This chapter aims to assess the financial soundness of the Kazakhstan banking sector and classify the banks into different groups based on the extent of their financial soundness. In order to achieve this aim, it is necessary to select a reliable statistical technique first. That can be done by classification tools such as data envelopment analysis (DEA), UTilite´s additives DIScriminantes (UTADIS), artificial neural network (ANN), classification and regression trees (CART), k-nearest neighbors (k-NN), ordered logistic regression (OLR), multiple discriminant analysis (MDA) (e.g., Bell, 1997, Alam et al., 2000, Gaganis et al., 2006, Ioannidis et al., 2010 and Paradi et al., 2012). Previous studies which used cluster analysis noted that it works even when there is little data and the requirements for the normalcy of the distribution of random variables and for other classical methods of statistical analysis are not fulfilled. Shuai et.al (2013; p.461) demonstrated that “Cluster analysis can be applied even when no performance result is available while logistic is characterised as simple result, small burden and propounding classification performance”.
Cecchetti, Kohler & Upper (2009) found that cluster analysis groups observations into clusters by minimizing differences within clusters and maximizing differences across clusters. These authors considered the costs of 40 systemic banking crises since 1980. Wolfson (2004) stated that the clusters identification is not a quest for the least number of variables that explain a result but the common features of similar groups are. Gutierrez
and Sorensen (2006) claimed that the results of cluster analysis may provide some
insights into the underlying interlinkages between a set of variables that other econometric techniques would not be able to detect.
Many studies use cluster analysis in finance and in particular in the banking sector. Table 3.1 provides a summary of relevant prior studies in this area. It shows that some research was devoted to the clustering of bank clients and creditors (e.g., Şchiopu, 2010, Amin et al., 2009, Tudor, Bâra and Andrei, 2012, Mäenpää, 2006, Kaynak and Harcar, 2005). Other studies used cluster analysis in the risk management of banks or to predict the likelihood of their bankruptcy (e.g., Dao and Khanh, 2014, Penikas et al., 2011, Shuai at.al, 2013). Furthermore, the IMF widely employs cluster analysis to determine groups of large complex financial institutions with common characteristics (IMF, 2010).
Table 3.1: Cluster Analysis in Key Prior Studies
Reference The Purpose of the Study Methods Used Country Number of Observations/ Time Period
Set of Variables Results
1 2 3 4 5 6 7 Alarm et al. (2000) Identification of potentially failing banks. Cluster Analysis USA 248 banks, 1991
Net income to total assets, Net loan losses to adjusted assets,
Nonperforming loans to total assets, Net loan losses to total loans, Net loan losses plus provision for loan losses divided by net income.
Both the fuzzy clustering and self- organizing neural networks seek to give classification tools for identifying potentially failing banks.
Peresetsky et al. (2004) Probability of default model development . Cluster Analysis, Logit and Probit Analyses Russia 1569 banks, 1998
Total assets, Bank reserves for possible losses, Loans to non- financial institutions, Government bonds, Equity, Liquid assets, Private customers’ deposits and accounts, Capital assets and other non- working assets, Non-government securities, Assets, Profit before tax, Amounts owed to credit institutions, Non-working assets, Overdue loans.
Developed model modifications that took into account the structural non- homogeneity of the set of banks. Proved that the bank probability of default models can be used for an EWS. Safdari, Scannell and Ohanian (2005) Developmen t of an alternative methodology for peer group determinatio n. Factor and Cluster Analyses Republic of Armenia 17 banks, 2001
Total assets, Average assets, Total liabilities, Loan investments, Total capital, Time deposits of physical entities, Total time deposits of physical & legal entities, Time liabilities, Demand liabilities,
Statutory fund, Securities, Loans to economy, Interbank loans.
Found that Bank Assets, measured in Weight Share (%) is the principal variable in explaining variation among the banks sampled in the study. Established cut-off points and methodically delineated peer
groupings.
Continuation of Table 3.1 1 2 3 4 5 6 7 Dardac and Boitan (2009) Assessment of risk profile and profitability of financial institutions Cluster Analysis Romania 16 credit institutions, from 2004 to 2006
Capital and reserves to total assets, Cash holdings, Securities holdings to total assets, Loans to deposits ratio, Loans to non-financial institutions. and households to total, Operational expenses to total, Return on assets ROA, Return on equity ROE, Profit margin, computed as net profit to total income, Customers’ deposits to total liabilities
Cluster analysis proves to be valuable not only for assessing homogeneous banking groups in terms of risk profile and profitability, but also it can identify groups sharing similar features of the
financial intermediation activity, large and complex banking groups, as a potential source of systemic risk, or the degree of financial integration in the euro area banking industry. Şchiopu (2010) Identification of Bank Customers’ Profile Cluster Analysis, PCA Germany 1000 records from German banks on 9 May 2010
Duration, Credit history, Purpose, Credit amount, Years employed, Payment rate, Personal status
Identified three groups of customer profiles using Two-Step cluster analysis as skilled customers with no bad credit history; middle class customers, unemployed, but with real estate; persons with unknown properties, mostly unemployed. Penikas et al. (2011) Modelling Risk Patterns of Russian Systemically Important Financial Institutions(S IFI) Cluster Analysis, Copula Models
Russia All the Russian banks, from 2004 to 2010
75 variables Proposed approach to SIFIs’
identification classifies the banking groups in terms of marginal risk distributions, and in terms of risk distribution copula shift moments. Five distinctive bank patterns
revealed comprise two SIFIs clusters of “too risky to fail” and “too many to fail” ones.
Continuation of Table 3.1
1 2 3 4 5 6 7
Abudu (2011) Bank Failure Prediction Cluster Analysis USA 326 failed banks and 324 non- failed banks, from 1989 to 2000
Assets size, Equity to assets Proposed the cluster based
approach to bank failure prediction with improved classification
accuracy. An important implication of the approach is that different clusters have different variable subsets and variables that distinguish them from banks in other clusters.
Paradi et al. (2012) Identifying managerial groups in a large Canadian bank branch network DEA and Cluster Analysis
Canada One bank with over 1000 branches, 2004
Sales, Service, Management, Day– day banking, Borrowing,
Investments, Transactions
Proposed a new grouping approach in a DEA framework designed to identify bank branch management groups. It groups branches based on their operational similarity and eliminates the impact of efficiency levels on the identification of a branch’s true operating
characteristics. Dao and Khanh
(2014) The ability of cluster analysis to recognize vulnerable banks, common characteristi cs. Cluster and PCA Vietnam 33 banks, from 2005 to 2007
ROA, ROE, Net interest margin, Net profit margin, Equity capital to assets, Net non interest margin, Noninterest income over non interest expense, Asset utilization, Reserve ratio, Operating efficiency ratio, Total loan over total deposit, Temporary investment ratio.
Found that cluster analysis helps identify the vulnerable banks in the crisis. ROA, ROE, and Equity capital to assets ratio can be the warning indicators.
It can be seen from Table 3.1 that, in most cases, cluster analysis is used in combination with factor analysis or PCA (Safdari, Scannell and Ohanian, 2005, Şchiopu, 2010, Abudu, 2011, Dao and Khanh, 2014). For example, Safdary, Scannell and Ohanian (2005), in their study of Armenian banks, use PCA and cluster analysis to allocate 17 banks to similar groups, based on 13 accounting based indicators. Division of banks into groups is usually made to specify their position in peer groups and the calculation of peer group ratio averages. Almost all presented studies use cluster analysis to produce final results such as a recognition of vulnerable banks or an identification of potentially failing banks. Only two research studies by Abudu (2011) and Peresetsky et al. (2004) used cluster analysis as preliminary step to improve the predictive power of their models. Abudu (2011) utilised two basis of clustering by asset size and time series of failed and non-failed banks a year prior to failure. Peresetsky et al. (2004) classified banks into three clusters by giving values to a bank parameter and used expert and automatic approaches.
This study employs two statistical techniques of PCA and cluster analysis following Dao and Khanh (2014) and Safdary, Scannell and Ohanian (2005) and defines the structure of the banking sector of Kazakhstan according to the degree of a bank’s financial soundness which was to similar Ioannidis et al. (2010). Banks are divided in three groups: unsound, risky and sound. This structure can be considered as the final result that gives an indication of the levels of soundness and the stability of bank activity and provides a clear picture for supervision bodies, bank managers, depositors and other decision makers. At the same time this structure can be considered as a preliminary step for determining samples of sound and unsound banks for the construction of a model to predict unsoundness.