DISCUSSIONPAPER SERIESIN ECONOMICSAND MANAGEMENT

Strategic Groups in European Commercial Banking

Andreas Hackethal

Discussion Paper No. 01-19

GERMAN ECONOMIC ASSOCIATION OF BUSINESS

ADMINISTRATION - GEABA

Andreas Hackethal*∗

Strategic Groups in European Commercial Banking

September 2001

Johann Wolfgang Goethe University Frankfurt/Main

* Please address correspondence to Andreas Hackethal, Wilhelm Merton Chair for International Banking and

Finance, Mertonstr. 17, P.O.-Box 111932, D-60054 Frankfurt/M., Germany. e-mail: Hackethal@em.uni-.frankfurt.de. The author thanks Ralf Elsas, Martin Höpfner and Reinhard H. Schmidt for valuable suggestions concerning both the substance of this paper and its exposition.

Strategic Groups in European Commercial Banking

Abstract

Despite contrary expectations of both researchers and practitioners, the European monetary

union has so far not led to a discernable convergence of the roles of commercial banks in

different European financial systems. This empirical study aims to shed more light on these

structural differences by analysing the strategic positions of 624 commercial banks from 12

European countries. Based solely on their market-based and resource-based profiles in 1999

and not on their respective country of origin, banks are classified into nine distinct strategic

groups. Applying discriminant analysis and logit models indeed reveals remarkable

differences concerning the strategies pursued by banks from different countries. Strongly

divergent market positions in terms of business risk exposure, asset growth patterns and the

relative importance of loan and fee businesses in conjunction with only minor differences in

terms of efficiency ratios and the skill levels of employees imply that banks face vastly

different national market environments. Our empirical results may explain why national

banking markets are still dominated by domestic players and indicate that performance

differentials are not so much driven by technical and scale efficiencies but much more by

(achievable) market positions.

JEL-Classification: G21, L1, P51

Keywords: financial systems, commercial banks, strategic groups, resource-based view

1

1 Introduction Recent research on the structural development of European financial systems

has shown that, despite wide reaching regulatory harmonisation, the roles of banks have evidently not converged during the last twenty years.1 Most continental European financial systems, that have been traditionally deemed bank-based, are still characterised by a dominant role of banks in the process of transferring funds from surplus units (eg households) to deficit units (eg enterprises and the government) and in exercising corporate control. In market-based financial systems like that of the UK, organised financial markets have been playing a much larger role in both respects. Figure 1 depicts for the three largest European economies the share of household claims on banks as a portion of their entire intersectoral financial assets (panel a) and the share of total non-financial firms’ intersectoral financial liabilities that are owed to banks (panel b). The level of bank intermediation differs strongly between the three countries and the differences are fairly stable over time. Bank disintermediation seems to be a common phenomenon concerning the liability sides of all banks. Only in the case of France, however, has the role of banks as financiers of enterprises strongly declined.

Figure 1: Intermediation ratios vis-à-vis banks

a) Claims of households b) Liabilities of enterprises

1981 1983 1985 1987 1989 1991 1993 1995 1997

0%

10%

20%

30%

40%

50%

60%

70%

80%

1981 1983 1985 1987 1989 1991 1993 1995 1997 Source. Schmidt/Hackethal/Tyrell (2001)

The explanation for the lack of structural convergence that is favoured by a growing number of observers rests on the perception of a financial system as a set of complementary elements. According to this view, it is an essential precondition for the functioning of a given financial system that such diverse

1 See for example Allen/Gale (1999), Schmidt/Hackethal/Tyrell (1999) and Schmidt/Hackethal/Tyrell (2001).

Germany

UK

France

UK

France

Germany

2

elements as the financing patterns of enterprises, the savings patterns of households, the prevailing mode of corporate governance, the regulation of capital markets and banks, the legal framework as evidenced for example by the insolvency code “fit together” and reinforce each other in a positive way. As a consequence of the large number of implied interdependencies, consistent configurations are characterised by a high degree of stability. Isolated changes to single elements lead to inconsistencies that in turn may urge the participants to reverse those changes in order to reinstall the old system. Systemic change can then only be achieved by altering many elements simultaneously. This might have happened in France during the late eighties and early nineties, transforming the French financial system from a bank based system with a dominant role of the state to a more market-based, Anglo-Saxon-style system.

Based on this conception of financial systems and given the empirical results on divergent bank intermediation ratios one should expect that commercial banks in different European countries have been pursuing dissimilar strategies. This paper addresses the question as to what extent and how those strategies differ. It thereby aims to extend the discussion on systemic complementarities towards the level of single financial sector institutions and also to shed some more light on the manifestations of different strategic positions of European commercial banks.

For that purpose we utilise the theoretical construct of strategic groups that was introduced by industrial organisation economists (Hunt, 1972). Generally speaking, firms form a strategic group within a competitive setting if they follow similar strategies with respect to product market combinations and/or to the resource bundles they employ (Porter 1979). Like the initial work by Hunt and Porter most of the studies in the strategic group literature have since been motivated by attempts to explain the diversity of demand and cost curves of firms within the same industry first discussed by Chamberlain (1932). Advocates argue that there exist persistent structural features not only on a firm-level but also on a group level that give rise to structural or strategic, asymmetric mobility barriers (Caves/Porter 1977) protecting a given group from the entry of potential rivals and thereby permitting persistent performance differences between groups and hence also between firms. Porter (1979, p 220) concludes that “the concept of strategic groups allows us to systematically integrate the differences in the skills and resources of an industry’s member firms and their consequent strategic choices into a theory of profit determination”.

3

Other authors like Hatten/Hatten (1987) and Barney/Hoskisson (1990) have raised concerns about the theoretical validity of the strategic group concept doubting that strategic groups do really exist and supposing that they are merely an analytical convenience lacking a theoretical underpinning. Much empirical progress has been made since the early studies surveyed in Thomas/ Venkatraman (1988) that do indeed only suggest a weak linkage between group membership and performance. Nath/Gruca (1997) by reconciling strategic groups formed through cognitive methods and through cluster analysing archival variables, Ketchen et al (1997) by means of meta analysis, Pitt/Thomas (1994) by again reviewing the literature and Dranove/Peteraf/Shanley (1998) by highlighting the role of true group effects provide strong support for the existence of mobility barriers. They conclude that firms may indeed find themselves locked into a low performance strategic group, unable to emulate the idiosyncratic strategies of firms in high-performance strategic groups.

Aiming to identify strategic groups and mobility barriers, studies have focused on a large variety of - albeit mostly US - industries, ranging from health care (Nath/Gruca 1997), information technology (Duysters/Hagedoorn 1995) and pharmaceuticals (Cool/Dierichx 1993) to the insurance industry (Fiegenbaum/Thomas 1993) and the banking sector (Mehra 1996, Amel/Rhoades 1988). Many have utilised panel data in order to corroborate the hypothesis regarding the existence of stable strategic time periods, in which the group structures are expected to be relatively constant and as a consequence, for which an analysis of strategic differences would be especially meaningful (Fiegenbaum/Thomas 1990, Fiegenbaum/Sudharshan/Thomas 1987). Others have analysed longitudinal data and were able to confirm Porter’s (1980) conjecture that firms in the same strategic group react similarly to environmental factors such as changes in regulation (Gruca/Nath 2000, Marlin/Lamont/ Hoffman 1992). Again others have identified reputation as an important mobility barrier (Ferguson et al. 2000) or have found evidence for strategic groups acting as reference points for their members in formulating competitive strategy (Fiegenbaum/Thomas 1995).

Our study is closest both in construct and in industry coverage to Mehra (1996). Based on interviews in which industry experts were asked to rate 45 US bank holding companies along ten key resources such as management quality, franchise value, placing power and risk management, the author derives five resource-based strategic groups. In addition, he utilises 1990 archival data for the same 45 banks on eleven market-based variables such as the ratio of time deposits to total deposits, the ratio of foreign deposits to total deposits, the

4

percentage of non-interest revenues over total revenues and the 5-year annual average asset growth rate to form four market-based strategic groups. The main thrust of the study is then to examine whether the resource-based clusters or the market-based clusters bear a higher explanatory power for individual performance differences as measured by the banks’ return on average assets, their net profit per employee and their adjusted price/earnings ratio in 1990. Based on the R2 measures from a simple ANOVA Mehra posits a “strong overall association between firm resource endowments and superior performance” (p 318) and as a consequence urges bank managers “to shift their strategic focus from privileged product-market positions as basis for competitive advantage to the underlying resource base supporting these positions” (p. 319).

Our study borrows heavily from Mehra’s own insights and from those quoted in his article in that we also conceive of resources and market strategies as two sides of the same coin (Wernerfeldt 1984) where resource groupings correspond to competition in factor and input markets while market groupings correspond to competition in output markets and where resources only have value if they are deployed in particular markets. We refrain, however, from merely transferring his research agenda and design into a wider and more recent European context. Instead, we intend to explore in more detail the linkages between the two complementary views of strategic firm behaviour in order to inductively derive insights into the structural differences among European commercial banks. We thereby subscribe to Porter’s (1996) answer to his self-addressed question “What is strategy?”, namely that superior strategies are creating fit among a firm’s activities. According to this view, which has been set on a more rigorous footing by Milgrom/Roberts (1995), competitive advantage and hence superior performance is not so much the result of a firm’s access to single resources or its specific choice of product-market combinations but rather it is the result of all of its activities being in perfect sync. Analogies between this firm-level notion of complementarity and the notion of complementarity as mentioned above in the context of financial system analysis are of course not coincidental but clearly deliberate. This paper can hence be perceived as an explorative attempt to extend in a top-down manner a promising recent strand of the comparative financial system literature towards a richer analysis of the European banking industry.

Our empirical evidence does indeed imply a close correspondence between complementary relationships on a system-wide level and those on a group-level. Clustering the entire sample, consisting of 624 European commercial banks, along both, six market-based variables and six resource-based variables yields

5

groupings that delineate neatly between banks from different countries. Moreover, five additional propositions are derived that allow for a better understanding of the workability of specific combinations of strategic levers chosen by European banks.

According to our knowledge, this is the first study attempting to bridge the gap between the comparative financial systems literature and the empirical literature on banking strategies by means of clustering analysis and subsequent application of discriminant analysis and logit modelling. A word of caution is therefore certainly warranted. It was our objective to derive initial but robust assessments from a simple and, as some serious econometricians might argue, not highly sophisticated approach. As this study marks a starting point there is clearly room for methodological as well as conceptual improvements and extensions. A more stringent selection of independent variables, thereby accounting more rigorously for endogeneity problems and the use of panel data to test for the robustness of the results over time are only two issues to proceed from should this line of research already prove fruitful at this preliminary stage.

The paper is organised as follows: The next section describes the fourteen variables along which banks and strategic groups are compared. Section 3 gives an introduction to the methodologies applied in this study. Readers familiar with clustering algorithms, discriminant analysis and logit models may without hesitation skip this section.2 Section 4 presents the empirical results, which are then discussed in section 5. The last section provides some tentative implications for bank management and concludes the paper.

2 The dataset The bank data was taken from Fitch-IBCA’s Bankscope dataset. This source

offers the great advantage in our context that financial data have been made as much comparable across countries as possible by applying so called Fitch country models to translate country specific reporting formats into a unified global format. We have thus reason to assume that it is not discrepancies in reporting requirements that drive our empirical results.

In order to include only those institutions that could have clearly been considered as commercial banks in 1999, a ratio of loans to assets of at least 60%, a ratio of total deposits to assets of at least 50% and a ratio of customer deposits to assets of at least 20% was required. After eliminating those observations with missing values for the twelve variables in Table 1, 624 banks

2 In a revised version of this paper, this section will be part of an appendix and discriminant analysis might be

skipped altogether.

6

from twelve European countries remained in the sample. Based on aggregated data from Eurostat, these 624 banks comprise of roughly a third of total loans to customers in the twelve countries. Sample shares range from below 10% for Denmark and Belgium to over 65% for Netherlands, Italy, Portugal, Finland and Sweden. 35% of German, 30% of Austrian, and approximately 20% of French, Spanish and British bank loans outstanding in 1999 are represented by the sample.

Table 1: Description of variables

Name Description Mean Min. Max. Stand. Dev.

Market-based-view variables

AVGRWTHA Average asset growth (1992-1999) 9,1% -9,2% 163,3% 11,4%

EQU_A Book value of equity over total assets 5,6% 2,2% 19,9% 2,5%

INTMRG Net interest revenue over earning assets 2,9% 0,4% 9,5% 0,9%

LOA_A Customer loans over total assets 68,7% 60,0% 98,1% 6,3%

OINC_INT Other operating income over net interest revenue 35,3% -9,2% 224,5% 23,3%

STDV_ROA Standard deviation of ROAA (1992-1999) 0,2% 0,0% 1,9% 0,2%

Resource-based-view variables

CST_INC Overhead cost over total operating income 68,0% 32,0% 151,7% 11,3%

DEP_A Customer deposits over assets 62,1% 20,6% 93,3% 13,3%

FIX_A Fix assets over total assets 1,7% 0,0% 9,5% 0,9%

INTBK Money lent to other banks over money borrowed from other banks

61,5% 0,0% 516,8% 65,7%

OFUND_A Other funding over total assets 7,5% 0,0% 34,3% 6,4%

PERSX_P Personnel expense per employee 45,9 19,0 119,2 10,7

Performance-based variables

ROAA Return on average assets 0,4% -1,6% 2,5% 0,41%

ROAE Return on average equity 7,3% -20,5% 31,3% 5,36%

The variables belong to one of three categories. The market-based category aims to capture the specific product–market combination chosen by a given bank, or in other words, the way a given bank deploys its resources in order to compete successfully in the output markets. Given the balance sheet- and profit/loss statement data provided by Fitch-IBCA we have selected those ratios that should be expected to serve as good proxies for generic strategic dimensions in banking. Growth and thereby past investment activity is reflected in the variable AVGRWTHA, which measures the average growth rate of total assets during the last seven years, or respectively, during a shorter period should data not be available back to 1992. There are two positively correlated (0,396) measures for the risk exposure of a bank’s businesses. STDV_ROA indicates the

7

standard deviation of returns on assets generated by the bank in its recent past. It is hence backward looking by definition, whereas EQUA_A (book value of equity over total assets) tends to be more forward looking, in that its value should be a positive function of future unexpected losses due to market-, credit-, and operating risk. Risk should also affect the ratio INTMRG, which measures net interest revenue from all operations over total assets (eg market risk due to the extent of maturity transformation). One should expect, however, that this ratio is even stronger driven by the types of borrower groups to which the bank is catering specific loan products and the types of depositor groups from which the bank is collecting funds in a specific form. Both, LOA_A (loans to non-banks as a portion of total assets) and OINC_INT (other operating income like fees, commissions and trading profits over interest income) indicate the diversity or scope of a bank’s businesses. A lower value for the former ratio implies that a given bank can be assumed to be more active in money markets, the securities business and/or the management of participations in other firms. The latter ratio should be increasing in the bank’s involvement in payment services, trading and investment activities and advisory services.

The resource-based category seeks to capture a bank’s competitive position on input or factor markets. DEP_A measures the strength of a given bank in terms of collecting deposits from non-banks by dividing the nominal amount of customer deposits by total assets. Higher values give us reason to believe that the respective bank possesses a strong core deposit base allowing it to refinance its asset business at favourable terms. “Other funding” as used in the nominator of the ratio OFUND_A comprises of long-term bank liabilities evidenced by certificates such as bank debentures, and various forms of subordinated debt. As to be expected, it is negatively correlated to DEP_A (-0,419); but not perfectly, as money market funding, reserves and provisions and equity are not included. OFUND_A indicates the reliance of a given bank on purchasing funds in the open market. As such this ratio is certainly subject to market access constraints the bank might face due to its size or its legal form. INTBK is computed as the ratio of interbank lending over interbank borrowing and as such circumscribes a given bank’s role in the money markets. Values greater than unity may indicate that a bank has funds at its disposal that exceed its short-term investment opportunities and vice versa. As this ratio is not too highly correlated to DEP_A (0,314) and hardly at all to OFUND_A it captures effects not controlled for by the other two funding-related variables. PERSX_P is defined as total personnel expense over the number of employees and is considered by us as a proxy for the average skill level of employees. By means of FIX_A (fixed assets over total

8

assets) we include a measure in our analysis that indicates a bank’s reliance on tangible assets such as office buildings and branch networks. Finally the cost-income ratio (CST_INC), which is defined as total operating cost over total operating income, provides us with a crude measure of how efficiently a bank has employed its resources.

Whereas we consider all of the above variables to be largely at management’s discretion (at least in the long-term) and therefore exogenous to our analysis we view both selected performance-based variables to be endogenous. Net income as a fraction of average assets (ROAA), and average equity (ROAE), respectively, are affected by the interplay of all market-based and resource-based variables. Hence, in clustering the banks into strategic groups we will only use the twelve strategy variables but not the performance measures.

3 Methodology

3.1 Clustering

Based on the MBV- and RBV-sets of strategy variables we assign each bank to one market-based group and to one resource-based group, respectively. Following previous research (Ketchen/Shook 1996), we apply the Ward algorithm for that purpose. This algorithm minimizes the within-group variance of observations and thus maximizes the homogeneity of groups. As group-homogeneity is a central feature of strategic groups the Ward algorithm is better suited for our purposes than alternative algorithms.3 Most clustering algorithms share in common that they combine observations and groups of observations, respectively, on the basis of their distances. As our dataset contains only metric data we can, after standardising all values on the range [-1;1], use the quadratic Minkowski metric in (1) as a measure for the distance between the variables j of two individuals, that is banks, m and n.

ä=

−=J

jnjmjnm xxd

1

2

, (1)

The Ward-algorithm belongs to the family of agglomerative methods and as such starts by treating every single bank as a group and than aggregates observations until all banks have been assigned to one of C clusters.

3 The single-linkage or “nearest neighbour” algorithm for example, is particularly useful in identifying and

isolating outliers. It hence tends to yield some very small and only a few large groups, which are, however, often quite heterogeneous.

9

In each step a new group containing at least two banks is formed in a way so that the increase in the sum of squared errors (SSE) over all groups is minimized. The SSE for the Ng banks in group g is defined by (2).

( )2

1 1ää

= =−=

gN

n

J

jjgnjgg xxSSE with ä

==

gN

nnjg

gjg x

Nx

1

1 (2)

The equation illustrates that at the start of the Ward-algorithm, when all groups have the size of one, the sum of all SSE is equal to zero and that those two banks are combined to a group that share the smallest distance in the dataset. After the first step, the total number of groups has been reduced by one and the total sum of SSE is equal to the SSE of the single group containing two banks. If, as in this study, quadratic Minkowski metrics are used, this total sum amounts to exactly twice the distance between the two combined banks. This relationship generalises to all subsequent steps: The increase in the total sum of SSE for each step is equal to twice the distance between the groups of banks that have been combined.

Several heuristic rules exist that help to determine the appropriate number of clusters at which to stop the Ward algorithm (Fiegenbaum/Thomas, 1990). The A greater number of clusters leads to a higher portion of explained variance over total variance but at the same time complicates the interpretation of results as the exposition becomes increasingly complex. In this paper we aim to strike a balance between both effects by opting for the smallest possible number of clusters, where jack-knifing procedures misclassify less than 10% of the sample banks.

3.2 Discriminant analysis

It is a well-known fact that discriminant analysis, which was introduced by Fisher (1936), rests upon fairly restrictive assumptions:4 To apply tests of significance on the estimators the assumption of normality of the explanatory variables is required.5 However, the method in deriving the estimators, ie the coefficients of the discriminant functions, is distribution-free. We have therefore pursued a two-pronged approach in this paper. We first apply discriminant analysis and then verify the robustness of the results by means of a logit model, which avoids the critical normality assumptions. Should both methods yield identical results with respect to the discriminating power of the variables we are

4 See for example Eisenbeis (1977). 5 Jarque-Bera tests reject the hypothesis of normal distribution for all twelve variables.

10

able to safely reapply discriminant analysis for the purpose of depicting the clustering results graphically. This would, in turn, allow for a quite intuitive way to assign out-of-sample banks to one of the clusters. Multinomial linear discriminant analysis seeks to find up to Min(J-1; C) orthogonal linear functions that provide the best discrimination between C groups with regards to J variables j. The coefficients β of the linear function in (3) are estimated to maximise the discriminant criterion Γ defined in (4).6

ä =+= J

j jj xy10 ββ (3)

( )

( )ää

ä

= =

=

−

−=Γ

C

c

N

nccn

C

ccc

c

yy

yyN

1 1

2

1

2

(4)

Γ is defined as the ratio of the variance of y’s between groups over the variance of y’s within groups. The nominator hence captures the explained variance of the y’s, whereas the denominator indicates their unexplained variance. The sum of the two is equal to their total sample variance.

The maximal value of the discriminant criterion Γ i for discriminant function i, γi , is called its Eigenvalue. The term (1+γi)

-1 is referred to as univariate Wilk’s Lambda and corresponds to the ratio of unexplained variance over total variance. Multiplying all univariate Wilk’s Lambdas yields the multivariate Wilk’s Lambda, which indicates the variance left unexplained after accounting for all discriminant functions. The lower its value, the more warranted it is hence to use the notion of strategic groups for the obtained clusters.

Because each discriminant function yi+1 seeks to explain the maximum portion of the variance left unexplained by discriminant function yi , the Eigenvalues decrease with increasing i. A measure for the relative importance of a single discriminant function yi is given by its Eigenvalue portion Ρi.

ä=Ρ

jj

ii γ

γ (5)

Eigenvalue portions are helpful in determining the sufficient number of discriminant functions that should be used in assigning out-of-sample banks to one of the clusters.

The binomial linear discriminant model corresponds to a linear ordinary least squares (OLS) regression model yi=ß´xi+ui with E(ui)=0, where yi is either zero

6 The index i has been suppressed in equations (3) and (4) for better readability.

11

or one. The estimated value iy , however, is equal to the vector product β ´xi and

corresponds to the estimated probability ip that the observation belongs to

group 1.7 Because the residuals ui can only take the value 1-β’xi or -β´xi, they are clearly not normally distributed. As a consequence of this heteroscedasticity problem, OLS estimators for β are not efficient. Moreover, the t-statistics for testing βi = 0 only have t-distribution only if the x’s are joint normal (Maddala 1983, p 21). These results carry over to discriminant analysis as follows: Although the discriminant-estimators are asymptotically more efficient than the logit maximum-likelihood estimators (MLE) under the condition of normality of independent variables, the discriminant estimator is not even consistent otherwise (Maddala 1983, p 27). The logit MLE is consistent in both circumstances and therefore more robust. As mentioned in the beginning of this section we will therefore also use a multinomial logit model with an unordered dependent variable to analyse the discriminatory power of the independent variables.

3.3 Logit models

Logit model analysis takes the heteroscedasticity problem of the linear probability model, which results from the restricted scale of the dependent model variable pi, as a starting point. It gets rid of the upper bound of one by defining pi’ as pi / (1-pi) and eliminates the lower bound of zero by defining pi’’= ln pi’. pi’’ is now defined on the range from –∞ to +∞ and is referred to as logit (Li). The basic, binomial logit model corresponds to (6):

( ) ii xyL β ′==1 , (6)

where β and xi are vectors with the number of components equal to the number of observations for bank i. Upon estimation of the model, the estimated probabilities ip can be derived by reversing the two transformations.

))ˆexp(1(

)ˆexp(ˆ

i

ii

x

xp

ββ

′+′

= (7)

In contrast to the coefficients in a linear probability model, the coefficients β in a logit model have thus no constant impact on pi. Changes in xi imply greater changes in pi when pi is close to zero or 100%, respectively, than in cases when pi is close to 50%.

7 The R2 of the linear probability model corresponds to the expression γ/(1+γ) in discriminant analysis.

12

The logit model builds upon maximum likelihood estimators (MLE) to estimate equation (7). The likelihood function to be maximised in the binomial model is given by (8). The MLE are hence those parameter values that imply the highest probability for the occurrence of the empirical data.

[ ]∏ä∏

=

=

−

′+

′=öö

÷

õææç

å′+

′öö÷

õææç

å′+

= n

i i

n

i ii

yn

i i

i

y

i x

yx

x

x

xL

ii

1

1

1

)exp(1

)exp(

)exp(1

)exp(

)exp(1

1

β

ββ

ββ

(8)

In the multinomial logit model there are C probabilities pc=p1, p2, …pC associated with C groups or clusters. These probabilities are expressed in binary form.

( )icC

ic xp

p β ′= exp (9)

Equations (10) and (11) follow directly from (9).8

ä −

=′+

′= 1

1)ˆexp(1

)ˆexp(C

c ic

icic

x

xp

ββ (10)

ä −

=′+

= 1

1)ˆexp(1

1C

c ic

iCx

pβ

(11)

The likelihood function to be maximised can now be written as

∏∏= =

=n

i

C

c

yic

icpL1 1

)( , or equivalently, ää= =

=n

iic

C

cic pyL

1 1

lnln (12)

where yic is equal to one if the i-th bank belongs to cluster c and yic is zero otherwise.

Most statistical software packages use the iterative Newton-Raphson procedure with the discriminant function coefficients as starting values to obtain the MLE. The estimators are consistent, asymptotically efficient and asympto-tically normal, so that significance tests can be carried out. Because asymptotic standard errors of the MLE are differentiable, their confidence intervals can also be computed.

The accuracy of the entire logit model can be derived from the indirect pseudo R2 in (13) proposed by McFadden (1974):

0

2

ln

ln1

L

LRpseudo Ω−=− , (13)

8 Because the pic’s add up to one, only C-1 binary logit equations have to be estimated.

13

where LΩ denotes the maximum of the likelihood function in (12) when maximised with respect to all parameters βc and where L0 denotes its maximum when maximised solely with respect to the constant coefficient β0c. McFadden’s pseudo-R2 hence measures the improvement in estimation quality of the complete model over the reduced-form model L(y) = β0c. A value of zero implies that variations in the characteristics xi of a given bank have no impact whatsoever on its classification.

Significance tests can be carried out on the likelihood-ratio G in (14) which is asymptotically χ2-distributed with degrees of freedom equal to the difference in the number of parameters in model LΩ and L0.

00 ln2ln2ln2 LL

L

LG −=−= Ω

Ω

(14)

The null-hypothesis to be tested is then H0: (β1g = β2g =…β2m) = 2 ln LΩ - 2 ln L0 = 0.

Another, quite intuitive, way to assess the accuracy of the model is to compare the observed classifications to the forecasted classifications based on the estimated probabilities. A corresponding measure is the percentage of banks in the sample for which these two match.

The statistical significance of each of the estimated effect coefficients across all C-1 logit equations can be tested in close correspondence to the likelihood ratio test on the entire logit model. Let LΩ-q be the maximum likelihood value of a model in which the single coefficient βqj is set to zero. Then the statistic Gq in (15) has χ2-distribution with C-1 degrees of freedom.

qq LLG −ΩΩ −= ln2ln2 (15)

In order to assess the power of a single variable in discriminating between a specific pair of the C groups the significance of the respective coefficient βjc from the respective logit equation has to be tested. Because the components of βc are asymptotically normal, the modified t-statistic in (16) can be used for that purpose.

2

2ˆ

jc

jc

set

β= (16)

sejc denotes the estimated standard error of the respective logit estimator. t is then asymptotically χ2-distributed with one degree of freedom. The null-hypothesis to be tested is H0: βjc = 0.

14

4 Empirical results

4.1 Strategic groups

MBV and RBV groupings were obtained from separately running the Ward-algorithm on the six market-based variables and on the six resource based variables. We then carried out both discriminant-analysis and logit-model jack-knifing procedures to determine the appropriate number of clusters for the study. Based on our rule of thumb, which requires that less than 10% of the banks be misclassified, we opted for three MBV clusters and three RBV clusters.

Tables A1 and A2 in the appendix and Table 2 show descriptive statistics on the nine strategic groups that result from entering the MBV and RBV clustering results into a 3x3 strategy space. The ratio of inner-group variance over total sample variance is far below one for most variables and most clusters implying a high degree of strategic homogeneity of banks belonging to the same cluster.

According to Table A1 the average bank in MBV group 3 is characterised by higher growth rates, higher interest margins and a larger reliance on fee-based businesses than its peers in MBV group2 and even more so than its peers in MBV group 1. Most strikingly, however, both variables that capture a bank’s risk exposure (ie EQA_A and STDV_ROA), take on considerably higher values for MBV group 3. Given these characteristics and presuming the existence of a general risk-return trade-off in the banking industry one should expect considerably higher means for the performance-based measures in MBV group 3. This is indeed the case. The return on average assets (ROAA) is almost thrice as high compared to the average bank in MBV group 1 and more than twice as high compared to the average bank in MBV group 2. The latter group stands out because it features the highest loans over assets (LOA_A) means and the lowest average growth rates.

Table A2 indicates large inter-group differences regarding the resources deployed by European banks. Compared to other banks in the sample the average bank in RBV group 1 relies heavily on customer deposits for refinancing and less on issuing debt on the open market. It acts as a net lender to other banks and possesses a higher portion of fixed assets than the average bank in RBV-groups 2 and 3. With the marked exception of group 3//3, cost/income ratios and personnel expenses per employee do not seem to differ very strongly across the groups.

In the next section, discriminant- and logit model analysis will indicate which of these differences are statistically significant and therefore which of the strategy variables do potentially form mobility barriers.

15

Table 2: Concentration of banks by number and assets

MBV 1 2 3 Country

MBV//RBV 1//1 1//2 1//3 2//1 2//2 2//3 3//1 3//2 3//3 Sum assets

AUS 1/1% 1/1% 3/42% 4/48% 2/1% 2/6% 1/1% 14 2,3%

FRA 1/1% 4/3% 3/2% 6/15% 15/15% 9/27% 18/39% 56 6,3%

GER 70/6% 163/28% 15/46% 12/1% 97/11% 36/5% 10/1% 20/3% 4/0% 427 34,0%

ITA 1/7% 1/0% 5/2% 59/91% 66 17,3%

SPA 1/4% 1/2% 8/30% 5/65% 15 3,6%

UNI 2/10% 3/22% 3/55% 4/14% 12 10,3%

FIN 1/19% 3/81% 4 3,9%

SWE 5/100% 5 7,8%

DEN 4/55% 2/32% 1/13% 7 0,2%

BEL 3/100% 3 0,1%

NET 1/73% 1/1% 1/6% 1/2% 2/18% 6 11,3%

POR 1/28% 1/2% 2/13% 2/5% 3/52% 9 2,8%

72/2% 165/10% 18/25% 20/2% 110/8% 49/9% 43/3% 53/10% 94/31% 624 100,0% All

255 / 38% 179 / 19% 190 / 43%

0,3 0,6 13,9 0,8 0,7 1,9 0,6 1,8 3,3 1,6 Size*

1,5 1,1 2,3

*Average bank size for each group as measured by the portion in sample total assets (in 0/00)

Table 2 contains the number of banks from a particular country and the portion of total sample bank assets for each cluster. It reveals a heavy concentration of banks from the same country in only a few clusters. Moreover, the size distribution of banks across the clusters seems to be highly heterogeneous. With just one exception (group 2//2), average bank size increases with RBV-group rank. For example, group 1//3 contains only 18 banks from four countries, that is less than 3% of the sample, but accounts for 25% of the total assets in the sample. The average bank in this group is roughly 25 times larger than the average bank in group 1//1 and 50 times larger than the average bank in group 1//2.

4.2 Discriminant analysis

Two orthogonal discriminant functions have been estimated for the three MBV-groups and for the three RBV-groups, respectively.9 Their multivariate Wilk’s Lambdas imply that only 23% (MBV) and 20% (RBV), respectively, of the total variance in observations is left unexplained.

9 The maximisation of Γ only determines the relation between the determinant function coefficients. The

coefficients in Table 3 have been normalised so that pooled within-group variances of the iy ’s are equal to

unity.

16

Table 3: Discriminant functions

MBV RBV

Discriminant function 1 2 1 2

Eigenvalue 1,423 0,776 2,517 0,395

Eigenvalue portion 64,7% 35,3% 86,4% 13,6%

Coefficients

Constant -0,56 -14,22 Constant -6,24 0,17

AVGRWTHA 1,84 -1,14 CST_INC -1,46 0,80

EQU_A 29,52 5,68 DEP_A 12,69 -3,94

INTMRG 44,39 -26,30 FIX_A -21,36 35,50

LOA_A -5,91 21,21 INTBK 0,33 1,83

OINC_INT 3,28 0,40 OFUND_A -3,14 -2,35

STDV_ROA 212,40 28,75 PERSX_P -0,01 0,00

Weighted standardised coefficients

AVGRWTHA 0,18 CST_INC 0,15

EQU_A 0,42 DEP_A 0,87

INTMRG 0,31 FIX_A 0,20

LOA_A 0,54 INTBK 0,29

OINC_INT 0,46 OFUND_A 0,17

STDV_ROA 0,28 PERSX_P 0,12

Because of different scaling, the non-standardised coefficients of the discriminant functions in Table 3 must not be interpreted in terms of the strengths of concerning variable effects. However, they can be readily used to classify out-of-sample banks into one of the nine groups.10 For that purpose, observations are entered into the discriminant functions given by these very coefficients to obtain the differences between the resulting yi and the three group centroids. The idea is to assign the bank to that particular group where the distance to the average bank – the centroid – is the smallest.11 These distances can also be used to obtain a quite intuitive measure for the goodness of fit of the discriminant functions themselves. Table 4 contains a comparison of the classification results from a jack-knifing procedure to the original classifications as implied by the Wald-algorithm. Both pairs of discriminant functions are able to classify roughly 90% of the 624 banks correctly.

10 Because the Eigenvalue portions for the first discriminant functions are only 60% and 80%, respectively,

it is necessary in most contexts to consider both functions.

11 The Box-M test clearly indicated dissimilar within-group variances of the six independent variables, Therefore modified Mahalanobis distances should be used. The required vectors and matrices are available from the author upon request.

17

Table 4: Classification matrix MBV RBV

Classification implied by discriminant functions

1 2 3 1 2 3

1 250 3 2 104 31 0

2 17 157 5 16 305 7 Ward clusters

3 26 14 150 2 16 143

Correctly classified 89,3% 88,5%

The standardised coefficients in Table 3 have been obtained by first multiplying the non-standardised coefficients of discriminant functions 1 and 2 by the standard deviation of the corresponding xi and by then computing the weighted sum of each pair based on the Eigenvalue portion of the respective function. Their absolute values indicate the relative discriminatory power of the variables across all three groups. Loans over assets discriminates best between the MBV-groups followed by “other income over interest income” and “equity over assets”. The overall discriminating power of “average asset growth rate” seems to be quite low. This also seems to be the case for four out of the six RBV variables except “deposits over assets” and the interbank ratio.

4.3 Logit model analysis

Not surprisingly, the two logit models are also characterised by a strong overall goodness of fit. The pseudo-R2 in (13) amounts to 82,2% for the MBV-model and 78,7% for the RBV-model.12 Jack-knifing procedures based on the probabilities as implied by the three logit functions assign 93% (90%) of the banks to the correct MBV (RBV) cluster.

A third measure for the goodness of fit, reported in the first row of Table 5, is the model’s G-statistic according to (14). It is highly significant. The other rows of Table 5 contain the G-statistics for the single variables (see (15)). They indicate the general importance of each variable across all three logit equations and as such indicate strategic levers that might constitute mobility barriers in European commercial banking. With the marked exception of OFUND_A and PERSX_P they are all highly significant. There exists a close correspondence between the relative importance of the twelve variables in the logit models and in the discriminant analyses. Both approaches rank the variables quite similarly

12 The pseudo-R2 declines drastically when the dependent variables are switched between the two models. A

logit model that tries to explain the MBV (RBV)-group membership based on the set of six RBV (MBV)-variables yields a pseudo-R2 of 0,163 (0,144) and classifies only 59% (63%) of the banks correctly.

18

in this respect. Exceptions are INTMRG that is more important in the MBV logit model and FIX_A that is more important in RBV discriminant analysis.

Table 5: Likelihood-ratio tests

MBV RBV

G G

Full model 1113,93 *** Full model 1000,18 ***

Gq Gq

Constant 527,92 *** Constant 227,98 ***

AVGRWTHA 64,09 *** CST_INC 39,89 ***

EQU_A 107,71 *** DEP_A 582,47 ***

INTMRG 55,89 *** FIX_A 56,26 ***

LOA_A 580,58 *** INTBK 164,21 ***

OINC_INT 177,26 *** OFUND_A 2,28

STDV_ROA 121,04 *** PERSX_P 4,52

***significant at the 1% level, **significant at the 5% level, *significant at the 10% level

These effects are, however, average effects across all three logit equations. Table 6 and Table 7 disentangle them for each group-wise comparison. In analogy to discriminant analysis, non-standardised logit-coefficients must not be compared across variables but should only be interpreted with respect to their signs and with respect to their absolute values across the three binary models13: As expected, almost all variables unfold their greatest discriminatory power in delineating the two groups that are “farthest apart”, namely group 1 and 3. The signs of all but one of the 30 significant coefficients are as implied by the group means in Table A1 and A2.14 In what follows, we will, however, only refer to variables with highly significant coefficients.

The modified t-statistics in Table 6 were obtained from (16) and indicate the relative strength of the variable-effects for each of the three binary models. Both, the weighted standardised discriminant coefficients and the logit G-statistics identified LOA_A as having a strong overall effect. The respective row in Table 6 reveals that this effect is particularly due to LOA_A acting as a potential mobility barrier between MBV-group 2 and the other two. Equivalently, the risk exposure as measured by EQU_A and STDV_ROA and

13 They can also be used in combination with (10) to obtain the classification probabilities for a specific bank.

Based on these probabilities, necessary changes in variables for this bank to be reclassified into another group, can then be analysed.

14 The coefficient of FIX_A in the equation relating to MBV-groups 1 and 3 has a negative sign, although the respective group mean of group 1 exceeds that of group 3.

19

OINC_INT seems to be considerably higher for MBV-group 1 than for the other two, thereby giving rise to other potential mobility barriers

Table 6: Non-standardised logit coefficients (MBV)

MBV-groups 1 vs. 3 2 vs. 3 1 vs. 2

coeff. mod. t-stat. coeff. mod. t-stat. coeff. mod. t-stat.

Constant 56,84 49,77 *** -45,68 52,75 *** 102,51 102,07 ***

AVGRWTHA -3,06 5,51 ** -20,56 58,02 *** 17,50 34,58 ***

EQU_A -170,83 37,78 *** -162,40 38,21 *** -8,43 0,08

INTMRG -145,70 10,76 *** -236,94 34,96 *** 91,24 3,16 *

LOA_A -52,54 25,25 *** 99,87 70,65 *** -152,42 102,60 ***

OINC_INT -12,49 35,15 *** -14,89 58,35 *** 2,39 1,06

STDV_ROA -2545,76 50,82 *** -1102,54 24,05 *** -1443,22 15,98 ***

***significant at the 1% level, **significant at the 5% level, *significant at the 10% level

Table 7: Non-standardised logit coefficients (RBV)

RBV-groups 1 vs. 3 2 vs. 3 1 vs. 2

coeff. mod. t-stat. coeff. mod. t-stat. coeff. mod. t-stat.

Constant -68,61 81,94 *** -36,57 33,60 *** -32,04 53,34 ***

CST_INC -9,85 10,81 *** -13,40 25,59 *** 3,54 3,16 *

DEP_A 127,02 68,93 *** 93,23 41,11 *** 33,79 46,53 ***

FIX_A -193,40 11,03 *** -269,58 25,79 *** 76,18 7,62 ***

INTBK 1,18 1,13 -3,90 15,42 *** 5,08 63,45 ***

OFUND_A 8,58 1,07 9,05 2,22 -0,47 0,01

PERSX_P -0,01 0,57 -0,01 1,88 0,00 0,02

***significant at the 1% level, **significant at the 5% level, *significant at the 10% level

This kind of reasoning carries over to the RBV-variables. RBV-group 1 seems to be isolated from the other two groups by very high ratios for DEP_A indicating mobility barriers due to a stronger core deposit base. Another barrier, albeit of much lower “height”, seems to exist between RBV-groups 2 and 3 due to lower CST_INC and FIX_A ratios of the banks in group 2. Finally, the difference in INTBK is most pronounced between groups 1 and 2, possibly indicating less abundant investment opportunities for the banks in the former group. We will discuss these observations in the next section.

A question that might occur at this point is whether the strategic groups differ significantly with respect to ROAA and ROAE. For a tentative answer, we first computed the serial correlations between each of the two performance measures for a bank and its group membership. Both ROAA (0,483) and ROAE (0,323) are noticeably and positively correlated to the MBV group ranks, whereas hardly any correlation could be found vis á vis the RBV group ranks (0,145 and 0,094

20

respectively). Secondly, we estimated three linear regression models with ROAA as the dependent variable and four/two dummy variables as regressors. A dummy variable was set to one if a bank belongs to the respective cluster. Table 8 shows the coefficients and their p-values in parentheses for all three models. Model “MBV and RBV”, in which all four dummies have been considered, has only a slightly higher R2 than the reduced model “MBV”, which merely features the dummies for MBV-groups 2 and 3. Not surprisingly, an F-test indicates that the MBV dummies have a significantly higher explanatory power than the RBV dummies.

Table 8: Regression results

Model R2 Constant MBV2D MBV3D RBV2D RBV3D

0,264 0,238% 0,082% 0,482% 0,023% 0,019% MBV and RBV

(0,000) (0,021) (0,000) (0,528) (0,651)

0,263 0,254% 0,085% 0,482% MBV

(0,000) (0,013) (0,000)

0,042 0,403% -0,037% 0,160% RBV

(0,000) (0,359) (0,001)

We therefore conjecture that, whereas there exists hardly any systematic relationship between the resource bundle chosen by a given bank and its ROAA, its market-based strategy determines performance to a considerable extent.15 The coefficients for MBV2 and MBV3 are economically significant, too. The model implies that the ROAA of the average bank in MBV-cluster 1, which amounts to 0,254%, increases by 0,085 (0,482) percentage points should the bank emulate the market-based profile of the average bank in MBV-group 2 (3). These incremental values (should) correspond exactly to the differences in ROAA group means as shown in Table A1.

5 Discussion In this section we will use the empirical evidence from above to derive and

discuss six general propositions on structural features of European commercial banks in 1999. For that purpose, the intermediation and securitisation ratios presented in the introduction serve as a natural starting point. In Figure 2 we have rearranged and summarised the data contained in Table 2.

The grey figures show the number of banks and their share in total-sample assets per cluster. Cluster 1//2 for example, contains 165 banks comprising of a tenth of total assets in the sample. Diamonds indicate the concentration of banks

15 Regressing ROAE on the dummies yields similar results with respect to their relative significance.

However, the R2 for the model with all six dummies is considerably lower at 0,116.

21

from a specific country and in their totality cover well over 90% of sample bank assets of each country. Two adjacent diamonds are linked by connectors if the banks are from the same country. Except for one Dutch bank in cluster 1//3 and one Portuguese bank in cluster 1//2 the great majority of banks from the same country are hence “neighbours”. In each chain we have labelled the diamond representing the largest number of banks with the corresponding country symbol.16

Figure 2: Distribution of banks across the nine clusters

NET

UNI

FRAITA

AUS

FINSWE

DEN

BEL

NET

POR

POR

SPA

1 2 3

1

2

3

722%

16510%

1825%

499%

1108%

202%

9431%

5310%

433%

GER

RBVMBV

NET

UNI

FRAITAITA

AUS

FINSWE

DEN

BEL

NET

PORPOR

PORPOR

SPA

1 2 3

1

2

3

722%

16510%

1825%

499%

1108%

202%

9431%

5310%

433%

GER

RBVMBV

An obvious observation from Figure 2 leads to our first proposition.

Proposition 1: There are remarkable differences regarding the strategies pursued by banks from different countries. Market based strategies differ more in this respect than resource based strategies.

Confirming the vast differences in roles played by French, British and German banks, most German banks belong to MBV-group 1, most British banks are members of MBV-group 2, and most French banks belong to MBV-group 3. Banks from seven other continental European countries have been assigned by the Ward-algorithm to MBV-group 3, with group 3//3 encompassing over 30% of total sample assets. Interestingly to note, 59 of a total of 66 Italian banks comprising 91% of total Italian sample assets and all 5 Swedish banks belong to this group.17 Market-based strategies and to some extent also resource based

16 In the few cases where the number of banks were about equal the cluster representing the larger asset base

was chosen. 17 For the sake of clarity, all Danish banks (0,2% of the total sample’s total assets) show up in cluster 3//1.,

although most of the banks actually belong to groups 3//2 and 3//3.

22

strategies hence seem to be more heterogeneous across banks from different countries than between banks operating in the same country.

For a closer examination of this country-wise delineation we extend Figure 2 by the results of the logit-model. Based on an observation from Tables A1 and A2, namely that average bank size clearly increases with RBV-group rank for all three MBV-groups but not so vice versa, we add a vertical axis to the left of the matrix indicating a growing asset base. The axis explains the vertical positions of the diamonds in the grid. For example, the labelled diamond for Germany indicates that the average German bank in group 1//2 is very small compared to other banks in RBV-group 2. It is located right on the border to group 1//1 because the average bank is indeed more comparable in size to the banks from groups 2//1 and 3//1.

The horizontal axis at the bottom of the matrix is a direct implication of the results of the regression model in Table 8: The higher the MBV-group rank the higher the return on assets of the average bank in the group. The horizontal position of the diamonds in a group has no meaning and was only chosen for the sake of a better exposition.

Figure 3: Mobility barriers between strategic groups

NET

UNI

FRAITA

AUS

FINSWE

DEN

BEL

NET

POR

POR

SPA

1 2 3

1

2

3

722%

16510%

1825%

499%

1108%

202%

9431%

5310%

433%

GER

RBVMBV

AS

SE

TS

ROAA

DE

P_

A

LOA_A

OINC_INT

INTMRG

EQU_A

CS

T_I

NC

INT

BK

FIX

_ASTDV_ROA

AVGRWTHA

NET

UNI

FRAITAITA

AUS

FINSWE

DEN

BEL

NET

PORPOR

PORPOR

SPA

1 2 3

1

2

3

722%

16510%

1825%

499%

1108%

202%

9431%

5310%

433%

GER

RBVMBV

AS

SE

TS

ROAA

DE

P_

A

LOA_A

OINC_INT

INTMRGINTMRG

EQU_A

CS

T_I

NC

INT

BK

FIX

_ASTDV_ROA

AVGRWTHA

23

Finally, the arrows on the top and to the right of the matrix reflect the results of the active-variable logit model. Arrowheads indicate that the corresponding variable was found to have significant discriminatory power to delineate two groups and as a result is on average significantly higher for the group towards which the arrow is pointing. If an arrow spans MBV groups 1 and 2, as is the case for three MBV variables, not only the banks in group 2 but also the banks in group 1 are differently positioned from the banks in group 3. It can argued that banks that wish to move from one group to another have to overcome the difference in the corresponding variable profiles. The separating lines can therefore be figuratively called mobility barriers. As has become obvious from Table 6, those barriers are generally “higher” between MBV-groups 1 and 3 and 2 and 3, respectively, than between the first two groups.

The second proposition rests upon the relationship of ROAA with the other variables:

Proposition 2: The return on assets of European commercial banks is primarily driven by their market based variable profile and hardly at all by their endowment with resources.

This proposition holds true whether one observes returns on average assets (ROAA) or a crude form of the Sharpe-ratio reported in the very last row of Table A1 (ROAA over STDV_ROA). Interestingly to note, the increasing returns generated on assets by banks in higher-rank MBV-groups seem to be over-compensated by the risk these very banks bear. As a consequence, the Sharpe ratio decreases (with one exception) with a higher MBV-group rank. In sharp contrast to these observations and thereby in contrast to Mehra’s (1996) results in the context of US bank holding companies, the membership to a particular RBV group does not seem to affect the performance measures in a discernible way.

As bank size does not seem to strongly affect bank performance our study provides indirect evidence on the existence of only minor economies of scale in commercial banking. It thereby confirms one of the few unambiguous results of the vast recent literature on measuring bank efficiency.18 Moreover, there is also some evidence concerning the existence of economies of scope. If we conceive of commercial banking as comprising interest revenue generating deposit and loan business on the one hand and some fee generating businesses like payment services and custody services on the other hand we can use the MBV variable

18 For an overview on this still growing strand of the econometric literature refer to Berger/Humphrey (1997)

and Goddard et al (2001).

24

OINC_INT, which measures the relation of commission income to interest income, as a proxy for a bank’s business scope. Low values indicate a strong concentration on traditional interest-based businesses, whereas higher values indicate an increasing degree of diversification into other, fee-based businesses. Combining the observation on the weak effects of bank size with the observation on the fairly strong effect of OINC_INT in discriminating between MBV-groups 1 and 3 and 2 and 3, respectively, directs us to the third proposition.

Proposition 3: Economies of scope seem to be more prevalent in European commercial banking than economies of scale.

Focusing on the growth patterns of the banks in the sample it can be observed that those banks having generated a much larger portion of their income from fees in 1999 and having assumed more risk as indicated by STDV_ROA and EQU_A tend to have been growing faster during the last couple of years. As a result, the average bank in MBV-group 3 indicating a value for OINC_INT almost twice as high as its peers in MBV-group 1 and 2, respectively and indicating a STDV_ROA about thrice as high has been growing on average by 10,7% per year. This compares to growth rates of 8,2% and 8,6% for its peers in the first two MBV-groups. The growth rates of banks in MBV-group 1 have thus actually been slightly higher than for the banks in MBV-group 2. However, the values for OINC_INC do hardly at all and the values for STDV_ROA do only slightly differ between these two groups. One could therefore conjecture that the high proportion of loans to customers on the asset side of the latter group’s average bank in combination with its low interest margin have been an obstacle to a stronger growth. Proposition 4 summarises these findings.

Proposition 4: Stronger growth in European commercial banking has generally been associated with the assumption of more risk and a diversification into fee generating businesses.

We now turn to the RBV-variables. A straightforward result implied by Figure 3 is that smaller banks possess a much stronger deposit base as compared to larger banks. The ratio “customer deposits over total assets” (DEP_A) is roughly 75% for the average bank in RBV-group 1 and only 65% and 45% for the average bank in RBV group 2 and 3, respectively. Associated with this result but not in unequivocal correspondence to it, do small banks with a large deposit base act as lenders in the interbank market, whereas the average (medium size) bank in RBV-group 2 acts as a borrower. As the variable INTBK does not discriminate well between RBV groups 1 and 3, the evidence on the different roles in the interbank markets played by small and large banks is mixed on an aggregate level. A look at Table A1, however, provides a possible explanation

25

for this phenomenon. The row containing the mean values of INTBK for each of the nine groups reveals that RBV-group 3 comprises of banks that engage in heavy interbank borrowing (2//3) and those that do only get a smaller portion of their funds from other banks (3//3). It is relatively safe to conclude that large banks in group 3//3, although they have a relatively small deposit base, prefer to tap the capital markets (as evidenced by a higher OFUND_A ratio) for the purpose of funding their asset business. The latter is considerably less dominated by customer loans than that of group 2//3, which has the highest LOA_A ratio of all groups. The heterogeneity in funding and lending-/investment patterns of banks in RBV-group 3 would hence require more detailed information to reconcile the relationship between deposit base and interbank activity but ultimately leaves proposition 5 unaffected.

Proposition 5: Smaller European commercial banks possess the largest customer deposit base and are in a position to lend excess funds over the interbank market.

The evidence on the differences between the strategic groups in terms of the other four RBV-variables is less clear. Proceeding from the market position of the average bank in group 3//3, one could argue that putting a very strong emphasis on fee based businesses (OINC_INT = 56,7%) and assuming very high business risks (STDV_ROA = 0,35%) entails the necessity to incur high personnel expenses (PERSX_P = 59), a high proportion of fixed assets (FIX_A = 1,9%) and a high cost/income ratio (CST_INC = 73,6%). However, as this kind of relationship can only be constructed for a small number of groups and as it is not directly implied by our empirical results we refrain from any further associated speculations. As a consequence, proposition 6 is formulated in a rather moderate tone.

Proposition 6: Except for a few polar cases comprising mainly large and diversified institutions from Scandinavia, Italy and Portugal, European commercial banks do not differ systematically in terms of their resource based strategies, ie in terms of their choice of personnel, the size of their fixed asset base and their cost/income ratio.

6 Implications and conclusion Banks that are more homogenous in terms of both their employment and

deployment of resources are less differentiated from their group peers but might be shielded from new entry because outsiders might find imitation very difficult. Porter (1979) therefore conjectures that competition should be more intense

26

between banks from the same strategic group than between banks separated by mobility barriers.

Translated into the context of our study this argument implies that competition between domestic commercial banks is more intense than across Europe. Moreover, the mobility barriers that have been identified might be part of the story explaining past failures of banks to successfully enter other European countries by means of organic growth. Another insight from the strategic group literature, namely that members of different groups tend to react differently to identical changes in their business environment, might explain the fact that regulatory harmonisation has so far not dismantled the mobility barriers between many countries. In the absence of a single (globally) superior strategy profile that all commercial banks believe worthwhile to strive for, it would have been pure coincidence if banks from structurally diverse financial systems had converged in terms of their market-product combinations.

This line of reasoning allows for two alternative, broad predictions concerning the future development of the European banking industry. Firstly, if consistency is indeed a central precondition for the well functioning of financial systems, we should expect path dependencies to furthermore impede structural convergence of their constituting elements. Because the role and thereby also the strategies of banks are without doubt an essential element of every financial system, this very element should neither be subject to dramatic change in the near future. Secondly, if complementarities between system elements were only a theoretical artefact, the observable differences between bank roles and strategies must be primarily due to different demand structures. The common currency and the ongoing integration and sophistication of financial and product markets is expected by many observers to diminish these differences in the short-term. As resource-based barriers do already seem to be lower across groups we should expect competition between banks from different countries to increase strongly.

Depending on which of the alternative conjectures is valid, different implications for bank management follow. One implication relates to the question of what banks should be considered as (potential) direct competitors. The first alternative implies that MBV mobility barriers will largely remain in place so that direct competitors of a given bank will still mainly come from the same strategic group. As a consequence, there would still be room in the future for small banks focussing on traditional banking products. According to the second alternative, dismantled mobility barriers would allow foreign competitors to compete more directly with domestic players. As smaller, purely

27

domestic banks might then suffer from a squeeze in their deposit base and/or deteriorating loan portfolios, consolidation in the European banking industry would probably intensify.

A second implication refers to the issue of optimal strategy selection. If Porter (1998) is right with his proposition that competitive advantage is a function of the - hard-to-observe and hence hard-to-copy - strategic fit of a firm’s strategic levers, then again, strategic diversity and performance differentials will probably remain in place. Banks “trapped” in a low-performing group would then face the almost insurmountable task to simultaneously change their strategic profile in many respects. According to Milgrom/Roberts (1995), a step-wise change would inevitably require a bank to endure a - possibly too long - period in which single activities do not fit together and performance is low.

In the other case, in which competitive advantage is a function of solely positioning a bank correctly along a few strategic dimensions (eg focus on fee based businesses and/or higher risk exposure), one should expect bank strategies to converge quickly across Europe. Recent trends in European banking in conjunction with the fact that roughly a third of the sample bank assets already belong to cluster 3//3, would then imply that the future of banks from other groups depends crucially on overcoming the respective mobility barriers.

28

References

ALLEN, F./GALE, D. (1997): Financial Markets, Intermediaries and Intertemporal Smoothing, in: Journal of Political Economy, Vol. 105, pp 523-546

Amel, D./Rhoades, S. (1998): Strategic Groups in Banking, The Review of Economics and Statistics, Vol. 70(4), pp 685-689

Barney, J.B./Hoskisson, R.E. (1990): Strategic Groups: Untested Assertions and Research Proposals, Managerial and Decision Economics, Vol. 11, pp 187-198

Berger, A.N./Humphrey, D.B. (1997): Efficiency of Financial Institutions: International Survey and Directions for Future Research, European Journal of Operational Research, Vol. 98, pp 175-212

Caves, R.E./Porter, M. (1977): From Entry Barriers to Mobility Barriers: Conjectural Decisions and Contrived Deterrence to New Competitors, Quarterly Journal of Economics, Vol. 91, pp 241-262

Chamberlain, E. (1932): The Theory of Monopolistic Competition, Cambridge, MA

Cool, K./Dierickx, I.(1993): Rivalry, Strategic Groups and Firm Profitability, Strategic Management Journal, Vol. 14(1), pp 47-59

Daems, H./Thomas, H., eds. (1994): Strategic Groups, Strategic Moves, and Performance, Oxford

Dranove, D./Peteraf, M./Shanley, M. (1998): Do Strategic Groups Exist? An Economic Framework for Analysis, Strategic Management Journal, Vol. 19, pp 1029-1044

Duysters, G./Hagedoorn, J. (1995): Strategic Groups and Inter-Firm Networks in International High-Tech Industries, Journal of Management Studies, Vol. 32 (3), pp 359-381

Eisenbeis, R.A. (1977): Pitfalls in the Application of Discriminant Analysis in Business, Finance, and Economics, Journal of Finance, Vol. 32 (3), pp 875-900

Ferguson, T.D./Deephouse, D.L./Ferguson, W.L. (2000): Do Strategic Groups Differ in Reputation, Strategic Management Journal, Vol. 21, pp 1195-1214

Fiegenbaum, A./Sudharshan, D./Thomas, H.(1987): The Concept of Stable Strategic Time Periods in Strategic Group Research, Managerial and Decision Economics, Vol. 8, pp 139-148

Fiegenbaum, A./Thomas, H. (1993): Industry and Strategic Group Dynamics, Competitive Strategy in the Insurance Industry 1970-84, Journal of Management Studies, Vol. 30(1), pp 69-105

Fiegenbaum, A./Thomas, H. (1995): Strategic Groups and Reference Groups: Theory, Modeling and Empirical Examination of Industry and Competitive Strategy, Strategic Management Journal, Vol. 16 (6), pp 461-476

Fiegenbaum, A./Thomas, H.(1990): Strategic Groups and Performance: The U.S. Insurance Industry 1970-84, Strategic Management Journal, Vol. 11 (3), pp 197-215

Goddard, J./Molyneux, P./Wilson, J.O.S. (2001): European Banking – Efficiency, Technology and Growth, New York et al

Gruca, T.S./Nath, D.(2000): The Effect of Regulation on the Strategic Group-Performance Link, mimeo, University of Iowa

29

Hatten, K.J./Hatten, M.L. (1987): Strategic Groups, Asymmetrical Mobility Barriers and Contestability, Strategic Management Journal, Vol. 8 (4), pp 329-342

Hunt, M.S. (1997): Competition in the Major Home Appliance Industry 1960-1970, PhD dissertation, Harvard University

Ketchen, D.J./Combs, J.G./Russell, C.J./Shook, C./Dean, M.A./Runge, J./Lohrke, F.T./Naumann, S.E./Haptonstahl, D.E./Baker, R./Beckstein, B.A./Handler, C./Honig, H./Lamoreux, S. (1997): Organizational Configurations and Performance: A Meta-Analysis, Academy of Management Journal, Vol. 40, pp 222-240

Ketchen,D.J./Shook, C.L. (1996): The Application of Cluster Analysis in Strategic Management Research: An Analysis and Critique, Strategic Management Journal, Vol. 17 (6), pp 441-458

Maddala, G.S. (1988): Limited-Dependent and Qualitative Variables in Econometrics, Cambridge

Marlin, D./Lamont, B.T./Hoffman, J.J. (1994): Choice Situation, Strategy and Performance: A Reexamination, Strategic Management Journal, Vol. 15, pp 229-239

Mehra, A. (1996): Resource and Market Based Determinants of Performance in the U.S. Banking Industry, Strategic Management Journal, Vol. 17 (4), pp 307-322

Milgrom, P./Roberts, J. (1995): Complementarities and Fit: Strategy, Structure, and Organizational Change in Manufacturing, Journal of Accounting and Economics, Vol. 19, pp 179-208

Nath, D./Gruca, T.S. (1997): Convergence Across Multiple Methods for Indentifying Strategic Groups, Strategic Management Journal, Vol. 18, pp 745-760

Pitt, M./Thomas, H. (1994): Industry Groups and Strategic Management: A Reappraisal of Strategic Group Concepts and Research Methodologies, in H. Daems and H. Thomas (eds.), Strategic Groups, Strategic Moves and Performance, pp 81-102

Porter, M. (1979): The Structure within Industries and Companies Performance, Review of Economics and Statistics, Vol. 61, pp 214-227

Porter, M. (1980): Competitive Strategy, New York

Porter, M. (1996): What is Strategy?, Harvard Business Review, Vol. 74, Nov.-Dec., pp 61-78

Schmidt, R.H./Hackethal, A./Tyrell, M. (1999): Disintermediation and the Role of Banks in Europe: An International Comparison, Journal of Financial Intermediation, Vol. 8, pp 36-67

Schmidt, R.H./Hackethal, A./Tyrell, M. (2001): The Convergence of Financial Systems in Europe, Working Paper Series: Finance and Accounting: No. 75, University of Frankfurt

Thomas, H./Venkatraman, N. (1988): Research in Strategic Groups: Progress and Prognosis, Journal of Management Studies, Vol. 6, pp 537-556

Wernerfeldt, B. (1984): A Resource-Based View of the Firm, Strategic Management Journal, Vol. 5(2), pp 171-180

30

Appendix

Table A1: Group means of the market-based and performance-based variables

MBV 1 2 3

MBV//RBV 1//1 1//2 1//3 2//1 2//2 2//3 3//1 3//2 3//3

8,6% 8,2% 10,7%

AVGRWTHA 9,4% 7,9% 11,4% 10,4% 7,8% 8,2% 12,6% 12,6% 8,7%

4,48% 4,85% 7,92%

EQU_A 4,7% 4,5% 4,1% 5,4% 4,7% 4,9% 7,1% 6,8% 8,9%

2,7% 2,7% 3,4%

INTMRG 2,7% 2,7% 2,2% 2,8% 2,8% 2,6% 3,7% 3,7% 3,1%

64,6% 74,5% 68,8%

LOA_A 64,4% 64,6% 64,9% 75,0% 73,3% 77,0% 69,6% 70,4% 67,6%

26,6% 29,6% 52,4%

OINC_INT 28,6% 25,5% 28,5% 34,8% 27,6% 32,0% 49,2% 47,2% 56,7%

0,08% 0,11% 0,31%

STDV_ROA 0,07% 0,08% 0,10% 0,13% 0,10% 0,13% 0,23% 0,31% 0,35%

0,25% 0,34% 0,74%

ROAA 0,24% 0,26% 0,23% 0,40% 0,32% 0,35% 0,67% 0,78% 0,74%

5,67% 6,85% 9,88%

ROAE 5,17% 5,91% 5,53% 7,29% 6,71% 6,99% 9,10% 11,59% 9,28%

3,4 3,0 2,4 ROAA / STDV_ROA 3,3 3,5 2,3 3,1 3,1 2,7 2,9 2,5 2,1

Table A2: Group means of the resource-based variables

RBV 1 2 3

MBV//RBV 1//1 2//1 3//1 1//2 2//2 3//2 1//3 2//3 3//3

68,3% 66,9% 70,2%

CST_INC 69,7% 69,7% 65,4% 67,3% 66,5% 66,1% 69,6% 63,8% 73,6%

74,9% 65,4% 44,5%

DEP_A 75,1% 74,0% 75,0% 65,7% 64,4% 66,6% 48,2% 45,1% 43,5%

1,9% 1,6% 1,8%

FIX_A 2,0% 1,6% 1,8% 1,6% 1,6% 1,5% 1,6% 1,7% 1,9%

133,1% 39,2% 46,9%

INTBK 106,1% 105,3% 191,0% 36,1% 36,9% 53,4% 38,7% 17,5% 63,8%

4,8% 6,4% 12,0%

OFUND_A 5,3% 5,5% 3,7% 6,5% 7,1% 5,0% 12,8% 7,1% 14,4%

45,9 43,6 53,6

PERSX_P 44,0 49,8 47,2 42,1 43,7 47,9 44,7 46,6 59,0