Using clustering ensemble to identify banking business models

• Published date: 01 April 2020
• Date: 01 April 2020
• DOI: http://doi.org/10.1002/isaf.1471
• Author: Carlos F. Alves,Bernardo P. Marques

RESEARCH ARTICLE

Using clustering ensemble to identify banking business models

Bernardo P. Marques

1,2

| Carlos F. Alves

2

1

Universidade Católica Portuguesa, Católica

Porto Business School, Porto, Portugal.

2

Universidade do Porto, Faculdade de

Economia, CEF.UP, Porto, Portugal

Correspondence

Bernardo P. Marques, Universidade Católica

Portuguesa, Católica Porto Business School,

Porto, Portugal.

Email: bmarques@porto.ucp.pt

Funding information

Fundaç~

ao para a Ciência e a Tecnologia, Grant/

Award Numbers: SFRH/BD/135939/2018

(Phd scholarship), UIDB/00731/2020 (CPBS),

UIDB/04105/2020 (CEF.UP)

Summary

The business models of banks are often seen as the result of a variety of simulta-

neously determined managerial choices, such as those regarding the types of activi-

ties, funding sources, level of diversification, and size. Moreover, owing to the

fuzziness of data and the possibility that some banks may combine features of differ-

ent business models, the use of hard clustering methods has often led to poorly iden-

tified business models. In this paper we propose a framework to deal with these

challenges based on an ensemble of three unsupervised clustering methods to iden-

tify banking business models: fuzzy c-means (which allows us to handle fuzzy cluster-

ing), self-organizing maps (which yield intuitive visual representations of the clusters),

and partitioning around medoids (which circumvents the presence of data outliers).

We set up our analysis in the context of the European banking sector, which has seen

its regulators increasingly focused on examining the business models of supervised

entities in the aftermath of the twin financial crises. In our empirical application, we

find evidence of four distinct banking business models and further distinguish

between banks with a clearly defined business model (core banks) and others

(non-core banks), as well as banks with a stable business model over time (persistent

banks) and others (non-persistent banks). Our proposed framework performs well

under several robustness checks related with the sample, clustering methods, and

variables used.

KEYWORDS

banking, business models, clustering ensemble, fuzzy clustering, self-organizing maps

1|INTRODUCTION

This paper deals with the special methodological requirements that

emerge from the task of business model identification—a task that has

gained particular relevance in the context of recent efforts to reform

the regulation and supervision of banks in Europe (EBA, 2014;

ECB, 2018). In particular, policymakers and researchers have become

increasingly focused on grouping banks based on the similarity of their

business model choices (such as size, types of activities, funding, and

diversification). However, in doing so, they have faced significant chal-

lenges in finding clearly separated and homogeneous clusters. This

occurs chiefly because business choices are likely to follow a fuzzy,

rather than a crisp, logic; for example, some banks may choose to

combine features of different business models following a merger or

acquisition (DeSarbo and Grewal, 2008).

In general, by applying clustering analysis to the business choices

of banks one may hope to achieve two main goals. First, to obtain an

objective and stable taxonomy of business model classifications,

which in turn may be used by supervisors to monitor the performance

of banks in each business model (e.g. by identifying outliers)—in line

with the guidelines for the supervisory review and evaluation process

(EBA, 2014). Second, to obtain a better insight into the competitive

structure of the banking sector, as banks with similar business choices

may be expected to compete more intensely among themselves

(Porter, 1979). The former goal (i.e. attaining an objective and stable

taxonomy of business models) seems particularly timely given that the

Received: 26 July 2019 Revised: 3 January 2020 Accepted: 5 March 2020

DOI: 10.1002/isaf.1471

66 © 2020 John Wiley & Sons, Ltd. Intell Sys Acc Fin Mgmt. 2020;27:66–94.wileyonlinelibrary.com/journal/isaf

method currently used by supervisors to identify business models

(expert judgement) may lead to inconsistencies and potential conflict

of interests. In particular, under the principle of proportionality, differ-

ent banking business models may be expected to entail different

degrees of monitoring effort for the supervisor. Hence, if the alloca-

tion of banks to business models is based on the subjective assess-

ment of supervisors, under first principles these may have the

incentive to allocate banks into business models that are easier to

monitor or subject to stricter regulatory requirements (e.g. higher cap-

ital requirements). The same rationale may be applied to business

model self-reporting by banks. In this context, we argue that finding

an objective and reliable method to allocate banks into business

models is paramount for the implementation of business model spe-

cific regulation and supervision.

The use of clustering analysis to identify banking business models,

however, bears significant challenges of its own, including those

related with the choice of method. For instance, a recent strand of

research has relied exclusively on hard clustering methods

(e.g. hierarchical clustering (HC)) to identify the business models of

banks, failing to apply methods that enable banks to have some affin-

ity with more than one business model, such as fuzzy c-means (FCM).

For instance, Mergaerts and Vander Vennet (2016) applied HC on

seven business model variables for a sample of European banks

(1998–2013) and reported an average silhouette width of 0.20 for a

partition in three clusters, a value that is below the threshold of 0.25

for minimum quality of clustering as proposed by Kaufman and

Rousseeuw (1990). Similarly, Martín-Oliver et al. (2017) applied HC on

six variables for a sample of Spanish banks and reported persistency

levels of business model classification across consecutive periods

(1999–2002 versus 2003–2007) that range from 10.4% (lowest) to

85.7% (highest). In our view, both studies raise some concerns regard-

ing the usefulness and reliability of results that are obtained by apply-

ing hard classification methods to the identification of banking

business models.

Conversely, by using fuzzy clustering to identify banking business

models one may be able to measure the similarity that each bank

holds with the prototypical models (i.e. percentage of cluster member-

ship). In turn, such a measure may be used in several empirical con-

texts in business model analysis, such as: (i) the identification of

whether a bank combines features of more than one business model,

and which models those are; (ii) the use of the measure in its original

format (i.e. continuous value from 0 to 1) as an explanatory variable in

performance and riskiness-related fixed effects regression (not possi-

ble when the business model assignment is stable and discrete); and

(iii) its conversion into a discrete measure, 0 or 1, based on the busi-

ness model with which the bank has the highest percentage of mem-

bership, enabling, for instance, a supervisor to identify peer groups of

banks based on their business model.

This paper contributes to the literature in several ways. First, we

provide a formal definition of ‘banking business model’grounded on

strategic management literature, namely the configurational approach

(Miller, 1986) and strategic groups theory (DeSarbo and

Grewal, 2008; Reger and Huff, 1993). Second, by applying principal

components analysis to an array of banking variables, we identify five

strategic dimensions along which banks assume a long-term position

relative to their peers (Galbraith and Schendel, 1983). Third, based on

the notion of consensus-based classification (Kuncheva, 2004), we

identify the business models of European banks using an ensemble of

three unsupervised clustering methods: FCM (Bezdek et al., 1984),

which allows us to handle fuzzy clustering; self-organizing maps

(SOMs; Kohonen, 1997), which yield intuitive visual representations

of the clusters; and partitioning around medoids (PAM; Kaufman and

Rousseeuw, 1990), which circumvents the presence of data outliers.

Fourth, we examine the level of similarity of banks operating with the

same long-term business model (core versus non-core banks). Finally,

we provide some evidence regarding the level of persistency of banks

in terms of their business model, as well as examine the factors that

influence the likelihood of non-persistency per business model.

Briefly put, our approach begins with the implementation of prin-

cipal component analysis with the aim of identifying a set of business

model components. This step allows us to perform clustering on a

space with orthogonal dimensions, as well as to focus on the most rel-

evant relationships between business model choices and, thus, hope-

fully mitigate the problem of data noisiness. The second step is to run

three clustering methods (PAM, FCM, and SOMs), combine their clas-

sification output and assign each bank to the business model (cluster)

with the majority of the ‘votes’(clustering ensemble). Next, we label a

bank as ‘core’in a given business model if (i) the ensemble is unani-

mous (i.e. if the three methods assign the bank to the same business

model) and (ii) the silhouette width using the clustering ensemble clas-

sification is above a threshold identified in literature. Finally, we look

for persistent banks by dividing the full sample period (2005–2016)

into four trienniums (2005–2007, 2008–2010, 2011–2013 and

2014–2016), identifying the business model of banks for each trien-

nium separately (using triennium average values) and looking for

banks for which the business model is the same in all the trienniums

in which the bank is present in the sample.

By applying our method to the context of the European banking

industry (2005–2016), we find evidence of four banking business

models: retail focused, retail diversified funding, retail diversified

assets and large diversified. Importantly, we test the stability of classi-

fication using alternative sub-sampling methods and find that the sta-

bility of classification is significantly higher when testing the samples

of core banks, and core and persistent banks when compared to tests

with the full sample. Also, we find that the mean values of key dimen-

sions of each banking business model change significantly when using

the sample of core and persistent banks when compared to other

banks. These results (stability and mean difference) may be seen as

evidence of the suitability of our approach to identify banking busi-

ness models.

This paper is structured in the following way. In Sections 2 and 3

we respectively survey applications of the ‘business model’concept in

banking regulation and recent literature on methods used to identify

banking business models. A conceptual framework for banking busi-

ness models is established in Section 4 . Section 5 provides an over-

view of key concepts in the ‘clustering ensemble’approach, as well as

MARQUES AND ALVES 67