Article body

To price the credit sought by firms, banks must estimate the quality of each applicant’s project accurately. They use different sources of information: accounting and financial data, credit history, firm characteristics, potential competitors, abilities of managers and so on. As Stein (2002) suggests, this information can be split into two parts. “Soft information” corresponds to qualitative details, such as the business skills or honesty of the firm manager. “Hard information” instead encompasses all quantitative data, such as accounting and financial data and credit scores. Using these two kinds of information, Berger and Udell (2006) distinguish two main categories of lending technology:

  • Transaction-based lending technology, primarily based on borrowers’ hard information.

  • Relationship lending technology, primarily based on borrowers’ soft information.

As Elsas (2005) notes, the choice between these two technologies depends on the characteristics of the borrowers, the bank and the market. Evidence clearly indicates that small and medium-sized enterprises (SME) and non-hierarchical banks benefit more from the relationship lending technology (Berger and Black, 2011; Berger et al., 2005; Berger and Udell, 1995), but the impact of banking competition on the choice between transactional and relationship lending technologies remains uncertain. Mayer (1988) proposes an initial answer, by arguing that competition and relationship banking cannot coexist, because firms can easily switch banks, so banks have no interest in developing relationships. Petersen and Rajan (1995) adopt a similar view, then show theoretically and empirically that a bank in a competitive sector restrains its credit and uses only transactional lending. Ogura (2010, 2012) and Fischer (2000) confirm this discovery empirically.

In contrast, with their theoretical model, Boot and Thakor (2000) show that the greater the competition between banks, the more banks engage in relationship lending. Their result reflects an informational advantage, such that accurate and private information about borrowers provides good protection against banking competition. Dell’Ariccia and Marquez (2004) go a step further and demonstrate that when competition increases, banks should extend their lending activities to include more “opaque” firms (i.e., the flight to captivity) (see also Hauswald and Marques, 2006). Schmeits (2005) also points out that competition is necessary to initiate relationship lending, because without competition, banks use the flexibility inherent in these contracts to demand high rates (i.e., holdup problem). Bonfim, Dai and Franco (2009), Montoriol-Garriga (2005), Black and Strahan (2002) and Memmel, Schmieder and Stein (2008) empirically validate these theoretical results.

How can we explain these discrepancies in empirical results? A first explanation is based on a non-monotonic approach to the relationship between banking competition and relationship lending. Even if relationship lending provides some protection against banking competition, this protection might weaken when competition increases: More lenders implies lower monopoly rents. The costs of achieving relationship lending are fixed though, so there is a threshold for competition, beyond which building a relationship is unprofitable and banks prefer to offer transactional lending. The relationship between competition and relationships then might not be monotonic but rather concave, prompting transactional lending when competition is low, relationship lending when it is medium and then transactional lending again when competition is high (Anand and Galetovic, 2006; Dinç, 2000; Yafeh and Yosha, 2001). Although Elsas (2005) and Degryse and Ongena (2007) empirically confirm this nonlinear relationship, they observe a convex, rather than concave, form. Finally, this nonlinear relation could be apparent rather than actual. As Presbitero and Zazzaro (2011) show, the nonlinear relationship between banking competition and the supply of relationship lending disappears when bank size is taken into account. Greater concentration causes banks to refocus on their core business, namely, relationship lending for small banks and transactional lending for large ones.

Another explanation relies instead on the measures of relationship lending that prior studies adopt. First, studies vary in how they measure relationship lending (Table A1 in the Appendix outlines the relationship proxies used in prior research). For example, Petersen and Rajan (1995) use firm age, whereas Bonfim et al. (2009) use the number of banks. These differences could affect the outcomes. The conflicting results obtained by Montoriol-Garriga (2005) and Petersen and Rajan (1995) illustrate this problem succinctly: Both use the same database (NSSBF 1988), but the former measures relationship lending according to the number of banks, whereas the latter relies on the age of the company. Second, measures of relationship lending may be problematic. For example, the size of a bank seems like an ambiguous proxy, and Uchida et al. (2012, p. 97) even show that though “loan officers at small banks produce more soft information than a large bank, large banks have the equivalent potential to underwrite relationship loan.”

Finally, most empirical studies distinguish relationship and transactional lending using what may be an overly simplistic categorization, with the assumption that the two technologies are mutually exclusive (e.g., Petersen and Rajan, 1995; Berger and Black, 2011). However, Uchida et al. (2006) and Bartoli et al. (2013) show that banks frequently rely on multiple lending technologies banks to finance SMEs. Assessing lending technologies without considering the potential complementarity among them thus might bias some results and explain the conflicting conclusions reached thus far.

The main goal of this article is to address the impact of banking competition on lending technologies by using a new measure of the relationship lending technology that captures the amount of soft information that the bank uses to price the loans requested by firms. We do not attempt to include all dimensions of relationship lending but instead acknowledge that banks and firms implement this form of financing for two main reasons, beyond increasing the information held by the former about the latter. First, some firms seek a protection against troubled times. With a longstanding relationship, a bank can support its customer, even during bad periods (e.g., by charging lower interest rates), and in return, the customer compensates for the loss when its situation improves (Sharpe, 1990). Second, such relationships help banks cross-sell other products or services to borrowers (Santikian, 2014).

To measure relationship lending technology according to the amount of soft information used by the banks in the loan process, we start at the same point as Cerquiero et al. (2011), who seek to explain the dispersion of loan rates offered by banks to small enterprises. In a frictionless world, such dispersion should not exist, and similar firms obtain similar rates. In reality though, “frictions in the credit market enable banks to price in a discretionary manner” (Cerquiero et al., 2011, p. 503). The greater these frictions, the less standardized is the lending technology used by a bank (loan officer). To go a step further, we seek to distinguish the use of soft information by the bank from other frictions in the credit market. Thereby we can build a measure of the use of soft information by banks, according to the level of standardization in the lending process. Moreover, our methodology disentangles transactional-based and relationship lending technologies without assuming that they are mutually exclusive. Rajan et al. (2015) propose a similar methodology in a different context, in that they study the behavioral changes exhibited by lenders in response to the boom in securitized subprime mortgages. With securitization, soft information becomes less valuable than hard information, so a lender’s incentive to produce the former information is weak, and interest rates become worse predictors of default. However, Rajan et al. (2015) do not explicitly measure relationship lending as we do; instead, they regress loan interest rates (mortgages) on some indicators of hard information to deduce (using R-squares) the level of hard information that the lender uses to price the loan.

With our measure built, we next can turn to the issue of the impact of banking competition on lending technologies and, more precisely, on the use of soft information by the bank when it prices a loan. We show that banks prefer to implement relationship lending technologies when competition is weak. This result has important consequences in terms of regulation. Indeed, it is now well-known that the main benefits of bank–firm relationships result from improved credit availability (Petersen and Rajan, 1994; Berger and Udell, 1995; Berger et al. 2005). Consequently, it could be inappropriate fostering banking competition in countries where firms are heavily dependent on bank credit.

With regard to the shape of the relationship between competition and lending technology, we also find, in accordance with extant theoretical conclusions, that the relationship is nonlinear and concave.

In Section 2, we describe the method we implemented to build our proxy for relationship banking and the econometric model that we use to measure the impact of banking competition on relationship lending. After we describe the database and variables in Section 3, we present the model results and the impact of competition in Section 4. Section 5 contains the robustness tests, and Section 6 reports on a panel analysis based on two databases (NSSBF 98 and SSBF 03). Finally, Section 7 concludes.

Models

Measure of lending technologies

To build our measure of relationship lending technology, we start with the methodology proposed by Cerquiero et al. (2011), which we introduced previously, and then go a step further. In particular, we split the non-standardized technology into two parts. The first is a pure discretionary technology, such as when the loan officer’s judgment might be affected by levels of bargaining power (for the bank or firm), experience, the gender of the applicant and so on. The second part is relationship lending, which accounts for the effects of this kind of financing on the rate charged by the bank. Our goal is to capture the amount of soft information the bank uses to price some loan, so we distinguish seeking better information from the other two central relationship lending objectives that we described in the introduction. Accordingly, our starting point is the following loan pricing equation:

In this equation, the interest rate (variable “Spread”) charged by a bank depends on the level of hard and soft information used by bank to value the quality of the firm’s project (variables “Hard” and “Soft”). But some discretion can add noise to the loan-pricing process (variable “Disc”). The establishment of relationship lending for reasons other than collecting soft information also affects the interest spread (variable “OtherRel”). Finally, the spread depends on contract characteristics, some macroeconomic variables and firm characteristics (variable “Control”). Because soft information by definition is non-quantifiable, it is not available in databases (e.g., SSBF 2003). We only have access to hard information, discretionary behavior and contract variables. Therefore, we use the following equation:

Equation 2 is the heart of our measure of relationship lending. Suppose a firm obtains a loan from a bank that resorts to mainly hard information in its risk assessment. In this case, the previous regression presents a weak error (small ε'i). In contrast, if the bank uses a great deal of soft information, the error will be high. We apply this idea to our sample. First, we regress the spread on variables measuring hard information (vector “Hardi”), discretionary behavior, contract variables and some other control variables. Second, we sort out all individual loans with high residuals. Because the residuals capture the quantity of soft information that banks take into account, we define our first measure of relationship lending technology as follows:

  • SOFT1: Continuous variable corresponding to the square of the residuals.

To check the results we obtain with this continuous variable, we also build three binary measures:

  • SOFT2: Dummy that takes a value of 1 when the absolute value of the residual of observation “i” is greater than 1 times the standard deviation of the regression’s residuals.

  • SOFT3: Dummy that takes a value of 1.1 when the absolute value of the residual of the observation “i” is greater than 1.1 times the standard deviation of the regression’s residuals.

  • SOFT4: Dummy that takes a value of 1.2 when the absolute value of the residual of the observation “i” is greater than 1.2 times the standard deviation of the regression’s residuals.

Our methodology differs from that used by Cerquiero et al. (2011), in that we do not use a regression with multiplicative heteroskedasticity (Harvey, 1976). This methodology may seem appropriate, because it allows residual variance to vary across different observations. But even if the variance equation in the heteroskedastic regression can measure the impact of some variables on residual variance, it does not provide an explicit measure of relationship banking.

Relationship lending technology and competition banking

The second, main step addresses the impact of banking competition on lending technologies. We test the following equation:

The variable Softi corresponds to our one of the four proxies of relational lending technologies (SOFT1, SOFT2, SOFT3, SOFT4) from the previous section. The vector Compi measures banking competition. In line with previous research (e.g., Degryse and Ongena, 2007; Ogura, 2010; Petersen and Rajan, 1995), we use the Herfindahl-Hirschman index[1] (HHI) for the commercial bank deposits of the metropolitan statistical area (MSA) or county[2] where the firm’s headquarters are located. Finally, Relationi is a vector of control variables related to relationship lending.

Data

The database

The 2003 Survey of Small Business Finances (SSBF), conducted by the Board of Governors of the Federal Reserve System, provides our data. The database contains information on 4240 SMEs, defined here as firms with fewer than 500 full-time-equivalent employees. For these firms, the detailed information includes balance sheets and income statements (e.g., liabilities, assets, income), firms’ and owners’ characteristics and relationships with financial service suppliers for a broad set of products and services (Mach and Wolken, 2006). This database often supports research on relationship banking (e.g., Berger and Black, 2011; Berger and Udell, 1995; Petersen and Rajan, 1994), though it contains sparse information about banks’ characteristics. Beyond the advantage of being accessible for free, this database is interesting for two main reasons. First, it contains a great deal of information about SMEs, which suffer substantial information asymmetry and for which the benefits of relationship lending are thus the greatest. Second, the survey underlying the database is renewed regularly, so we can compare our model and results over time.

Because accounting data in SSBF 2003 are available only for 2003, we only retain firms that received credit during 2003 or 2004. In the sample of 1502 firms that negotiated credit in 2003 or 2004, we removed 76 finance, insurance and real estate firms (so-called FIRE firms), due to their specificities, as well as 185 firms that did not obtain loans from commercial banks. Of the remaining 1241 firms, only 688 provided all the needed variables (e.g., spread, credit score, maturity). Finally, we excluded 12 firms that had been in business for less than two years, because establishing a strong relationship takes time, and it is difficult for very young firms to implement such a relationship.

With the remaining 676 observations, we built a data set with five types of variables: firm characteristics, bank characteristics, loan characteristics, bank–firm relationships and market characteristics. Table A2 in the Appendix details the data set.

Variables used to measure relationship lending

Recall that we obtain a proxy of relationship lending from Equation 2:

The dependent variable is based on the spread, defined as the percentage over the index of the loan. Banking competition clearly influences the spread (Degryse and Ongena, 2005). We need to control for this influence without using banking concentration, a variable that is central to our regressions (Equations 4) for the impact of banking competition on lending technologies. To resolve this issue, we decided to subtract the spread of a firm by the mean spread of the zone of competitiveness in which the firm is located. Our dependent variable (SPREAD2) is the result. Yet we also recognize that this subtraction could affect our measure of relationship banking[3]. Indeed, if the mean spread of the zone of competition included a lot of soft information, then subtracting spreads by the mean spread would remove a significant portion of soft information. Thus, our measures of soft information constructed from residuals[4] would be biased[5]. To test this possibility, we regressed the mean spread of the zone of competitiveness on two variables (often used as proxies of soft information):

  • PERSONAL: Dummy equal to 1 if the most frequent method of conducting business with the bank offering the credit is personal.

  • DREL: Dummy that takes a value of 1 if the firm has only one bank and if the length of the relationship with this bank at the time of application is at least two years. This variable integrates two classic dimensions that characterize a strong relationship between banks and firms: duration and exclusivity.

Neither of these variables correlates with our variable of interest (Table A3).

Next, we split our vector of hard information into five variables: [6] the firm’s rating on Dun & Bradstreet Rank Credit Score (D&B); the firm’s leverage (LEVERAGE), as measure of its creditworthiness (D’Auria et al., 1999); the owner’s experience, which offers a proxy of the owner’s rating (EXP); the previous firm’s bankruptcy (BANKRUPTCY); and an interaction (dummy) that combines D&B ´ DSIZE to control for ratings according to firm size.

As we noted previously, the “Disc” vector corresponds to discretionary variables that are not formally linked to the relationship lending. It comprises four subgroups of variables. The first group contains variables that capture the reason the firm applied for credit (CAPT1, CAPT2, NOBANK, CHANGEBK); a second group of variables pertains to firm manager characteristics, including gender (FEMALE) and ethnicity (WHITE, BLACK, HISP, ASIAN); the third group captures the influence of the size and the structure of the firm (SIZE, OWNER AGE, CORPORATE, SUBS); and the fourth features bank characteristics (BHC).

Our strategy for separating the non-standardized technology into two parts (pure discretionary and relationship banking) becomes an issue for the variable that measures the physical distance between the firm and its main bank office (DISTANCE). On the one hand, the distance between a firm and its bank increases information asymmetry and thus implies more noise in loan pricing (Cerquiero et al., 2011). On the other hand, this variable appears connected to relationship lending, such that a shorter physical distance might facilitate soft information gathering by the loan officer and help establish a lending relationship (Berger et al., 2005). We decided to follow Cerquiero et al. (2011) and integrate DISTANCE into the “Disc” vector.

As explained previously, we also control for the other type of relationship lending, for which we include seven dummy variables that reflect why the firm applied for credit from this institution: PRIORRL, LGPOLICIES, PREVLOAN, PDTQ, PDTA1, PDTA2 and PDTA3.

Finally, we include several groups of control variables. Loan characteristics might explain some variability in the spread, such as loan maturity (MATURITY), its amount (AMOUNT), its type (CREDIT LINE, LEASING CAPITAL, MVE LOAN), potential partial credit rationing (RATIONING) and the amount of guarantee required (GARANTY). We also include the cost of the loan (COST) to the applicant, because sometimes banks decide to offer a low spread but compensate for it with high fees. We control for the value of the original index[7] to which the credit is tied (INDEX1, INDEX2). A second set of control variables integrates industry specifications (five dummies), the year (one dummy) and the area, as represented by two sets of variables: eighteen dummies (one by geographical area) and URBAN (equal to 1 if the firm is located in a rural county). Finally, similar to Ogura (2010), we include the default premium (DEFAULT PREM) and term premium (TERM PREM) of the market when credit is applied. For a complete description of each variable, see Table A2.

Variables in the analysis of banking competition

To study the impact of banking competition on the choice of lending technology, we test:

The dependent variable is one of the four proxies of relationship lending (SOFT1, SOFT2, SOFT3, SOFT4). As noted previously, we measure banking competition using the Herfindahl-Hirschman index (HHI) of banks’ market shares. Because the SSBF 2003 divides the concentration variable by 3, we created two dummies: HHI1, equal to 1 if HHI < 1000, and HHI2, equal to 1 if 1800 ≤ HHI. The vector Conc integrates these two dummies. The “Relation” control variables refer to relationship lending (DREL, PERSONAL) (Table A2).

Using the Herfindahl-Hirschman index (HHI) of banks’ market shares as proxy of banking competition raises two issues[8]. Firstly, it is not a direct measure of banking competition. But it is currently well-established that the concentration measure could be a good approximation for bank market power (Fischer, 2000; Elsas, 2005; Boot and Thakor, 2000). Secondly, we use the concentration in the deposit market as the measure of the concentration in the market for SME. However, Petersen and Rajan (1995) explain that the concentration in the deposit market is a correct proxy of the concentration in the SME credit market, if the firms in the sample used borrow mainly from local markets. In our database (SSBF), we observe that half of our firms have their furthest bank maximum five miles away, and seventy-five percent below fifteen miles away. In addition, half of our firms have their main bank maximum three miles away and seventy-five percent below nine miles away. Consequently, we presume that the condition of Petersen and Rajan (1995) is respected and that our measure of concentration is a good proxy for the concentration in the SME credit market.

Results

Measure of relationship lending

Table A4-1 in the Appendix presents the results of the spread equation (Equation 2). In all regressions, we winsorize all our variables at 1%, to avoid potentially spurious outliers.[9] For the hard variable results, as expected, the coefficient of the D&B (rating) variable is negative and significant, in support of our intuition that a higher rating means a lower spread. In addition, LEVERAGE is negative and significant. Firms that choose their bank depending on their lending policies have a better spread than others.

Regarding the discretion variables, captive firms must pay a higher spread than others, and both HISP and WHITE are (highly) significant, such that the spread is higher if the manager is Hispanic or White, which suggests a surprising outcome. A firm in an urban area also has a higher spread than a firm in a rural zone. Regarding loan characteristics, we recognize a potential endogeneity problem between these variables and our dependent variable (SPREAD2), so we do not interpret these results. We do not rely on either measure or interpret the possible relation of causality between these variables though, so even if the problem is relevant, endogeneity does not affect our measures of the management of soft information (Introduction to Econometrics, 3/e Stock and Watson)[10].

Following the method we described in Section 2.1, we use residuals from Equation 3 to build our four indicators of relationship lending. Specifically, we determine our continuous variable (SOFT1) and the three binary proxies (SOFT2, SOFT3, SOFT4) when the absolute value of the residual of a given observation is greater than (respectively) 1, 1.1 or 1.2 times the standard deviation of the regression’s residuals. Table A4-2 displays the results of each sort. For example, in the case of SOFT2, there are 151 observations (22% of our sample) for which the bank mainly used soft information to price the loan.

Concentration and relationship lending

Table A5 in the Appendix reports the results of our analysis (Equation 4), related to the impact of concentration on relationship lending. The first column corresponds to our continuous proxy (SOFT1), and the three others reflect our binary proxies of relationship lending (SOFT2, SOFT3, SOFT4). From the continuous proxy SOFT1, we determine that HHI1 is negative and significant, such that low concentration in the banking sector diminishes the probability that banks use soft information. Therefore, our results validate Petersen and Rajan’s (1995) conclusions, rather than those proposed by Boot and Thakor (2000), regarding the impact of banking competition on the lending technologies that banks implement. For the dummy variables, we find that HHI1 is significant and negative for SOFT3 and SOFT4 but not for SOFT2. Perhaps banking concentration affects the important use of soft information by the bank (SOFT 3 and 4) rather than its mean use (SOFT2).

If HHI1 is almost always negative and significant, HHI2 is never significant. This result indicates that the link between the use of soft information and banking market concentration is nonlinear. To confirm this result, we test Equation 4 after replacing our Conci vector variables with the HHI and its square (see Table A6). Coefficients of HHI and HHI² are, respectively, positively and negatively significant. Such results validate the anticipated nonlinear link and support theoretical findings by Dinç (2000), Anand and Galetovic (2006) and Yafeh and Yosha (2001): when banking concentration is weak this latter positively influences the use of relationship lending but this link is reversed when the competition is high.

By extrapolating results obtained for HHI and HHI2 (Table A6, model 3) in the event that the bank concentration index fluctuates continuously between 1 and 3, we are able to determine the maximum reached by the variable Soft and find the probability that the bank will use relationship lending technology to price the loan increases to a maximum of almost 20% (19%) for a theoretical concentration of 2.68[11]. We interpret this result in the following way. In geographical areas characterized by a weak or medium banking concentration (HHI index equals 1 or 2), an increase of the banking concentration facilitates the establishment of relationship lending technology; this effect is reversed when concentration is strong (HHI index equals 3). Table A7 displays the distribution of our observations according to their location in one of the nine geographical divisions[12] and the average banking concentration in the division where the firm is based. We observe that 44% of our observations belong to a division[13] whose banking sector is very concentrated.

Finally, regarding the variables that characterize bank–firm relationships, we note that the personal contact between the firm and the bank does not seem to affect the use of soft information. The variable measuring the strength and duration of this relationship (DREL) reveals a positive and significant result though, in support of the accuracy of our approach.

Panel analysis

Panel sample

In this section, we construct a new sample by merging the 1998 NSSBF (224 observations of credit provided from 1996 to 1998) and the 2003 SSBF (676 observations of credit provided from 2003 to 2004). We thus obtain a new sample of 900 observations of credit provisions from 1996 to 2004 but use the same variables, with the exception of the dummy BHC, which is absent from the 1998 NSSBF. Table A2 contains the statistical description of the 1998 NSSBF.

Methodology

We follow the same methodology and first run our spread equation,[14].

Then we use the residuals to construct our continuous variable (SOFT1) and the three dummy variables (SOFT2, SOFT3, SOFT4) for use in the second equation:

Results

Table A8-1 in the Appendix displays the results for our first equation. As before, we do not interpret the loan characteristic variables, due to the endogeneity with our dependent variable. Table A8-2 details the number of soft variables further, according to this analysis.

Using these variables in our second equation, we obtain results for the linear impact of competition on the use of soft information (Table A9 in the Appendix). As these results show, HHI1 always exerts a negative, highly significant impact. Moreover, HHI2 is positive and significant, in support of our previous results. A high level of concentration leads to a preference for the use of soft information. Moreover, the results seem to indicate that concentration exerts a linear impact on the use of soft information, as supported by the evidence in Table A10. With a dummy variable that equals 1 if the credit was granted in 2003 rather than 1998 (Y2003), we estimate the same regression, splitting the impact of HHI1 and HHI2 according to this dummy. Whether in 1998 or 2003, the impact of concentration on the use of soft information by the bank remains the same: Low (high) concentration leads to a decrease (increase) in the use of soft information (Table A11). However, the nonlinearity is only significant in 2003 (Table A12). Finally and interestingly, we note that the variable DREL is always positive and significant, which confirms the appropriateness of our measure.

Conclusion

Questions about the impact of banking competition on the choice between transaction-based and relationship-based technology have persisted for decades (both empirically and theoretically), without any clear resolution. With this study, we seek to address these questions by using a new measure of relationship lending technology. Starting from the methodology developed by Cerquiero et al. (2011), we build an accurate measure of lending relationship technology that reflects the actual level of soft information a bank uses at the time of loan pricing, as precisely as possible. This new approach permits us to conclude that banks prefer to implement relationship lending technology when competition is weak, in support of Petersen and Rajan’s (1995) findings. We also can specify the shape of the relationship between competition and relationship banking; in accordance with theoretical predictions from Dinç (2000), Anand and Galetovic (2006) and Yafeh and Yosha (2001), we find that the relationship between competition and relationship banking is nonlinear and concave: when the banking concentration is weak or medium, an increase in it facilitates the establishment of relationship lending technology, but this effect is reversed when the concentration is strong.

A first managerial implication of our results concerns firms issuing a lot of “soft information” as SMEs. As these firms are heavily dependent on relationship lending, they should favour banks located in areas with weak competition in order to maximize their access to credit.

The second implication concerns the organizational structure of banks. In following the theoretical work of Stein (2002), Berger et al. (2005) show that relationship and transactional banks do not have the same organizational form: the former is bigger and more centralized than the latter. Agarwal and Hauswald (2010b) go further by observing that “delegating real authority” to loan officers “provides strong incentives for collection, transmission, and strategic use of soft information”[15]. Hence, banks located in areas where the banking competition is weak should promote the development of relationship lending by a strong delegating authority and less turnover for loan officers, so that they can build strong relationships with firms.

Our conclusion also highlights a “dark side” of banking competition. Indeed, fierce competition could drive banks to focus mainly on hard information penalizing small and medium sized firms. This “dark side” appears all the more important as we observe an increase of banking competition all over the world (Mirzaei and Moore; 2014[16]). In that perspective, developing countries, whose economies are mostly characterized by small firms, should think twice before fostering banking competition.

Finally, at the European level, the non-linearity between bank competition and relationship lending should induce the regulator to take into account the heterogeneity of banking competition in each country. Table A13 displays banking concentration in the European Union in 2016 measured by the Herfindahl index of banks’ market shares. We can observe a significant disparity in the banking structure: some countries have a highly competitive banking industry (Germany, Luxembourg, Austria), while others are characterised by a low level of competition in this sector (The Netherlands, Greece, Estonia). Consequently, a European policy seeking to improve business credit conditions (availability and cost) using competition in the banking sector as a lever could have opposite effects depending on the level of banking competition in each country. For instance, an increasing of the increase in banking competition would ease the use of relationship lending technology in The Netherlands and Greece but at the same time would be a brake on this kind of financing in Germany and Luxembourg.