The Relationship between Financial Structure and GDP

The Relationship between Financial Structure

and GDP.

A Worldwide Perspective.

A clear lesson of history is that a 'sine qua non' for

sustained economic recovery following a financial crisis

is a thoroughgoing repair of the financial system.

Janet Yellen

But Which System?

APPLICATIONS FOR ECONOMICS, MANAGEMENT AND FINANCE

Applied Research Project and Report

In submitting this assignment:

1. We declare that this written assignment is our own work and does not include (i)

material from published sources used without proper acknowledgment or (ii) material

copied from the work of other students.

2. We declare that this assignment has not been submitted for assessment in any other

course at any university.

3. We have a photocopy and electronic version of this assignment in our possession.

Stefano Valeri 1672146 ________________________

Edoardo Filippo Ferraresi 1641836 ________________________

Georgina Young Bloom 1806694 ________________________

1 | T h e R e l a t i o n s h i p b e t w e e n F i n a n c i a l S t r u c t u r e a n d G D P

Abstract

This paper is a dissertion analysis on how different financial structures foster

economic growth and development at the aggregate world level. We identified and analyzed three different financial systems - Bank-Based, Market-Based and Government-Based -

whose differences were derived from five common factors to the worldwide financial realm: Solvency, Profitability, Market Efficiency, Foreign Presence and Core Revenue Cost Structure. The different levels of these five underlying factors characterizing each financial

structure led to our conclusion that only a more market oriented system is strictly related to the economic development in terms of contribution to the “added value” measured by the

Gross Domestic Product.

Introduction

Our interest to uncover and detect the linkage between the financial system and the

global level of economic and social development steems from the recent Great Recession that characterized the world economy. The brutal changes brougth by the current financial turmoil have affected major aspects of our everyday life. They took the stage scene in any random

conversation entering the popular knowledge while determining the day-to-day decision making process of every person. During the past years the major economies were devasted

with no less power than what would have happened in the case of epidemic deseases or war conflicts, experiencing the gradual fading of the middle class, the disruption of wealth and the loss of thousand of jobs.

Financial systems are critical to the way the world functions today. They not only help in the formation of capital but also have a responsibility to mobilize the savings in the form of

money and invest them a productive manner. According to the Indian economist Manmohan Singh, “if you don’t have a functioning financial system the world economy won’t be revived. All the major economies have their responsibility to assist at a pace which is required to clean

up the balance sheet of the banking system and to ensure that credit flows are resumed. i” Financial systems refer to a necessary set of components and mechanisms such as monetary

policies and insurance that allows economic transactions to occur. For these reasons we wanted to acquire further knowledge and insights over the

linkage between the financial sector of the economy we live in and the level of social and economic development. However, we extended our research to a more global perspective

trying to detect any possible difference between the various financial sectors and the way they may shape the level of Gross Domestic Product of the countries in which they are implemented.

According to our research findings, in line with the literature we analyzed, there are

three main types of financial systems: Bank-Based financial systems, Market-Based financial systems and Government-Based financial systems. In Bank-Based financial systems, banks play a leading role in mobilizing savings, allocating capital, overseeing the investment

decisions of corporate managers and in providing risk management vehicles. In Market-Based systems, securities markets share center stage with banks in transfering investors’ savings to

firms, exerting corporate control and easing risk managementii. Government-Based financial systems have a double role in preventing buildup of systemic risk and in creating a framework in which financial failure is managed in an orderly and cost effective manneriii. Financial

systems and GDP (growth domestic product) go hand in hand as the basic assumption is that financial systems foster growth and economic development. This research paper investigates


if and how the three main types of financial systems influence GDP during the years prior the 2008 Great Recession. We decided not to consider the data from the post-crisis years because we believed that the values of the financial ratios would have been distorted by the crisis and

because we also encountered an increase in missing values.

Various data from the World Bank Financial Development and Structure Database and form the Penn World Tables have been used to perform the questions. Factor analysis, a method to reduce or summarize many variables into one or more indexes based on the

variables' correlations, will be used to develop five latent variables/factors that characterize financial systems; solvency, profitability, market efficiency, foreign presence and revenue

cost structure. Following this, a cluster analysis, a method of identifying groups within a population based on measured characteristics in a sample, will be used to identify the three types of financial systems that affect GDP. Finally, we developed a linear regression using

“PPP GDP” - a GDP index measured based on the purchase power parity- as the dependent variable and three dummy variables without intercept (to avoid the dummy variable trap) of

the three clusters “Bank-Based financial systems, Market-Based financial systems and Government-Based financial systems” as the independent variables.. This model will demonstrate if there is any relationship between the five latent variables representing bank-

based financial systems, market-basked financial systems and government-based financial systems and the GDP level of a certain country(s). Thus, there have been three hypothesis that

have been tested using this model:

Hypothesis 1: There is a relationship between Bank-Based financial systems and the GDP

level. Hypothesis 2: There is a relationship between Market-Based financial systems and the GDP

level

Hypothesis 3: There is a relationship between Government-Based financial systems and the GDP level.

This paper has been devided into two sections. The first one presents a brief background of vast past literature on the most relevant arguments and topics treated in this analysis

concerning both the financial systems and its relation to GDP, with the aim of helping the understainding of the research question and the factors utilized. The second section,

describes the data that has been used in the analysis along with the model developed.

1. Background

Financial systems are of doubtless importance for productivity growth and economic

development. Financial structures consist of firms that provide financial services and advice to clients and their main role is to act as an intermediary between the capital market and debt

market. The ultimate goal of the economy is to design a financial system that facilitates the optimal allocation of scarce economic resources to promote economic growth that improves living standardsiv. Our research identifying three main types of financial systems, bank-based,

market-based and government-based, produced similar results to the three types of financial systems stated in “Market-Based Banking and the International Financial Crises” by Hardie

and Howarth (H&H)v.

With regard to our dependent variable, Growth Domestic Product (GDP) is one of the

most widely used measures of an economy’s output or production. It is the monetary value of all the finished goods and services produced within a country’s borders in a specific period.


GDP includes all of private and public consumption, government expenditures, investments and exports less imports that occur within a specific country or region. GDP plays a key role in today’s financial market since it is commonly used as an indicator of the economic health

of a country and as a way to assess a country’s standard of living. GDP enables policymakers and central banks to judge whether the economy is contracting or expanding or on the verge

of a recession or inflation. Because of GDP’s essential and wide role, we chose to use it as the dependent variable to be regressed on the different types of financial systems. As countries rise along the scale of wealth and income, their financial structures are usually becoming

increasingly rich in financial institutions, assets and markets.

Solvency, profitability, market efficiency, foreign presence and the core revenue cost structure all characterize financial systems in different and unique ways. Profitability is the primary goal of all business ventures and without profitability a business would not survive in

the long run since Profitability is a condition yielding a financial gain. An Efficient Market is one where the market price is an unbiased estimate of the true value of the investment. It

requires that errors in the market price be unbiased (i.e. the prices can be greater than or less than the true value as long as the deviations are random)vi. Foreign Presence refers to foreign direct investment, controlling ownership of an asset in one country by an entity based in

another country. Core revenue cost structure relates to cost structure and revenue streams. Core cost structure is the expenses that a firm must take into account when manufacturing a

product or service such as transaction costs, sunk costs and marginal costs. Core revenue streams, on the other hand, is a method that a company, organization or individual uses to collect money from users of their product or service. Solvency represents a necessary

operating condition of the financial system and refers to a company’s long run financial viability and ability to cover long-term obligationsvii. If a company is unable to meet its obligations, it is said to be insolvent and must undergo bankruptcy to either liquidate or

restructure. According to Stiglitz in “The Design of Financial Systems for the Newly Emerging Democracies of Eastern Europe”, solvency is strictly related to the financial

system. Indeed, the government based financial systems provide insurance to depositors to restore confidence in banks and to prevent bank runsviii.

Ross Levine in her article, “More on Finance and Growth: More Finance, More Growth?” explicitly declares that while the theory suggests that financial systems influence

growth by easing information and transaction costs and thereby improving the allocation of capital, corporate governance, risk management and financial exchanges, the empirical measures do not directly measure these financial functionsix. On the other hand, Robert King’s

“Finance, Entrepreneurship and Growth” research paper states that better financial systems improve the probability of successful innovation and thereby accelerate economic growthx.

The purpose of this paper is to solve this controversy and examine how the financial systems affect economic growth, development and GDP, while analyzing the unique characterizing factors of a financial system.

2. Data Description and Initial Data Screening

The data considered are available on the World Bank Financial Development and

Structure Database as well as on the Penn World Tables. From the latter, we considered only the GDP adjusted for the Purchasing Power Parity of each country. Using the World Bank Financial Development and Structure Database, we extracted all the indicators of financial

development and structure across countries and over time that measured the size, the activity, and the efficiency of financial intermediaries and markets. We then carried out an extensive


initial data screening in order to address all the relevant issues with these measures (suitability, inclusion, descriptive statics, correlations), an activity described in the following paragraphs.

Firstly, we divided the data from the World Bank database into one smaller subgroup,

pre-financial crises (before 2008) with a range of five years. We chose these years because financial crises’ have had significant impact on the historical development of financial systems. We then calculated the means of each variable for each country (our cases) for our

sub-period. We did this in order to disentangle the chaos that was brought about by the financial crisis, especially to our main observations and variables (countries and financial

ratios) since they were most likely the ones that suffered and changed.

The main issues encountered were related to the basic structure of our analysis and the

data available, as we had to develop a model based on both factor analysis and cluster analysis. The issues arose due to conceptual issues rather than statistical considerations. As a

consequence, the sample size was of primary importance to our research because of a trade-off between sample size representativeness and number and heterogeneity of variables considered, i.e. the risk of overfitting the data (with regards to factor analysis) and of

describing the random error in the variables instead of the underlying relationship.

Regardless, during the beginning stages, some variables (ofagdpp, bdgdp, prbond, pubond,

intledebt, intledebt net) were completely eliminated before progressing with the analysis because of an excessive number of missing values in the range of 70% to 90% and their superficiality. These variables were more meaningful but also similar indicators. Following

this, we removed all the countries that had too many empty variables (more than 10) and that were meaningless for our analysis, ending up with approximately 153 cases of countries.

At this point, we believed that 30 were too many for two main reasons. Firstly,

conceptually some of them were really close in their meanings and a problem of multicollinearity could easily arise. Secondly, a higher case-to-variable would be desirable

with 153 cases. With this in mind, we decided which variables were to be eliminated based on our

knowledge about their intrinsic meaning and on our research on the topic. For example,

Intldebtnet (loans from non-resident banks (net) to GDP%) is the same as nrbloam (loans from non-resident banks (amount outstanding to GDP%). We retained the outstanding as

more representative of the influx of capital from non-resident financial institutions. Other variables were disregarded because they did not address our specific research question, such as inslife and insnonlife, which provided the amount of life insurance premium volume to

GDP (%). Other indicator variables were encompassed into more general, and superior, indicators like bdgdp (demand, time and saving deposits in deposit money banks as a share of

GDP) into fdgdp (financial system deposits to GDP %). We then considered the correlation matrix for the remaining variables and check the

assumptions of linearity, homoscedasticity and possibly normality, and their descriptive

statistics. Because these variables are ratios, they ranged over different areas on the number line and had different variances and distributions, even though the majority of variables were

slightly positively skewed. We decided to tackle the problem straight away instead of anticipating its arrival during cluster analysis, our second step in the analysis procedure. We converted each variable to standard scores (Z-scores) by subtracting the mean and dividing by

the standard deviation for each variable. All standardized variables retained their properties with a mean of zero and a standard deviation of one. Also, thanks to the Central Limit

theorem we could worry less about violations of variables’ normality because this principle


states that as the sample size increases over 30 cases a variable’s distribution can be well approximated by a normal curve. Moreover, low correlations were definitely not a problem.

Next, the missing values were replaced with the variables means. Even though we were fully aware of the problem that this choice might create (i.e. depressing the standard

error and the correlation with the other variables), we believed that this solution was the most appropriate considering our sample data. Since we already deleted various ratios, we tried to retain as many core values as possible. Fortunately, we were able to confirm that this decision

did not have any significant impact on the variable correlations due to the subsequent tests for multivariate adequacy. .

The diagnostic tests that we performed in order to check for multivariate adequacy demonstrated interesting values in all parts one must consider when approaching a factor analysis. During the first stage of design and data screening, both Barlett’s Test of Sphericity

and Kaiser-Meyer Olkin (KMO) Measure of Sampling Adequecy (MSA) produced satisfactory results showing that there is enough correlation in the data set to form factors,

meaning the indicator variables could be reduced to latent variables.

The idea behind Barlett’s Test of Sphericity is to check that the correlation matrix is

statistically significant and different from an identity matrix, whereas KMO MSA is useful in determining if there is satisfactory correlation in the data. This measure shows the proportion of variance in the dataset that is common (initial estimate) with respect to the partial

correlation among the variables thus providing a logical basis for the creation of the factors if the value is over and above .7.

We then examined the determinant of the correlation matrix to look for the absence of multicollinearity. Before analyzing the results, we knew this would be an issue because of the analysis rational of the research. All of our indicator ratios were taken from financial

institutions that were well interlinked and therefore, we expected that some variables would follow similar patterns across the countries analyzed.

Indeed, as seen above, the determinant was different then zero, showing no perfect multicollinearity existed. Rather, the low correlation matrix determinant revealed that a high correlation between the variables existed. To obtain a factor analysis solution, the correlation

matrix must be invertible and this could only be achieved if the correlation matrix was not singular with a determinant of zero originating from multicollinear variables. As a result, both

diagnostic tests for multivariate adequacy suggested that the correlation matrix was factorable having P-value barlett’s equal to zero (probability of obtaining the observed sample results (or a more extreme result) when the null hypothesis is actually true), KMO MVA higher than 0.7

and a determinant close but different from zero. The Measure of Sampling Adequacy (MSA) is a measure of prediction of each

variable by the others and is obtained from the anti-image matrix. It quantifies the degree of

KMO and Bartlett's Test

Kaiser-Meyer-Olkin Measure of Sampling Adequacy. .755

Bartlett's Test of Sphericity

Approx. Chi-Square 1459.336

df 231

Sig. .000

Table 1

Correlation Matrixa

a. Determinant = 5,03E-8

Table 2

http://en.wikipedia.org/wiki/Null_hypothesis


inter-correlation among the variables in the dataset. MSA can be used to see whether any variable must be removed due to its conceptual non-correspondence with other variables. We produced the anti-image correlation matrix and only considered variables with a value higher

than 0.50, meaning that at least 50% of the variance was explained by other indicator ratios. For example, bank concentration (assets of three largest banks as a share of assets of all

commercial banks) presented a value close to 0.5 (0.404 from one of the previous trials) but lower which we believe was unacceptable and as a consequence that indicator was removed.

Once the factors were initially extracted, their Communalities (the proportion of

common variance explained by the extracted factor model) with the variable kept in were all well above the threshold level of .40 except for one which, although acceptable, was dropped from the analysis to make the factor interpretation easier and clearer and about which we

write in later paragraphs.

Concerning cluster analysis, fewer problems were encountered especially by means of fewer statistical data requirements. Considering our final goal, we believed that a sample composed of 153 countries was sufficiently large to uncover the theoretical typologies in the

data, i.e. three clusters well diversified representing market, bank and governmental conditions. Having already discarded irrelevant variables in the pre-stages of factor analysis,

we focused on identifying outliers but due to their importance in our research project, we prefer to explain our decisions in the multivariate analysis part.

Finally, we began our multivariate analysis with the following indicator variables:

3. Multivariate Analysis

Our final purpose is to test the hypothesis that different financial structures do, or do not, differently affect a country’s development.

Our analysis is divided into three intercorrelated steps: first, we begin with a factor analysis in order to reduce a number of highly correlated variables into few significant factors representing the main characteristics of a financial structure; second, we employ a cluster

analysis in order to group the countries considered into different financial structures that show relevant similarities according to the latent variables; finanlly, with a regression model

(although insignificant per se) we see which model actually differs and has a statistically impact via a linear equation formed only by dummy variables and no intercept (to avoid the so called dummy variable trap) as follow:


𝑌 = 𝛽1(𝑋) + 𝛽2(𝑊) + 𝛽3(𝑍)

in which:

X represents a MARKET-BASED FINANCIAL STRUCTURE

W represents a GOVERNMENT-BASED FINANCIAL STRUCTURE Z represents a BANK-BASED FINANCIAL STRUCTURE

Y represents the Total Real Gross Domestic Product, current price milions

3.1 Factor Analysis

After the initial data screening, we argued in favor of a principal component analysis (PCA)

method of factor extraction given our research question and the type of data we considered. PCA considers the total variance and derives factors that contain small proportions of unique

variance or error variance. For pure data reduction purposes, this method is the most widely used in the industry and it is also in line with our factor analysis objective of creating factors that take into account as much of the variance of the variables as possible by not partitioning

the variance into common and unique variance before extraction of factors. In addition, provided that our indicators are ratios of financial measures taken from official reports, we

can fairly assume that there is no error in the measurements. In fact, with PCA, we were able to extract from a set of possibly correlated variables, a set of uncorrelated (before oblique rotation) principal components i.e. factors. We began our analysis by forming factors with the

remaining indicator variables and produced the following problematic communalities outputs. Communality can be defined as the variance each variable shares with other variables in the

data set.

Because we were already considering several different indicators for market efficiency

and profitability, we decide to remove these two indicators from our project. Following this, we re-ran the factoring procedure and evaluated if all criterions were met. As suspected, the MSA from the ANTI-IMAGE MATRIX spotted a very low value (0.399) for

bcbd (bank credit to bank deposit); this measure was highly correlated with zscore and the latter’s exclusion was the driver for the former’s elimination.

To determine the number of factors, we opted for a percentage of variance criterions to choose the numbers of factors to extract and assessed 80% as sufficient. We were confident that with 15 variables, a scree plot criterion (one that plots the eigenvalues, or variance

explained by each factor, against every factor) is difficult to interpret and to decide which factor’s eigenvalues are different from the flat portion of the line, whereas on the other hand,

a latent root criterion (one that retains only factors whose eigenvalues are higher than one) risks of being too restrictive and of extracting too few factors.

Communalities

Initial Extraction

BANK Z-SCORE 1.000 .586

NO. OF LISTED COMPANIES PER 10K POPULATION 1.000 .424

Table 3


Table 4 By chance, latent root and percentage of variance suggested both five factors.

Even though for subsequent analyses it would be more fitting to have independent factors, especially for cluster analysis, we used an oblique rotation to provide more readily

interpretations of factors solutions. Oblique rotation methods are best suited to the goal of obtaining several theoretical meaningful factors or constructs because realistically few

constructs in the real world are uncorrelated. We expected the factors found to be correlated. The solution that was initially extracted was the one that maximized the amount of variance explained by the first factor, and subsequent factors are based only on the residual

unexplained variance. Thus, by creating mathematically equivalent but more interpretable solutions in a multivariate space via a simple rotation of axis, rotated solutions assisted our

research to give factors more appropriate labels. As these labels were fundamentals for us to identify and name target clusters, we prefer an oblique rotation that does not retain the orthogonality of the initial extracted factors but that (1) provides us with a more meaningful

factor structure (low loadings on almost all but few variables for each factor), that (2) is more representative of true concepts as we expect these theoretically meaningful factors to be

correlated because they will explain related ideas and that (3) summarizes the data better. The following solution was acceptable except for stock market capitalization to GDP

variable loadings, which were high on two factors. We also tried different rotations, such as

orthogonal varimax and equimax, but both loadings remained high.

Table 5

We finally obtained our final 5-factor solution (with total variance explained now equal to

87.080%) and name the factors as follows based on the variables that loads high on each.

Total Variance Explained

Component

Initial Eigenvalues Extraction Sums of Squared Loadings

Total % of

Variance

Cumulative

% Total

% of

Variance

Cumulative

%

1 4.896 37.663 37.663 4.896 37.663 37.663

2 2.168 16.678 54.341 2.168 16.678 54.341

3 1.661 12.775 67.117 1.661 12.775 67.117

4 1.186 9.126 76.243 1.186 9.126 76.243

5 1.142 8.783 85.026 1.142 8.783 85.026

6 .550 4.227 89.253

7 .460 3.538 92.791

8 .306 2.353 95.144

9 .247 1.899 97.042

10 .160 1.234 98.277

11 .114 .873 99.150

12 .081 .627 99.777

13 .029 .223 100.000

Pattern Matrixa Component

1 2 3 4 5

Zscore(fdgdp) .987

Zscore(llgdp) .957

Zscore(dbagdp) .678

STOCK MARKET

CAPITALIZATION

to GDP (%)

.562 .522

Zscore(roe) .924

Zscore(roa) .876

Zscore(stvaltraded) .941

Zscore(stturnover) .910

Zscore(overhead) .824

Zscore(costinc) -.555 .765

Zscore(netintmargin) -.238 .410 .659

Zscore(offdep) .905

Zscore(nrbloan) .211 .801


Core Revenue-Cost structure

Cronbach's Alpha N of Items

,762 3

Table 6

We formed these factors scores instead of computing summated scales primarily

because of the factor loadings distribution. As the concepts on which the factors are created

are interrelated, we could see from the table that not only the variables with high loading have

an influence on the factors. Although generally this is something that should be avoided, i.e

meaning contamination, it is normal for example that a variable such as net interest margin,

which loads primarily on core revenue-cost structure, loads highly and positively on

profitability which is increasingly related to the management of financial institutions. By

considering the relative contribution of each observed variable to each factor and computing

the regression factor scores, we were able to maximize factor scores validity obtaining the

highest possible correlation of the latter with the intended factors and to account for more

variance in the data than with other methods. Doing this, we were also allowed to retain just 2

variables for each factor.

To confirm we were aiming at the correct targets, we obtain cronbach’s alpha measures for each factor. This is a correlation-based measure of internal consistency among

the variables used to create the new factor. We believe that all suggest good validity for the factor scores and, given the high within variables correlations, high reliability.

Tables 7

Market Inefficiency


,881 2

Solidity Solvency


,932 3

Foreign Presence


,652 2

Profitability


,838 2


3.2 Cluster Analysis

With our latent variables identified, we proceeded to the second step of our analysis

and tried to find the clusters specified in our research of the literature. With cluster analysis, we tried to discover sub-groups in a set of cases based on multiple characteristics (factors), i.e. polythetic analysis. We will try to maximize the distances (dissimilarities) between

groups, while at the same time minimize them within each cluster. As our research question is complicated we think that both data reduction and taxonomy/typology description are the

reasons why we deploy it. On the one hand, we face a huge number of cases that are useless unless classified into manageable groups; on the other hand, given our previous literature research we would like to find empirically that a 3-group financial structure is indeed present

in our data. While our sample is large enough to assure us that we could find all the relevant clusters, we do also take into account that with such a large sample data it is difficult to

decide what is a good representation of our hypothetical clusters with a simple hierarchical method of agglomeration. Thus, we decide to deploy a new method (two step) that is able to deal with large sample sizes by first clustering cases into pre-clusters and then groups those

pre-clusters hierarchically obtaining meaningful results ia a tree. Also, this choice allows us to choose a dissimilarity measure (log-likelihood) for

clustering purposes that takes into account outliers and reduce their impact on the final solution (Assume outliers or noises follow a uniform distribution. Calculate both the log-likelihood resulting from assigning a record to a noise cluster and that resulting from

assigning it to the closest non-noise cluster. The record is then assigned to the cluster which leads to the larger log-likelihood). This is fundamental to us as we explain hereby; in fact we

decide not to discard outliers but to keep them in our analysis.

Cluster analysis is extremely sensitive to outliers (cases different from the others) and

irrelevant variables. We have already dealt with the latters extensively during our FA, whereas the formers may still have a negative impact on our analysis. Thus, we necessary

check for their presence and deal with them consistently with our research question and literature findings. Because we look for outliers in a multivariate sense (different from other cases across several variables), we calculate the Mahalanobis’Distance (D^2) which estimates

how far each case is from the center of all the variables’distribution, i.e. Centroid, and test whether each case is part of the multivariate distribution. If not (PrD^2<0.05), we label it as

an outlier. Specifically, they are: Bahamas, Cyprus, Ghana, Hong Kong SAR, Luxembourg, Pakistan, Paraguay, Samoa, Saudi Arabia, Switzerland, Syrian Arab Republic. Even though some of them are not fundamental for the analysis, we strongly believe that removing such

countries as HK or Saudi Arabia or even Switzerland would bias our groups because those are probably one of the main countries financially speaking of their respective areas of influence

(Orient, Far east and Europe respectively). As a consequence, we decide to keep them in the analysis at first and check their impact by comparing the results without them later. Note that this is in line with our purpose of clustering the cases according to (factors)

characteristics to obtain only the main financial structures present around the world, which would be impossible if we chose not to include some of the most influential countries.

Nevertheless we check which of the two solutions (with and without outliers) is the most appropriate to see whether the expected structures are actually present in our sample and conclude that, indeed, (1) the cluster memberships are the same (being based on similar

values for each factor) and (2) they produce the same results regardless of their inclusion with respect to our final step, a dummy regression model. Thus, by keeping multivariate outliers in


the sample we think we are able to make our research better organized and our clusters more meaningful and representative, given similar results.

We have already dealt with the problem of different units of measure across different variables: we standardized all the variables (mean=0 and std dev=1), obtaining factor scores

with equal properties, in order to obtain results that do not depend more on characteristics with larger measurement scales which would bias the results.

Thanks to the creation of factors prior to the cluster analysis we can be fairly sure that no multicollinearity problem would arise acting as an implicit weighting of variables and

distorting the results. True, our factors are correlated due to the oblique rotation applied in the previous step but their correlations are no higher than 0.2 meaning no multicollinearity problems.

The followings are the results from a two-step clustering procedure that gives a 3

group solution, that we then checked with a k-means analysis that showed a similar 3 group solution as well. The k-means is also the best method to be used when seeking a cluster solution with no outliers. It is a procedure that starts from a set of values that act as initial

seeds for the cluster centroid (mean scores on all factors used in CA) and reiteratively reassigns cases to clusters according to the evolution of the centroid value (as it changes every

time a new case is assigned to a cluster) until no further improvement is reachable. (Tables 8)

Tables 8

Number of Cases in each Cluster

Cluster 1 58.000

2 22.000

3 62.000

Valid 142.000

Missing .000

Final Cluster Centers

Cluster

1 2 3

Solidity_Solvency -.58968 1.02703 -.00127

Profitability .21766 -.61056 -.02586

Mkt_Inefficiency .39713 -1.52832 .30602

Foreign_Presence -.10627 -.18933 -.11546

Core Revenue_Cost_Structure .87806 -.67023 -.61822


When we compare their composition the two solutions seem fairly similar, although some minor differences arise.

For example, cluster number 2 from k-means and 1 from two-step can be clearly

named Market-Based financial systems due to their low (high) level of market inefficiency (efficiency), their negative general profitability due to competitive pressure and their negative

core revenue-structure because of their well known dependence on exotic operations in the financial markets.

Reasoning similarly, cluster 3 from two-step and cluster 1 form k-means can be named

Bank-Based financial systems; both show similar structure across factors with positive profitability, positive revenue-cost structure and insolvency. Considering the low level of

market efficiency (high inefficiency), this result makes sense: exploiting the inefficiency and leveraging highly their structures, the financial institutions in these systems are able to earn positive returns.

Finally, the Government-Based financial system is as expected: negative results across all indicators meaning that these systems are driven not with a profit oriented management

style but rather with one directed to the community sacrificing solidity and profitability. In the end, the sample size are similar between the two with- and without-outliers methods. We then create 2 clustering variables, one for each method.

It is interesting to notice which factors actually shaped the clusters more, namely Market inefficiency, Core revenue cost structure and Solidity (from two step analysis). This is in line

with our expectations of different structures: every cluster (type of structure) differs under one of these aspects and positively lead us to our final step of checking for actual differences among them.

3.3 Dummy Regression

In our final part, we test whether the aforementioned groups have any impact on a country’s development. We find similar results both with and without outliers: only the market-based system actually has an impact on the GDP of each country, whereas the other

two systems do not show statistically significant positive effects.

To do this, we transform the clustering variable into 3 dummy variables and regress GDP onto these 3 requiring spss not to include an intercept so that we both avoid the dummy trap (for n dummies, only n-1 dummies are needed because the nth will be the intercept) and

show each dummy’s slope more easily. Dummies are variables that take the value of 1 when the country is part of the dummy’s group and 0 otherwise. As a last note, it is to say that we

transformed our dependent variable using a log transformation in order to avoid any bias in our estimations due to its non normality of errors. Tables 9

Consequently, in this analysis it is not important to look at the R^2 because we are not

interested in the actual predictive power of our dummies, but only on their slopes and significance. In fact, if a dummy slope’s p-value is lower than 0.05 we can be fairly certain (at a level of 95%) that also in the true population that system will influence the development.

From both analysis (two clustering variables), the results are very similar. Only the Market-Based system has an impact on the level of GDP of a country.

Coefficientsa,b

K-MEANS MODEL

Unstandardized Coefficients

Standardized

Coefficients

t Sig. B Std. Error Beta

1 MARKET_BASED 1793300.958 263062.736 .498 6.817 .000

GOVERNMENT_BASED 166762.391 156702.104 .078 1.064 .289

BANK_BASED 139424.726 162015.542 .063 .861 .391

Coefficientsa,b

TWO-STEP MODEL

Unstandardized Coefficients

Standardized

Coefficients

t Sig. B Std. Error Beta

1 MARKET_BASED 1602957.644 247549.203 .465 6.475 .000

BANK_BASED 136873.341 156563.863 .063 .874 .383

GOVERNMENT_BASED 185966.694 145996.555 .091 1.274 .205


Conclusion

Starting from our sample of characterizing financial ratios of the all countires of the world, thanks to the Factor Analysis, we were able to extract five factors that shape the

financial structure: solvency, profitability, market efficiency, foreign presence and revenue cost structure.

These factors are common to all financial systems and by executing a Cluster Analysis

we were able to group countries with similar factor’s values in three different financial structures: Bank-Based financial systems, Market-Based financial systems and Government-

Based financial structure. After creating clustering dummy variables we regressed the three financial structures on the dependet variable GDP finding that only the Market-Based financial systems is

statistically significant in explaining the level of growth and the economic development. Unsurprisingly and definetely in line with the recent development of the world

economy during the last decade, we conclude that a more market efficient economy contributes greately in terms of “added value” to the global wealth and the social economic development. Unfortunately the contribution is symmetrical, which implies it could be either

in positive or negative terms, as the striking example of the 2008 Great Recession. Started overseas in what it is considered to the perfect Market-Based economy, the recession spread

rapifly in the Old Continent before affecting and destroying the world economic as a whole. Furthermore, our conclusion that the GDP is linked to a Market-Based financial

system perfectly reflect the consolidated Optimal Capital Allocation Principle that ultimately

links with the growth of wealth, and that applies in economies that follow the Efficient Market Theory.

Other References

Data Sources: Penn World Tables, Version 7.1

World Bank Financial Development and Structure Database

Text Books: Carlson, Newbold, Thorne: Statistics for Business and Economics

Hair, Multivariate data analysis

Norusis; SPSS Statistics 17.0, Statistical Procedures Companion

Web links references: i http://www.brainyquote.com/quotes/keywords/financial_system.html ii http://www.nber.org/papers/w9138.pdf iii http://fic.wharton.upenn.edu/fic/papers/01/0115.pdf ivhttps://books.google.it/books?id=V3RpAgAAQBAJ&pg=PA60&lpg=PA60&dq=types+of+national+financial+

systems&source=bl&ots=TOL3Hfy8DG&sig=9fPS1Z6WAcSY2x4K0vl0onfhOyw&hl=it&sa=X&ei=SRqLVM

jXDousPfrcgRg&ved=0CF8Q6AEwBw#v=onepage&q=types%20of%20national%20financial%20systems&f=f

alse vhttps://books.google.it/books?id=V3RpAgAAQBAJ&pg=PA60&lpg=PA60&dq=types+of+national+financial+

systems&source=bl&ots=TOL3Hfy8DG&sig=9fPS1Z6WAcSY2x4K0vl0onfhOyw&hl=it&sa=X&ei=SRqLVM

jXDousPfrcgRg&ved=0CF8Q6AEwBw#v=onepage&q=types%20of%20national%20financial%20systems&f=f

alse vi http://pages.stern.nyu.edu/~adamodar/New_Home_Page/invemgmt/effdefn.htm vii http://my.liuc.it/MatSup/2007/F84382/Lecture%209-Capital%20Structure%20&%20Solvency.pdf viii https://www0.gsb.columbia.edu/faculty/jstiglitz/download/papers/1992_Design_of_Financial_Systems.pdf ix http://research.stlouisfed.org/publications/review/03/07/Levine.pdf x http://www.sciencedirect.com/science/article/pii/030439329390028E

Documents

The Relationship between Financial Structure and GDP