1
FINANCIAL REGULATION VERSUS SUPERVISION: EFFECTS
ON INDIVIDUAL BANKING RISK
University of Groningen Faculty of Economics and Business
Thesis for MSc Economic Development & Globalization
Student: Erik Baarda
Student ID number: 2782502 Email: e.r.baarda@student.rug.nl Date: 16-06-20
Supervisor: Dr. A.C. Steiner
2
Abstract:
This paper analyzes financial regulation and supervision, and their effect on an individual bank’s risk. Literature suggests that regulation and supervision may have ambiguous effects on risk. With the use of the latest BankFocus data, this research quantifies the 2019 World Bank Bank Regulation and Supervision survey with the formulas from Barth et al. (2013). The models use observations from 95 countries from 2011 to 2017. Across fixed-effects panel data models, the findings suggest an opposite effect between regulation and supervision on risk between 2011 and 2017. In the preferred model, regulation positively correlates to equity relative to total assets, while supervision has a negative correlation, for banks within high-income countries. Models that disaggregate regulation and supervision suggest heterogeneous effects on risk within these categories. Accordingly, the policy recommendations underline a likelihood of false monitoring and disproportionate development-rate within the supervisory framework relative to regulatory reform.
Key words:
3
Table of Contents
1 Introduction ... 4
2 Literature review ... 7
3 Data and method ... 10
3.1 Data ... 10 3.1.1 Dependent variable ... 10 3.1.2 Explanatory variables ... 12 3.1.3 Control variables ... 16 3.2 Method ... 17 4 Baseline Results ... 20
5 Further Detailed Analysis ... 23
5.1 Second Model: Four Disaggregated Measures ... 23
5.2 Third Model: Disaggregated Regulation and Supervision ... 24
5.3 Fourth Model: Two Data Points ... 26
5.4 Fifth and Sixth Model: Low-income versus High-income ... 27
5.5 Seventh Model: US Banks ... 29
5.6 Eighth Model: Dynamic Effects ... 30
6 Conclusions ... 34
References ... 38
Appendix A Quantification of Survey ... 45
Appendix B Descriptive Statistics ... 51
Appendix C Additional Models ... 68
4
1. Introduction
The global financial sector has gained increased scrutiny since the 2008 Financial Crisis. This period started with the sub-prime crisis, continued through the Lehman-failure, and extended to the European Debt Crisis. The scrutiny originates from a change in the perception of the financial sector. Namely, the crisis highlighted the excessive profits, risk-taking, and misconduct within the sector. Accordingly, some economists argue that the risk-taking is disproportionate to the financial sector’s GDP-contribution (Haldane and Madouros, 2011; Vives, 2016: 75). Numerous countermeasures followed: increased capital requirements, bank structure reform, and remuneration control (Vives, 2016: 1). Thus, questioning the regulatory framework resulted in extensive reform.
Arguably, this reform has distortionary potential. For example, entities now deemed too big to fail (TBTF) experience significantly lowered capital costs. Therefore, these reforms call for an analysis of their effectiveness and prudence. Reforms may have been sweeping, short-term crisis measures with unintended side effects. Moreover, reform may affect banks heterogeneously due to individual characteristics or country-aggregate fixed effects. All-encompassing regulation intended to lower banking risk faces challenges such as regulatory prudency, supervisory transparency, and side-effect estimation (Borio, 2004). This paper focuses on one objective of financial regulation and supervision: ensuring financial stability by reduction of banking risk (Quinten and Taylor, 2003: 6). Klomp and de Haan (2012) argue that there is no general definition of banking risk (3208). As its definition, this paper uses “banking risk” as that which decreases a bank’s financial soundness according to the IMF’s core set of financial soundness (CAMEL) indicators (IMF, 2019). Henceforth, “regulation and supervision” refers to financial regulation and supervision within this definition focused on banks. A disaggregated view on financial regulation and supervision shows which dimensions have been effective in decreasing risk.
5 to the economy. These phenomena coincide with the unique role of banks within the economy, summarized in Table 1. In contrast to other industries, banks are exceptionally fragile through contagion and systemic risk mechanisms, which makes bank failure socially costly (Vives, 2016: 5). Banks serve to provide payment services, risk-sharing from turning illiquid assets into liquid liabilities, and financing initiatives that otherwise suffer from asymmetric information, moral hazard, and adverse selection (Vives, 2016: 37). Table 1, therefore, illustrates that banks are unlike other industries as intermediators. Moreover, the fragility and inherent risk of the banking sector validates analysis of regulatory and supervisory effectiveness.
Table 1. Functions of a bank and its consequences.
Functions of a bank
Intermediation across:
1. Maturity: short-term funding versus long-term loans 2. Liquidity: liquid funding versus illiquid loans
3. Size: many small depositors versus large investment projects
4. Risks: (un)bundling of financial risks (risk allocation through diversification)
Causes of risk
5. Large weight of short-term debt, which is widely dispersed and held by small investors 6. Increased risk of failure that the high amount of short-term debt entails
7. Limited institutional ability and lack of incentives of dispersed investors to monitor bank activity; moral hazard problem that is compounded by opacity and long maturity of bank assets
8. Intermediation requires high leverage, with maturity and liquidity mismatches 9. Fragility with high social cost of failure and potential systemic impact
Source: Berk, 2020; Holthausen and Rønde, 2003; Morrison and White, 2009; Vives, 2016: 44.
Table 1 further connects banking’s unique role with risk. Many works highlight the deterioration of economic fundamentals, a solvency issue, as a source for many financial crises (Calomiris and Mason, 2003). Bryant (1980) and Diamond and Dybvig (1983) show that banking crises work in a self-fulfilling manner, with a confident equilibrium and a panic equilibrium. Illiquidity and insolvency threaten banks inherently and continuously. Over 100 crises have occurred since the 1970s, which at least partially reflect on regulation and supervision (Barth, Caprio Jr., and Levine, 2013: 5). Moreover, the 2008 Financial Crisis serves to expose contemporary regulatory shortfalls.
6 regulation and supervision is, therefore, clear. However, there are issues with the attainment of a global regulation and supervision overview. Within countries, the measurement of the intended effects requires assembling hundreds of laws and regulations. Between countries, the diversity in laws and regulations create an opaque abundance of incomparable frameworks. The capture of the variance in intra and inter-country regulation and supervision within meaningful statistics is challenging. Barth et al. in the World Bank Bank Regulation and Supervision Surveys forms a unique solution (2001; 2004; 2008; 2013; 2019).
This paper aims to evaluate whether more stringent country-aggregate regulation and supervision decrease bank risk. Central to this approach is a move beyond aggregate-level analysis of regulation and supervision with its relation to banking risk. Demirgüç-Kunt and Detragiache (2011), among other works, find an inconclusive relationship between aggregate banking risk and regulation. Specifically, using compliance with the Basel-framework’s core principles and its relation to country-aggregate bank Z-scores, the authors find no significant relationship. Bank Z-scores compare capitalization and returns with the volatility of those returns. Moreover, Barth et al. (2013) find near-unanimous conformity to capital requirements for otherwise heterogeneous countries. Anginer, Bertay, Cull, Demirgüç-Kunt, and Mare (2019) further suggest an increasing disparity between regulation stringency and supervisory capabilities. These works suggest that aggregate risk, regulation, and supervision measures may disregard effects within each variable.
Pasiouras, Gaganis, and Zopounidis (2006) serve as a template for inclusion of disaggregated regulation and supervision into seven categories. Moreover, Pasiouras et al. (2006) and Barth et al. (2013) serve as a template for the quantification of data from Barth et al.’s surveys (2013; 2019). These categories are activity restrictions, capital regulation, supervisory control, deposit insurance, private sector monitoring, liquidity regulation, and entry restrictions (Klomp and de Haan, 2012; Pasiouras et al., 2006). As another contribution, this paper uses Barth et al.’s 2013-quantification technique to gain comparable indices from the 2019-survey. The 2019-survey has not yet been used by literature that examines the relationship between regulation and risk, as this paper verified.
As illustrated, the financial sector contributes to systemic risk and has a unique position as an industry. The need for prudent regulation is, therefore, paramount. However, this regulation requires a fundamental understanding of what determines risk within the financial sector. The traditional hypothesis is that more stringent regulation and supervision leads to lower risk-taking. However, these regulations and supervision may also lead to profit-seeking in other bank activities that may be less regulated or inherently riskier. The literature review illustrates the ambiguous effects of regulation and supervision. Therefore, this paper aims to challenge this approach of only bank Z-scores and capital requirements. Namely, the models use four risk indicators including the return on average assets (ROAA) on which the bank Z-scores are based. Moreover, this paper includes a maximum of seven categories of regulation and supervision, including capital requirements. Accordingly, this approach analyzes whether the four risk types respond heterogeneously to the types of regulation and supervision.
7 three supervision categories serve to answer this question. Incorporation of the Barth et al. survey (2019) contributes insight into the latest trends within the BankFocus dataset. Numerous robustness checks verify the assumptions and method. The conclusion draws from two main findings. Within the preferred models, regulation has an opposite effect to supervision on risk. While replicated in other models, especially equity relative to total assets in high-income countries suggests this finding. Moreover, the models of disaggregated regulation and supervision find heterogeneous effects within each category on risk.
Following this introduction, the literature review section outlines the theoretical basis for this paper’s approach. The subsequent data and method section outlines the data collection process, the variables and their justification, and several tests for the correct model selection. Afterward, the results section outlines the econometric findings. A further detailed analysis section adds robustness with the implementation of various alterations to the baseline model. The last section discusses limitations and consideration to the models and approach. The conclusion section then answers the research question and interprets the results and robustness section. Lastly, the drawn conclusions lead to policy and future research recommendations.
2. Literature
The literature review supports the research question of the potentially heterogeneous correlation between regulation, supervision, and banking risk. The subsequent data section reveals the arguments behind the selected variables to test this theory.
This paper’s explanatory variables are quantified regulation and supervision indices. The St. Louis Federal Reserve serves as this paper’s definition of regulation and supervision. Bank regulation defines acceptable behavior and conduct of financial institutions, as written rules. Regulation is often prescriptive, inflexible, and quantitative, such as NY Federal Reserve vice-president Patrikis argues (1997). The rules may prohibit certain activities or aim to prevent them. This may be a total prevention, or a limited one in proportion to capital (Patrikis, 1997). Therefore, regulation requires formulation and issuance of specific rules that control and conduct banking.
Supervision, in contrast, is often qualitative. Bank supervision addresses the enforcement of the rules. Namely, this form involves an examiner or inspector. Supervision involves the safety and soundness of the financial institution within its scope (Patrikis, 1997). Generally, there is a continuous review of banking’s activities to ensure compliance to law and regulation. Connected is oversight, which is at a higher level and less intrusive than direct supervision of banking (Patrikis, 1997). Most of the models approach regulation and supervision as separate entities due to the inherent differences between these measures. The method section outlines the different effects of regulation and supervision. Moreover, the third model aims to address potential differences within each category.
8 section and references, while the data section uses the actual survey-years of 2011 and 2016. The two surveys state that each country report describes bank regulatory and supervisory policies as of 2011 and 2016, respectively. Barth et al. provide the first cross-country database on supervisory and regulatory frameworks within the financial sector. The database provides survey results on supervisory practices and regulatory features. The surveys have formed the basis of numerous papers on regulation and supervision. As the data section delves further into the characteristics of these surveys, Barth et al. provide a relatively unique dataset for quantifying regulation and supervision.
Barth et al. (2019) provide their latest dataset and paper, which reviews the global supervision and regulation of banks. Notably, their capital requirements findings indicate near-unanimous conformation to the Basel-guidelines. Therefore, following Barth et al. (2001), a near-unanimous finding of capital-adequacy warrants further disaggregated inquiry of other regulatory and supervision indicators. A homogeneous finding for otherwise heterogeneous countries means that analysis of only capital requirements is limitedly useful. Therefore, while including country aggregate level controls, this paper aims to contribute to the literature by analyzing individual bank’s risk exposure. Moreover, the inclusion of complementary variables to capital requirements helps understand the multi-faceted way regulation and supervision are implemented and affects bank risk. As a result, this paper presents a broader view of banking risk and regulation and supervision than works using unilateral relationships.
When a crisis occurs, banks experience incentives to decrease their asset base instead of increasing their capital: the debt-overhang problem. Another aspect of systemic risk has been the securitization markets, in which regulations have increased the risk-weights of securitized assets. The cross-section aspect of macroprudential regulation aims to decrease contagion and interconnectedness, while the time-series aspect aims to address volatility (Hanson, Kashyap, and Stein, 2011; Freixas, Laeven, and Peydró, 2015; Vives, 2016: 56). The Basel III framework provides the most prominent initiative to decrease systemic risk-taking within the banking sector. A time-series between 2011 and 2017 enables analysis of macroprudential reform and bank balance sheet effects.
The literature further illustrates the considerations and transformations that regulation and supervision have undergone since the Financial Crisis. Regulation and supervision have aimed to stimulate financial stability while protecting small investors. Arguably, there is a microprudential and macroprudential aspect of regulation and supervision (Borio, 2003). As Vives (2016) argues, the Financial Crisis has illustrated the failure of Basel II’s three pillars: the macroprudential dimension. Particularly the effects on systemic risk were underacknowledged within the Basel II framework (ESRB, 2014). Risk-weighted capital ratios alone were not successful in predicting threatened banks, as Haldane and Madouros (2012) propose more effective leverage ratios. The fulfillment of capital requirements has not always taken these macroeconomic dimensions into account. Banks’ incentives were to hold short-term debt and to maintain too little capital (Vives, 2016: 55). Therefore, these works indicate increased reform in regulation.
9 complexity and extent of new regulations (Anginer et al., 2019: 19-20). Supervisory resources and risk monitoring capabilities require proportional growth to banks that increase in size and complexity. The new rules after the Financial Crisis put additional strain on the required information generating, processing, and disseminating (Anginer et al., 2019: 20). Lastly, the new regulations including bank resolution require significant discretion from the supervisors. Thus, these works suggest significant post-crisis regulatory reform in the timespan between 2011 and 2017. Meanwhile, supervision may not have experienced a proportional increase. The literature illustrates that regulation and supervision are relatively under-researched. Barth et al. (2001) highlight the relative scarcity of regulation and supervision quantification. Similarly, Klomp and de Haan (2012) underline the rarity of studies in bank regulation and supervision’s impact on bank fragility. Following Klomp and de Haan (2015), this paper will differentiate itself from “most studies” in the body of literature that use singular risk indicators (5). For example, Klomp and de Haan (2012) use multiple measures of banking risk. Pasiouras et al. (2006) use Barth et al.’s surveys to argue that in a disaggregated view of banking supervision and regulation, only some dimensions have a significant correlation with the soundness of banks. Klomp and de Haan (2012), Klomp and de Haan (2015), and Pasiouras et al. (2006) find ambiguous evidence on banking risk and the banking’s fragility. Their approach provides guidance for inclusion of specific disaggregated risk, regulation, and supervision dimensions.
Klomp and de Haan (2012) further examine the impact of regulation and supervision on banking risk. The authors highlight that “most previous research” only finds inconclusive evidence for this relation (3). Their findings suggest that regulation and supervision lower banking risk for high-risk banks. However, the findings suggest no effect on low-risk banks. Klomp and de Haan (2012) highlight differences in effects for individual banks. Following this approach, this research has selected data on individual banks. Moreover, inclusion of low and high-income countries and their development allows insight into high versus low-risk bank effects. Lastly, Klomp and de Haan (2012) utilize factor analysis for both banking risk and regulatory measures, as a template for this paper’s principal component analysis.
A growing body of literature links individual bank characteristics to systemic risk contribution (Van Oordt and Zhou, 2019). For example, Anginer, Demirgüç-Kunt, and Zhu (2014b) find a negative relationship between bank competition and systemic risk. Vallascas and Keasey (2012) find that bank leverage ratios and liquidity ratios indeed improve resilience against systemic risk (7). However, the authors find that increases in bank size and asset growth have an increasing effect on systemic risk contribution. Brunnermeier, Dong, and Palia (2012) find that banks with a higher non-interest income have a higher contribution to systemic risk. Girardi and Ergün (2013) find that leverage, size, and equity beta primarily explain systemic risk contributions. Lopez-Espinosa, Moreno, Rubia, and Valderrama (2012) find that excessive short-term wholesale funding triggers systemic risk. These works justify multidimensional bank size and bank cost-to-income ratio as control variables for the models.
10 Detragiache (2011) use the Basel Committee of Banking Supervision’s core principles on banking supervision as their measure of banking regulation and supervision. Similar works that use a single measure of risk are González (2005), Demirgüç-Kunt, Detragiache, and Tressel (2008), and Fonseca and González (2010). Notably, these works that use singular risk measures find inconclusive results. These four works prompt a multidimensional approach to measurement of risk. Accordingly, this paper contributes to the literature with broader dimensions of risk measurement. The use of single measures, such as bank Z-scores, may disregard heterogeneous effects between risk indicators. The method section identifies the IMF’s CAMEL indicators as suitable risk categorization.
The literature section outlines the primary works that have driven this paper’s approach. As a result, this paper uses multiple risk, regulation, and supervision indicators. The subsequent data and method section delves further into the specific variables and data sources. Moreover, the data section reveals the ambiguous effects of regulation and supervision on risk. The method section clarifies the tests and descriptive statistics that identify and transform the variables appropriately for the whole paper. Lastly, the method section reports the tests and formulation of the baseline range of models.
3. Data and Method
3.1 Data
This paper uses panel data from 95 countries between 2011 and 2017. These countries, while dependent on data availability, include both low-income and high-income countries. The dataset includes 1870 individual banks. Stata econometric software enables the calculations and tests for the models and assumptions. Table B3 in the Appendix illustrates a variable observation range between 15032 for total banks and 12151 for the real interest rate.
3.1.1 Dependent variable
Bureau van Dijk, through its Orbis BankFocus database, provides the primary dataset for this paper.1 The BankFocus dataset consists of 1879 banks within 95 countries. Observations span between 2011 and 2019, as the most recent data. US banks relatively have the most observations. As this chapter illustrates, the availability of country control variables limits the years to 2017.
1 Initially, this paper aimed to utilize data from before and after the Financial Crisis. This enables full advantage
11 Figure 1. Bank supervisory criteria for their bank’s assessment of systemic risk contributions.
Source: Barth et al., 2013: 45.
For the dependent variable, this paper aims to quantify “risk” within the banking sector. Studies on banking risk are often one-dimensional (Klomp and de Haan, 2012: X). However, studies such as Agoraki, Delis, and Pasiouras (2011), Gaganis and Pasiouras (2007), Klomp and de Haan (2012), and Varotto and Zhao (2018) find that banking risk is multidimensional. Multiple indicators measure bank risk, as Table 2 shows. The selection of four risk measures follows both Evans (2000) and Klomp and de Haan (2012). These studies construct proxies based on the IMF’s CAMEL indicators. As Figure 1 illustrates, bank supervisors participating in the Barth et al. (2013) survey indeed report these categories among their primary risk indicators. This paper selects the indicators based on comparable studies and data availability from BankFocus. As Table 2 reports, CAMEL indicators include capital adequacy, asset quality, management, earnings and profitability, and liquidity (IMF, 2019). Due to data availability, the models leave out management. Namely, there were no management indicators available from BankFocus with more than 300 observations. The Evans (2000) and Klomp and de Haan (2012), the IMF’s CAMEL indicators, and observation maximization determine the selection. This paper selects total equity divided by total assets as a measure for capital adequacy. Liquid assets relative to deposits and short-term funding provides a liquidity measure. Furthermore, loan loss reserves relative to impaired loans indicate banks’ asset quality. Lastly, the ROAA serves as an earnings indicator. These indicators’ values are higher when there is less associated risk. Namely, each position refers to a value where the numerator increases as the associated bank risk decreases (IMF, 2019).
0 20 40 60 80 100 120
Bank capital ratios Bank liquidity ratios Bank loan portfolio composition Bank credit growth Bank non-performing loan ratios Bank profitability ratios Bank provisioning ratios Bank leverage ratios FX position of banks Housing prices Stock market prices
number of countries reporting yes for each factor
12 Table 2. Selected risk measures and alternatives.
IMF’s CAMEL Indicators Capital
Adequacy Liquidity/Funding Asset Quality Earnings
Selected Dependent
Variables
Total equity / total assets
Liquid assets / deposits &
short-term funding Loan Loss Reserves / Impaired loans ROAA Banks with observations 1879 1871 1879 1879 Alternatives Common Equity / Core Tier 1 Tier 1 Ratio Total capital adequacy ratio Basel III leverage
ratio
Gross loans & advances to customers / customer deposits Interbank assets / interbank liabilities Customers loans & advances / total
assets Net charge-offs / Average loans Unreserved impaired loans / equity ROAE Net Interest Margin Net interest income / operating revenue Banks with observations 384 – 1810 430 – 1722 195 – 1632 1192 – 1829 Source: Evans, 2000; Klomp and de Haan, 2012; IMF, 2019.
3.1.2 Explanatory Variables
Key to this paper is the quantification of regulation and supervision within the banking sector. Inherently, regulation and supervision cannot easily be quantified or even measured. Several papers within the literature limit arguments to strictly theoretical hypotheses without empirical evidence (Wagner, 2010). While this paper faces the same challenge, it aims to implement quantified measures for regulation and supervision. De Haan, Jin, and Zhou (2019) outline two potential sources for proxy construction of regulation and supervision. On the one hand, numerous studies use an index based on the Basel Committee’s Core Principles of Effective Bank Supervision (Demirgüç-Kunt and Detragiache, 2011). This IMF measure is not freely available. On the other hand, the World Bank publishes the surveys of Barth et al. These survey-datasets are freely available.
13 2013, and the fifth in 2019. The data from Barth et al. consists of surveys as a unique source for banking regulation and supervision, globally. The survey uses bank regulatory agencies from more than 180 countries. However, as the participation changes per edition, the full panel of countries is smaller than 180 between 2011 and 2017.
After a request on Ross Levine’s Berkeley-website, this paper has gained access to the survey’s indexation approach (Levine, 2020). Barth et al. (2013) use this indexation in the working paper that accompanied the publishing of the survey. However, for the 2019-survey, no quantification was available and reaching out to the authors was unsuccessful. As a solution, this paper has quantified the 2019-survey using the 2013-formulas. This approach enables the use of recent BankFocus data. Table A1 and A2 illustrate the quantification of the survey by Barth et al. (2013) and this paper’s corresponding quantification of the 2019-survey. Notably, the surveys’ question order is different. Accordingly, this paper constructs a similar indexation as Barth et al. (2013) for the most recent data.
Thus, this paper quantifies the 2019-survey with the formulas used by Barth et al. in their 2013-paper. Table A2 in the Appendix shows the formulas and the necessary changes made between both surveys. The survey itself consists of hundreds of questions regarding permissible bank activities, capital requirements, the power of supervisory agencies, information disclosure requirements, external governance mechanisms, deposit insurance, barriers to entry, and loan provisioning (Barth et al., 2013: 3). Therefore, Barth et al. construct summary indices to handle the complexity of the underlying questions. This approach facilitates cross-country comparison and policy-changes over time (3). This paper has identified the questions used most commonly in the papers of Barth et al. (2001; 2004; 2008; 2013), Pasiourias et al. (2006), and Klomp and de Haan (2012). Where necessary, this paper has made inclusions of newer regulation, such as compliance to Basel III in the 2019-survey. A small number of countries participated in only one survey, which have been removed from the dataset to ensure observation variance. Table A2 in the Appendix illustrates the process from raw survey data to indexation.
This paper has used the questions in Table A1 in the Appendix and 2013-quantification for its 2019-quantification. Using Microsoft Excel, this paper identified the changed question order per measure. Afterward, this paper quantified string variables where necessary, for instance, equaling “yes” to a value of 1. Several indices require a threshold, for instance of 100%, which again this paper implements in Excel formulas. Table A2 further depicts specific criteria, such as outright survey answers below nine counting as an index-value of zero that requires further implementation.
14 Table 3. Regulation and supervision variables, and expected effects with an increase in each index.
Decreasing banking risk
-
+
Increasing banking riskRegulation Smaller activity range simplifies
enforcement
Limits non-interest income for banks, linked to financial shock
sensitivity
Activity Restrictions
Limits diversification as a risk-averse strategy
Decreases risk of (unregulated) entities affecting (regulated) banking
activities
Financial Conglomerates Restrictiveness
Limits ability to compensate losses with other conglomerate activities Decreases mergers, which maintains
diversity
Bank Entry
Requirements Lowers competition, bank efficiency
Buffer for losses
Aligns bank incentives with those of creditors
Capital Requirements
Misalignment of requirements with actual risk
Uniform portfolios result in increase of aggregate shock severity Tail risk increases with higher equity Supervision
Ability to restructure or penalize malfunctioning bank Incorporation of long-term societal
incentives
Supervisory Control
Misaligned incentives in the supervisory framework may lower effectiveness, which may obscure actual risk-taking
Reduces moral hazard and risk-taking
Private Sector Monitoring
Insufficient quality induces false monitoring
Requires fair risk evaluation, externally funded or otherwise independent Prevents bank runs Deposit
Insurance Moral hazard, induces risk-taking
15 that rely on non-interest income cause financial instability. Thus, restrictions on these activities lead to lower risk-taking. Similarly, other works find that non-interest income levels positively correlate to financial shock sensitivity (Brunnermeier et al., 2012; Vallascas and Keasey, 2012; Van Oordt and Zhou, 2019).
Secondly, overall financial conglomerate restrictiveness addresses the extent of regulation on bank ownership. This measure defines the extent to which banks and nonbanks may combine to form conglomerates (Barth et al., 2013: 11). Nonbanks cannot accept deposits, which differentiates this category from deposit-taking banks. After 1999, the US has joined Europe and Japan in permitting banks to merge with other financial enterprises (Herring and Litan, 2003). This phenomenon problematizes capital requirement measurement, for instance, for a conglomerate with a bank and life insurance company. Among others, the European Banking Authority works on “identification and measurement of systemic risk of financial conglomerates and assessing its cross-sector implications” (EBA, 2012: 12). Particularly for the US, Europe, and Japan, these restrictions aim to minimize systemic risk contributions. Namely, supervisors lacked power during the Financial Crisis to simultaneously apply banking, insurance, and supplementary supervision at the level of the same parent entity (EBA, 2012: 7). On the one hand, lower restrictions lead to large, complex financial institutions. These cross-sector enterprises problematize single-sector regulation and supervisory power (EBA, 2012: 7). As evidenced by the Financial Crisis, non-regulated entities took risks that impacted the overall conglomerate, including regulated entities (EBA, 2012: 8). On the other hand, a financial conglomerate may compensate losses within its banking sector with unaffected external activities, which lowers its risk (EBA, 2012: 9).
Thirdly, this paper includes bank entry requirements. Keeley (1990) and De Nicolo and Kwast (2002) find that entry restrictions lead to a decrease in bank mergers. Consolidation leads to more similarities within the sector, which reduces the effect of idiosyncratic shocks (De Nicolo and Kwast, 2002). A decrease in mergers protects the diversity of banks. In contrast, Beck et al. (2006) find that bank entry requirements increase the probability of a financial crisis due to competition decreases within the sector. The decrease in competition may decrease bank efficiency, which increases systemic risk.
Fourthly, capital regulation serves as an indicator of regulation. The majority of theories emphasize that capital regulation reduces banks’ risk-taking (Fernandez and González, 2005; Dewatripont and Tirole, 1994). Namely, capital serves as a buffer for losses that banks may experience. Capital regulation functions to align bank owner incentives with those of creditors (Barth et al., 2004; Berger, 1995). However, Perotti et al. (2011) and Rochet (1992) find ambiguity between capital and banking risk. Namely, the selection of risk weights in regulation possibly misaligns with actual market risk (Perotti et al., 2011). Furthermore, the tail risk may increase as banks have more equity funding (Rochet, 1992). Zhou (2013) further finds that capital regulation possibly increases systemic risk, as portfolios align and diversification of the aggregate sector decreases. Response to shocks may therefore become more uniform and severe on aggregate.
16 banks (de Haan et al., 2019: 8). Under-monitored banks may have incentives for high risk profit-seeking. Without monitoring, these banks may not take into account long-term, systemic risk contributions of this profit-seeking. Moreover, supervision possibly limits contagion when a crisis occurs (de Haan et al., 2019: 9) While Fernandez and González (2005) find that supervisory control indeed negatively relates to banking risk, Barth et al. (2004) find inconclusive evidence. Barth et al. (2013) argue that misaligned incentives in the supervisory framework were central to the Financial Crisis. This misalignment may obscure the actual risk-taking.
Sixthly, private sector monitoring implements a measure of auditors and rating agencies as a supervisory body in the banking sector. The literature has ambiguous findings, as Fernandez and González (2005) find that increased monitoring reduces moral hazard and risk-taking. In contrast, Barth et al. (2004) underline that these monitoring bodies require sufficient quality and fair risk evaluation incentives. Otherwise, false monitoring may increase risk.
Seventhly, deposit insurance provides the last supervision measure. Following Demirgüç-Kunt and Detragiache (2002), deposit insurance has an ambiguous theoretical effect. Namely, the insurance prevents bank runs, which lowers risk to the banking sector. In contrast, deposit insurance may induce bank risk-taking through moral hazard. Both Barth et al. (2004) and Anginer et al. (2014a) find evidence for the latter, where deposit insurance increases financial instability.
3.1.3 Control Variables
BankFocus provides the individual bank control variables. Firstly, a multi-dimensional employee indicator controls for bank size. Namely, the employee-variable is an index for country-rank, size, and activity of the individual banks. Banks with more employees and higher activity may be associated with lower risk, due to higher diversification (Demsetz and Strahan, 1997; Van Oordt and Zhou, 2019). In contrast, banks perceived as “too-big-to-fail” may use their implicit guarantees to increase their portfolios’ risk (Penati and Protopapadakis, 1988; Van Oordt and Zhou, 2019). Therefore, this paper does not expect a specific sign for the employee variable. Lastly, the cost-to-income ratio serves as an indicator of variance in banks’ technical efficiency. This paper expects a negative sign for this efficiency measure, as more efficiency is likely to cause lower risk (Van Oordt and Zhou, 2019).
17 and positions, which leads to profit-seeking in higher-risk activities. Thirdly, the size of the current account relative to the GDP-levels serves to include variability within countries’ trading behavior. The World Bank provides the data for this indicator (2019a). This paper expects a positive sign for this variable, as a trade shock likely indicates a crisis where banking risk increases. Furthermore, unemployment serves as a fourth control variable that controls for country-specific employment levels. The IMF publishes the data for this variable. This paper expects a negative sign for unemployment, as higher unemployment likely indicates a crisis in which banking risk is high. The real effective exchange rate serves as a fifth control variable, again provided by Beck et al. (2019). This paper expects a negative sign, as a decrease in exchange rate poses a threat to a bank’s capital and earnings (Klomp and de Haan, 2015. Lastly, the real short-term interest rate further attributes variance within the model due to monetary environment characteristics (Agoraki et al., 2011). This paper expects a negative sign for this variable, which Beck et al. (2019) provide. Namely, low short-term interest rates may signify a monetary policy response to a crisis, in which bank positions are likely higher in risk.
3.2 Method
RISKit = β0i + β1REGULATIONit + β2SUPERVISIONit + β3EMPLOYEESit +
β4BANKCOSTTOINCOMEit
+ β5GDPGROWTHit + β6GDPDEFLATORit + β7CURRENTACCOUNTGDPit +
β8UNEMPLOYMENTit + β9REALEFFECTIVEEXCHANGERATEit +
β10REALINTERESTRATEit + uit
(1)
The literature review reveals that expected regulation and supervision effects are ambiguous. The data section delves further into the expected effects of each measure. Based on the theoretical body, this paper formulates its baseline equation. In equation (1), RISKit refers to the dependent variable, which the subsequent method section formulates as the first and second principal component of risk. Furthermore, β0i represents the intercept with the subscript that signifies individual specificity accordingly to the fixed effects model (Adkins and Hill, 2011: 446). Moreover, β1 to β2 represent the coefficients of the explanatory variables. Then, β3 to β4 represent the coefficients of the bank control variables. Subsequently, β5 to β10 represent the coefficients of the country-aggregate control variables. Lastly, uit represents the within-entity errors. The following section outlines the data collection process and the selection of variables. Moreover, Section 3.2 reports the tests that lead to equation (1)’s fixed effects model.
18 variable’s factor loading, eigenvalues, and the determination per component. The coefficient value illustrates the magnitude of that variable’s importance in the component’s calculation (Minitab, 2020). Following the Kaiser criterion, an eigenvalue equal to or more than one is significant (De Haan et al., 2019: 14). The criterion results in a reduction from four banking risk categories to two components, henceforth “first principal component” and “second principal component.”
This paper then aims for a similar minimization of variable dimensions with regulation and supervision. Initially, this paper aimed to follow the PCA-approach of Klomp and de Haan for these variables (2012). However, using Stata econometric software, the eigenvalues and factor loadings for regulation and supervision did not lead to meaningful exclusions of components. This finding suggests that the indices lack sufficient information in their values for PCA. Notably, the models use a smaller timespan than Klomp and de Haan (2012). Instead, this paper divides the variables between regulation and supervision. Table B3 clarifies that “regulation” contains activities restrictions, overall conglomerate restrictions, bank entry requirements, and overall capital stringency. Namely, these measures refer to a compliance criterium formed by a supervisory institute or sovereign itself (Patrikis, 1997). The paper sums the four categories to form the regulation variable. Accordingly, the “supervision” variable contains supervisory power, private sector monitoring, and deposit insurance. The first two variables have a clear connection to supervision. Deposit insurance requires a supervisory entity that verifies insurance claims, which justifies its inclusion. The supervision variable is the summation of these three indices. Lastly, this paper uses the weights of Barth et al. (2013) for each variable’s contribution to the summed indices, which Table A2 in the Appendix reports. The limitation section reports the considerations on the use of these weights. Figures B15 until B18 in the Appendix depict the levels in index-scores per country between 2011 and 2017. The figures illustrate a relatively high increase of regulation in Brazil and a high decrease in the US. Similarly, the figures depict a significant increase in supervision for South Africa and a decrease for Denmark. Due to the nature of the non-annual surveys, this paper assumes a constant value for regulation and supervision between 2011-2015 and 2016-2017. This approach maximizes observations. Moreover, this assumption allows inclusion of risk and control variables for each year. Lastly, this approach enables a dynamic system-GMM analysis that requires at least three observations. Section 5.3 and the conclusion analyze the correctness of this constant value assumption.
Subsequently, this paper aims to analyze any violations of the estimation assumptions for the variables and first range of models. Table B3 in the Appendix reports the descriptive statistics for the variables: the mean, standard deviation, minima, maxima, N, and total number of banks. The included variables display relatively comparable means and standard deviations. The following paragraphs report several transformations that lead to Table B3’s values. These include first-differences, outlier exclusions, and natural logarithms.
19 changes from units to percentages for these variables. Table B4 in the Appendix reports the results. Unemployment, real effective exchange rate, and real interest rate have unit-roots. Accordingly, this paper implements the first differences of these variables into its models. As compensation for first-differencing, this paper uses a timespan of 2010 to 2017 for these three control variables.
The correlations in Table B5 in the Appendix warrant analysis of the possibility of multicollinearity issues. Through overinflation of standard errors, multicollinearity makes statistically significant variables insignificant. Firstly, the different regulation and supervision indicators display relatively strong collinearity in the table. Therefore, a later model omits overall capital stringency and supervisory power from the seven indicators in Table C1 in the Appendix. Namely, Baum (2006) suggests a threshold of 0.800 for collinearity issues. Multicollinearity tests for Table 4 find VIF-scores between 4.86 and 8.34 for the PCA models. This score confirms that there is no reason for assuming multicollinearity issues as below the accepted threshold of 10 (Baum, 2006: 34). This paper does expect that the VIF-scores may be higher throughout the paper, as regulation and supervision change in the same year. The tests measure whether other explanatory variables predict an individual explanatory variable linearly and accurately. This finding is repeated throughout the paper, as there is an absence of high R2 -values with low significance as an indicator of high collinearity. Notably, every linear model uses adjusted R2-values that penalize addition of regressors. The Akaike information criterion (AIC) and Bayesian information criterion (BIC) benefit goodness-of-fit assessment. These two penalized-likelihood criteria use different penalty weights. This paper uses the adjusted R2, AIC, and BIC for its preferred model selection as complementary quality indicators.
Inquiry into heteroskedasticity and serial correlation requires a further test. The test finds heteroskedastic panels, but no serial correlation within the model. Serial correlation signifies a variable in a time series that has significant correlation with a lagged version of itself. Heteroskedasticity refers unequal variance across units (Baum, 2001). As a result, standard errors will be incorrect and hypotheses may be wrongfully rejected. The scatterplots of Figure B1 until B8 in the Appendix confirm heteroskedasticity. This paper implements robust standard errors as a solution. Namely, this approach obtains unbiased standard errors for the model (Hoechle, 2007: 284).
Furthermore, this paper analyzes outliers within the dataset. Significant variance and outliers may occur due to the inclusion of banks with heterogeneous sizes and nationalities. Moreover, this paper aims to exclude extreme values or outliers from the dataset. This paper considers winsorizing. This technique excludes a specific percentile of the data, for instance, all values below the fifth percentile and above the 95th percentile. Scatter plots identify a number of extreme values, which have been excluded manually. The remaining scatter plots confirm an absence of outliers in Figure B1 until B8. Therefore, manual exclusion avoids winsorizing and unnecessary removal of information.
20 distributions. In the Appendix, Figures B9 until B14 illustrate the risk indicators’ distribution. Equity relative to total assets and the ROAA resemble a normal distribution closely. Liquid assets relative to deposits and short-term funding and loan loss reserves relative to impaired loans display a skewness to the left. Therefore, these findings imply non-normal distributions. This paper considers a natural logarithm for achieving normality. Appendix D depicts the mathematical transformation of these variables to their natural logarithm and log-linear coefficient interpretation. As a result, the four variables in Figure B9 until B14 show distributions that resemble normality.
Accordingly, the models with loan loss reserves relative to impaired loans have a log-linear instead of linear-linear interpretation of its coefficients. Similarly, models with liquid assets relative to deposits and short-term funding require a log-linear interpretation. This paper standardizes the two non-logarithmic risk categories. Appendix D clarifies the calculation of standardized variables. In doing so, the two linear variables have a near-zero mean and a standard deviation of one. Standardization makes coefficients between the linear risk models more comparable without influencing the variance or significance. Another transformation scales the employees control variable. A division by 100000 allows the coefficients to become non-zero, while this scaling does not affect the significance.
Lastly, Breusch-Pagan Lagrange multiplier tests confirm that pooled OLS regressions are inapplicable. This paper argues its choice between a random and fixed effects model as follows. The fixed-effects model assumes that individual-specific effects are correlated with the independent variables. As a result, this approach analyzes the relation of the predictor and outcome variables within an entity (Hill, Griffiths, and Lim, 2015: 543). The random effects model also includes a unit-specific error-term, but this error-term is assumed to be uncorrelated to the independent variables (Bartels, 2018; Kohler and Kreuter, 2009: 245). Therefore, the random effects model can provide estimates for time-invariant variables. However, the models include variables that vary over time, where this paper assumes correlated individual-specific effects with the independent variables. A Hausman-test indeed confirms that a fixed effects model is appropriate. The estimated coefficients are significantly different between the random and fixed effects models, which signifies a fixed effects approach (Bartels, 2018). Equation (1) depicts the resulting model for the first range of models. Table B6 in the Appendix reports these tests concisely for the subsequent models.
4. Baseline Results
Table 4 depicts the most parsimonious regressions as baseline results. Following the Principal Component Analysis in the method section, two aggregate measures for banking risk remain: the first and second principal components of bank risk. Moreover, these models implement the two summed regulation and supervision indices. Columns (1) to (3) report the findings for the first principal component, while columns (4) to (6) report the findings for the second principal component. The models progressively add control variables to each column.
21 correlation between the supervision variable and the risk measures in (1), (2), and (6). These findings align with this paper’s expectations that regulation and supervision are two separate dimensions.
Table B2 in the Appendix reports that total equity relative to total assets and ROAA positively determine the first component primarily. Liquid assets relative to deposits and short-term funding and loan loss reserves relative to impaired loans positively determine the second component primarily. Table B2 reports that both components are positively correlated to each risk variable pair. Thus, a positive coefficient signifies a decrease in banking risk, and a negative coefficient signifies an increase. In column (1), this paper interprets the findings as follows. One unit change in supervision, as the aggregate of indices, leads to a 0.048 decrease in the first principal component.
In contrast to the subsequent models, using principal components limits the interpretation of the coefficients. The models use two natural logarithm transformations, which problematizes further interpretation. Both linear variables primarily determine the first component, while the logarithmic variables determine the second component. Pooling the risk measures causes arbitrary coefficient values. A one unit increase in ROAA does not signify the same change in risk as a one unit increase in liquid assets relative to short-term funding. Therefore, this paper limits the interpretation of the coefficients in this first range of models. The R2-values, AIC, and BIC report that the model with two logarithmic variables and all control variables, column (6), has the best fit.
22 Table 4. First regression results for the baseline model.
1st Regression results
Baseline Fixed effects
1st Principal Component Bank of Risk
2nd Principal Component Bank of Risk
(1) (2) (3) (4) (5) (6) Regulation 0.005 (0.133) 0.008* (0.024) 0.008 (0.155) 0.008** (0.001) 0.011*** (0.000) 0.011** (0.004) Supervision -0.048*** (0.000) -0.049** (0.001) -0.025 (0.186) -0.012 (0.226) -0.013 (0.275) -0.030* (0.015) Employees -0.526 (0.071) -0.231 (0.300) -0.106 (0.355) 0.015 (0.855)
Bank Cost-to-income Ratio 0.007
(0.125) 0.018** (0.002) -0.005 (0.247) 0.004 (0.206) GDP Growth 0.012 (0.263) 0.003 (0.694) GDP Deflator 0.025* (0.021) -0.005 (0.654) Current Account / GDP -0.003 (0.754) 0.001 (0.632) Unemployment -0.006*** (0.000) 0.006* (0.033)
Real Effective Exchange Rate 0.001
(0.789)
0.001 (0.523)
Real Short-term Interest Rate 0.007
(0.459) -0.012 (0.350) constant 0.766*** (0.000) 5.552 (0.082) 17.091** (0.002) -0.062 (0.730) -5.961 (0.650) 3.771 (0.218)
Country Fixed Effects Yes Yes Yes Yes Yes Yes
Time Fixed Effects Yes Yes Yes Yes Yes Yes
R2 within 0.035 0.042 0.100 0.005 0.008 0.019 R2 between 0.078 0.126 0.127 0.020 0.125 0.167 R2 overall 0.002 0.096 0.074 0.005 0.129 0.141 AIC 20124 17529 6949 9876 9155 -935 BIC 20183 19695 7752 12208 11313 -125 N 13052 12248 9505 13052 12248 9505 VIF 4.86 7.54 8.32 4.89 7.61 8.34 * p<0.05, ** p<0.01, *** p<0.001
p-values in parentheses, robust standard errors adjusted R2-values
23
5. Further Detailed Analysis
5.1 Second Model: Four Disaggregated Measures
RISKit = β0i + β1REGULATIONit + β2SUPERVISIONit + β3EMPLOYEESit +
β4BANKCOSTTOINCOMEit + β5GDPGROWTHit + β6GDPDEFLATORit +
β7CURRENTACCOUNTGDPit + β8UNEMPLOYMENTit + β9REALEFFECTIVEEXCHANGERATEit
+ β10REALINTERESTRATEit +uit
(2.1)
ln(RISK)it = β0i + β1REGULATIONit + β2SUPERVISIONit + β3EMPLOYEESit +
β4BANKCOSTTOINCOMEit + β5GDPGROWTHit + β6GDPDEFLATORit +
β7CURRENTACCOUNTGDPit + β8UNEMPLOYMENTit + β9REALEFFECTIVEEXCHANGERATEit
+ β10REALINTERESTRATEit +uit
(2.2)
For clarity, this paper uses the same steps for specification of each additional model. Equation (2.1) represents the second model. This model incorporates the four risk measures. Namely, while the PCA benefits clarity, each risk measure has a unique value. Disaggregation of the four risk measures enables meaningful coefficient interpretation. RISKit now represents four disaggregated dependent risk variables. The included independent variables are the same as in the first model. The linear-linear models use equation (2.1) in columns (1), (2), (7), and (8). Accordingly, the log-linear models with liquid assets relative to deposits and short-term funding use equation (2.2) for columns (3) and (4). Similarly, equation (2.2) reports the log-linear model for loan loss reserves relative to impaired loans of columns (5) and (6).
Table B6 in the Appendix reports the assumptions for this model. Robust standard errors compensate for heteroskedasticity, while the models indicate no autocorrelation. Multicollinearity tests report no reason for variable exclusion, which Table C1’s VIF scores confirm in the Appendix. Hausman-tests confirm that fixed effects models are appropriate. Table C1 in the Appendix reports the findings for this second model with two regressions for each risk measure. The columns report the findings in pairs for each risk measure, where the second column for each adds all control variables. The regulation variable has a significant coefficient in column (1), (2), (4), (5), (6), and (8). In column (1), a one unit change in the summed regulation-index leads to an 0.009 increase in a bank’s equity/total assets.
The summed supervision index has significant coefficients in columns (1), (2), (3), (4), (6), (7), and (8). These coefficients are negative. Column (1)’s coefficient signifies that a one unit change in the summed supervision index leads to a 0.029 decrease in a bank’s equity relative to total assets. Furthermore, for column (6), the supervision coefficient signifies that one unit increase in the supervision index results in a -7.9% change in a bank’s loan loss reserves relative to impaired loans.
24 represents a 5.6% increase, on average, in Belfius Banque’s loan loss reserves relative to impaired loans. Notably, these interpretations disregard the other variables’ effects.
The R2, AIC, and BIC indicate the explanatory power and quality of each model. Column (4), using the natural logarithm of liquid assets relative to deposits and short-term funding, has the highest R2 with lowest AIC and BIC values. Thus, this paper selects the model of column (4) in Table C1 as its preferred model.
5.2 Third Model: Disaggregated Regulation and Supervision
As the third table of regressions, this paper analyzes the broadest disaggregation that the dataset allows. Table C2 in the Appendix reports the findings with the four risk measures and seven explanatory variables. This incorporation enables a comparison of the significance and signs of the variables between the main and additional models.
RISKit = β0i + β1BANKINGACTIVITIESit + β2CONGLOMERATERESTRICTIONSit +
β3BANKENTRYREQUIREMENTSit + β4CAPITALSTRINGENCYit + β5SUPERVISORYPOWERit +
β6PRIVATEMONITORINGit + β7DEPOSITINSURANCEit + β8EMPLOYEESit +
β9BANKCOSTTOINCOMEit + β10GDPGROWTHit + β11GDPDEFLATORit +
β12CURRENTACCOUNTGDPit + β13UNEMPLOYMENTit +
β14REALEFFECTIVEEXCHANGERATEit
+ β15REALINTERESTRATEit +uit
(3.1)
ln(RISK)it = β0i + β1BANKINGACTIVITIESit + β2CONGLOMERATERESTRICTIONSit +
β3BANKENTRYREQUIREMENTSit + β4CAPITALSTRINGENCYit + β5SUPERVISORYPOWERit +
β6PRIVATEMONITORINGit + β7DEPOSITINSURANCEit + β8EMPLOYEESit +
β9BANKCOSTTOINCOMEit + β10GDPGROWTHit + β11GDPDEFLATORit +
β12CURRENTACCOUNTGDPit + β13UNEMPLOYMENTit +
β14REALEFFECTIVEEXCHANGERATEit
+ β15REALINTERESTRATEit +uit
(3.2)
25 cause coefficients to change significantly with small changes in the variables. Moreover, multicollinearity overinflates standard errors. The models omit overall capital stringency and supervisory power due to an accepted threshold of 0.8 for excessive collinearity (Franke, 2010). The post-exclusion VIF-scores in Table C2 confirm there are further multicollinearity issues within the models, as the values are higher than 10. Therefore, the models likely suffer from inflated standard errors and associated decrease in significance. Hausman-tests further confirm fixed effects as the appropriate models.
Accordingly, Table C2 reports the results of this third range of models. Activity restrictions has significant coefficients in columns (1), (2), (3), (5), (7), and (8). Overall financial conglomerate restrictiveness has significant coefficients for columns (1), (2), (3), (4), and (8). Furthermore, bank entry requirements has significant coefficients for columns (1), (3), (4), (5), (6), and (8). Lastly, private monitoring has a significant coefficient for column (8), while deposit insurer power has no significant findings. In contrast to Tables 4 and C1, the signs for these coefficients are heterogeneous across risk, regulation, and supervision variables. This paper interprets the coefficients as follows. In column (1), a one unit increase in the activity restrictions index results in a 0.645 increase in equity relative to total assets. In column (5), a one unit increase in the bank entry requirements index results in a 5.7% increase in a bank’s loan loss reserves relative to impaired loans.
Within the dataset, the Netherlands experienced an activity restrictions decrease from five to four, while Singapore experienced an overall conglomerate restrictiveness increase from seven to eight. Accordingly, one can interpret from Table C2’s column (7) that, on average, ING Nederland experienced a 0.244 decrease in its ROAA from 2015 to 2016. Similarly, in column (2) on average DBS Group Singapore experienced a 0.901 decrease in its equity relative to total assets from 2015 to 2016. Notably, these interpretations disregard the other variables’ effects.
These models report the disaggregated effects of each regulation and supervision index. The five included variables illustrate more dispersion in direction of effect than the summed regulation and supervision variables in the earlier models. Therefore, the findings suggest that categories within regulation and supervision have both increasing and decreasing effects on an individual bank’s risk. Moreover, this finding suggests compensation effects within the summed regulation and supervision variables. For example, a positive coefficient of activity restrictions and negative coefficient of overall financial conglomerate restrictiveness may count out each other’s effect in summed regulation.
26
5.3 Fourth Model: Two Data Points
Thus, this paper views the combination of four risk measures, regulation, and supervision as its preferred approach. Another robustness check employs the years 2011 and 2016 as the only data points. While significantly reducing the observations, this approach only uses the data points of the survey rounds itself. In the previous models, this paper has assumed a constant level of regulation and supervision in between surveys. The literature indeed confirms that regulation and supervision levels change non-gradually between policy changes (Barth et al., 2013; Klomp and de Haan, 2012). Despite this argument, this approach may be viewed as data manipulation. Policy change in 2012 affects the supervision index only in 2016. Potentially, the findings are distorted by these filled-in data points for the years 2012, 2013, 2014, 2015, and 2017. Table C3 in the Appendix reports the findings of a model that analyzes this potentially distortive characteristic. The limitations section delves further into the issues inherent to non-annual survey data.
The models in Table C3 use equation (2.1) and (2.2), where the latter refers to the log-linear regression with loan loss reserves relative to impaired loans. Table B6 reports the assumptions for these models using 2011 and 2016. A finding of heteroskedasticity requires robust standard errors. The models in Table C3 do not indicate autocorrelation. As the VIF-values in Table C3 and a multicollinearity test confirm, there are no multicollinearity issues. Lastly, Hausman-tests confirm fixed effects models as appropriate.
The summed regulation variable has significant positive coefficients in column (1), (2), (5), and (8). In contrast, the supervision variable has significant negative coefficients in column (1), (2), (4), (7), and (8). This paper interprets the coefficients as follows. In column (1), a one unit increase in the regulation variable results in a 0.011 increase in a bank’s equity relative to total assets. Furthermore, in column (5), a one unit increase in regulation results in a 3.5% increase in a bank’s loan loss reserves relative to impaired loans. The R2, AIC, and BIC report the best fit for column (4).
27
5.4 Fifth Model and Sixth Model: Low-income versus High-income
Additionally, this paper analyzes differences between low-income and high-income countries. The models use the same equations as those in Table C1: (2.1) and (2.2). The World Bank provides the data and definitions of the income groups (2019b). The models pool low-income and lower middle income to form a “low-income” group. Furthermore, the upper middle income and high-income group form the “higher-income” group. Table C4 and 5 report the findings for these models. The model specification and assumptions are the same for the low-income and high-low-income models, which Table B6 reports. Accordingly, the models implement robust standard errors. Hausman-tests confirm fixed effects as the appropriate approach. Table C4 reports the findings of the low-income models. These models have relatively few observations between 652 and 134. Therefore, this paper attributes only tentative conclusions to these findings. Regulation has a significant negative coefficient in column (1). Supervision has a significant negative coefficient in column (1) and (3). Column (8) reports a significant positive coefficient for supervision. Table C4 reports relatively low significance across the models. This paper interprets the findings as follows. A one unit increase in the regulation index results in a decrease of 0.279 in the equity relative to total assets of a low-income country’s bank.
Table B15 to B18 report that Ghana experienced an increase in regulation from 27 to 34, while Nigeria experienced a decrease from 22 to 21 in supervision between 2015 and 2016. Accordingly, Table C4’s column (8) suggests that, on average, Societe Generale Ghana experienced a 3.717 increase in its ROAA between 2015 and 2016. Similarly, in column (3) Ecobank Nigeria experienced a 33.4% increase in liquid assets relative to deposits and short-term funding. Notably, these interpretations disregard the other variables’ effects.
The low-income group’s findings suggest that regulation has an increasing effect on the risk measure in column (1). Supervision has both an increasing effect on risk in columns (1) and (3), with a decreasing effect on risk in column (8). These findings suggest that banks within low-income countries may experience opposite effects of regulation and supervision than in the broader sample. However, while Barth et al.’s surveys include numerous low-income countries, BankFocus provides only limited data. Therefore, this paper returns to the specification of Table C1 as its preferred approach.
Table 5 reports the results for the high-income models. The results report an increase in significance for the regulation and supervision variables over the results in Table C1. Regulation has significant positive coefficients in column (1), (2), (4), (6), and (7). Supervision has significant negative coefficients in column (1), (2), (6), (7), and (8). This paper interprets the coefficients as follows. In column (1), a one unit increase in the regulation index results in a 0.013 increase in the equity relative to total assets of a high-income country’s bank. In column (6), a one unit increase in the supervision index results in a 8.9% decrease in the loan loss reserves relative to impaired loans of a high-income country’s bank.
28 to impaired loans between 2015 and 2016. Similarly, Table 5 column (2) suggests that, on average, Bayerische Landesbank experienced an increase of 0.084 in its equity relative to total assets. Notably, these interpretations disregard the other variables’ effects.
This paper now prefers the high-income models, due to the increase in the explained variance that Table 5’s R2-values report. Moreover, the AIC and BIC-values are lower than in Table C1, which confirms a better fit. The explanatory power of the models increases with the exclusion of low-income countries. In particular, column (2) with equity relative to total assets has the best fit. As Table C4 suggests, banks within low-income countries may experience different effects of regulation and supervision on their bank risk.
Table 5. Sixth regression results for the high-income countries
6th Regression
results High-income Fixed effects
Equity/Total Assets
Liquid Assets/Deposits & Short-term Funding
(natural logarithm)
Loan Loss Reserves / Impaired Loans (natural logarithm) ROAA (1) (2) (3) (4) (5) (6) (7) (8) Regulation 0.013*** (0.000) 0.007* (0.043) 0.006 (0.102) 0.057** (0.006) 0.009 (0.047) 0.032*** (0.000) 0.008** (0.002) -0.002 (0.267) Supervision -0.036** (0.002) -0.028* (0.034) 0.001 (0.917) -0.015 (0.272) 0.016 (0.368) -0.089*** (0.000) -0.038*** (0.000) -0.032*** (0.000) Employees -0.225 (0.254) 0.063 (0.238) -0.238 (0.098) -0.005 (0.913) 1.335* (0.019) 0.748 (0.064) -0.119 (0.224) 0.059 (0.326) Bank Cost-to-income Ratio 0.005 (0.136) 0.052 (0.192) -0.006 (0.190) -0.005** (0.008) -0.012* (0.015) -0.026*** (0.000) -0.001 (0.691) -0.005* (0.019) GDP Growth 0.013 (0.217) 0.228* (0.031) -0.049 (0.010) 0.017** (0.001) GDP Deflator -0.015 (0.860) 0.291** (0.003) -0.025 (0.338) 0.001 (0.832) Current Account / GDP 0.007 (0.481) 0.026* (0.017) -0.023 (0.319) 0.004 (0.331) Unemployment -0.015 (0.371) 0.052* (0.027) 0.160*** (0.000) -0.029*** (0.000) Real Effective Exchange Rate 0.002 (0.221) 0.026** (0.007) 0.008* (0.033) -0.001 (0.116) Real Short-term Interest Rate -0.005 (0.513) 0.365* (0.015) -0.029 (0.092) -0.000 (0.985) constant 3.850 (0.105) 5.226 (0.166) -4.745 (0.164) -11.842*** (0.000) 8.137* (0.024) -8.306 (0.211) 1.227 (0.555) 4.555* (0.016)
Country Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes
Time Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes
R2 within 0.024 0.023 0.008 0.076 0.073 0.106 0.049 0.109 R2 between 0.007 0.013 0.056 0.123 0.095 0.071 0.011 0.014 R2 overall 0.055 0.015 0.062 0.087 0.067 0.010 0.008 0.035 AIC 2568 459 4739 2118 18532 15696 6641 3892 BIC 2641 372 4798 2558 18605 15808 6714 4005 N 12082 9510 12046 9508 11702 9376 12082 9510 VIF 7.99 8.38 7.95 8.36 7.52 8.59 8.00 8.34 * p<0.05, ** p<0.01, *** p<0.001
p-values in parentheses, robust standard errors adjusted R2-values
