The Effects of Banking System Reforms in Pakistan. Emilia Bonaccorsi di Patti* and Daniel C. Hardy** November 2003 Preliminary and incomplete

Transcription

1 The Effects of Banking System Reforms in Pakistan by Emilia Bonaccorsi di Patti* and Daniel C. Hardy** November 2003 Preliminary and incomplete JEL Classification: G21, G28, O16, O53. Keywords: Banking, Efficiency, Deregulation, Pakistan. *Banca d Italia, Research Department; Via Nazionale 91, 00184, Rome, Italy. **IMF, Monetary and Financial Systems Department. The opinions expressed do not necessarily reflect those of the Banca d Italia, the IMF or their staffs. s: Bonaccorsidipatti.emilia@insedia.interbusiness.it; dhardy@imf.org.

2 1. Introduction A large number of industrialized, developing and transition countries have undertaken extensive reforms of their banking systems over the past two decades (see Swary and Topf (1992), Zahid (1995), and Fanelli and Medhora (1998) for surveys). These reforms have included such measures as the lifting of barriers to competition, the privatization of public financial institutions, the introduction of market-based securities, the liberalization of interest rates and the removal of quantitative controls on lending, The principal aims of these reforms has generally been to raise both the level and the efficiency of the allocation of investment, and to enhance the provision of financial services to all sectors of the economy. Given the importance of the financial sector in itself and its influence on other sectors, evaluating the degree to which these aims were achieved is important. However, due to the time needed to implement structural reforms such as those in the banking sector and for them to take effect, in many countries convincing evidence regarding this issue is only now accumulating. There are two main approaches to evaluating the effects of financial market reforms, including large scale privatization. The first concentrates on macroeconomic variables that are closely related to the ultimate objectives of reform such as savings and investment rates, real growth, and interest rates and attempts to discern changes in their behavior induced by the financial sector reforms. 1 The main drawback of this approach lies in the difficulty in isolating the effects of financial market reforms from those of other institutional or macroeconomic developments. The difficulty is made more acute by the small number of observations typically available for any one country, and the heterogeneity of experiences across countries. An alternative approach is based on an analysis of the financial sector itself, and in particular on an assessment of changes in the structure and performance of the sector following reform. Where data at the microeconomic level are available, the performance of banks and other financial institutions can be observed and compared to that prior reforms. In particular, it is possible to estimate the effects of reforms on productivity, costs, revenues and 1 Examples include Gelb (1989), Johnston and Pazarbasioglu (1995), Demetriades and Luintel (1996), Jbili, Enders, and Treichel (1997), Bandiera, Caprio, Honohan, and Schiantarelli (2000), and, specifically regarding Pakistan, Khan and Aftab (1994) and Khan and Hasan (1998). 2

3 profitability of the financial institutions operating in one country. In the case of privatization, changes in individual performance of institutions can be related to changes in their ownership and governance. The approach based on econometric methods to estimate efficiency and productivity changes has been followed by Grabowski, Rangan, and Rezvanian (1994), Berger and Mester (1997) and Humphrey and Pulley (1997), for example, in looking at the U.S. experience with bank deregulation, and by Lozano-Vivas (1998) for Spanish banks. Fewer econometric studies of the effects of reform on the financial sector are available for developing countries (e.g. Gilbert and Wilson (1998), Leightner and Lovell (1998)), mainly because of the lack of microeconomic data. Therefore, Pakistan s experience in transforming its financial sector may therefore be of general relevance, besides being of importance in itself for the functioning of one of the larger and more populous developing economies. Given that reforms have been extensive and have affected the entire banking system, not only the privatized institutions, we focus on systemic changes in productivity, focusing on the concepts of profits, cost and revenue productivity following the approach proposed by Berger and Mester (2003). The data set is sufficiently rich that it will be possible to assess the effects of the reforms on the banking system as a whole, and on sections of the industry, such as stateowned banks and privatized banks. A novel aspect of our analysis is that we estimate efficiency taking into account the fact that the data do not behave as required by the assumption of normality underlying ordinary least squares estimation. We apply the Distribution Free methodology estimating the cost and profit frontiers both with Feasible Generalized Least Squares and the least absolute deviation estimator (LAD). LAD may be more appropriate because the number of observations is very small and the method is less sensitive to extreme values. We compare the results obtained with the standard and non standard estimators. The next section outlines the recent evolution of the Pakistani banking system and the reforms introduced starting in the late 1980s. The following section explains the estimation procedures used and defines the data set. In the fourth section estimation results are presented, and the final section concludes. 3

4 2. The Evolution of the Pakistani Banking System In the pre-reform period, the activities of the financial sector in Pakistan were largely directed by government as a means to implement its development strategy. The domestic banks had all been nationalized in 1974 and were consolidated into six major national commercial banks, besides which several specialized credit institutions and household savings schemes were established. The nationalization had the objective to direct bank credit towards specific sectors and to ensure government funding. High domestic borrowing by the government together with an administratively set yield structure resulted contributed to financial repression. Banks were required to hold 30 per cent of their deposits in the form of government securities and 5 per cent as a cash reserve requirement at the central bank. This constraints resulted in low returns on banks portfolios About 25 foreign banks operated, but their market share was relatively small. Banks were given detailed instructions on the allocation of credit to specific sectors, and a plethora of administrative interest rates were set for various purposes. Fees were also regulated in detail. Yet prudential regulation, and in particular capitalization and provisioning requirements, were weak. In this environment, banks had little incentive and scant means to mobilize additional savings, reduce operating costs, or make lending decisions based on creditworthiness. The financial market reforms were initiated in the late 1980s (more information is available in Khan and Aftab (1994), Ul-Haque and Kardar (1995), and Ul-Haque (1997), and in various issues of the Annual Report of the State Bank of Pakistan). The system of administered interest rates was streamlined and loosened starting in The share of credit directed to particular sectors was reduced in 1989, and bank-by-bank credit ceilings eliminated in New prudential regulations were introduced in 1989 and strengthened in 1992, while the State Bank of Pakistan (the central bank) enhanced its supervisory capacity. A system of auctioning government securities was established, and regular auctions for sixmonth bills and longer term bonds began in 1991 (see Hardy, 2000). Concurrent with these efforts at deregulation, the authorities sought to liberalize access to the financial sector by licensing private banks starting in Ten new Pakistani banks were licensed in 1991; later, other eleven banks started to operate. In addition, two stateowned banks (Muslim Commercial Bank and Allied Bank of Pakistan) were privatized in 4

5 By 1997 there were still four major state-owned national commercial banks, but they faced competition from 21 domestic banks (including the two privatized banks) and 27 foreign banks. In addition, the branching policy was eased for the domestic private banks, whereas the state-owned banks were prevented to open new branches. As part of the reforms, the state-owned and privatized banks introduced restructuring plans to reduce employees and close unprofitable branches (these banks reduced their employees from 99,900 to 81,000 by the end of 1999). Another structural break occurred in Effective from the end of 1997, banks were required to maintain an 8 per cent capital and reserves of risk weighted assets and to have a minimum paid-up capital. In addition, the central bank (State Bank of Pakistan) launched a plan for recovery of sub-standard and non-performing loans and reformed the guidelines for loan classification. The tighter approach to risk management and emergence of loan losses had a major impact on banks in 1997 and following years. State-owned banks experienced a substantial increase in non-performing loans. Although the transformation of the Pakistani financial system has been profound, the role of the state remains important, not only through the national commercial banks, but also the specialized credit institutions and sundry measures promoting the allocation of credit to certain favored sectors and facilitating the financing of the government. In addition, since 1997 poor macroeconomic performance, high costs of restructuring plans and political instability put the banking system under stress. The estimation of bank performance below accounts for differences in these factors over time. 3. Empirical Design 3.1. Measurement and sources of changes in cost and profit productivity The many regulatory and other structural changes affecting the financial sector in Pakistan during , and the reforms of 1997, can be expected to have had a significant impact on the behavior of banking institutions there. In particular, the entry of new, private banks and the privatization of two among the largest credit institutions should have given all banks stronger incentives to cut costs and generate revenue, even as intensified competition might have driven down monopoly rents. One problem in evaluating the effect of 2 Plans to privatize more state owned banks were under discussion but not put into effect. 5

6 the shift from an almost fully state-controlled banking system to a mixed system in Pakistan is that such shift has occurred together with more general deregulation of interest rates and credit controls, and the lifting of restrictions on banking operations. One might hope that these reforms led to an improvement in the efficiency of the banking system, where such an improvement can be achieved both by the initially less efficient institutions catching up with those using best practices, and by a general advance involving even the most efficient operators. It will be particularly important to identify effects on pre-existing institutions (and especially those that were privatized), which were subject to different constraints than those on new entrants. Nevertheless, allowance must also be made for changes in exogenous conditions, such the prevailing macroeconomic situation, together with the imposition of capital requirements and the adverse effects of the emergence of past bad loans in the books of the largest banks. 3 With respect to studies that address only privatization and its effect on the institutions involved, in the case of Pakistan it is more appropriate to evaluate the systemic effects of the reforms, including the privatization and the entry of new private banks, rather than focusing solely on the effects of changes in governance. The approach followed below is to estimate the effects of the reforms on the productivity of the banking system, taking into account the relevant exogenous variables. Since the reforms could have quite different effects on costs, revenues and and profitability, we follow Berger and Mester (2001) and measure the performance of banks using the concepts of profit, cost and revenue productivity. Only two studies have used, to our knowledge, change in profits to study the evolution of bank performance over time (Berger and Mester, 2003 and Kumbhakar et al., 2001). Following Berger and Mester (2003), the gross changes in profits, costs and revenues at the industry level can be separated into: i) changes in productivity, ii) changes in exogenous business conditions. The change in productivity can be divided into two components: a) changes in the best-practices adopted by banks, b) changes in efficiency (the dispersion from the best-practice). Reforms may had an impact on the change in productivity, beyond the direct effect on exogenous conditions such as interest rates, competition (concentration) and foreign entry. 3 Because under this approach the effects of different exogenous factors such as output prices or structural variables are removed, it is more informative than a mere comparison of financial ratios. 6

7 The latter factors are directly included as exogenous business conditions, as explained below. In addition, privatization may have affected directly the efficiency of the privatized banks, increasing their relative performance with respect to the best-practice banks. Taking the case of the profit function, standard profit maximization is based on the assumption that banks choose their outputs in response to relative output prices and other exogenous variables. In each period, the profit function is: ln(π b + θ) = f π (p, w, z b, v) + lnu πb + lnε πb (1) where π b is the variable profits of the bank and theta is a scalar that is added to every bank s profits before taking the logarithm to correct for negative values of profits (theta varies over time). The term ln u indicates the deviation of each firm from the best-practice frontier, defined by the bank or the group of banks for which lnu is higher. 4 Variable profits are a function of a vector of input prices p, a vector of variable input prices w, a vector of semifixed netputs z b, and environmental and structural variables v. The term ε πb represents a random error that is assumed to be multiplicatively separable from the rest of the profit function. The term u πb is a bank-specific efficiency term that represents the reduction in bank b s profits due to persistent X-inefficiency, and is thus constant across the estimation period. The bank that has the highest u πb is presumed to be located on the profit efficiency frontier. The profits of the industry are subject to change over time. In particular, the financial sector reforms could shift the frontier for the entire industry and affect the dispersion in efficiency levels of banks. Some of the reforms also affect directly the business conditions, for example interest rates. The profit function can be estimated for different periods and compared over time. Denoting two periods by s and t, the change in the profit function is given by: where Πˆ ˆ ˆ s ( X s,ln ut ) Π s ( X s ) Π s ( X t ) Total Change Π = = (2) Πˆ ( X,ln u ) Πˆ ( X ) Πˆ ( X ) t t s Πˆ τ is the predicted profit function in period τ (τ = s, t; s>t) setting the explanatory variables X τ at their average value for the industry in each period and for the average efficiency level of banks. As shown in equation (2), the change in the predicted s t t t 4 Berger and Humphrey (1997) provide a survey of a large number of studies of bank efficiency based on this approach. 7

8 profits for the average bank can be decomposed into two components. The first is due to the change in the exogenous business conditions (X τ ); the second is due to the change in productivity resulting from the changes in the estimated parameters of the function and in the average efficiency of the industry. Reforms should affect both terms because deregulation of interest rates and entry of private banks have influenced some of the exogenous variables directly, but any other effect on bank choices should be captured by the productivity change. Considering the two periods t and s (s>t), the effect of the change in exogenous conditions (DEXCOND) is given by the predicted profit from the average practice function estimated for period s evaluated at the exogenous conditions of period s, divided by the predicted profits from the average practice function estimated for period s evaluated at the exogenous conditions of period t. The change in productivity (DPROD) in the case of the profit function is computed as the predicted profit from the average-practice function in period s applied to the business conditions from period t, divided by the predicted profits from the average practice function from period t applied to business conditions in period t. (in the analysis below t is a panel and we employ the predicted value for the average bank in the period). The intuition is to derive the share of the total change that can be attributed to the change in the parameters of the function. Berger and Mester (2003) decompose the change in productivity into a shift in the profit frontier and a change in inefficiency. They first compute the gross change and then the change in productivity (equivalent to the product of the change in best practice and change in inefficiency). They obtain the change in business conditions as the ratio of the total change to the change in productivity. Finally, the compute the best-practice frontiers and obtain the change in inefficiency as the change in productivity divided by the change in the bestpractice. In the case of Pakistan, given the limitation of the data and the potential of large measurement errors at each step, we prefer to estimate total change only and the average efficiency, for all and for groups of banks, without carrying out all the decomposition. The standard profit function is based on the assumption that firms are price takers in all markets, output prices are well measured, and output levels are readily variable (see Berger and Mester, 1997). When these conditions are not met, it is more appropriate to concentrate on the alternative profit function, which can be defined as: ln(π bt + θ) = f aπ (y bt, w t, z bt, v t ) + lnu aπb + lnε aπbt (4) 8

9 where the output vector substitutes the output prices of bank b. Financial sector reform might have a modest impact on banks profitability if, for example, greater pressure to achieve cost efficiency were offset by more intense competition and lower margins. The welfare gain from financial sector reform may accrue not so much to bank shareholders, as to the users of bank services. Hence, it is worthwhile to consider the effects of reform on costs in addition to bank profits. 5 The cost function can be specified as: ln(c bt ) = f C (y bt, w t, z bt, v t ) + lnu Cb + lnε Cbt (5) The cost function is based on the assumption that output levels and input prices are given and the firms set output prices and input levels. Efficiency is measured relative to û Cmin, that is, the lowest value of u Cb obtained in the sample from the bank with the best practices in cost minimization. The change in alternative profits and costs can be decomposed in the same way as for the standard profits. Likewise, the indirect revenue function: ln(r bt ) = f R (y bt, w t, z bt, v t ) + lnu Rb + lnε Rbt (6) can be estimated, so that relative revenue efficiency can be assessed, and the change in revenue decomposed accordingly Data and definitions of variables Data were obtained from various editions of the publication Money and Banking Statistics issued by the State Bank of Pakistan, which contains annual information on the main balance sheet entries and revenue and expense items for all banks operating in Pakistan. The full sample covered the period from 1981 through We divide the sample into three sub-periods, the first covering the years before privatization and removal of 5 Studies of banking sector efficiency have tended to concentrate on the cost function rather than the revenue function; Berger, Humphrey, and Pulley (1996) is an exception. It is worthwhile looking at both in order to assess the various effects of financial sector reform, even if the profit function is also estimated. 9

10 constraints on bank operations, a second period, by when the privatization and liberalization have taken effect, and after other reforms. 6 We separate the second from the third period because the third period is characterized by the reforms in loans recovery and tighter loan classification procedures, and by the enforcement of capital requirements. In addition, macroeconomic and political difficulties led to further changes in the regulation 7 and may have interfered with potential improvement of banks conditions. Data were collected on licensed banks for which adequate time series were available (33 between and 40 for the later period); several specialized credit institutions were excluded from the sample, as they were subject to different regulations and may have operated in distinct markets. The number of banks is not the same across periods. For the sample includes seven state-owned banks, of which two were privatized in , 22 foreign banks, and 4 private domestic banks, for which reliable data begin in For the number of private domestic banks in the sample increases to17. The published data were corrected for a number of manifestly typographical errors (for example, when total assets did not equal total liabilities, or when one observation on a series different by an order of magnitude from the values observed in preceding and subsequent years); even so a small number of outliers were identified and deleted, leaving a total sample of 547 observations. A question that arises in the study of banks concerns how to identify bank output. Banks generally provide a variety of intermediation and transaction services, and often these activities are bundled together. For example, depositors provide funds as an input to a bank s lending activities, but also use the bank to make payments. Thus, the volume of deposits is possibly a good indicator of the bank s output of transaction services. Furthermore, when the return on the deposited funds consists primarily of low-cost transaction services, the cost to the bank of this input may appear in large part as non-interest operating expenses. On the asset side, some banks may be relatively specialized in lending to enterprises, others in retail 6 The first sub-period could be further divided into a pre-reform period , and a reform period The results of interest turned out not to be qualitatively very different when this refinement is adopted, and for the sake of concision not reported. 7 The nuclear detonation and the freeze of foreign deposits among other relevant events. 8 Data on some private banks begins in 1992, but in most cases the banks were clearly not yet operational for the full year, and in particular their interest revenue and expenses accrued for only part of the year. 10

11 business, and some may be engaged mainly in interbank lending and the holding of securities. We follow the intermediation approach (Sealey and Lindley, 1977), in which liabilities are considered as inputs and assets are considered as outputs. Labor and physical capital are considered as inputs that generate costs. In recent applications of this approach (e.g. Berger and Mester, 2003) financial equity capital is considered as a fixed input because it changes slowly and because its price is difficult to measure. Similarly, premises and other fixed physical assets are considered as fixed. Given the limitations of the available data, 9 we use a fairly comprehensive indicator of bank output. We define a bank s output y as the sum of earning assets, which comprise loans and advances, holdings of government securities and bills purchased and discounted, balances with other banks, and other investments. The price of banks output (denoted by PEARN) is defined as total income from interest receipts and fees, divided by earning assets. In an alternative specification we separate loans from other earning assets. Other earning assets are defined by the sum of government securities, bills purchased and discounted, balances with other banks and other investments. The two outputs are employed in the alternative profit function, cost and revenue functions, whereas the comprehensive output measure is employed in the standard profit function because we cannot compute separate prices for loans and other earning assets. Similarly, we use a comprehensive indicator of financial inputs (purchased funds), defined as the sum of deposits and amounts due to other banks. The unit price of purchased funds (PPURCHF) is defined as total interest expenses and fees divided by total purchased funds. A further distinction between different financial prices is not possible because interest and fees are aggregated on both the revenue and expenses sides in the reported data. In addition, we are not able to account for the quality of output because loan loss provisions are not separated from other non-specified costs for the earlier data. 10 We modify the standard intermediation approach because of the unavailability of data on the number of employees to measure labor costs per employee. We assume that non- 9 Revenues are aggregated into broad categories. In addition, the published classification of assets and liabilities and income statement items has changed over time so that further aggregation was necessary to have comparable data between and In any case banks may take charges for loan losses in a lumpy manner, and therefore it may be difficult to assign loan losses to particular years. When regulations were tightened in Pakistan, banks had to made a stock adjustment in provisions for losses incurred over many past years. 11

12 interest operating costs are incurred both in acquiring earning assets and in obtaining funds as inputs, including providing services to depositors. Therefore, we include non-interest unit costs, given by wage and administrative expenses divided by the sum of earning assets and payable liabilities (these other costs are denoted by OTHC). We consider two fixed netputs. The first is the sum of equity capital and reserves (CAPRES), since capital is an alternative to deposits and interbank borrowing in the financing of the bank s operations. Financial capital is considered as fixed because it changes relatively slowly over time and because its price is difficult to measure (see Berger and Mester 1997); the second netput is given by the value of fixed assets. For each bank the prices of earning assets and purchased funds are computed as exogenous market prices, given by the weighted average of prices of other banks, where the weights are given by the shares of the relevant input or output. 13 All financial variables are in 1995 national currency (deflated by the CPI). The variables used are summarized in Table 1. We employ two definitions of profits. The variable PROF1 is defined as profits before tax, given by total revenues minus total costs. A second definition, PROF2, is given by total interest and fee receipts minus total variable costs (the sum of interest costs and fees, and other wage and non-wage operating expenses). Total costs are given by the sum of interest costs and fees, wage and non-wage operating expenses plus other expenditures. The latter includes in the later period provisions for loan losses, which are instead not reported separately in the earlier periods. For this reason we include as variable costs the item other costs although we cannot determine if they are variable or fixed. Revenues are defined as total revenues (interest and fee receipts plus other revenues). In the cost function, costs and prices are normalized by the money market interest rate to impose linear input price homogeneity. Although not necessary, the constraint is imposed 11 For the earlier period administrative costs are given by: wage costs, rent, depreciation, directors and auditors fees, and sundry operating expenses; in the later period they are reported as a total. 12 One might want to exclude depreciation and rent from variable costs, but in the data set used here these items are conflated with other items such as lighting and telephone charges, which are variable costs. All these items are relatively small. 13 In an alternative specification we allowed for the possibility that banks may be facing segmented market for inputs or output and reran regressions using bank-specific prices. 12

13 on the profit function as well. This normalization is also an implicit way to control for differences in the regulated interest rates firs, and in market conditions later, over time. In all specifications costs, profits, revenues, outputs and fixed netputs were divided by total assets in order to account for differences in size between institutions and control for heteroskedasticity. In an alternative specification (not shown) we normalized the quantities by equity capital. In addition to reduce potential scale bias, the normalization yields dependent variables that can be roughly equated with common measures of bank performance; for example, the variable PROF1 is close to the concept of return on assets (ROA), when normalized by assets, and to the return on equity when normalized by equity (ROE). We prefer the normalization by total assets because in Pakistan capital requirements and provisioning regulations were tightened considerably during the sample period. In particular, the state-owned banks were severely undercapitalized in the earlier years, and over time were required to meet capitalization standards in line with international norms. Normalizing by capital would conflate these institutional changes with changes in bank behavior, and overstate profitability in the early years. We included a number of variables that proxy for environmental factors that affect bank performance; in particular we included the real rate of growth of GDP, the Herfindahl index of concentration of deposits and the market share of foreign banks as a measure of foreign entry. Unfortunately, information on non-performing loans is not available for the entire period so we cannot control for shifts in aggregate risk. 14 The evolution over time of some of the main variables is illustrated in the figures. Figures 1 and 2 show that the average unit price of earning assets PEARN and the average unit cost of funds PPURCHF were both fairly stable at around 11 percent and 6 percent, respectively, for the period before and during reform, but began to rise thereafter, though the spread was roughly constant. Variation in unit non-interest costs (operating costs divided by intermediated funds) increased substantially since 1987, which may make it difficult to determine accurately the role of this variable in the profit, cost, and revenue functions (Figure 3). 14 In any case, the aggregate level of non-performing loans would be problematic because it would be influenced strongly by the reforms implemented after 1997, that forced banks to reclassify their loans. 13

14 Figures 4 and 5 depict the paths of profits (both definitions) for different categories of banks, relative to their total assets. 15 The most striking trend is that the financial market reforms are associated to a sharp decline in profits, after a long period of growth. However, besides the different macroeconomic conditions, it should be noted that the average profitability is strongly influenced by the pattern of foreign banks, whose profitability declined after The profitability of the privatized banks increased after 1992 but less than the remaining state-owned institutions; however, their performance remained relatively stable whereas the state-owned banks experienced a drop to negative profits between 1996 and Figures 6 and 7 illustrate the pattern of revenues and costs for all banks and for each category. The period after is characterized by a rise in both costs and revenues. The foreign banks initially had relatively low costs, but seem to have converged following the reforms. The banks that were privatized do not appear to have been very different from the other public sector banks before the reforms, but subsequently their revenues rose relatively quickly. The new private banks, for which data exists starting in 1993, had comparatively low costs and low revenues in first two years of activity, but quickly caught up. 16 Concerning other variables of interest, Figures 8 and 9 show the composition of assets of different types of banks. The share of loans of public and the to-be-privatized banks was substantially larger than for the foreign banks until the late 1980s; since the early 1990s all banks seem to hold a similar proportion of assets in loans and advances. One exception is given by the state-owned banks, whose share of loans drops in , most likely because of the reclassification of loans that brought about an increase in bad loans and writeoffs. Finally, the low level of capitalization of the state-owned banks is shown in Figure 10. In 1980s capitalization was low for all categories of banks but it started to increase for the foreign and state-owned banks in the early 1990s. The private banks started with high levels of capital and reserves, but the public sector banks and in particular the foreign banks increased their capitalization significantly in the last sub-period. The data suggest that the 15 Each observation is the unweighted arithmetic mean of the relevant ratios in that period for the banks belonging to the category. 16 Ul-Haque and Kardar (1995) contain a further discussion of the balance sheet and revenue and expenses ratios of banks in Pakistan. 14

15 privatized banks did not participate in this tendency and remain substantially below the capitalization of all other banks Model specification and estimation The available data impose a particularly parsimonious specification of the cost, profit and revenue functions since we estimate the functions for each of the three periods separately (respectively, , and ). We use the translog functional form. 17 Thus, for the full sample, the standard profit function f π (p t, w t, z bt, v t ) is assumed to take the form ln[( π / w 3 z ) + θ ] = α i=1 β ln( w / w ) β ln( w / w 1 + γ 1 ln( p1 / w3 ) + γ 11 ln( p1 / w3 )ln( p1 / w3 ) δ 1 ln( z1 / z2 ) + δ11 ln( z1 / z2 )ln( z1 / z2 ) ηik ln( wi / w3 )ln( pk / w3 ) i=1 k= 1 i i 3 2 i=1 2 j= 1 ij i=1 r= 1 )ln( w / w ρ ln( w / w )ln( z τ kr ln( pk / w3 ) ln( zr / z2 ) + υ1 lngdpgr + υ 2 ln FOREIGN 2 k=1 r= υ3 ln HERF + υ11(lngdpgr) + υ 22 (ln FOREIGN) υ33 (ln HERF) + lnuπ + lnε π 2 i 3 ir j i 3 ) 3 r / z 2 ) where p 1 =PEARN, w 2 =PPURCHF, w 2 =OTHC, w 3 =MMRATE, z 1 =FIXA, z 2 =ECAP. The environmental variables are the real growth of GDP (GDPGR), the Herfindahl index of deposit concetration (HERF), and the share of foreign banks (FOREIGN); time and bank subscripts have been dropped for clarity. The right-hand sides of the equations for the alternative profit function, revenue, and costs are similar, except that output quantities replace output prices (y 1 =LOANS, y 2 =OTHINV). The scalar θ is added to profits of each bank to be 17 The limited number of observations available preclude the adoption of a more flexible specification such as the Fourier-flexible functional form, that would have used more degrees of freedom. 15

16 able to take logs even when profits were negative and was set equal to the minimum profit observed in the sample, plus unity. We estimate the profit, cost and revenue functions for three sub-periods including all banks, irrelevant of their efficiency levels. This way we obtain average-practice profit, cost and revenue functions for the three periods. The change in the average-practice functions across any two periods s ad t reflects both the change in productivity and the change in exogenous business conditions (the prices, outputs and environmental variables). The average practice functions are used to separate the effect of changes in productivity from change in exogenous business conditions. The change in productivity in the case of the profit function is given by the predicted profits computed applying to the function estimated in period s the exogenous conditions of the two periods. In practice, in the case of the profit function we take into account in the computation the scalar that is added to the profits before taking the log, that has to be subtracted from the numerator and denominator. We evaluate the numerator and denominator using the estimated parameters from period s and t for each pair of the sub-panels. The change in productivity obtained this way can be used to compute the effect of the change in exogenous business conditions (DEXCOND) as the ratio of total gross change in profits, costs or revenues to the change in productivity. In our application we modify the multiplicative approach followed by Berger and Mester (2003) in computing DEXCOND and DPROD. We compute the changes between each pair of period as a difference in percentage points rather than an index. Thus, our total change is given by: DTOTAL = DEXCOND + DPROD We prefer to show our results this way because in many cases one of the two components is negative and a ratio of the two would not be easy to interpret. In addition, our predicted profits, costs and revenues are as a percentage of total assets, hence it is more intuitive to present the results as differences Computation of efficiency scores and ranking The change in profit (cost or revenue) productivity is the result of changes a component due to firms getting closer to [or farther from] the efficient frontier, a component due to shifts in the best-practice frontier. To assess the relative contribution of these two 16

17 components we would need to define the best-practice frontier for each period, i.e. a set of banks that have zero inefficiency. In the case of Pakistan it could be the case that average productivity is strongly influenced by the state-owned banks even if the best-practice has improved. Rather than decomposing the change in productivity into its two components we prefer to compute the average efficiency score and the dispersion in the three periods. A full decomposition could be significantly affected by measurement error, given that the frontier is not perfectly identified. 18 Since we have panel data for each sub-period we follow the distribution free method (Berger 1993, Berger, Bauer etc. 1993; Bauer, Berger, Ferrier, and Humphrey, 1998). This method assumes that efficiencies are relatively stable over time whereas random error averages out. 19 The efficiency component can be estimated either as the individual effect from a panel regression or as the average of the residual from a pooled estimation. 20 Including a dummy variable for each bank to measure efficiency as fixed effect would reduce substantially the degrees of freedom. For this reason we prefer to estimate the efficiencies by averaging the residuals from the pooled estimation of the profit, cost and revenue functions for each subsample over which efficiency and the regulatory environment are assumed to be stable. 21 In this way a reasonable number of degrees of freedom are available to estimate each equation, and the distributional assumptions required are relatively weak. In the case of the profit function estimation, in each period the efficiency scores are computed as: 18 Berger and Mester (2003) employ the think frontier to obtain the shift in the frontier, and then derive as a residual component the shift in inefficiency. 19 The term distribution free indicates that no strong assumption is made concerning the distribution of the error and the efficiency terms, even though in principle the deviations from an efficiency frontier should be asymmetric. The distribution free approach is, however, fully parametric as opposed to nonparametric methods such as Data Envelopment Aanalysis. 20 Berger (1993) compares the results from the two methods and finds that efficiencies estimated with fixed effects are significantly lower than those estimated averaging residuals; however, the variables in the cost function were not normalized and fixed effect efficiencies were probably affected by scale bias. 21 In studies of the U.S. banking system, such as those cited above, the equation is often estimated cross-sectionally, and the error term for each bank is averaged across periods to obtain an estimate of that bank s efficiency. The pooled estimation was dictated by the relatively small number of banks in Pakistan. 17

18 πˆ i SPEFF = πˆ i max = i i i i i { exp [fˆ π (w, p, z, v )] exp[ lnû π ]}- θ i i i i i max { exp [fˆ (w, p, z, v )] exp [lnû ]}- θ π π For cost efficiency the scores are obtained as the exponential of the minimum average residual divided by the average residual of the bank. We also rank the efficiencies to compare the changes of the ranking between periods. Specifically, for each sub-period we compute the average efficiency of all banks and the average for classes of banks. 22 We then compare the change in efficiency accomplished by the banks that are state-owned to that estimated for the privatized banks and the foreign banks. For the second and third sub-period we can also compare the efficiency change for the private domestic banks. 4. Main results 4.1. Changes in profit, costs and revenues The various profit, cost and revenue functions were estimated separately for each subperiods to allow different coefficients. For the sake of brevity we do not report the estimated functions but only the results from the computations of the gross changes in profits, costs and revenues. We estimate the pooled model with OLS, that is standard in the estimation of Distribution-free efficiency, and by FGLS because the observations are likely to exhibit substantial autocorrelation and heteroskedasticity. One important issue with the estimation of Equation (1) (or the other analogous equations) is that since it represents a frontier it could therefore give rise to non-normal error terms. 23 In addition, the distribution of banks in Pakistan is strongly skewed in favor of a small number of large institutions. The least absolute deviation (LAD) estimator could be better suited to estimate the profit, cost and revenues functions. The LAD estimator is based on minimizing the sum of absolute deviations. The fitted regression line in this case is equivalent to fitting through the sample median rather than the mean hence LAD is also referred to as median regression. The results obtained with LAD 22 Rather than estimating bank-specific fixed effects, it is possible to estimate the average deviation from the efficiency frontier of the sub-set and test for significance by including a dummy variable identifying the banks of interest in the relevant time period. 23 This problem would be particularly relevant in the estimation with fixed effects. 18

19 are compared to those obtained by GLS and OLS because there are no previous applications of LAD to efficiency estimation. In the computation of the predicted values of profits, costs and revenues when the functions are estimated by LAD we employ sample medians instead of means. Overall, the fit is satisfactory for the first two periods and less so for the third period; in general, higher R 2 statistics are achieved for the cost and revenue functions than for the profit functions. The standard profit function seems to be somewhat better estimated than the alternative profit function, perhaps because the outputs are measured with error because quality cannot be accounted for. The top parts of the each panel in Table 1 (A-B-C) report the predicted values of standard profits, alternative profits, costs and revenues based on each of the estimation methods. The predicted values are computed at the sample mean (medians for LAD) of the exogenous variables. These values are in per cent of total assets. For example, panel 1A shows that overall predicted profitability is around 3 percent of total assets and is higher in the later period. Profitability based on the second definition (excluding other income and other expenses) deteriorates over time and is negative in between The results from LAD are substantially different from the others because they reduce the impact of the few large state-owned and privatized institutions. To make the results more clearly readable, we compute the changes as differences in predicted profits (costs or revenues) as a percentage of total assets. The total change and its decomposition is reported for each pair of periods. In all cases the results should be interpreted qualitatively because the predicted values of profits, and to a minor extent, of costs and revenues, are substantially different from the average profits, costs and revenues of the industry in each period. 24 This fact produces in some cases very large and implausible variations in productivity and external business conditions. For standard profits (definition 1) the results show that performance has increased on average by 50 per cent between the first and the third period (the predicted ROA is 1.4 percentage points higher in than in ). LAD shows a similar improvement but the time pattern is different because profitability increase substantially between the preand post-reform period, and then declines. As shown in Table 1B, the change is smaller for 24 The non linearity of the log transformation is such that the predicted value at the sample mean is not equal to the sample mean of profits, costs and revenues. Since we do not carry on the decomposition between productivity change and change business conditions, we do not rescale the predicted value as suggested by Berger and Mester (2003). 19

20 much larger for the alternative profits model but these results are strongly influenced by the poor fit of the regression by all methods. Since the results for the alternative profits model appear to be affected by a large bias as the predicted profits are different from the actual profits, we focus on standard profits. The change is not homogeneous across the two subperiods; between and there was a deterioration in profitability according to OLS and GLS estimates. LAD, that is more influenced by the behavior of private and foreign banks that are more numerous in the sample shows instead a pattern of sharp increase followed by decline. For the average bank the improvement was concentrated between the second and the third sub-period, although the results for the second definition of profits indicate that the origin of the change is not in the standard intermediation business but in extraordinary items. The pattern for costs is consistent across methods of estimation. Costs increased on average two to three times between and , but most of the growth occurred in the years after reform; a similar pattern occurred for revenues. The change is larger for both costs and revenues for the OLS and GLS estimates. The total change in revenues has more than compensated the change in costs. The same Tables (1A, 1B and 1C) report preliminary results for the decomposition of the total change into DPROD and DEXCOND for the three sub-periods. The exercise tends to yield extreme changes in some cases because the predicted values from the exogenous variables of other periods are unstable; this fact points to a limitation of the translog functional form when one moves away from the sample mean and should be further analyzed. The results for the standard profit model based on the first definition suggest that the total change that occurred between the pre-reform period and the period is the outcome of two opposite effects: a sharp decline in profit productivity and a substantial positive impact of the change in exogenous conditions (the function estimated for the third period applied to the initial conditions yields negative predicted profits). A similar results holds also for the shift between the and Further analysis should be carried on to verify which factors are more relevant among the exogenous conditions, in particular what is the contribution of the prices and of the environmental variables. The conclusions are different for the narrower definition of profits because productivity shows an increase, whereas the exogenous conditions have a negative impact. This finding is not in contrast with the available evidence because this profit measure captures mainly the effect of the spread between interest revenues and costs and the effect of operating costs. 20

21 The decomposition of costs suggest an increase in costs (a decline in cost productivity), mainly driven by changes in the exogenous variables. The deterioration of cost productivity has occurred in both sub-periods, whereas exogenous conditions had a favorable impact in the period immediately following the reform. Finally, the predicted changes in revenues (per unit of total assets) are very large, which suggest instability of the estimates; the pattern points to a decline in productivity and beneficial effects of exogenous conditions. The decline in revenue productivity occurred in the most recent period; between the pre- and post-reform period instead revenues moderately increased due to productivity, even if exogenous conditions had a depressing effect. Preliminary estimates of the frontier shift suggest that adverse movements in profit productivity were driven mainly by shifts in the best-practice; the analysis of average efficiency below tends to support this indication Efficiency We estimate the average deviation from best practice (X-inefficiency) in terms of profits, costs and revenues, as explained above (see equation (2)). In particular, the average efficiency based on the residuals from the estimated cost, revenue and profit functions can be calculated for the sub-periods and categories of banks. 25 The results are shown in Tables 3-5. The average deviation is of the order of percent for profit efficiency, pee cent for cost efficiency and per cent for revenue efficiency in the earlier period. In the second period average inefficiency slightly decreases for profits, and increases for costs and revenues, reflecting the greater dispersion in the data shown in the figures. Profit efficiency is the highest in the later period, whereas cost and revenue efficiency is slightly lower. The divergence in profitability was indeed somewhat larger in the second, post-reform sub-period (based on either the profit function or the alternative profit function), due to greater relative inefficiency in generating revenue; cost inefficiency declined between the first and second sub-periods, so that average cost X-inefficiency became less pronounced than revenue X-inefficiency The dispersion in the efficiency scores increased in , consistent with the intuition that the liberalization of interest rates and credit ceilings gave 25 In a previous version of this analysis, we also estimated efficiency scores from fixed effects obtained directly in the estimation of the cost, profit and revenue functions for each subperiod. The average inefficiency obtained was significantly larger, consistent with prior evidence (Berger, 1993) but the qualitative results did not change. 21