3.1 Research procedure
We collect bank-specific data from the Compustat database, choose nine financial indicators, and employ factor analysis and cluster analysis to analyze the data. The research procedure is shown in Figure III-1.
Figure III-1 Research procedure
Cluster analysis
To categorize the eighty-four commercial banks into strategic groups and then
examine their financial performance to find out which cluster possesses competitive advantage
Factor analysis
To uncover a smaller set of salient and uncorrelated factors is able to replace the original set of correlated financial
indicators and thus extract resource configurations
Nine indicators of the banks’
financial condition Compustat database
Eighty-four commercial banks in the five advanced economies during the years 2004-2008Nine financial indicators are based on balance sheet and income statement data
22
3.2 Data collection
The sample analyzed in this paper consists of the eighty-four commercial banks whose SIC Code is 6020 in the five advanced economies during the period 2004-2008.
The reasons are as below: first of all, the global financial crisis broke out at the end of 2007 and reached a critical stage in September 2008. Second, taking into consideration the usefulness of financial indicators, the degree of financial liberalization, and the extent of public disclosure of banks‟ financial positions, we collect data for commercial banks in the United States, the United Kingdom, Germany, France, and Japan.
All bank-specific data is year-end and is from the Compustat database. Banks are excluded if they failed, were acquired under duress, or were subject to government takeover in the late 2000s financial crisis. A bank is also excluded if its return on average assets is outlier by more than three standard deviations from the industry mean.
3.3 Methods of analysis 3.3.1 Factor analysis
Factor analysis is a statistical method exploited to describe the interdependences among a larger number of observed variables in terms of a smaller number of unobserved variables called factors. There are two types of factor analysis: exploratory factor analysis and confirmatory factor analysis.
Exploratory factor analysis seeks to uncover the underlying structure of a relatively large set of observed variables without prior theory. Conversely, confirmatory factor analysis seeks to test and verify whether the number of factors and the loadings of measured variables on them are consistent with the expectation by pre-established theory.
This paper applies exploratory factor analysis to the financial data of the banks which survived in order to find out the resource configurations causing them
23
to continue to exist after the financial tsunami. Besides, a total of nine financial indicators are adopted in this step. On the strength of factor analysis, a smaller set of salient and uncorrelated factors is able to replace the original set of correlated financial indicators. What‟s more, those banks on the factor scores of the financial indicators are utilized in the following cluster analysis.
3.3.2 Cluster analysis
Cluster analysis is a statistical technique employed to assign a set of observations into subsets called clusters so that observations in the same cluster tend to be similar to each other and dissimilar to observations in the other clusters.
Cluster analysis is most frequently employed as a classification tool. It is also used by some researchers as a means of discovering and representing the structures in similarity data through the construction of dendrograms (Hartigan, 1967) or overlapping clusters (Shepard and Arabie, 1979). Whereas classification is concerned with the identification of discrete categories, structural representation is concerned with the development of a faithful representation of relationships (Punj and Stewart, 1983).
There are two types of clustering algorithms: hierarchical clustering algorithms and nonhierarchical clustering algorithms. Hierarchical algorithms find successive clusters by previously established clusters. These algorithms can be agglomerative or divisive. Four primary hierarchical clustering procedures are available: single linkage method, complete linkage method, average linkage method, and Ward‟s minimum variance method. In addition, two variants of the average method, centroid method and median method have very undesirable properties (Aldenderfer, 1977; Sokal and Sneath, 1973) which recommend against their use (Punj and Stewart, 1983). Of the hierarchical clustering procedures,
24
average linkage method and Ward‟s minimum variance method have been shown to perform better than the other procedures.
These are more varieties of the nonhierarchical methods, though all work on similar principles. These iterative partitioning methods begin by dividing observations into some predetermined number of clusters. Observations are then reassigned to clusters until some decision rule terminates the process. These methods may differ with respect to the starting partition, the type of reassignment process, the decision rule used to terminate clustering, and the frequency with which cluster centroids are updated during the reassignment process. These methods include K-means method, hill-climbing method, and combined K-means and hill-climbing method (Punj and Stewart, 1983).
Punj and Stewart (1983) drew a conclusion from the empirical findings on the performance of clustering algorithms: iterative partitioning methods are preferable to hierarchical methods. However, this holds only when a nonrandom starting point can be specified. Moreover, iterative partitioning methods require prior specification of the number of clusters desired. Hierarchical methods require no such specification. Thus, the users are confronted with the determination of both an initial starting point and the number of clusters in order to employ the methods that have demonstrated superior performance. Information about determining starting points in the form of a priori descriptions of expected clusters may be available. In the absence of such information, a means of obtaining starting points and an estimate of the number of clusters is required. A two-stage procedure may be employed to deal with this problem.
In the first place, one of the hierarchical methods which has demonstrated superior performance, average linkage method or Ward‟s minimum variance
25
method, may be adopted to obtain a first approximation of a solution. By examining the results of this preliminary analysis, one can determine both a candidate number of clusters and a starting point for the iterative partitioning analysis. Furthermore, this preliminary analysis can be used to examine the order of clustering of various observations and the distances between individual observations and clusters. This provides an opportunity for the identification of outliers which may be eliminated from further analyses. The remaining cases may then be submitted to an iterative partitioning analysis, such as K-means method, for the refinement of the clusters.
To make the results of this cluster analysis more ideal, this paper applies two-stage clustering to the financial data of the banks which survived the financial tsunami. The nine financial indicators are also adopted in this step. To begin with, we make use of Ward‟s minimum variance method to obtain a candidate number of clusters. Secondly, K-means method is applied to those banks on the factor scores of the financial indicators to attain a meaningful and useful set of clusters. Lastly, each cluster is interpreted and given a name through the examination of the cluster centroids. This analysis enables us to divide the banks which survived the global financial crisis into strategic groups.
26