Discussion In this section, we discuss our results. We first examine the economic impact of the network technology adoption and then relate our findings to previous work in this area. 5.3.1. Evaluation of the Technological Impact. Our hypothesis of a negative sign for the stand-alone effect was rejected. The coefficient of the stand-alone variable (SA) is significant but with an unexpected positive sign. Before the adoption of the new network technology (when SA=0), there was no effect on OPC. However, after the new network technology was adopted by all banks (when SA=1), the results show an increase in OPC equivalent to 30.5%. Thus, the initial stand-alone adoption effect reflects decreased perfo rmance, i.e., higher OPC. A potential explanation for this finding is that the initial benefits of the technology adoption are more than offset by the need to use two technologies (old and new) concurrently. In other words, when a bank adopts the electronic inter-bank payment network, the new technology improves the level of aggregation by producing a summary of the bank’s operations. However, this improvement is not enough to compensate for the operational burden of dealing with both old and new

27

technologies. That is, during the transition period, the bank has to work with the information generated in the new system as well as with the paper-based information generated by banks not yet in the network. As Srinivasan, Kekre and Mukhopadhyay (1994) have shown, integrated information systems lead to performance improvements. In this case, the introduction of the information system created a lack of integration between the existing printed reports and the electronic queries produced by the new network. Thus, the need to work with two technologies and the lack of integration between them could have produced an operational burden that overshadows the initial improvement in the information set. We find strong support for the network externalities hypothesis. The coefficient of the network externalities variable (NE) is significant and with the expected negative sign. When the ratio of electronic transactions to overall transactions is equal to zero, there is no change in cost. When the ratio is different from zero, the coefficient indicates a reduction in cost of 0.404% for every 1% of electronic transactions in the system. When all banks have joined the network, and the ratio equals one, the overall impact indicates a 40.4% reduction in OPC. Our results imply that ind ependent evaluation of technology adoption effects provides an incomplete picture of the impact of adopting a new network technology. The stand-alone effect (SA) and the network externalities effect (NE) exist simultaneously, and their impact should be evaluated together. As we explained in an earlier section, in our empirical setting, there are three stages of the implementation process: the first when the electronic network was not adopted, the second when the banks were gradually adopting the system, and the third when all banks had subscribed to the electronic network. During the first stage, when the electronic network was not adopted, the technology impact function A(•) was equal to one (or equal to zero after taking logs) and no effect existed. During the third stage, when all banks adopted the system, the standalone effect and the network externalities effect were present. In this stage, the combined effect on OPC was a 9.9% reduction. This result is derived from a 30.5% increase for the stand-alone effect and a 40.4% reduction attributed to the network externalities. During the second stage, when the system was gradually adopted, the banks experienced the full stand-alone effect and an incremental network externalities effect. The first banks that joined the network had the full set of electronic outgoing flows but a low ratio of electronic incoming transactions to overall incoming transactions. As more banks joined the network, the ratio

28

increased. Table 5 provides an illustration of the ratio of ele ctronic incoming flows relative to the percentage change in OPC experienced by each bank on the first day it joined the network. Table 5 Bank 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19

Ratio of electronic incoming flows to total incoming flows 0.00 0.12 0.16 0.35 0.45 0.62 0.83 0.84 0.87 0.88 0.91 0.92 0.93 0.93 0.93 0.94 0.96 0.98 1.00

Percentage change in cost the first day of network adoption 30.5% 25.9% 24.0% 16.3% 12.4% 5.4% -3.2% -3.5% -4.7% -5.1% -6.2% -6.7% -6.9% -7.1% -7.2% -7.4% -8.2% -9.1% -9.9%

As we mentioned earlier, the transaction ratios provide a measure of network externalities. When the level of externa lities is low, the stand-alone effect is greater than the network externalities effect, and OPC increases (see the first 6 banks in Table 5). However, when the ratio of electronic transactions is high, the stand-alone effect is lower than the network externalities effect, and the initial increase no longer dominates. A graphical illustration of the simulated combined effect of technology adoption on typical banks can be seen in Figure 5. As can be seen in Figure 5, given the combined existence of the stand-alone effect and the network externalities effect, the increase in OPC is only transitory. The later a bank joins the network the less susceptible it is to suffer a reduction in performance. This observation is consistent with the results from our post hoc analysis of adoption sequence showing that later adopters of the network had lower OPC than early adopters. Evaluations of the impact of information technology have attempted to determine the return on investment (Brynjolfsson and Yang 1996; Barua et al. 1995). In this study, the aggregated savings in OPC for all banks equaled $5.3 million dollars for the 6 months after the network was installed. The estimated total cost of implementing the inter-bank payment network was $680,000 dollars. 14 Transaction processing savings from layoff of clerical personnel were approximately $1.5 million dollars on an annual basis. 15 If all benefits accrued for another ten

29

years at an interest rate of 7.5%, the net present value (NPV) of the transaction processing savings would be $10.3 million dollars, and the NPV of the savings in OPC would be $72.7 million dollars. Thus, an NPV evaluation based only on transaction processing impact can miss the overall effect of the network. At our research site, the benefits realized from improved decision- making about reserve account balances are 7 times larger than the benefits realized in transaction processing. Overall, the inter-bank payment network project yields a significant positive NPV of $82.3 million dollars. 16 Figure 5 Percentage Change in Cost by Typical Banks (The number of each bank reflects the sequence of adoption)

400

0

100

200 300 time

400

100

400

400

500

200 300 time

40 20 0

400

500

400

500

40 20 0

% Changes in Costs

40 20

100

500

100

200 300 time

Bank 19

0 0

400

0

500

-20

% Changes in Costs

40 20

% Changes in Costs

0

200 300 time

200 300 time

200 300 time

-20

% Changes in Costs

40 20 0

Bank 16

-20

100

100

Bank 10

0

500

Bank 13

0

0

400

500

40

100

500

-20

% Changes in Costs

40 20 0 0

400

Bank 7

-20

% Changes in Costs

Bank 6

200 300 time

-20

% Changes in Costs

40 20 0

500

20

200 300 time

0

100

-20

0

Bank 5

-20

0

20

40

% Changes in Costs

Bank 4

-20

% Changes in Costs

Bank 1

0

100

200 300 time

5.3.2 Relationship of Our Findings to Prior Work. Traditional models define the effect of network externalities as a function of the number of users in the system (Farrell and Saloner 1986, Kauffman et al. 2000). However, our work proposes the characterization of network externalities as a function of the number of transactions. We tested the specification of the number of users by redefining the network externalities variable as the proportion of the number of banks in the system. We did not get a good fit in the model. The coefficient of the stand-alone

30

variable SA is not significant and we can reject the network externalities specification (p>0.05). This result confirms that relaxing the assumption of user homogeneity was an important step in the development of our model. To determine whether our results are consistent with previous empirical findings, we tested whether large banks adopted the network before smaller banks. Saloner and Shepard (1995) in a study of ATM network adoption found that the greater the number of branches the earlier the technology adoption. Kauffman et al. (2000) also found similar results. We performed the test by regressing the variable indicating the adoption sequence to a measure of size given by the number of branches. The relationship is negative and statistically significant (p<0.01) implying that larger banks adopt earlier. Thus, the sequence of adoption followed by the banks analyzed in this study is consistent with previous empirical findings. 6. Conclusions Our work makes several important contributions to the literature on network externalities and technology adoption. First, our analysis provides much needed empirical support for previous theoretical work and corroborates the network externalities hypothesis by determining the economic value of externalities. Prior work has mainly focused on the theoretical aspects of network externalities and network adoption. The few empirical papers that measure economic value (e.g., Gandal 1994; Brynjolfsson and Kemerer 1996) have focused on standards. Similarly, the work on network adoption (Saloner and Shepard 1995; Gowrisankaran and Stavins 1999; Kauffman et al. 2000) has not measured the economic value of the externalities. Our work is the first attempt to analyze the value of externalities in an electronic payment network. Second, we have shown that the initial impact of network technology adoption on firm performance can be negative. This is because there can be significant implementation costs associated with network adoption, particularly when there is concurrent use of old and new network technologies. Finally, our work extends the typical model used to depict technology adoption and network externalities (Farrell and Saloner 1986). We have enhanced the model of technology adoption to account for network externalities when the users of a network are not homogeneous. We have shown that the number of transactions rather than the number of users can better characterize network externalities when users are not homogeneous.

31

Our general model and approach can be adapted to other contexts where researchers and managers are interested in determining the business value of externalities from network technology adoption. However, in our study, the Central Bank focus on monetary policy rather than shareholder value could be a limiting factor in the extrapolation of our results. The reserve accounts of the banks are important components of the monetary aggregates. The development of systems that contribute to better control of these aggregates is far more important to a central bank than any reduction in operational costs. Future work in the area should consider networks in which the sponsor is a for-profit firm in a competitive environment to determine whether the results of this work hold. Likewise, the work should be extended to dynamic models of network effects. In summary, the importance of networked technologies demands further theoretical and empirical work that scrutinizes the economic impact of networked applications and network externalities in different contexts. This study takes a first step in that direction. 7. References Ahituv, N., “Assessing the Value of Information: Problems and Approaches,” Proceedings of the 10th International Conference on Information Systems, Boston, MA., (1989), 315-325. Barua, A., Kriebel, C.H. and Mukhopadhyay, T., "MIS and Information Economics: Augmenting rich descriptions with analytical rigor in Information Systems Design,” Proceedings of the 10th International Conference on Information Systems, Boston, MA., (1989), 327-339. Barua, A., Kriebel, C.H., Mukhopadhyay, T., (1995) "Information Technologies and Business Value: An Analytic and Empirical Investigatio n," Information Systems Research, 6(1), 3-23. Bauer, P.W. and Ferrier, G.D., "Scale Economies, Cost Efficiencies, and Technical Change in Federal Reserve Payment Processing." Journal of Money Credit and Banking, 28, (1996), 10041039. Brynjolfsson, E. and Kemerer, C., “Network Externalities in Microcomputer Software: An Economic Analysis of the Spreadsheet Market,” Management Science, 42,12, (1996), 16271647. Brynjolfsson, E. and S. Yang, ”Information Technology and Productivity: A Review of the Literature,” Advances in Computers, 43, (1996), 179-214. Economides, N. “The Economics of Networks,” International Journal of Industrial Organization, 14, 6, (1996), 673-700.

32

Economides, N. and Himmelberg, C. H., “ Critical Mass and Network Size with Application to the U.S. Fax Market”, Working Paper, NYU, EC-95-11, (August 1995). Farrell, J. and Saloner G., “Standardization, Compatibility, and Innovation”, Rand Journal of Economics, 16, 1 (Spring 1985), 70-83. Farrell, J. and Saloner, G., “Installed Base and Compatibility: Innovation, Product Preanouncement, and Predation”, American Economic Review, 76,5, (1986), 940-955. Furfine, C. “Interbank Payments and the Federal Funds Rate”, Working Paper, Board of Governors of the Federal Reserve System (1998). Gandal, N. "Hedonic Price Indexes for Spreadsheets and an Empirical Test for Network Externalities", RAND Journal of Economics 25, (1994), 160-170. Greene W.H., Econometric Analysis, Prentice Hall (1997). Gowrisankaran, G., Stavins, J., "Network Externalities and Technology Adoption: Lessons from Electronic Payments", Working Paper 99-5, Federal Reserve Bank of Boston (1999). Hancock, D., Humphrey, D.B., Wilcox, J.A., "Cost Reduction in Electronic Payments: The roles of Consolidation, Economies of Scale, and Technical Change", Journal of Banking and Finance, 23, (1999) 391-421. Katz M., and Shapiro C., “Network Externalities, Competition, and Compatibility”, American Economic Review, (June 1985), 424-440. Katz M., and Shapiro C., “Technology Adoption in the presence of Network Externalities”, Journal of Political Economy, 94,4, (1986), 822-841. Kauffman R., J. McAndrews J., Y. Wang, “Opening the “Black Box” of Network Externalities in Network Adoption”, Information Systems Research, 11,1, (2000), 61-82. Rogers, E.. Diffusion of Innovations, 4th Edition, The Free Press (1995). Saloner, G. and Shephard A. "Adoption of Technologies with Network Effects: An empirical Examination of the adoption of Automated Teller Machines." RAND Journal of Economics, 26, 3, (Autumn 1995), 479-501. Solow, R., “Technical Change and the Aggregate Production Function.” The Review of Economics and Statistics, 39, 3, (August 1957), 312-320. Srinivasan, K., Kekre, S., Mukhopadhyay, “Impact of Electronic Data Interchange Technology on JIT Shipments,” Management Science, 40,10, (1994), 1291-1304.

33

White, H., "A Heteroskedasticity Consistent Covariance Matrix Estimator and a Direct Test for Heteroskedasticity". Econometrica, 48, (1980), 817-38. 1

Production Economics distinguishes between neutral technological changes and biased technological changes. The former affect all inputs equally (leave the marginal rate of technical substitution unchanged). The latter affect only some inputs (change the marginal rate of technical substitution). 2

In an attempt to improve the readability of the notation for the Stand Alone effect and the Network Externalities effect, we modify the notation used by Farrell and Saloner (1986) and Kauffman et al. (2000) from a+b(n) to sa+ne(n). 3

Furfine (1998) describes the problem of uncertain operations affecting the reserve account of U.S. banks.

4

During the period of study, the discount rate or penalty rate was 3 times the rate of the last auction of government T-bills (Cetes). 5

Inter-bank operations were in general low in volume and high in value. A large bank would have an average of 150 fund transfers in a day, whereas a small bank would have an average of five. The mean value of a transaction was equivalent to one million dollars. 6

The settlement of the clearinghouse is the result of the process where a bank interchanges the checks of other banks received during the day in the form of deposits. At the end of the day, every local clearinghouse sends the final balances to the Central Bank so that the Central Bank can credit or debit the reserve account of each bank. 7

Repurchase Agreements (REPO) of government bonds are fixed term loans that have government bonds as collateral. 8

The data sources include: 1) Boletin Estadistico, Comis ion Nacional Bancaria y de Valores – Contains monthly analytical statements of commercial banks. These statements were used to obtain financial and operational figures such as: financial profits, operating cost, number of branches, number of checking accounts and number of employees. 2) Sistema de Informacion Economica (WWW), Banco de Mexico – Includes economic indicators such as inflation and monetary aggregates. 3) Modulo Banca, Banco de Mexico – This system contains archives that keep historical information of bank operations such as balances of the reserve account and daylight and overnight operations. We had more than 2 million inter-bank transactions captured for this study. 4) Boletin Informativo, Bolsa Mexicana de Valores - Includes daily liquidity indicators used to create the rate references for opportunity and penalty costs. 9

The use of either the amount of operations or the number of operations is practically the same. The correlation between the two variables is 0.98. 10

Correspondent banks are those that operate a vault cash facility on behalf of the Central Bank.

11

We use the real rate because we are using deflated amounts. To convert the nominal rate to the real rate we used the actual inflation price index. The alternative way to compute the real rate with the expected inflation was not possible because there are no public time series for expected inflation in our sample period. 12

Ibid.

13

Bauer and Ferrier (1986) and Hancock et al. (1999) use a similar function to evaluate the technology impact in the cost function of an electronic payment network. 14

The Central Bank estimated spending about $500,000 dollars on the project for additional hardware for the Bank’s mainframe, the transactional software, and the labor cost of programmers and systems analysts for the new system. For the commercial banks in the network, the total costs were on the order of $180,000 dollars as the banks only had to purchase an average of 3 PCs each and a telephone line. Hence, the total cost of the system was approximately $680,000 dollars. 15

On the Central Bank side, 50 clerical workers were made redundant by the implementation of the new system. These layoffs resulted in an approximate annual savings of $1.5 million dollars. On the commercial banks’ side, there were no signs of personnel redundancy, and therefore we assumed no transaction processing savings. 16

This assumes that benefits accrue for another ten years at an interest rate of 7.5%.

34