Why high productivity growth of banks preceded the financial crisis. Alfredo Martín-Oliver (Universitat de les Illes Balears)

Transcription

1 Why high productivy growth of banks preceded the financial crisis Alfredo Martín-Oliver (Universat de les Illes Balears) Sonia Ruano (Banco de España) Vicente Salas-Fumás (Universidad de Zaragoza) ABSTRACT April 2013 The high levels of operating efficiency, profs, and market values for banks in the years before the financial crisis raise reasonable doubts about the accuracy of the assessments of the efficiency of banking intermediation. We examine the productivy growth in Spanish banks in the pre-crisis period by separating out the contributions to productivy growth from business practices and from industry-wide technological progress. We find that more than two thirds of the estimated productivy growth in the years 2000 to 2007 is attributed to banks practices, such as the expansion of cred in the housing market, the high recourse to securization and short-term finance, the reduction in liquidy holdings, and the leveraging process of banks balance sheets, that the lerature claims are the ultimate causes of the crisis. We estimate that the remaining cumulative annual growth rate is 2.8% for the industry s technical progress, which is similar to that in the period of 1992 to JEL classification: D24, G21 Keywords: Productivy of banks, financial stabily production function, IT capal, simultaney bias. Alfredo Martín-Oliver acknowledges the financial support from project MCI-ECO Vicente Salas-Fumás also acknowledges financial support from project ECO of the Spanish MICINN. Address for correspondence: Alfredo Martín-Oliver; Universat de les Illes Balears, Ctra. Valldemossa km. 7.5, Palma de Mallorca, Islas Baleares, Spain. Tlf: ; alfredo.martin@uib.es. This paper is the sole responsibily of s authors and the views represented here do not necessarily reflect those of the Banco de España 1

2 1. Introduction Banks and other financial intermediaries perform the economic functions of providing liquidy, transferring funds from savers to investors, and collecting and diffusing information (Diamond, 1984; Diamond and Dybvig, 1983; Gorton and Winton, 2003; Merton, 1995). These functions involve value-added activies that facilate payments and manage cash, select and monor borrowers, and provide advice and consultation services. Banks use labor, capal, and other inputs to perform these activies and earn revenues from interest-rate differentials and fees. The level of efficiency when performing banking intermediation activies is a key factor in economic development (Buera et al. 2011; Greenwood et al. 2010; Mehra et al. 2011). Further, the changes in the costs of intermediation have important macroeconomic consequences for investment and growth (Bernanke et al. 1999; Christiano and Ikeda, 2011; Hall, 2011). In conventional competive markets, profs are the reward for providing services demanded by costumers at the lowest cost. The expansion of banks balance sheets around the world and the record-high growth rates of profs and productivy up to 2007 should have been an indicator of substantial efficiency gains in financial intermediation. However, the outburst of the severe financial crisis in 2007 showed that at least in the case of banks, the usual indicators of performance might have failed to disclose the true economic results. Potential explanations for this paradox might be as follows. Managerial incentives can exist that distort reported profs (Rajan, 1994). Also, financial innovations for regulatory arbrage (Achayra et al. 2011) and the measurements of profs and output not adjusted for risk can exist (Haldane et al. 2010). Further, innovations of business models that change the nature of the banks output over time (Philippon, 2012) can occur, such as the originate to distribute model and the market-based intermediation or shadow banking. In this paper, we rely on bank-level productivy estimations to quantify the productivy growth of the Spanish banking industry in the years before the financial crisis ( ) and examine s determinants. The Spanish example is a good case study for a better understanding of why the usual measures of efficiency and profabily for banks might not be informative about the true efficiency gains in 2

3 financial intermediation. First, the estimated productivy growth of the country s banking industry before the crisis was one of the highest among developed countries. Second, Diamond and Rajan (2009) consider what occurred in Spain as representative of the proximate causes of the crisis: (i) investors perceived a permanent reduction in interest rates when Spain joined the Euro zone, (ii) there was an unprecedented expansion of the housing industry, and (iii) banks financed a good portion of the loans wh wholesale financing and short-term debt. However, Spain also has different features from the Uned States and other countries in two main aspects. First, securized loans remained on the balance sheets of banks, and they were subject to capal requirements; and, second, savings banks compete on an equal basis wh commercial banks. We derive the estimates of the bank-level total factor productivy (from now on productivy) from the estimation of the banks production function. We model the production and sales of bank services at the branch level assuming a Leontief technology (Martín-Oliver and Salas-Fumás, 2008) wh two variable inputs: labor and services from information technology assets (IT capal) and a quasi-fixed input (the physical capacy of the branch). Then, we aggregate the branch-level production function to obtain the bank-level production function, which is the function that we empirically estimate wh the Spanish banks data. The estimation of the technology parameters follows the methodology developed in Olley and Pakes (1996) and extended by Levinsohn and Petrin (2003) to control for the potential simultaney bias between the unobserved productivy shock and the management decisions on input quanties in response to the shocks. Next, we explore what is behind the estimated productivy. For this purpose, we isolate the factors that can determine the estimated productivy values because of reasons different from true economic efficiency. These factors comprise, on the one hand, differences in the operating characteristics of the banks in the sample (Berger and Mester, 1997; Frei et al. 2000); on the other hand, they comprise factors related wh the proximate causes of the crisis, which is the focus of this paper. In other words, we aim to explore whether certain business decisions by banks improve the short-term private performance of banks but incur the social cost of future financial instabily. Examples of these business decisions are the concentration of loans in the housing market, the 3

4 issuance of securies and short-term debt to finance loans, or the reduction of liquidy holdings and the increase of leverage. Our empirical results show that the growth rate of productivy in the Spanish banking industry more than doubled during the years after the Euro, a result that is consistent wh other productivy estimates obtained wh other methodologies and wh aggregate industry data (O Mahony and Timmer, 2009). However, we also find that an important part of this productivy growth in the pre-crisis years is explained by business decisions that, ex-post, have been identified as drivers of the crisis. When removing these and other operational factors from the estimated productivy, the productivy residual grows at a similar rate in the years before and after the introduction of the Euro. Therefore, we show that the high growth rates of raw productivy estimated for the banking industry during the years prior to the crisis can not be taken as indicators of efficiency gains. The paper is related to the long list of published papers on productivy and the efficiency of banks. 1 We are the first to estimate the total factor productivy (TFP) from a Leontief-type production function formulated at the branch level. Most of the productivy estimates in banking are obtained wh cost or prof functions (Kumbhakar and Lovell, 2000). 2 This paper is also, to the best of our knowledge, the first to consider IT capal as a productive input, which seems essential in one of the most IT-capal intensive industries. This paper is also one of the few (together wh Bunch et al and Nakane and Weintraub, 2005) that corrects for the simultaney bias in the estimation of the production function, following the methodology in Olley and Pakes (1996) and Levinsohn and Petrin (2003). Our paper is also related wh the growing lerature that is interested in measuring the cost efficiency of financial intermediation, eher through a more accurate measurement of the output of banks and market-based (shadow banking) intermediaries (Philippon, 2012) or through the calculation of risk-adjusted measures of productivy (Basu et al. 2011; Haldane et al. 2010). Our contribution regarding this lerature is 1 Some reference papers in this lerature are Berger and Humphrey (1991, 1992, 1997), Berger and Mester (1997), Fixler and Zieschang (1992), and Sealey and Lindley (1977). Hughes and Mester (2008) and Berger (2007) provide updated lerature reviews. 2 There have been several papers published on the measurement and determinants of productivy and efficiency of Spanish banks (Grifell-Tatjé and Lowell, 1996; Lozano-Vivas, 1997; Maudos et al. 2002) but they all use a different method and explanatory variables so that their results are not comparable wh those in this paper. 4

5 twofold: First, we estimate the production function and the productivy values using bank-level data, whereas the previous papers use aggregate industry data for their estimations. Second, our analysis goes beyond the scope of these papers, obtaining a more accurate estimation of the contribution of technical progress to the productivy growth of the banking industry and provides an empirical test for some of the theories on the causes of the financial crisis. The rest of the paper is organized as follows. Section 2 describes the production technology of banks and the methodology used in the estimation of productivy. Section 3 shows the results of the estimation of the production function and the average productivy from the bank-level data for the Spanish banking industry. Section 4 contains an analysis of the determinants of the observed productivy of Spanish banks in the context of the banking practices that are related wh the causes behind the recent financial crisis. The conclusion summarizes the main results of the paper. 2. Production function estimation In this section, we describe the production function of banks proposed for the analysis as well as the econometric method used in the estimation of the technology parameters, including the productivy one. We adopt the production function approach instead of the intermediation one because is realistic to consider that banks consume labor and capal inputs to produce both loans and deposs, as well as related services such as payments, liquidy selling and bookkeeping of saving products, screening of potential borrowers and so on. Our starting point is based on the empirical fact that retail banks services are produced at branches that provide the physical space for employees, computer terminals, and other infrastructure. If the branches of one bank are relatively similar, the output (inputs) of the bank can be computed as the output (inputs) per branch, multiplied by the number of branches. Once in the branch, customers receive services that are produced by combining labor and IT capal inputs, which is the branch s capacy as an indivisible and fixed input. Because bank services are not directly observable and measurable, we rely on the assumption that the variabily across banks 5

6 in services produced can be approximated by the sum of loans and deposs at constant prices The production function We assume that the representative bank collects deposs, D, and grants loans, L, that require physical capal (branches, B), IT capal (IT), and labor (N). Because the services attached to these loans and deposs are provided in each branch, the inputs and outputs of banks are first defined at the branch level and then aggregated to the bank level. Each branch has a given capacy q. The number of workers per branch (N b ) and the IT capal per branch (IK b ) can be substuted wh each other, but not wh the physical capal. For a given number of branches (B), the total output of the bank that is defined as the sum of the loans (L) and deposs (D) can be wrten as follows: [ min { q, F ( )}] L + D = B N, IK b b. (1) Therefore, the branch production technology is a Leontief-type function wh a given investment in fixed capal that lims the total capacy of the branch. The function F( ) is assumed to be increasing and concave for the two variable inputs, labor and IT capal. Equation (1) assumes constant returns to scale at the bank level (i.e., the output of the bank is a scale factor of the output per branch). 4 If the function F( ) at the branch level is linear homogeneous, then equation (1) can be wrten as [ min { B q, F( N IK )}] L + D =, (2) where N=B N b and IK=B IK b denote labor and IT inputs at the bank level respectively. We assume that the capacy q is nonbinding for the standard branch, so the observed level of output is determined by the function F(N, IK). For the rest of the paper, the actual specification of the constant returns to scale production function will be wrten as, L + D = e N IK f ( ω) α 1 α (3) 3 Other papers that use the sum of loans and deposs as a measure of the single output of banks are Humphrey (1992), Prasad and Harker, (1997), and Tirtiroglu, Daniels, and Tirtiroglu et al. (2005). 4 We formally test this assumption and find empirical evidence supporting. 6

7 where f (ω) e is the term for the total factor productivy that increases wh the productivy shock ω Estimation of the production function: Methodology The estimation of the parameters of the production function (3) follows the methodology proposed by Olley and Pakes (1996) and extended by Levinsohn and Petrin (2003). In both cases the concern is to correct for the endogenous bias in the estimation of the elasticy of the output wh respect to labor and capal caused by the fact that the quanty of the labor input used in production might self be determined by the value of the productivy shock. The estimation procedure, adapted to this particular case, proceeds as follows. Let y = β 0 + β n + β ik + ω + ε (4) n k be the log-transformation of the production function in (3) where ε is a pure stochastic component. The term ω is a state variable in the firm s decision problem and, therefore, affects the demand for inputs. This variable is observable by the firm s manager but not to econometricians. We do not impose the condion of constant returns to scale, which will be empirically tested. Let the variable τ be one observable variable that depends on the two state variables ω and ik. This proxy variable τ is required to be monotonic in ω for all the values of ik because in doing so makes possible the inversion of the function and yield ω into a function of τ and the level of capal ik: 5 t ( τ ik ) ω = h,. By replacing this expression for the productivy in (5), the possibily exists to control for ω in the estimation of y n t ( τ ik ) ε = β n + ϕ, + (5) 5 Olley and Pakes (1996) propose to use capal investment as the proxy variable τ. Levinsohn and Petrin (2003) extend the list of observable variables correlated wh the productivy shock that can be used in the estimation procedure to eliminate the potential estimation bias. They argue that adjustment costs could imply that firms decide not to invest even though the productivy shock exists. To overcome this limation they propose to use intermediate inputs as proxy variables for productivy shocks. 7

8 where ϕ ( τ, ik ) β + β ik + h ( τ, ik ) t on (τ ιt, ik ιt ) from (5) we obtain: y 0 ik t. Subtracting the expectation in (5) condional ( y τ, ik ) = β n ( n E( n τ ik )) + ε E,. (6) This equation can be estimated using the nonparametric approach proposed by Levinsohn and Petrin (2003) that consists of first estimating the condional moments ( y t τ ik ) and E ( n t, ik ) E, t t τ by using a locally weighted quadratic least squares t t approximation and then by using a non-intercept OLS to get a consistent estimate of β n. 6 Parameter β ik that is associated wh IT capal is estimated in a second stage. The estimation begins wh the assumption that the productivy ω t follows a first-order Markov process: ( ω ω 1 ) ω = E + ξ (7) where ξ is the innovation term. Following Olley and Pakes (1996), β ik is identified by assuming that neher IT capal nor the lagged number of workers, respond to the innovation in productivy. The function E( ω ) ω can be estimated by locally t t 1 weighted least squares relying on the estimations of ω t and ω t-1 obtained from the firststage results and one candidate value for the coefficient associated to IT capal denoted as β. Thus, for a given candidate value, parameter β ik can be estimated from the * ik following equation: y * [ ω ˆ ω ] = β ik + ξ ε ( β ) ˆ β nn E 1 ik + (8) where the residuals are expressed in terms of the candidate β * ik. A set of orthogonaly condions is defined as follows: ( '( + ε )) 0 where z t is a vector that comprises {ik t, ik t-1, n t-1 }. Then, E( Λ) = E ξ = (9) z βˆ ik is estimated by minimizing the GMM crerion function defined from the orthogonal condions of the population: 6 Alternatively, parameter n β in Equation (5) can be estimated by using an OLS that comprises some approximation for function ϕ t (.). Olley and Pakes (1996) approximate this function wh a polynomial expansion in τ t and ik t 8

9 Q H * ( β ) = min β * Λ'( β ) Λ'( β ) h= 1 * * (10) where h indexes the H instruments. To measure the precision of our estimates, we use bootstrapped standard errors of the coefficients. 7 Finally, the GMM estimator of β ik is chosen for a grid search as in Levinsohn and Petrin (2003) because this estimator is more robust than using the starting candidate value estimator). * β ik (as for example the OLS 3. The productivy estimation for Spanish banks 3.1. Database and variables We draw the bank-level data from the nonconsolidated confidential balance sheets and income statements as well as from the complementary files reported by banks to Banco de España. The sample period spans from 1992 to 2007 and contains information on commercial and savings banks. We exclude cred cooperatives because they do not provide all the information that is needed in the analysis, as well as the banks whose market share of assets is smaller than 0.1%. When two banks merge, we consider that a new bank brand is created. In 2007, banks in our sample represent 89.25% of the Spanish banking industry in terms of assets. This coverage is similar in terms of other variables, such as the number of employees, and remains fairly stable across the time period. We consider three different inputs that enter into the production of banking services: (i) the input services from the physical capal (i.e., number of branches of a given capacy), (ii) the services from IT capal, and (iii) the services from workers. The fixed capacy, q i, per bank branch i is obtained by dividing the replacement cost of all of the bank s buildings by the number of branches owned (i.e., we assume homogeneous branches for each bank). The total capacy of the bank is equal to B i q i, where B i is the total number of branches (owned and rented) of bank i. 7 Petrin et al. (2004) provides an estimation command that implements this method into Stata. This command allows the estimation of the production function by using one or two proxies for productivy and one variable of capal (non-variable inputs). In this paper, we will need the inclusion of two variables for capal, that is, branches and IT capal. 9

10 The banks report their total IT capal on the asset side of the balance sheet, and they also report an annual flow of IT expendures in the income statement. We define the total IT capal of bank i in year t, IK, as the sum of the IT capal in the balance sheet at book value plus the estimated capal stock accumulated from annual expendures assuming a perpetual inventory model wh a depreciation rate of 35%. In the calculation of the total of IT capal at replacement cost, we assume that incorporated technical progress compensates for the price inflation. 8 Also, the labor input services of bank i in year t, N, is measured as the average number of employees of bank i during year t. The output of the banking firm is approximated by the sum of the loans and deposs. The balance of loans and deposs in year t is calculated at homogeneous current prices by applying the permanent inventory model. Following this scheme, the estimated volume of deposs (loans) in year t is equal to the deposs (loans) in t-1 valued at year t prices using the general inflation price index plus the flow of new deposs (loans) in year t. After their estimation at current prices, all inputs and outputs in monetary uns are deflated and expressed in the prices of year Table 1 presents the time evolution of the descriptive statistics of the inputs, outputs, and the number of branches per bank. The data shows that the number of banks in the Spanish industry over time has fallen from 140 in 1992 to 90 in For the rest of the variables in the table, the mean is substantially above the median that indicates that the distribution of the variable is highly skewed. The growth rate of the average number of employees per bank is smaller than minus the growth rate of the number of banks, which implies that the total number of industry employees decreases in the period. On the other hand, the stock of IT capal increases over time. This increase suggests that labor is substuted by IT capal in the input mix of banks. More concretely, the average stock of IT capal per employee increases 68% during the sample period. The average rates of growth of output per bank and IT per bank are similar (around 7.7% of average annual growth rate) and much higher than the growth rate in the average number of workers per bank (around 1.5%). This increase in output implies an important increase in labor productivy during the period (around 130% in 2007 compared wh 1992), as well as an increase in total productivy. 8 Martín-Oliver et al. (2007) explain in detail the method used in the calculation of physical and IT capal of Spanish banks. 10

11 3.2. Estimation of the production function Table 2 presents the estimates of the parameters of the production function. The upper part of the table shows the estimates of the parameters of the production technology Equation (4) that ignores the simultaney bias (i.e., the OLS). The lower part reports the estimates that control for the simultaney between the efficiency shocks and the labor input decisions that use the investment in IT as a proxy for the productivy shocks (in line wh Olley and Pakes, 1996). We use the investment in IT capal as the proxy variable τ after comparing the results wh those obtained wh other alternatives, mainly externally supplied intermediate inputs. Levinsohn and Petrin (2003) pos that intermediate inputs are an alternative proxy to investment in capal because of the existence of adjustment costs that could result in firms that do not invest in some periods. This existence implies a large proportion of zero-investment observations in the sample that might not be usable in the estimations. In our application, all of the banks in the sample invest a posive amount in IT in all the years of the period. In banking firms, the link between the productivy shocks and the external purchases of intermediate inputs might be weaker than in industrial firms because in banks the elasticy of output to variations in the inputs is expected to be small compared to that in industrial firms. In addion, investment in IT is the variable that is better fted to the three specification tests posed by Levinsohn and Petrin (2003) for the selection of the proxy variable: (i) monotonicy, that is, the higher levels of investments are associated to higher values of productivy for any level of IT capal (see Figure A1 in the Appendix A); (ii) correction of the bias in which the estimated coefficient of labor (IT capal) is lower (higher) in the correction of the simultaney than in the OLS estimation; and (iii) the orthogonaly of the freely variable input (labor in this case) and the innovation in productivy in t+1 [the estimated correlation (-2.07%) is not significant at 10%]. The two columns in the left-hand side of Table 2 correspond to the estimation when the condion for the constant returns to scale at the bank level is not imposed. The total inputs of each bank are wrten as the product of inputs per branch times the number of branches. In this specification, the hypothesis of constant returns to scale at the bank level is satisfied if the coefficient associated to (the log of) the number of 11

12 branches is equal to one. The right-hand panel exhibs the estimation results when the constant returns to scale at the bank level are imposed: inputs and outputs are all defined at the bank level and the number of branches is excluded from the right-hand side of the equation. The p-values associated to the null hypothesis of constant returns to scale at the bank level (i.e., the hypothesis that the coefficient of the number of branches (in logs) equals one) and at the branch level (i.e., the hypothesis that β n + β ik = 1) are also reported. Considering the OLS estimation results reported in the upper part of Table 2, the null hypothesis of constant returns to scale is rejected at both the bank and branch levels. The respective estimates of the elasticy of output wh respect to labor and IT capal are 0.69 and 0.22, and the coefficient of ln(branches) is Therefore, the estimation that uses the OLS implies that the production technology of banks has decreasing returns to scale, both at the bank and the branch levels. Results are different when the estimation is performed taking into account the potential simultaney between the firm s input decisions and s productivy when using the IT investment as the proxy for productivy (lower part of Table 2). Now, the null hypotheses for constant returns to scale at the bank and the branch levels cannot be rejected at any standard significance levels. That the OLS estimation overestimates (underestimates) the elasticy of output to labor (capal), compared to the values from the Levinsohn-Petrin estimations confirms the simultaney bias affecting OLS estimations of the production functions. If the null hypothesis of the constant returns at the bank level is imposed, the estimated elasticy of output to labor and IT capal is and respectively (compared to and 0.207). This elasticy of output to IT capal implies that if the stock of IT capal per employee doubles then the output per worker increases by 42.1%. In the period 1992 to 2007 the stock of IT per employee increased from 10.2 to 16.7 thousands of Euros (64 %). Therefore, the deepening of IT capal increased the labor productivy of Spanish banks by 26.9% (0.421 multiplied by 64) from 1992 to Robustness tests Table A1 of Appendix A reports the estimates of the production function wh a list of alternative intermediate inputs used in solving the simultaney bias. Our final 12

13 choice is the investment in IT capal because the other potential candidates do not fulfil the minimum requirements stated by Levinsohn and Petrin (2003). First, the use of external administrative services, in spe of providing similar results, drastically reduces the sample because of information is only available from 1999 on. Next, neher total operating expendures nor office stationary (and, to a lesser extent, electricy supply) satisfy the monotonic condion for the low levels of IT capal: Figure B1 in Appendix B shows a decline or stagnation in the productivy at high levels of IT capal when the intermediate input increases. Also, all of the proxy candidates fail to have a correlation between labor and the productivy shock t+1, wh the exception of external administrative services (-5.3%, non-significant at 10%). These reasons can explain why electricy and other supplies return similar estimates to investment in IT capal, but wh decreasing returns to scale at the branch level (against the findings of Martín- Oliver and Salas-Fumás, 2008). These reasons also help explain why office stationary and operating expenses provide the estimated elasticy in the output to the inputs that are to low compared to the rest of the estimates (labor is below 0.3 and that IT capal is even smaller than in the OLS). Further, following Levinsohn and Petrin (2003), we estimate the production function for different time periods to allow for variation in the coefficients that measure the contribution of the inputs to the output. In particular, we estimate the production functions (at both the bank and branch levels) by distinguishing among three subperiods: , , and The differences among the estimated coefficients that correspond to the different periods are not statistically significant. Therefore, we present only the estimates for the whole sample period and focus on these coefficients to measure and further analyze the productivy level of Spanish banks Productivy in the banking industry Using the estimations in Table 2, the productivy level of bank i in year t, p, can be estimated as 9 : ln p = ln( L + D) ln N ln IK. Then, relying on the bank- 9 As in Bunch et al. (2009), Nakane and Weintraub (2005), and Olley and Pakes (1996), we estimate the productivy as a residual from the difference between the observed and the predicted output of the bank in time period t, not the productive efficiency (distance to an efficient frontier) to be consistent wh the general methodology. Implic to the paper is the assumption that the elasticy of output wh respect to labor and IT capal are the same for banks of different characteristics; in other words, the heterogeney of banks only affects the constant of the production function. 13

14 level productivy estimates, we construct the indicator of industry-wide average productivy as the weighted average of the banks productivy by using the shares of the banks in terms of output as weights. Olley and Pakes (1996) distinguish between two sources that might explain the evolution of the industry s productivy, the evolution of the (un-weighted) average productivy of the firms in the industry, and the differences in productivy that are associated wh differences in the sizes of banks: p t = Nt N t i= 1 s p = p t + ( s st )( p pt ) i= 1 (11) where and p t is the industry s productivy at time t, s is the share of bank i at t, and s p are the un-weighted means of the bank s productivy and output shares respectively. A posive (negative) value of the second term of the right-hand side indicates that larger banks tend to be more (less) productive than smaller ones. 10 Figure 1 shows the evolution of industry productivy (p t ) and s two components. The productivy of the banking industry shows an increasing trend over the whole time period that is attributable to both the productivy gains of the average bank and to a posive reallocation effect. The fact that bigger enties are more productive and increase their output share in the industry results in a higher growth rate in the average productivy of the industry, compared to the suation where all of the banks are equally productive. This reallocation effect is clearly manifested after the years 1999 to 2000 when two different mergers (Banco Santander-Central Hispano and BBV-Argentaria) created the two biggest banks in the Spanish banking system. These mergers enhanced the contribution of the size effect to the productivy of the industry. Focusing on the evolution of productivy in the Spanish banking industry, Figures 2A and 2B show the annual and the cumulative growth rates respectively of the estimated productivy of the banking industry p t. The weighted average productivy in 2007 is 2.8 times the value in 1992, which implies an increase of 180% (Figure 2B). Most of the increase in the aggregate productivy occurs in the second part of the period ( ) when the average annual rate of growth was 10.01%, compared to the 3.85% average annual rate during the 1992 to 1999 period (Figure 2A). 10 For a similar decomposion applied to Spanish manufacturing industry see Fariñas and Ruano (2004). 14

15 4.What is behind the observed productivy? Our measure of productivy for each bank and year is the residual obtained from the difference between the actual total output of the bank (loans plus deposs) minus the output predicted from the quanties of labor and IT capal used in production. One of the determinants of the productivy differences that matters most for welfare analysis is the underlying differential in intermediation efficiency. We measure the intermediation efficiency of the banking industry as the Hicks-neutral technical progress that determines the time trend in productivy after removing other sources of crosssectional and time variabily from the pooled data of banks productivy. Among these other sources of productivy differences, we particularly focus on the changes in the balance sheet of banks that have been pinpointed as potential factors for the outburst of the financial crisis: increasing risk taking, excessive growth, and higher illiquidy risk. In this section, we first show the changes in the assets and liabilies of Spanish banks that preceded the financial crisis. Next, we hypothesize how these changes might affect the estimated productivy of banks. Then, we estimate an empirical model on the determinants of the banks productivy for the double purpose of testing the hypotheses and estimating the residual technical progress in the banking industry The behavior of Spanish banks in the pre-crisis period The suation for Spanish banks in the pre-crisis period can be summarized as follows: (i) Spain joined the Euro zone and benefed from low interest rates and the financial integration of the monetary policies by Central Banks around the world, including the ECB, that were extremely accommodative; (ii) the prices and demand of real-estate related assets, including houses, experienced a high increase and banks provided the cred to fund this expansion; (iii) Spanish banks used a combination of mortgage-backed securies (MBS) and short-term funding to finance the lending to real-estate developers and house buyers; (iv) the banks were able to comply wh regulatory capal and sustain the high level of cred growth by issuing hybrid financial products that were cheaper than pure equy. The behavior of Spanish banks during these years is close to what Diamond and Rajan (2009) describe as the proximate 15

16 causes of the crisis. 11 Table 3 shows how this suation translated into important changes in the banks balance sheets. We present the mean and median values of the variables that capture the mortgage and real-estate lending activies, the securization intensy, the importance of short-term finance, the evolution of liquidy, and the equy capal ratio (inverse of leverage) of Spanish banks in three successive time periods: the pre-euro years of 1993 to 1997, the consolidation period of 1998 to 2002, and the period of high growth in 2003 to More complete definions for the variables are in Appendix B. According to Table 3, the average proportion of real-estate loans increased over time from 12.7% in 1993 to 1997 to 21.0% in 2003 to 2007, while the proportion of mortgages rose from 32.1% of all loans to 51.8%. These percentages confirm the expansion of the lending activy in real estate and housing markets after Spain joined the Euro zone. During the years previous to the Euro, Spanish banks did not participate in securization activies, but afterwards the number of banks issuing MBS increased steadily: in the period of 2003 to 2007 these securies represented 15.2 % of the total assets for the banks that issued them. The matury of the wholesale finance for the banks that got funds from these markets is measured by the weighted matury of wholesale financing (variable Duration) and by the net posion of banks in the interbank market. The mean value of Duration increases over time that indicates that Spanish banks issued securies of longer matury in wholesale markets at the same time that the weights of the long-term loans in real estate and housing were also increasing. Thus, from this point of view, the banks liquidy posions were then maintained. Spanish banks also increased their dependence on interbank finance over time, both in terms of a larger number of banks wh a net borrowing posion in the interbank market (variable IdB) and in terms of a higher value of the ratio for borrowing over lending in this market (variable IB borrow/ib lend). Comparing the time periods of 1992 to 1997 and 2003 to 2007, the proportion of banks wh a net borrowing posion grew from 20.6% to 53.2% and the ratio of the amount borrowed and the amount lent in 11 Regulatory arbrage through securization (Achayra et al. 2011) was not possible among Spanish banks. Banking regulation forced banks to keep the issued securies in their balance sheet unless banks were effectively transferring the risk out of the bank, which did not happen in most of the cases (banks kept worse tranches or granted cred enhancements). Thus, securized and non-securized loans had the same impact in terms of regulatory capal requirements. 16

17 the interbank rose from 0.86 to 2.56 respectively. Assets liquidy holdings also decreased. The proportion of liquid assets including cash, reserves in the Central Bank, and bonds issued by the government wh respect to deposs decreased from 25% during the period of 1992 to 1997 to 13.4% in 2003 to As wh banks elsewhere, Spanish banks responded to the banking integration by reducing their liquid assets and shifting their portfolios towards investments wh higher returns but lower liquidy (Castiglionesi et al. 2010; Rajan, 2005). Further, Table 3 shows a decreasing trend in the equy capal ratio during the 1990s from 9.9% in the period of 1998 to 2002 to 6.8% in the period of 2003 to However, the regulatory capal ratio stayed relatively stable during these time periods (16.15% in and 14.29% in ), which means that the regulatory capal requirements derived from the high growth of the years 2003 to 2007 were fulfilled wh debt-like instruments Implications for productivy and productivy growth We now explain why the changes in the banks balance sheets can have posive effects on productivy and productivy growth over time Real estate and mortgage loans. We identify two reasons why the concentration of loans in real estate and mortgages can have posive effects on productivy: lower screening and higher loanto-value ratios. On the one hand, the incentives of banks for screening the cred qualy of the borrower in collateralized loans are milder than in loans whout collateral (Manove et al. 2001). Mortgages and real-estate loans are collateralized so lower screening implies more loans granted (or produced) per employee. During the years previous to the crisis, screening incentives might have been even milder because the escalation in house and real estate prices was perceived as an addional guarantee for lenders. On the other hand, the expectation of the permanent increase in house prices could have encouraged bank managers to grant loans wh higher loan-to-value ratios than in periods wh flatter expectations. This higher loan-to-value ratio per average loan will be translated into higher productivy because, wh the same amount of inputs and the 17

18 same number of loans, the total balance of loans granted will be larger because the average size of the loans granted is also larger Securization hypothesis We expect that banks that tradionally financed their loans wh deposs and, at some point in time, started to obtain finance by issuing securies backed wh loans will experience an increase in productivy for reasons related wh changes in the business model rather than wh efficiency gains. In tradional banking, there is a relation between the use of inputs (labor and IT capal) and the total output of banks in such a way that an increment in the bank s output (loans plus deposs) has to be accompanied by an increase in the use of inputs. However, in the originate to securize model, banks have become intermediaries, not producers of deposs and loans. The issuance of MBS provided Spanish banks wh funds for loans directly from the market and outside the network of branches. Therefore, banks could get funds to grant new loans whout opening new branches and/or increasing the use of inputs in old and new ones, something that they could not do if they were to obtain the funds from deposs. Securization also changed the lending practices of banks. The credworthiness of securized loans is based on the qualifications of rating agencies and not so much on the detailed soft information collected by bank officers. These agencies do not have privileged information about the beneficiary of the loan (homeowner, for example), so they only process hard information such as the cred score of the borrower and the loan-to-value ratio. As a consequence, bank officers stop collecting this information and focus only on lending to borrowers who have good cred scores and observable adequate loan-to-value ratios (Diamond and Rajan, 2009). Collecting the useful soft information on the cred qualy of the borrowers is more time consuming than collecting the hard information of the cred score. This is another reason why issuing MBS banks might have increased the ratio of output relative to inputs Short-term wholesale funding and liquidy Securization is not the only way banks can raise funds in financial markets. Under the umbrella of the Euro, Spanish banks became net borrowers in the interbank market, and they issued bonds and other debt instruments to fill the gap between loans and deposs. 18

19 Ideally, long-term loans should be financed wh long-term debt instruments but, under the expectation of future low interest rates, leveraged banks have incentives to finance wh short-term debt; thus they become more illiquid (Diamond and Rajan, 2009b). Moreover, investors are keen to provide banks wh short-term finance because facilates the option to ex if things go wrong and banks get into trouble (Diamond and Rajan, 2001). Spanish banks were probably affected by these incentives and took advantage of the cheaper and practically unlimed short-term finance through the interbank and the financial markets. Therefore, banks will tend to adopt a financial structure more oriented towards short-term leverage, even though their loans have a long-term matury. Banks hold liquid assets to provide liquidy to financial markets. Liquid assets act as slack resources that smooth the transmission of financial shocks to the real sector of the economy (Rajan, 2005). The decreasing trend observed in the asset liquidy ratio (Table 1) will then have a posive effect on productivy growth Regulatory capal and leverage ratio Regulatory capal standards are intended to restrain the growth of bank cred and to lim leverage ratios by preventing excessive risk taking (Kim and Santomero, 1988; Morrison and Whe, 2005; Rochet, 1992). Over the years leading up to the financial crisis, there is evidence that banks changed the composion of their regulatory capal to a higher proportion of subordinate debt and preferred stocks (Acharya and Schnabl, 2009; Acharya et al. 2012; Khorana and Perlman, 2010). The high rates of the cred growth in loans among Spanish banks together wh stable dividend policies increased the regulatory capal requirements above the retained earnings. In order to comply wh capal requirements, the banks issued hybrid instruments instead of issuing equy 12 because the former is less costly (i.e., interest from hybrid instruments are tax deductible). As a result, we can observe the apparent paradox of banks keeping their regulatory capal ratio at constant levels while they become more leveraged. 12 Savings banks do not have equy on their balance sheet, so they cannot issue common shares. 19

20 4.3. Empirical model on the determinants of banking productivy The full econometric model on the determinants of the productivy differences in banks is formulated as follows: ln p 2007 = j j γ 0 + γ j x + ϕ j z + θ tdt + ν, j j t= 1993 (12) The dependent variable is the log of the productivy of bank i in year t obtained as a residual. There are three sets of explanatory variables. The first one, j x, comprises the variables in Table 3 that account for the presumed posive effect on productivy because of the cred growth in real estate and mortgages, securization, short-term finance, and leverage. The second block, j z, corresponds to the control variables such as ownership, market scope, size, qualy of the inputs, priced services, growth, merger activy, risk, and so on that are relevant to explain the productivy differences among banks in previous studies (Berger and Mester, 1997; Frei et al. 2000). The precise definion of each of these variables appears in Appendix B and the descriptive statistics appear in Table 4. The third block of explanatory variables is the time dummy variables d t, that equal one when the observation belongs to year t and zero otherwise. The parameters associated to the time-dummy variables, θ t, capture the time effects on productivy common to all banks in the industry. We estimate a variation of model (7) where the time-dummy variables are replaced by macroeconomic variables for the Spanish economy (i.e., inflation, interest rates, and business cycle) together wh a timetrend variable. In this specification, the coefficient for the time-trend variable is our estimate of the industry s technical progress. Finally, ν is the random error term. The model is estimated by using an OLS wh standard errors clustered at the bank level, and Table 5 presents the results. The column Specification I shows the results from estimating the model as is formulated in (12). The column under Specification II shows the estimation results when the time dummy variables in (12) are substuted wh the macroeconomic variables (growth of GDP, inflation, and interbank interest rate), and the time-trend variable that are common to all banks. Also, the results in the column under Specification III correspond to model (12) wh bank fixed effects (bank dummy variables of ownership, size, and geographic market are excluded). The overall conclusions about the determinants of the cross section and the over time 20

21 evolution of the productivy in Spanish banks in the sample period are similar for the three specifications. In the fixed-effect estimation, the magnude and sign of the main variables do not vary, though the statistical significance of some variables changes (i.e., liquidy becomes statistically significant, and salary loses s significance) Productivy and proximate causes of the crisis Our estimation results confirm that the changes in the asset composion of the banks and the proximate causes of the crisis can to a large extent explain the differences in the observed productivy of banks. As expected, the proportion of real estate and mortgages in the loan portfolios of banks is posively correlated wh productivy. This result confirms the hypothesis that specialization in these types of products in the precrisis period can explain part of the observed increase in gross productivy of banks over time. The coefficient for the variable proportion of the securized assets is also posive and statistically significant. This coefficient confirms that generate-to-securize is more productive in terms of the labor and IT capal services consumed than tradional banking is. The negative coefficient of the variable Deposs/Loans correlates to this result, although is not statistically significant once we control for the other finance instruments. The explanatory variables on the short-term sources of the funds that finance the gap between loans and deposs are also posively related to the banks productivy. The estimated coefficient of the variable IdB IBborrow/IB lend is posive and statistically significant at 1%, which implies that the higher the net borrowing posion in interbank markets is, the higher the bank s productivy level is. 13 The rest of the coefficients related to the short-term finance are non-significant. Therefore, what matters for productivy is the net lending posion in the interbank market, not the net borrowing one. As expected, the assets liquidy ratio contributes negatively to productivy, although the estimated coefficient is statistically significant only in the fixed-effect estimation. Furthermore, the estimated coefficient for the equy capal ratio is negative and statistically significant. More leveraged banks have higher productivy than the less leveraged banks because the latter are expanding their balance sheets wh non-core capal instruments. Nonetheless, these banks manage to keep their 13 On the contrary, the coefficients associated to the lending posion in the interbank market are not significant. 21

22 regulatory capal ratios relatively constant over time (see Table 3) because they issue hybrid and debt-like instruments that count as regulatory capal. Therefore, the posive effect of leverage on banks productivy is because of the increase in leverage whin the regulatory capal, that is, the increasing weight of hybrid instruments is detrimental to the core capal Control variables The ownership, size, and market scopes of banks affect their productivy. Saving banks are 25% less productive than commercial banks, while foreign subsidiaries are almost 20% more productive than national commercial banks. 14 The size of the banks has a posive effect on productivy, as well as their concentrations in local and regional markets. The productivy of banks also varies wh the qualy of the productive inputs. The posive association between the average salaries of banks and their productivy suggests that higher salaries go together wh more productive workers. Next, the posive (although not statistically significant) coefficient for human capal from training and the negative coefficient for the proportion of temporary employees also point to a posive effect for human capal on the bank s productivy. In addion, a higher proportion of advertising capal in the total operating capal of banks also has a posive effect on productivy. Banks collecting more revenues in the form of net commissions (relative to total assets) are less productive, possibly because commissions are associated wh services that banks provide to their customers, and these services are not properly captured by the output measurement used in this paper. Because the inputs involved to produce these services are effectively accounted for, we find that banks wh a business profile that is more oriented to services and that charge more commissions are penalized in our productivy measure. 15 Table 5 also shows that a higher annual growth rate in the number of branches has a negative effect on productivy, possibly because of the 14 Berger (2007) reviews the lerature on productivy comparisons between foreign subsidiaries and national banks; broadly, foreign subsidiaries tend to be more productive than national banks. The results of the comparison might be affected by the differences in the portfolios of services and markets served by each group of banks. 15 We compute the net present value of the flow of the commissions assuming a permanent annual flow equal to the current value of the net commissions by using as the discount factor the current value of the 12-month Interbank interest rate. When we recalculate the productivy measure and estimate the parameters in model (7), the variable commissions over the total assets are no longer statistically significant, which confirms our interpretation of the results. 22