The eectiveness of (macro)prudential policies in mitigating credit and house price growth in Asia

The effectiveness of (macro)prudential policies in mitigating

credit and house price growth in Asia

Jesse Huijts

Student Number: 10479988

Supervisor: K. Mavromatis

University of Amsterdam (UvA) and De Nederlandsche Bank (DNB)

Master Thesis MSc Economics

Word count: 13725

Abstract

Using a recent quarterly (macro)prudential database compiled by Cerutti et al. (2016) and ex-panding on Zhang and Zoli (2016) and Cerutti et al. (2017) this thesis studies the effectiveness of prudential tools in mitigating real credit growth and real house price growth for 14 Asia-Pacific economies over the period 2000Q1 to 2014Q4. Macro panel regressions testing two baseline models for credit and house prices find evidence for the effectiveness of the prudential stance in curbing credit expansion, and the cumulative prudential stance in curbing procyclicality of house price appreciation. Several individual (macro)prudential instruments and cumulative (macro)prudential instruments pro-vide epro-vidence of effective mitigation of credit and house price growth and procyclicality. One key insight, inspired by Azisa and Shin (2015), is that more attention should be paid to how global liquid-ity affects financial stabilliquid-ity in Asia, as well as to the interrelation between credit crises and currency crises in Asia. Future research studying the effectiveness of (macro)prudential policies in Asia should take into account noncore bank liabilities and capital flows, as well as capital flow management.

Contents

Statement of Originality 4

Acknowledgements 5

1 Introduction 5

2 Literature Review on (Macro)prudential policy effectiveness in Asia 6

2.1 (Macro)prudential policy framework: determining its effectiveness . . . 6

2.2 (Macro)prudential policy effectiveness in Asia . . . 9

3 Selecting Variables for Empirical Analysis 11 3.1 Dependent variables . . . 11

3.2 Database on (macro)prudential policies . . . 13

3.3 Control variables . . . 16

4 Methodology and Empirical Analysis 16 5 Results of the Empirical Analysis 18 5.1 Credit growth . . . 19

5.1.1 OLS methods and endogeneity . . . 19

5.1.2 Aggregate prudential indexes . . . 19

5.1.3 Prudential instruments . . . 21

5.1.4 Procyclicality in credit growth . . . 25

5.2 House price growth . . . 30

5.2.1 Aggregate prudential indexes . . . 30

5.2.2 Prudential instruments . . . 30

5.2.3 Procyclicality in house price growth . . . 35

6 Robustness 40

7 Conclusions 45

Dedication 48

Appendix: country coverage Asia-Pacific 49

Statement of Originality

This document is written by Jesse Huijts who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Acknowledgements

I am very grateful for the recommendation for the choice of topic and the many helpful comments from my supervisor, Kostas Mavromatis. Also, I would like to express a warm word of gratitude to Gabriele Galati for the helpful advice on empirics of prudential policy effectiveness and to Maurice Bun for helpful comments on panel econometrics.

1

Introduction

Macroprudential policies have the objective of improving stability in the financial system with prudential means in order to avoid or reduce macroeconomic costs from financial distress. This thesis will focus on evaluating the effectiveness of (macro)prudential policies in addressing the vulnerabilities of Asian financial systems via the time varying and the cross-sectional dimension of systemic risk. Hence, this study will take a positive approach, in documenting and evaluating which (macro)prudential tools have been established in 14 Asian economies and whether the policies have been successful in mitigating the vulnerability of the financial sector. Although there are several studies concerning (macro)prudential policy use and effectiveness in Asia, the analysis of their effectiveness on a long time period with higher than annual frequency data has been scarce. And related, few analyses exist on what policies are most effective in providing a reduction in procyclicality in financial markets and associated systemic risks. This thesis aims to fill these two gaps.

Regarding the methodology, typically it is quite challenging to access the appropriate data for macro-prudential policy actions. However pioneering recent advancements have been made with the launch of a new database which documents prudential tools spanning 64 economies with quarterly data for the period 2000Q1 through 2014Q4 (Cerutti et al. 2016). Five types of (macro)prudential instruments are included in the database, namely capital buffers, concentration limits, interbank exposure limits, concentration limits, loan-to-value (LTV) ratio limits and reserve requirements. Further decompositions deliver nine prudential tools. Here we expand on using the term (macro)prudential with macro in brackets. Namely, while the main aim of this thesis is to explore how macroprudential policies, as listed above, affect target variables, the database also lists microprudential policies. Therefore formally we are studying the effectiveness of prudential policies in general in mitigating systemic risk, whereby the large majority of instruments are in fact macroprudential instruments. In fact, as has been shown by Cerutti et al. (2016), all the pruden-tial instruments have impact on systemic risk, showing the predominantly macroprudenpruden-tial nature of the database.

testing for the model of Zhang and Zoli (2016) with more sophisticated insights from Cerutti et al. (2017). Specifically, this thesis employs cross country two-stage least squares first-differenced dynamic panel re-gressions (FD2SLS) using the Anderson-Hsiao (1981) estimator with quarterly data on 14 Asia-Pacific advanced and emerging economies (see appendix for country coverage) in the period 2000Q1 through 2014Q4. The focus of the study is to analyse how effective the documented prudential tools, as well as other tools, have been in containing developments in aggregate and sectoral credit growth and house prices growth.

The thesis proceeds as follows. Section II will provide a literature review of the (macro)prudential policy effectiveness studies focusing on cross-country panel regressions, whilst first highlighting some es-sential features of studies regarding (macro)prudential policy effectiveness in general, as well as several key issues related to its effectiveness in practice. Section III shall discuss the data selection, as well as issues concerning this task, together with a discussion of the identification of relevant (macro)prudential objectives and policies. Section IV will provide the methodology and the empirical analysis. Section V will provide a discussion of the results. Section VI will check the robustness of the results using Arellano-Bond (1991) GMM estimation. Lastly, Section VII concludes.

2

Literature Review on (Macro)prudential policy effectiveness

in Asia

2.1

(Macro)prudential policy framework: determining its effectiveness

After the Great Financial Crisis, increasing consensus in academia and among policymakers emerged that financial institutions had taken on excessive risk in the boom of the economy, given perceptions of low risk during a robust and a low interest rate environment. In hindsight, strong beliefs existed that the accumulation of debt and leverage in the economy, as a result from the credit and asset prices surges, would have been able to adjust itself to stable levels. However, the impact of the increasingly opaque and complex financial system, reflected by the increase in financial innovation and regulation, on the boom in the economy and the unwinding of financial imbalances had not been accurately identified. Altogether, these developments have led to an increased focus on studying a macro-based approach to financial regulation and supervision besides the widespread traditionally used micro-based approach (Galati and Moessner, 2011).

Theoretical research on macroprudential policies is divided into two strands. One strand underlines that individual price-taking agents do no internalise the general equilibrium impact of their decisions by taking on too much credit. Given this feature, macroprudential policy shows to be useful and valuable

in providing a way to incentivise agents to internalise this negative externality (Korinek, 2009; Bianchi and Mendoza, 2010; Jeanne and Korinek 2010, and Bianchi, 2011). The second strand emphasises the existence of procyclicality in the financial system as a result of financial frictions. Financial frictions are either modelled from a demand-side approach of credit by usage of the financial accelerator mechanism conceptualised by Bernanke et al. (1999) such as in studies by Kannan et al. (2012), Unsal (2013) and Medina and Roldos (2013), or modeled from a supply-side approach of credit which includes a general equilibrium framework with a banking sector such as in studies by Angeloni and Faia (2013) and Angeloni and Marco Lo Duca (2013). In this analytical framework with financial frictions macroprudential policy could provide a dampening of procyclicality and a mitigation of cyclical macroeconomic shocks.

Given developments towards the introduction of an increasing number of macroprudential policies over time, there has been an increasing volume of studies conducted regarding the link between macroprudential policies and financial stability. However, these empirical analyses are constraint by the fact that there is a relative absence of established models concerning the interaction between the financial system and the real economy, as well as the relative lack of data required to conduct these studies. In general, there are four main approaches: assessments of authorities and outside observers, event studies, reduced-form regression analysis based on micro data, i.e. individual bank balance sheets, and lastly, cross-country panel regressions. Given the scope of this thesis, the literature review will be limited to the latter two approaches.

Regarding the empirical approaches, there are three notable challenges in the analysis of macropruden-tial policy effectiveness (Galati and Moessner, 2017). In the first place, there is the issue of endogeneity, which refers to the identification problem of the impact of macroprudential tools on macroeconomic and financial variables as well as the related issue of distinguishing between correlation and causation. In brief, it is clear that in response to the development of macroeconomic and financial variables, macroprudential policy tools could be adopted. Secondly, there is the issue of distinguishing between the effects of macro-prudential policy and other policies, including monetary policy and fiscal policy, which are frequently used in conjunction with macroprudential policies. And lastly, there is the issue of the requirement to control for certain global and local factors. Below, previous literature related to the empirical approach of this period shall be critiqued on the basis of resolving the aforementioned challenges, and the strategy in this thesis for dealing with these issues shall be elucidated.

The first group involves micro-level evidence based research. Research which employs panel regressions with micro data mitigating the endogeneity problem, controlling for local and global factors and controlling for other policies is considered promising. In comparison to the cross-country approach, it would be an appropriate method to resolve the endogeneity issue given that macroprudential policies are less likely to change in response to individual bank level than to aggregate country level developments. However,

on the contrary, these regressions typically suffer from the micro aggregation problem, given that it may be difficult to present individual bank developments as being representative for the country level. There are two drawbacks of using this reduced-form regression analysis. Firstly, the interrelationship between macroeconomic, financial and policy variables is not captured well using reduced-form models. Secondly, there are data limitations since on average policymakers have adopted macroprudential policies only in recent years. One strategy for overcoming the endogeneity issue in panel regressions is by using general method of moments (GMM) estimation, by which explanatory variables are included lagged once and instrumental variables are used (Cerutti et al. 2015; Zhang and Zoli 2016; Claessens et al. 2013). An example of a relevant study that employs a panel regressions with GMM estimation for individual balance sheet data of roughly 2800 banks in 48 countries covering the period 2000 to 2010 is by Claessens et al. (2013). They also further address local and global factors in their regressions by including variables such as real GDP growth, degree of openness, exchange rate regime and financial structure measures for specific countries, e.g. related to the dominant type of financing. Additionally, they also control for the monetary policy stance by adding the changes in interest rates, as well as for fiscal policy by adding the public debt to GDP ratio.

The second group is concerned with cross-country research studying the relation between macropruden-tial policies and credit growth and other financial indicators. Subgroups involve cross-country research on the relation between macroprudential policies and risks of a financial crisis and banking and international finance developments, as well as studies focusing specifically on real estate market developments. One of the first studies in this group was conducted by Lim et al. (2011) and analysed the relation between macro-prudential policies and developments in credit and leverage. They found evidence that the policies such as LTV and DTI limits, reserve requirements (RR), dynamic provisioning rules and ceilings on credit growth are conjuncted with declining procyclicality of credit and leverage. Furthermore, Dell’Ariccia et al. (2012) find that credit surges could be reduced by macroprudential policies. Also, the likelihood that a credit boom goes bust could be significantly decreased by these policies. Using the database constructed from the IMF survey, Akinci and Olmstead-Rumsey (2015) analyse macroprudential policies in 57 advanced and emerging economies which cover the period 2000Q1 to 2013Q4 recording loosening and tightening episodes for 7 macroprudential tools. Their findings are that tightening macroprudential policy episodes tend to decrease bank credit growth, housing credit growth, and house price inflation. Additionally, it appeared that the extent of the targeting of the policies mattered for the effectiveness level. Also, capital flow restrictions focusing on inflows in emerging market economies tend to have complementary roles Aysan et al. (2015). Using the database compiled using the IMF survey named Global Macroprudential Policy Instruments (GMPI) conducted by the Monetary and Capital Market Department of the IMF in the period 2013-2014, Aysan et al. (2015) study 6 macroprudential tools in 18 emerging economies. Their findings

are that macroprudential policies that target borrowers tend to be effective in reducing credit growth, and that the macroprudential policies that target financial institutions tend to aid to reduce the impact of capital flows on domestic credit, specifically for foreign exchange related measures. Additionally, they find that the effectiveness of policies comes after 2 to 3 lags, a larger impact on more evident financial cycles and some evidence for complementarities among tools. Furthermore, using the same database, Cizel et al. (2015) find evidence that policies may stimulate substitution from bank to non-bank credit, specifically in advanced countries. Also, it appeared that quantity-based measures tend to have more pronounced effects than price-based measures, yet at the same time also leading to more substitution to non-bank credit in advanced economies.

Furthermore, there are several cross-country studies that limit their scope to real estate markets. Namely, Crowe et al. (2011) and Cerutti et al. (2015) find that policies including maximum LTV tend to be most effective in mitigating real estate price booms. In a similar fashion, IMF (2011) finds LTV tools tend to be most effective in mitigating spillover between credit and asset prices and in reducing the shocks to prices. Moreover, a study that focuses on real estate developments and macroprudential policy implementation in Asia-Pacific economies is conducted by Kuttner and Shim (2012). Constructing a database of the macroprudential policy actions by 57 economies taken over a period of 30 years, as well as constructing a database of the housing finance system’s structural characteristics, they were able to establish a clear link between interest rates and macroprudential policy actions as well as movements in real housing prices and real housing credit. Macroprudential tools, including maximum LTV and DSTI ratios, provisioning requirements, real estate exposure limits and risk weights were all found to be jointly significant in the regression. However, in comparison, the only policies that were able to impact house price appreciation were housing-related taxes, as in line with Vandenbussche et al. (2015).

2.2

(Macro)prudential policy effectiveness in Asia

The focus of this thesis is on the Asian region in studying the effectiveness of macroprudential policies. This would not be a trivial choice given the fact that macroprudential instruments have been used more frequently and intensely in Asia than in other regions, especially targeting the housing markets, which assigns a pioneering role to Asia in this perspective.

There are several studies on single countries employing micro-level evidence, frequently for a limited number of macroprudential policies for Asian countries. Igan and Kang (2011) use sectoral data to find that LTV and DTI limits dampened housing credit growth in Korea. For Hong Kong, Wong et al. (2011) find that real estate cycles are reduced by the usage of macroprudential policies which target real estate borrowing.

Cross-country literature for the Asian region, delivers several key studies with regards to this thesis. In the first place there is the comprehensive and solid research by Cerutti et al. (2017) which reports the use of macroprudential policies for 119 countries over the 2000-2013 period covering a wide range of instruments. They employ a panel regression using Arellano-Bond GMM estimation to study the effects of macroprudential actions on credit growth alongside several control variables, several extensions from a base model, including robustness tests. The specified models deliver useful reference outcomes with respect to studying the effectiveness in Asia. Regressions that differentiate by level of income deliver the the result that the statistically significant negative association between the macroprudential index and credit growth is largest for emerging and developing economies, compared to advanced economies. The authors ascribe a number of reasons to this observation. In the first place, emerging economies tend to have employed macroprudential policies more frequently than advanced economies. And in the second place, effectiveness in advanced economies may be constrained by the alternative sources of finance and regulatory arbitrage practices that characterise a developed financial system. Altogether, emerging and developing economies are concluded to have been able to employ macroprudential policies more effectively. Given the large set of emerging economies in our sample of Asian countries, this would be a hypothesis to bear in mind for the analysis in this thesis. Regressions that differentiate by the level of capital account openness deliver the result that macroprudential policies are more effective for relatively closed eocnomies and less effective for relatively open economies. A number of reasons are ascribed to this result by the authors. In the first place, open economies face more circumvention of macroprudential polices. Another reason may be that closed economies have less liberalised financial systems, thus reducing the leakage of macroprudential policies. Given the relatively large share of closed economies in Asia, this hypothesis will be taken in consideration in this thesis. Additionally, there is the finding that the coefficients for the lagged credit growth variable are largest for advanced economies compared to emerging and developing economies, which may reflect a higher level of stability in credit developments in advanced economies. Also, the coefficient on GDP growth is smallest and not statistically significant for advanced economies, which may reflect that credit developments are less related to the real economy in these economies. For the interest rate variable, we note that monetary policy seems to be more effective in curbing credit growth in developing economies, compared to advanced and emerging economies. And, in open economies there similarly seems to have been a smaller effect of the policy rate on credit growth. Lastly, the coefficients of the global risk variable, the banking crisis dummy, is larger in emerging markets and financially closed economies compared to advanced economies. A reason for this may be that emerging and developing economies are largely dependent on bank finance, as well as may be constrained the lacking effectiveness of fiscal or monetary policies in overcoming crises.

macroprudential policies and capital flow measures in 13 Asian economies and 22 economies in other regions from 2000Q1 to 2014Q1. The constructed indices are used to analyse whether the overall macroprudential policy stance and its interaction with monetary policy has been effective in mitigating credit and house prices through an event study, cross-country panel regressions, and bank-level micro panel regressions, which in the literature has been named the ’suite of models’ approach. Limiting our scope to the cross-country panel regression results for Asia, we note that with the usage of Arellano-Bond GMM procedure, their findings are that housing related measure have dampened private credit growth in Asia, however, no significance of other policies was found. On average in Asia, tightening of house-related macroprudential tools has been estimated to reduce credit growth by 0.7 percentage point after one quarter and by 1.5 per-centage points after one year. An additional panel regression has shown that on average macroprudential policy has lowered procyclicality in credit by 27 percent Asia, compared to 13 percent in other regions. Furthermore, housing-related macroprudential policies have mitigated house price appreciation Asia, with a tightening found to have lowered house price appreciation by 2 percentage points after one quarter.

3

Selecting Variables for Empirical Analysis

This section will describe the data that is used and documents the macroprudential policy actions which will be studied. A list of the main summary statistics of all the variables is given in table 1.

3.1

Dependent variables

We employ dynamic panel regression with quarterly real credit growth and quarterly real house price growth as dependent variables. Real credit growth is defined as the quarter on quarter real credit growth in percentages. For the Asian economies data is obtain from the BIS where available. Specifically, data is obtained from the entry named Adjusted BIS Domestic Bank Credit to Private non-financial sector, with the exception of the Philippines, Taiwan and Vietnam for which we used the entry Depository Corporations Domestic Claims on Private Sector from the IMF IFS database. These numbers were then deflated with quarterly CPI growth data. This data was obtained from the FRED database for China, Japan, India, Korea and New Zealand, from the BIS database for Hong Kong, Singapore, Malaysia, Philippines and Thailand, and ultimately, for Taiwan the data was obtained from the National Statistics Office and for Vietnam from the General Statistics Office. The second dependent variable is real house price growth, which is defined as the quarter on quarter house price growth in percentages. Data for this variable is obtained from the BIS database ’Long series on nominal residential property prices’ under the specific entry named the quarterly residential property price index series for Hong Kong, Australia, Japan, Korea,

Table 1: Summary Statistics

mean sd min max

credit .0438812 .1001308 -.1426254 .8584469 house .0025957 .0356652 -.1588 .1479415 sscb res .0130952 .1417411 -1 1 sscb cons .0011905 .0597852 -1 1 sscb oth .0011905 .0771884 -1 1 sscb .0154762 .2150444 -3 2 cap req .025641 .1581633 0 1 concrat .0198511 .1396619 0 1 ibex .0333333 .1805111 0 1 ltv cap .0696056 .3338006 -1 1 rr foreign .002381 .1760216 -2 2 rr local .0202381 .3100539 -2 1 cum sscb res .3511905 .8180027 0 4 cum sscb cons .0345238 .3164139 -1 1 cum sscb oth .0345238 .3417664 -1 2 cum sscb .4202381 1.266915 -2 6

cum cap req .2833333 .5759218 0 2

cum concrat .7593052 .8488397 0 3 cum ibex 1.511111 .502677 1 2 cum ltv cap 1.296984 2.510437 -3 8 cum rr foreign -.0035714 .8630885 -3 3 cum rr local .5142857 2.15043 -4 13 PruC .0988095 .4239533 -1 1 cum PruC 2.385714 4.826307 -5 25 PruC2 .0988095 .4239533 -1 1 cum PruC2 2.52619 4.871584 -4 25 GDP .0201001 .0427696 -.1975039 .1434524 policy rate 4.140297 2.90678 .1 17.62 DVIX -.1098333 6.483358 -13.6 33.53 Observations 840

Malaysia, New Zealand, Thailand and Singapore. For the remaining countries data has been obtained from Datastream.

3.2

Database on (macro)prudential policies

There are several constraints regarding the availability of data concerning actual macroprudential policy actions. The fact that this information has been limited is in part due to the fact that the use of tools has not been correctly defined as macroprudential. Namely, several countries have clearly defined a macro-prudential framework, while many still have not. Previous notable efforts have been made in collecting information on macroprudential policy actions. This has been performed for a relatively small set of 42 countries by the IMF in the study by Lim et al. (2011). However, more notable are relatively recent efforts in compiling a database, namely in the comprehensive IMF survey, named Global Macroprudential Policy Instruments (GMPI), which was set up by the Monetary and Capital Market Department of the IMF during 2013-2014. Another groundbreaking effort has been the new database compiled and described by Cerutti et al. (2016) as part of the 2015 International Banking Research Network (IBRN) initiative. This database highlights the changes in the intensity and the implementation of several widely performed prudential tools focusing on both macroprudential and microprudential objectives. The database covers quarterly data for the period 2000Q1 through 2014Q4 spanning 64 countries. Given that the database provides exclusive high frequency coverage for the sample of 14 Asia-Pacifc economies (see the appendix for the country coverage) which were chosen for this thesis, this thesis will adopt the IBRN 2015 database in its empirical analysis.

To elaborate on the content of the database, the database covers five types of macroprudential instru-ments, namely capital buffers, interbank exposure limits, concentration limits, loan-to-value (LTV) ratio limits, and reserve requirements. Within these five types, there is a total of nine prudential tools that have been constructed according to the following breakdowns. In the first place, capital buffers can be divided into four sub-indices, namely general capital requirements, real estate credit related specific capital buffers, consumer credit related specific capital buffers, and other specific capital buffers. Secondly, reserve requirements can be subdivided into domestic currency reserve requirements and foreign currency reserve requirements. It should be emphasised that the database covers prudential instruments independent of their microprudential or macroprudential objective. The database is aggregated with information from primary sources, such as central bank reports, as well as secondary sources such as the Global Macro-prudential Policy Instruments (GMPI) survey. The database represents a further progress in measuring the use of prudential tools. The consistency of the database is ensured by the direct feedback from the regulatory authorities of the countries which are included on the correctness of the policy changes which

are recorded in the database. Furthermore, the database reflects ongoing progress in efforts to measure prudential policy across on a large cross-country basis (Cerutti et al. 2016).

We shall now elaborate on the construction of the prudential instrument indices, as well as on the details on specific prudential instruments that are included in the empirical analysis below. There are several representations of the index data to be distinguished. In the first place, there is the form in which changes in a policy instrument are recorded with a 1 in case a prudential tool was tightened and -1 in case a prudential tools was loosened, and a 0 in case no change occurred in a given quarter. The advantages of this method is that the intensity of a policy change can be captured, which can measure the direction of the policy change. The second class of index, named the ’cumulative index’ is defined as the sum since 2000Q1 of the total of changes in the policy index before and during the quarter of interest. The objective of this cumulative index is to capture the ’tightness’ or ’looseness’ level of a certain instrument at a given point in time. However, one should cautiously use this instrument if one is conducting cross-sectional comparisons, given that differences in the policy stance across countries are difficult to disentangle from the data.

We shall now delve into the details of the specific prudential instruments. The first instrument is the general capital requirements index, which documents the changes introduced in the Basel Accords through the four revisions: I, II, II.5, and III. The index takes the value 1 when capital regulation is introduced or tightened, or 0 when no changes in the capital regulations occur. In the empirical analysis the change in the capital requirements, i.e. Basel capital agreement implementations, is denoted by cap req, and the cumulative change in capital requirements is denoted by cum cap req. The second instrument is the sector specific capital buffer index, which is an indicator that documents changes in regulation targeting a reduction in bank claims to specific sectors of the economy. Innovations in this type of prudential instruments are proxied by adjustments to risk-weights of specific bank exposures, as these are tightened or loosened with the financial cycle. Three categories of credit are distinguished according the type of borrower: real estate credit, consumer credit, and other credit. Furthermore, the aggregate sector specific capital buffer index is defined by the sum of the changes of a prudential instrument across the three categories of credit, such that is is allowed to take on values greater or lower than 1 and -1 in a given quarter signalling changes in capital buffers for multiple sectors simultaneously. In the panel regressions, the change in real estate credit capital buffers is denoted by sscb res, similarly for consumer credit and other sectors these are denoted by sscb cons and sscb oth respectively. Cumulative indexes are again denoted by cum sscb res, cum sscb cons and cum sscb oth respectively. The aggregate sector specific capital buffers index is denoted by sscb, and the cumulative version is denoted by cum sscb. Thirdly, there is the reserve requirement instrument. These instruments are typically reported numerically, but may be applicable to different account types. Namely, there are several subcategories such as demand and

savings accounts included in the deposit accounts for example. At the same time, reserve requirements may be tailored to deposits that differ in maturities. A numeric index is used which can also take values above 1 and below -1 to document the intensity in the changes. Furthermore, one should note that changes in reserve requirements for deposit accounts are documented separately: those denominated in domestic and those denominated in foreign currency. In the empirical analysis, the change in the reserve requirements on local currency and foreign currency denominated accounts are denoted by rr local and rr foreign respectively. Their cumulative indexes are denoted by cum rr local and cum rr foreign respectively. In the fourth place, we have the concentration limits and interbank exposure limits. These instruments are characterised by a wide set of policies that affect claims between banks and their borrowers and the instruments can be changed by adjusting five elements that characterise these exposures: the definition of large exposures, the level of the limit, the differentiation across counterparties, aggregate limits and sectors and assets that are covered by the regulation (see Cerutti et al. 2016 for further details). With the usage of the GMPI survey changes in the index were recorded relying on several assumptions explored in Cerutti et al. (2016). In the empirical analysis, changes in concentration limits are denoted by concrat and changes in interbank exposure limits are denoted by ibex. Similarly, cumulative indexes are denoted by cum concrat and cum ibex respectively. The last instrument type is the loan-to-value (LTV) ratio limit, which is defined as a type of restriction of the amount that an individual or firm can borrow against their collateral, most commonly applied to real estate transactions. The related index documents changes in the LTV ratio limits that affect transactions in real estate. Additionally, we should note that for one country in our sample of 14 Asian countries, namely Hong Kong, also changes related to the maximum amount insured in transactions involving real estate are included. In the empirical analysis, the change in the loan-to-value (LTV) ratio cap is denoted by ltv cap and equivalently the cumulative index is denoted by cum ltv cap. Ultimately, there are several variables that focus on the total of all the prudential instruments that should be highlighted. Two aggregate prudential indexes are used in the empirical analysis. Firstly, there is the PruC index, which documents country indexes by time t and country c equaling 1 in case the sum of the nine instruments is great than or equal to 1 and equaling -1 in case the sum of the instruments in lower than or equal to -1, and 0 otherwise. Secondly, there is the PruC2 index which documents country indexes by time t and country c with the where the index takes values in a similar manner as PruC, but in this case all the individual instruments are modified to take maximum and minimum changes of 1 and -1. The connected cumulative indexes are denoted by cum PruC and cum PruC2 respectively.

3.3

Control variables

We include several control variables to proxy for macroeconomic and financial factors both local and global. In the first place, in line with studies such as Zhang and Zoli (2016), Cerutti et al. (2017) and Bruno et al. (2017) we include the growth of real GDP, which is defined as the year on year real GDP growth in percentages, in order to proxy for credit demand. We obtained the data from the iOECD database for Japan, Korea, New Zealand, India, Indonesia and Australia. The remaining data was obtained from various country level authority sources. Again in line with previous research by Zhang and Zoli (2016) and Cerutti et al. (2017) the second control is the quarterly policy rate, which is the quarterly monetary policy rate in percentages or the quarterly money market rate in percentages, in order to capture the monetary policy stance. The objective was to obtain data on the Central Bank Policy Rate from IFS when available, and otherwise the Discount Rate or Repurchase Agreement Rate or the Money Market Rate, all from the IFS. For Australia, Korea, New Zealand, Singapore, Malaysia, Indonesia and Vietnam we were able to obtain data on the Central Bank Policy Rate. For China, Hong Kong, Japan, India and Thailand we were able to obtain data on the Discount Rate. For the Philippines we obtained the quarterly rediscount rate from the Bangko Sentral ng Pilipinas. Lastly, for Taiwan we were able to find the quarterly money market rate data from the Central Bank of the Republic of China (Taiwan). Ultimately, in line with Zhang and Zoli (2016), we control for global factors by including the VIX index (in logs) as well as its change. The VIX index is the implied volatility of the S&P500 index derived from option prices and according to research the VIX can proxy for the leverage of global banks (Bruno and Shin, 2015) as well as for the risk appetite of global investors, particularly in bond markets (Ahmed and Zlate, 2013). Quarterly data on the VIX and its change are obtained from the FRED database. Lastly, in line with Zhang and Zoli (2016) and Cerutti et al. (2017) we also control for local factors by including a country fixed effect to capture any country specific conditions which do not vary over time, namely part of a country’s economic and financial development, the degree of concentration of its financial system, the shares of market versus bank related financial intermediation in the financial system as well as numerous other, often institutional, features.

4

Methodology and Empirical Analysis

For the empirical analysis we shall conduct fixed-effect dynamic panel regressions. One should note that we resort to alternative methods given that OLS fixed effect is likely biased due to the presence of a lagged dependent variable and a country fixed effect in our regression. Furthermore, two notable issues should be resolved or mitigated. In the first place, as stated above, if we apply within transformation the demeaned lagged dependent variable is correlated with the demeaned error term. This issue requires

therefore the need for a different estimator than standard OLS. Secondly, macroprudential policies are endogenous variables in the empirical analysis. More specifically, there is a high likelihood of attenuation bias as a result of reverse causation. Namely, it is expected that a tightening of macroprudential policies dampens credit growth and house price growth. Simultaneously, the fact that credit and house prices grow is expected to increase the likelihood of a tightening of macroprudential policies. Referring to this bias, Zhang and Zoli (2016) argue that for the macroprudential policy index variable the existence of a negative significant coefficient should be interpreted as the lower bound for the absolute value of the population coefficient. In this methodology section, we shall analyse how the MPI and other indexes affect the growth of real credit and real house prices in the following base regressions, which were inspired by Zhang and Zoli (2016):

Yi,t = β0+ β1Yi,t−1+ β2M P Ii,t−1+ β3GDPi,t−1+ β4P olicyi,t−1+ β5V IXi,t+ µi+ i,t (1)

Yi,t = β0+ β1Yi,t−1+ β2cum M P Ii,t∗ GDPi,t+ β4P olicyi,t+ β5V IXi,t+ µi+ i,t (2)

In these specifications Y stands for the dependent variable, so either quarterly real credit growth or quar-terly real house price growth. The M P I stands for aggregated prudential indexes, individual instruments or cumulative indexes and instruments in subsequent regressions. Furthermore, GDP is the quarterly real GDP growth, P olicy denotes the quarterly monetary policy rate to capture the monetary policy stance, V IX stands for global risk and is captured by the quarterly change in the VIX index. Lastly, µ denotes the country specific fixed effects variable.

The first regression is formulated to analyse the extent to which macroprudential policies affect real credit growth and real house prices growth while controlling for local and global factors. The second spec-ification is formulated to analyse the extent to which the macroprudential stance affects the procyclicality of credit and house prices growth. This specification in macroprudential policy effectiveness studies has been introduced firstly by Lim et al. (2011) and was subsequently adopted in later research (Zhang and Zoli, 2016; Bruno et al. 2017). The idea behind the interaction term of GDP with the cumulative macro-prudential policy index denoted by cum M P Ii,t∗GDPi,t is to isolate the effect of the response of credit and

house prices growth to innovations in GDP growth. This specification is motivated by the concept of the financial accelerator mechanism. In an economy financial frictions can make the result of macroeconomic shocks to credit, asset prices and leverage larger, and hence it is conjectured that the tightness level of macroprudential policies may reduce leverage, and hence mitigate the procyclicality of credit growth.

In contrast with previous research that aimed to resolve the aforementioned issues with the use of Arellano-Bond GMM estimation, we shall use dynamic panel regressions using IV methods, more specif-ically we shall use the Anderson-Hsiao (1981) estimator, which is an adaptation of the two-stage least-squares first-differenced estimator (FD2SLS) made suitable to fit a panel data model with a lagged

depen-dent variable. The reason that GMM estimation is not valid in our empirical analysis is that the property of our data involves a dataset with few panels and many time periods (i.e. small N and large T), whereas the opposite would be required, namely many panels and few time periods (i.e. large N and small T). For this reason we use IV methods. A following consideration is whether to use standard IV using excluded IV’s, or alternative IV using lagged values of endogenous regressors as IV’s, although the last option is only allowed in case we have no serial correlation in the error term. Given the absence of suitable external instruments, the latter method is used. The following procedure will be followed. First differenced IV models will be formulated for the two base panel regressions of the following form.

∆Yi,t = β1∆Yi,t−1+ β2∆M P Ii,t−1+ β3∆GDPi,t−1+ β4∆P olicyi,t−1+ β5∆V IXi,t+ ∆i,t (3)

∆Yi,t = β1∆Yi,t−1+ β2∆cum M P Ii,t∗ GDPi,t+ β3∆GDPi,t−1+ β4∆P olicyi,t−1+ β5∆V IXi,t+ ∆i,t (4)

Secondly, we shall include internal instrument IV’s, namely lagged values of the following regressors: credit growth, the macroprudential policy index, policy rate and GDP growth given their endogeneity as given by the endogeneity test. And we shall use the second lag of the same instrumented variable as an instrument. For example, for ∆M P Ii,t−1 we use ∆M P Ii,t−2 as an instrument and so forth. Next, we

tested whether this procedure has been valid by checking whether the error term is not serially correlated, which was confirmed. Tests for residual autocorrelated are given by for example Arellano and Bond (1991). Furthermore, we tested for the possibility of exogeneity of the instruments by using the overidentifying restrictions statistic. Lastly, a crucial consideration here is that including too many regressors may lead to overfitting, with as an accompanied result that this will lead to a generation of many instruments which may lead to bias in the IV estimators as well as the estimated standard errors. Equivalently, the J test will exhibit low power. In this thesis this issue will be avoided by aiming to formulate parsimonious regressions with only subsets of near lags as instruments. Furthermore, in order to address the previous concerns and in order to explore the implications for our model, we shall include period specific IV’s in adapted regression specifications, which will improve the efficiency of our regression models. Also, we shall include a set of quarter dummies, which are not included in the regression tables for reasons of brevity, in order to enrich the model and to improve the validity of the IV’s.

5

Results of the Empirical Analysis

This section will introduce the main results of the empirical analysis, by presenting the outcomes of the various panel regressions. Table 1 presents the most important summary statistics for the variables in the panel regressions. We observe the mean and standard deviation of all the variables, as well as their minimum and maximum values.

5.1

Credit growth

5.1.1 OLS methods and endogeneity

Firstly, we estimate equation 1 using OLS with three subsequent specifications. In the first column depicted in table 2 we have listed the output from a pooled OLS, where we ignored the panel characteristics of the data. We note biased results mainly due to the presence of a lagged dependent variable and country fixed effects. The coefficient on the lagged real credit growth variable is highly significant, with an excessively large p-value. Furthermore, the policy rate seems to be significant, but with the wrong sign, the coefficient on the aggregate prudential policy index variable is small and not significant. In the second column we have the output from an OLS specification with country fixed effects. We note that the bias in the lagged credit variable is decreased slightly, and also it improves the goodness of fit overall, however these are meagre improvements. Lastly, the third column displays the output from a first-differenced OLS. This time we do see a large decrease in the bias of the credit growth variable, although the p-value is still quite large. Further improvements are the correct sign and larger p-value of the prudential policy index variable as well as the policy rate variable. Both fe OLS and fd OLS classify as panel date data techniques, given that they permit invariant unobserved heterogeneity (individual specific effects). Furthermore, all the regressions above include clustered standard errors at the country level, given that their validity holds in case of i.i.d error terms. This provides indication that the strategy of first-differencing provides the most useful direction for the methodology.

5.1.2 Aggregate prudential indexes

After having observed the useful properties of first-differencing we shall now employ the empirical strategy which entails the aforementioned Anderson Hsiao (1981) estimator, which is the two-stage least-squares first-differenced estimator (FD2SLS) adapted to fit a model with a lagged independent variable. All regressions for the Anderson-Hsiao (1981) estimator are performed in STATA using the innovative xtivreg2 wrapper. The endogeneity test is used to check whether included instrumented variables are endogenous and hence are usefully included in the IV model. Unless states otherwise, all the instrumented variables are the lagged credit growth, lagged prudential policy index, the lagged policy rate and the lagged GDP growth. Moreover, we use the lagged variable of the instrumented regressor as internal instrument. Next, we also generated and included a range of period specific internal IV’s for the instrumented variables in order to improve estimation efficiency. Unless stated otherwise, these cover the full time period from 2000Q1 until 2014Q4, depending on the outcomes of the overidentifying restrictions test, the underidentification test and the weak identification test, with the most essential emphasis lying on the overidentifying restrictions test as this indicates whether instruments are valid. Lastly, unless stated otherwise, a full range of quarter

Table 2: OLS aggregate prudential index

(1) (2) (3)

crediti,t crediti,t ∆crediti,t

crediti,t−1 0.804∗∗∗ 0.718∗∗∗ (39.00) (29.49) ∆crediti,t−1 0.419∗∗∗ (7.55) P ruCi,t−1 0.000902 0.000163 (0.19) (0.03) ∆P ruCi,t−1 -0.00365 (-0.58) GDPi,t−1 0.0940∗∗ 0.00596 (1.96) (0.11) ∆GDPi,t−1 0.0174 (0.29) policy ratei,t−1 0.00146∗∗ -0.0000484

(2.07) (-0.04)

∆policy rate_i,t−1 -0.00244

(-0.44) V IXi,t 0.000336 0.000447∗ (1.07) (1.44) ∆V IXi,t 0.000149 (0.41) cons 0.00134 0.0132∗∗ 0.0439∗∗∗ (0.36) (2.30) (12.90) N 826 826 812 z statistics in parentheses ∗ _{p < 0.10,}∗∗ _{p < 0.05,}∗∗∗ _{p < 0.01}

dummies covering the entire time period are included as control variables in order to improve on the validity of the IV’s.

In table 3 we have the results from applying the Anderson-Hsiao (1981) estimator to the equation 1. We note that the prudential policy index is significant at the 5 percent level and has the correct sign. This implies that a quarter on quarter effect of a change in the prudential index which is a 100 percent change, so from 0 to 1 for example, is associated with a reduction of 6 percent in quarterly real credit growth. Furthermore, lagged credit growth is negative, -0.492, indicating ambiguous persistence in country level credit growth developments. As expected, the policy rate has a negative sign, however it has a small and insignificant coefficient. Similarly, GDP and the VIX have the right sign, namely positive and negative respectively, however they also have small and insignificant coefficients.

5.1.3 Prudential instruments

Next, we shall move on to study more individual instruments and their effect on credit developments, One of these are the capital requirements. In our regression table 4 we note that the cap req is not statistically significant, but has a negative sign as we would expect. It is rather surprising that capital requirements appear not significant. There are several reasons that could be proposed to explain this observation. In the first place, as outlined previously, given the high likelihood of attenuation bias we could expect the actual coefficient to be larger. Secondly, given the lack of quarterly changes of capital requirements, the quarterly change may have a lesser effect on credit growth than for example an annual change. Lastly, as has been stated before, there seems to be a high persistence in credit developments, which appear highly significant in almost all regression specifications, which points to the dominating effect of lagged credit. The procyclicality of credit in response to prudential policy changes will be tested further below.

We shall now study the regression results of the concentration limit and interbank exposure limit respectively. Firstly, the concrat variable is not statistically significant, and has the wrong sign. Again, the lagged credit developments appear to be highly significant at the 1 percent level, with other control variables all appearing wth small z statistics and small coefficients. This seems to point to the fact that quarterly changes in concentration limits do not provide any evidence of curbing quarterly credit growth. There are two side notes that should be made. In the first place, given the unbalanced nature of our panel dataset for quite some countries no data was available on concentration limit changes either because these tools did not exist in the respective country or because they have not been implemented in our sample. Secondly, even in those countries where concentration limits have been introduced we see very little changes being made on a quarterly basis, making the impact of changes in concentration limits on credit growth less meaningful. Furthermore, we should study the ibex variable, which has the correct

Table 3: Anderson-Hsiao (1981) FD2SLS aggregate prudential indexes

(1) (2) (3) (4)

∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t

∆crediti,t−1 -0.492∗∗∗ -0.466∗∗∗ -0.570 -0.543 (0.145) (0.127) (0.419) (0.378) ∆P ruCi,t−1 -0.0635∗∗ (0.0264) ∆P ruC2i,t−1 -0.0274∗ (0.0152)

∆cum P ruCi,t−1 -0.0905

(0.145)

∆cum P ruC2i,t−1 -0.0851

(0.136)

∆policy rate_i,t−1 -0.00473 0.00208 0.00333 0.00240

(0.0148) (0.0124) (0.0525) (0.0501) ∆GDPi,t−1 0.0265 -0.00868 0.130 0.148 (0.147) (0.127) (1.308) (1.250) ∆V IXi,t -0.0895 -0.0601 -0.108 -0.100 (0.107) (0.0954) (0.188) (0.175) cons 0.223 0.153 0.279 0.259 (0.272) (0.243) (0.487) (0.455) N 811 811 811 811

Standard errors in parentheses

sign, but no statistical significance is to be observed. This may be for similar reasons as listed for the concrat variable.

The loan-to-value (LTV) ratio has the expected negative sign, however the coefficient of -0.00341 is not statistically significant and rather low. In this specification we yet again note a rather strong persistence in credit developments. All other control variables poorly explain quarterly credit growth.

In contrast, the reserve requirements on foreign currency rr foreign is highly statistically significant at the 1 percent level. This points to evidence that reserve requirements on foreign currency already have a relatively large mitigating effect on real credit growth on a relatively high frequency, namely quarterly, basis. There could be several reasons for this observation. One essential empirical observation is that banks behave in a procyclical manner which influences capital flows. In case of rapid expansion of credit, banks require additional pools of retail deposits to fuel lending, and will therefore resort to other funding sources. These other sources are predominantly other banks which operate in the capital market as wholesale lenders, which brings us to the interrelation between currency and credit crises. Procyclicality in the behaviour of banks stimulates credit growth through capital inflows which enter via the banking sector. Thus, as is concluded from several empirical investigations, a rather consistent indicator of vulnerability for both currency and credit crisis is a high degree of foreign bank liabilities (Azisa and Shin, 2015). Given that foreign exchange denominated banking liabilities are a harbinger for the diagnosis of financial stability issues, addressing this issue via implementing reserve requirements on foreign currency may be a powerful and potent tool. Namely, this tool ensures building a buffer on the asset side to cover the foreign-exchange denominated liabilities side, and therefore is expected to hamper foreign liabilities’ inflows, and therefore reduce real credit growth. For similar reasons, reserve requirements for local currency rr local may be less effective since there is a currency mismatch between the credit fuelling foreign exchange liabilities and the local currency assets.

Sector specific capital buffers Furthermore, we now arrive at the aggregate sector specific capital buffers variable. In table 5 we have the regression output. In table 4 we noted that the sign of the aggregate sscb variable is correct, but there is no statistically significant coefficient, and the coefficient is rather small. We noted again a large, statistically significant at the 1 percent level, persistence in credit developments, with a negative sign. Zooming in on individual sector-specific capital buffers in table 5, we obtain more meaningful results. In the first place, the sscb res variable is statistically significant at the 20 percent level and has a large negative coefficient. In contrast, the sscb cons variable has a wrong sign and is not statistically significant, indicating that a change in the consumer credit capital buffer is not associated with a mitigating effect on credit developments. Partly, this result may be biased by the limited innovations of consumer credit capital buffers in our sample. However, instead the sscb oth,

Table 4: Anderson-Hsiao (1981) FD2SLS prudential instruments

(1) (2) (3) (4) (5) (6) (7)

∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t

∆crediti,t−1 -0.195∗∗∗ -0.158∗∗∗ -0.489∗∗∗ -0.366∗∗∗ -0.525∗∗∗ -0.196∗∗∗ -0.366∗∗∗

(0.0527) (0.0528) (0.120) (0.0752) (0.157) (0.0521) (0.115)

∆sscbi,t−1 -0.00968

(0.00865)

∆cap req_i,t−1 -0.000144

(0.0135) ∆concrati,t−1 0.0123 (0.0293) ∆ibexi,t−1 -0.00308 (0.0116) ∆ltv cap_i,t−1 -0.00341 (0.0134) ∆rr f oreign_i,t−1 -0.0306∗∗∗ (0.0101) ∆rr locali,t−1 -0.0122 (0.0160)

∆policy rate_i,t−1 -0.000937 0.000764 0.0000993 -0.00170 0.00541 0.000830 0.00327

(0.00541) (0.00652) (0.00813) (0.00468) (0.00891) (0.00539) (0.0117) ∆GDPi,t−1 0.0644 0.0211 -0.0291 0.966∗∗∗ -0.0114 0.0460 -0.0474 (0.0479) (0.0435) (0.0732) (0.148) (0.189) (0.0474) (0.122) cons -0.00210 -0.00571 -0.00793 -0.0000834 -0.0100 0.0429 -0.00280 (0.0152) (0.0145) (0.0197) (0.00249) (0.0148) (0.217) (0.0158) N 811 752 385 86 405 811 811

Standard errors in parentheses

indicating other credit capital buffers, is statistically significant at the 10 percent level, with a clear negative, mitigating effect on credit developments.

5.1.4 Procyclicality in credit growth

In this section we shall study the second regression specification of equation 2 where we test for the effect of cumulative indexes on the procyclicality of credit growth. In previous research, and notably in the case of Asia referring to Lim et al. (2011) and Zhang and Zoli (2016), the general macroprudential policy index seems to have been effective in dampening the procyclicality of credit growth in a whole country sample. However, this evidence seems to be absent in our case where we study the effect of cumulative prudential indexes on the procyclicality of credit growth, referring to table 6. Although the signs of the interaction terms are negative, which seems to be in line with previous research and our hypothesis, they are not significant. It seems that persistence in credit growth developments trumps the results of other variables, although several variables such as the policy rate, GDP and the VIX seem to provide correct signs.

Secondly, we shall study the effect of individual cumulative prudential instruments on the procyclicality of credit growth in table 7. After gauging the list of regression results, we note that the only interaction term which is significant involves the cumulative interbank exposure limit, along with the GDP growth variable in this specification. This variable is statistically significant at the 1 percent level with the following interpretation. While a 1 percentage point increase in GDP growth implies a 0.806 percentage point increase in credit growth, the cumulative interbank exposure limit mitigates the increase of GDP by -0.241 percentage points. The cumulative interbank exposure limit seems to implicate a rather large effect on dampening credit growth procyclicality. Let us analyse the possible reasons for these findings. Interbank exposure limits are typically implemented to increase the resilience of the banking system. Part of the answer may lie in the increasing amount of noncore liabilities in the Asian banking systems. Given that Asia is highly dependent on bank financing, the volatility of noncore liabilities entails risks of procyclicality of credit. Given that the definition for noncore liabilities is given by the sum of the following elements as documented by Shin and Shin (2010): foreign exchange denominated banking liabilities, certificates of deposits (CDs), repos, promissory notes and bank debt securitis that are funded via foreign credits, this builds up vulnerabilities to foreign credit deleveraging in the Asian financial systems. Given these interrelations, interbank exposure limits reduce the financial instruments that flow between foreign creditors and the banking sector, this immediately lowers the noncore liabilities of banks and therefore has the potential of reducing incentives to expand credit. Namely, the results from table 7 seem to be the result of the interbank exposure limit being rather effective at reducing procyclicality of real credit growth.

Table 5: Anderson-Hsiao (1981) FD2SLS sector-specific capital buffers

(1) (2) (3)

∆crediti,t ∆crediti,t ∆crediti,t

∆crediti,t−1 -0.353∗ -0.197∗∗∗ -0.184∗∗∗ (0.203) (0.0549) (0.0533) ∆sscb resi,t−1 -0.501∗ (0.285) ∆sscb consi,t−1 0.00227 (0.0324) ∆sscb othi,t−1 -0.0431∗ (0.0245)

∆policy rate_i,t−1 0.00859 0.000295 -0.00125

(0.0239) (0.00547) (0.00543) ∆GDPi,t−1 0.0104 0.0577 0.0619 (0.229) (0.0482) (0.0479) ∆V IXi,t 0.000213 0 0 (0.000569) (.) (.) cons 0.000272 -0.00212 -0.00209 (0.00416) (0.0152) (0.0151) N 811 811 811

Standard errors in parentheses

Table 6: Anderson-Hsiao (1981) FD2SLS procyclicality cumulative prudential index

(1) (2)

∆crediti,t ∆crediti,t

∆crediti,t−1 -0.202∗∗∗ -0.204∗∗∗

(0.0513) (0.0510)

∆cum P ruCi,t∗ GDPi,t -0.00144

(0.00350)

∆cum P ruC2i,t∗ GDPi,t -0.00211

(0.00352)

∆policy rate_i,t−1 -0.000808 -0.000563

(0.00518) (0.00517) ∆GDPi,t−1 0.0563 0.0531 (0.0517) (0.0518) ∆V IXi,t -0.000748 -0.000744 (0.00420) (0.00421) cons -0.000112 -0.0000432 (0.0108) (0.0108) N 811 811

Standard errors in parentheses

Table 7: Anderson-Hsiao (1981) FD2SLS procyclicality cumulative prudential instruments

(1) (2) (3) (4) (5) (6) (7)

∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t ∆crediti,t

∆crediti,t−1 -0.240∗∗∗ -0.0841∗ -0.142∗∗ -0.436∗∗∗ 0.0326 -0.171∗∗∗ -0.185∗∗∗

(0.0490) (0.0442) (0.0641) (0.0800) (0.0640) (0.0498) (0.0495)

∆cum sscbi,t∗ GDPi,t -0.0685

(0.0821)

∆cum cap reqi,t∗ GDPi,t -0.0358

(0.116)

∆cum concrati,t∗ GDPi,t 0.0187

(0.0778)

∆cum ibexi,t∗ GDPi,t -0.241∗

(0.138)

∆cum ltv capi,t∗ GDPi,t -0.00339

(0.00811)

∆cum rr f oreigni,t∗ GDPi,t 0.00138

(0.0224)

∆cum rr locali,t∗ GDPi,t -0.00149

(0.00696)

∆policy rate_i,t−1 -0.00171 0.00467 -0.00376 -0.00295 0.000217 -0.00145 -0.000351

(0.00525) (0.00559) (0.00493) (0.00471) (0.00486) (0.00514) (0.00513) ∆GDPi,t−1 0.0590 -0.00586 0.0109 0.806∗∗∗ 0.0356 0.0628 0.0426 (0.0477) (0.0443) (0.0540) (0.169) (0.0451) (0.0507) (0.0504) cons 0.000981 -0.00215 -0.00909 -0.00121 -0.00601 -0.00210 -0.00204 (0.0109) (0.0103) (0.0136) (0.00263) (0.00943) (0.0151) (0.0151) N 811 752 385 88 416 811 811

Standard errors in parentheses

Table 8: Anderson-Hsiao (1981) FD2SLS procyclicality cumulative sector specific index

(1) (2) (3)

∆crediti,t ∆crediti,t ∆crediti,t

∆crediti,t−1 -0.173∗∗∗ -0.222∗∗∗ -0.107

(0.0512) (0.0513) (0.0650)

∆cum sscb resi,t∗ GDPi,t -0.105

(0.155)

∆cum sscb consi,t∗ GDPi,t -0.310

(0.355)

∆cum sscb othi,t∗ GDPi,t -0.303

(0.254)

∆policy rate_i,t−1 -0.00140 -0.00146 0.000262

(0.00583) (0.00527) (0.00354) ∆GDPi,t−1 0.0246 0.0644 0.0263 (0.0377) (0.0478) (0.0375) ∆V IXi,t -0.00103 -0.000694 -0.000798 (0.00419) (0.00422) (0.00416) cons 0.00113 0.000472 -0.000269 (0.0108) (0.0108) (0.0107) N 811 811 811

Standard errors in parentheses

Lastly, we should study the effect of the cumulative sector specific capital buffers on the procyclicaliy of real credit growth. The results are in table 8, and we conclude that significant cumulative variables are absent. However, the coefficients are large and negative, indicating that there is some possibility for a mitigation, only it seems that the quarter on quarter effect is small.

5.2

House price growth

5.2.1 Aggregate prudential indexes

We shall now shift our focus to real house price growth. In table 9 we display the results from a regression with aggregate prudential indexes. Each of the prudential indexes, from PruC to cum PruC2 has a negative sign, but no significance occurs. This may either be because the credit persistence trumps the effects, or because of the few quarterly changes of the indexes in the sample. Noteworthy and important is the observation that in the specification with PruC and PruC2 the quarterly change in the VIX is significant at the 10 percent level and have the negative sign, which we hypothesised. This provides indication for the fact that the change in the VIX, the proxy for global risk, dampens house price growth on a quarterly basis. Hence, house prices are quite susceptible to changes in global risk perception. The results are quite in line with findings from Zhang and Zoli (2016). Namely, macroprudential policy indexes are found to be insignificant for Asia. Although this study also confirms the above result, it is still quite surprising given that various studies have established clear results relating macroprudential instruments to stabilisations in house price growth in different regions (Kuttner and Shim, 2012; Cerutti et al., 2015).

5.2.2 Prudential instruments

Next we study the effect of prudential instruments on real house price growth as documented in table 10. One immediately observes a rather strong heterogeneity in results, although none of the individual instruments are significant. Namely for sscb, cap req and ltv cap we document a negative coefficient. One reason these variables have a negative relationship with real house price growth could be that they would classify as being housing-related prudential measures. However, no significance is found for the variables, which is in line with aforementioned results from Zhang and Zoli (2016). In contrast, for ibex, concrat, rr foreign and rr local we document a positive coefficient, which would point to an ambiguous relationship of these variables in controlling real house price expansion. One reason for the interbank exposure limit and the concentration limit to appear to induce house price appreciation could be that when excessive liquidity is constrained banks may cut back on lending, but given persisting property price expansion, invest in real estate securities to fuel asset increases to make profits even when lending is constrained. This argument is strengthened by observing that GDP is positive and significant at the 5

Table 9: Anderson-Hsiao (1981) FD2SLS aggregate prudential indexes

(1) (2) (3) (4)

∆housei,t ∆housei,t ∆housei,t ∆housei,t

∆housei,t−1 -0.468∗∗∗ -0.483∗∗∗ -0.916 -0.930 (0.171) (0.144) (0.676) (0.679) ∆P ruCi,t−1 -0.0174 (0.0143) ∆P ruC2i,t−1 -0.0116 (0.00891)

∆cum P ruCi,t−1 -0.0584

(0.100)

∆cum P ruC2i,t−1 -0.0610

(0.101)

∆policy rate_i,t−1 -0.00370 -0.00409 -0.0283 -0.0281

(0.00698) (0.00650) (0.0348) (0.0345) ∆GDPi,t−1 -0.0386 -0.0373 0.101 0.116 (0.0561) (0.0527) (0.505) (0.512) ∆V IXi,t -0.0779∗ -0.0740∗ -0.0915 -0.0914 (0.0448) (0.0435) (0.0996) (0.0997) cons 0.194∗ 0.184∗ 0.230 0.229 (0.114) (0.111) (0.257) (0.257) N 693 693 693 693

Standard errors in parentheses

Table 10: Anderson-Hsiao (1981) FD2SLS prudential instruments

(1) (2) (3) (4) (5) (6) (7)

∆housei,t ∆housei,t ∆housei,t ∆housei,t ∆housei,t ∆housei,t ∆housei,t

∆housei,t−1 -0.252∗∗∗ -0.324∗∗∗ -0.0218 -0.238∗∗∗ -0.271∗∗∗ -0.323∗∗∗ -0.243∗∗∗

(0.0685) (0.0776) (0.213) (0.0892) (0.0694) (0.0608) (0.0663)

∆sscbi,t−1 -0.00283

(0.00558)

∆cap req_i,t−1 -0.00432

(0.00601) ∆ibexi,t−1 0.00169 (0.0101) ∆concrati,t−1 0.0108 (0.0109) ∆ltv cap_i,t−1 -0.00192 (0.00405) ∆rr f oreign_i,t−1 0.00292 (0.00490) ∆rr locali,t−1 0.00340 (0.00567) ∆policy rate_i,t−1 -0.00727∗∗∗ -0.00757∗∗ -0.0164∗∗∗ -0.0143∗∗∗ -0.00380 -0.00717∗∗∗ -0.00660∗∗∗

(0.00228) (0.00299) (0.00495) (0.00448) (0.00456) (0.00204) (0.00221) ∆GDPi,t−1 -0.00155 -0.00103 0.738∗∗ -0.00976 -0.000984 -0.0156 -0.00197 (0.0169) (0.0168) (0.293) (0.0214) (0.0250) (0.0205) (0.0168) ∆V IXi,t -0.000303∗∗∗ -0.000284∗∗∗ -0.000339 -0.000531∗∗∗ -0.000358∗∗ -0.000330∗∗∗ -0.000291∗∗∗ (0.000102) (0.000110) (0.000233) (0.000193) (0.000173) (0.000108) (0.000102) cons -0.000330 -0.000246 -0.00111 -0.000695 -0.000406 -0.000356 -0.000307 (0.000991) (0.00110) (0.00220) (0.00188) (0.00166) (0.00109) (0.000988) N 693 635 56 269 329 693 693

percent level in the ibex specification, which indicates that GDP growth is highly positively correlated with property price expansions, which indicates that banks are eager to invest in real estate sectors, which again fuels house price appreciation. The story could appear to be similar in case of the reserve requirements, rr foreign and rr local, given that lending is constrained by the usage of reserve requirements, banks invest in securities such as the real estate sector. However this argument may be ambiguous given the negative coefficient of GDP in both specifications.

Furthermore, besides the results for the prudential instruments, what stands out from table 10 is the observation that both policy rate and VIX are highly significant at the 1 percent level for all the specifications except for ltv cap which does not have a significant policy rate variable and ibex which dos not have a significant VIX variable. This points to evidence that house price appreciation is highly affected by global risk, as proxied by the VIX index, and the interest rate, as proxied by the monetary policy rate or close alternatives. For policy makers this is quite a remarkable result, since it indicates that the interest rate tool is quite powerful and potent in dampening house price growth on a relatively high frequency basis, i.e. quarterly in our case, as opposed to changes in prudential instruments. Zhang and Zoli (2016) conclude for their sample of 11 Asian economies in the period 2000-2013 that the interest rate is not significant. However, for their sample of 5 advanced Asian economies, they conclude the opposite result in line with our findings, that the interest rate is highly significant at the 1 percent level with a negative coefficient. Also, the fact that the VIX affects house price appreciation, although the coefficients are quite small, indicates that global financial conditions are quite important determinants of regional and country level asset prices such as house prices. Equivalent to the interest variable, for the VIX, Zhang and Zoli (2016) find that the VIX is not significant for the sample of 11 Asian economies in the period 2000-2013, although with a large negative coefficient. But for the sample of 5 advanced Asian economies they find that the VIX is highly significant at the 5 percent level with a large negative coefficient.

Sector specific capital buffers Next, we shall analyse the sector specific capital buffers and how these affect house price growth. Looking at the results in table 11 we note that the coefficients on sscb res and sscb oth have the expected sign, but are not significant. And sscb cons has a positive sign, which was not conjectured, and is also insignificant. This tells us that it seems that sector specific capital buffers are not particularly effective in mitigating house price expansion on a quarterly basis. However, we should note that this may also be caused by the relatively few changes of these buffers. One thing that stands out, is that in all the specifications the coefficients on the policy rate and the VIX are highly significant at the 1 percent level and all have a negative sign, which matches our expectations. Although, the significance of the global risk variable has been observed, it is noteworthy to observe that policy rate has a strong dampening impact on house price growth on a quarterly basis. This may provide evidence that on a

Table 11: Anderson-Hsiao (1981) FD2SLS sector specific capital buffers

(1) (2) (3)

∆housei,t ∆housei,t ∆housei,t

∆housei,t−1 -0.250∗∗∗ -0.314∗∗∗ -0.240∗∗∗ (0.0693) (0.0618) (0.0707) ∆sscb resi,t−1 -0.00665 (0.00849) ∆sscb consi,t−1 0.0225 (0.0167) ∆sscb othi,t−1 -0.0235 (0.0241) ∆policy rate_i,t−1 -0.00711∗∗∗ -0.00725∗∗∗ -0.00712∗∗∗

(0.00222) (0.00210) (0.00223) ∆V IXi,t -0.000339∗∗∗ -0.000296∗∗∗ -0.000336∗∗∗ (0.000107) (0.000103) (0.000107) cons -0.000368 -0.000317 -0.000369 (0.00108) (0.00100) (0.00108) N 693 693 693

Standard errors in parentheses