Bank regulatory arbitrage via risk weighted assets dispersion

G Model

ARTICLE IN PRESS

JFS-491; No. of Pages 15

Journal of Financial Stability xxx (2016) xxx–xxx

Contents lists available at ScienceDirect

Journal of Financial Stability journal homepage: www.elsevier.com/locate/jfstabil

Bank regulatory arbitrage via risk weighted assets dispersion夽 Giovanni Ferri a,∗ , Valerio Pesic b a b

Department of Economic, Political Sciences & Modern Languages, LUMSA University, Via Pompeo Magno, 22, 00193 Rome, Italy Department of Management, Sapienza University, Via del Castro Laurenziano, 9, 00161 Rome, Italy

a r t i c l e

i n f o

Article history: Received 1 December 2015 Received in revised form 10 June 2016 Accepted 19 October 2016 Available online xxx JEL Classification codes: G2 G21 G28

a b s t r a c t Increased dispersion of Risk Weighted Assets (RWA) troubles regulators as potentially undermining prudential supervision. We study the determinants of RWA/EAD (Exposure-At-Default) on data painstakingly compiled from Basel Pillar-Three for 239 European banks over 2007–2013. We improve on most previous studies, which consider instead RWA/TA (Total Assets). Indeed, Internal-Rating-Based (IRB) models allow lawful capital-saving Roll-Out effects which RWA/TA analyses disregard and likely misidentify as regulatory arbitrage. Instead, encapsulating Roll-Out effects, RWA/EAD avoids false positive identification. We find that regulatory arbitrage: (i) was present; (ii) likely materialized via risk weights manipulation with IRB models; (iii) was stronger at Advanced-IRB vs Foundation-IRB banks. © 2016 Elsevier B.V. All rights reserved.

Keywords: Regulatory arbitrage Internal rating based models Risk weighted assets dispersion

1. Introduction Regulatory arbitrage occurs when a bank engages in practices that, while being formally legitimate, end up in reducing (eluding a rise of) regulatory capital while risk doesn’t decrease (increases). This would lead to a reduction of the Risk Weighted Assets (RWAs) density, as given by the ratio of RWAs to Exposures At Default (EADs). The issue has become topical as evidence mounted of sizable dispersion in RWA density across otherwise similar banks. By endangering fair treatment and raising systemic risk, this could prove particularly nasty for regulators. Calculating RWAs largely remains, in fact, with a bank’s regulatory accounting choices. If two otherwise equivalent banks show different RWA density, this might imply that one of them underrates risk and artificially reduces its capital requirements. Moreover, should that be widespread across banks in a country, that country would be prone to high systemic risk.

Deplorably, though, we know little on the true size of this phenomenon and its causes remain largely unexplored. In spite of Basel II third pillar’s obligations, micro data is still lacking. This paper has two main aims. First, we provide fresh evidence on RWA dispersion extending the analysis to a relatively ample number of European banks. Second, we assess how much RWA dispersion stems not just either from a “roll out” effect – i.e., shifting larger EAD shares from Standard to Internal-Rating-Based (IRB) model – or from a business specialization effect but is liable to the suspicion of regulatory arbitrage. To this end, we also compare Foundation-IRB (F-IRB) to Advanced-IRB (A-IRB) banks, the latter having more latitude for risk weights manipulation. To accomplish our task, given that during the observation period Eurozone countries underwent the asymmetric euro sovereign crisis, we also control for the macroeconomic conditions of the country where a bank is established. In the remainder of the paper, Section 2 summarizes the existing literature on the topic. Section 3 presents the data that we painstakingly compiled. In Section 4 we report and comment the results of our econometric estimates. Finally, Section 5 recaps and discusses policy implications.

夽 We are particularly thankful to two anonymous referees for providing key insights that helped us to greatly improve the paper. We also acknowledge the useful comments that we received from participants in presentations at the 5th FEBS International Conference and at the University of Cagliari. ∗ Corresponding author. E-mail addresses: [email protected] (G. Ferri), [email protected] (V. Pesic). http://dx.doi.org/10.1016/j.jfs.2016.10.006 1572-3089/© 2016 Elsevier B.V. All rights reserved.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

2

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

2. Balancing stability and profitability of banks in the economics literature Prudential supervision of banks considers an adequate level of capital as a necessary, though not sufficient, condition to reach financial stability of a single bank and of the whole banking system (BCBS, 2012). However, how to determine an adequate threshold of capital able to ensure banking soundness and stability is still quite an unresolved issue. In particular, as the level of capital to comply with the regulatory framework can affect banks profitability, by enlarging the denominator of their return on equity (ROE), since the inception of the Basel Accord (BCBS, 1988) supervisors tried to minimize the negative effects of their requirements on profitability. The tools supervisors used to that end varied over time. At first, supervisors allowed including in regulatory capital resources besides common shares and retained earnings, granting the option to alternatively comply with capital requirements without issuing too much capital, which would depress financial performance (Ayadi et al., 2016). Later on, supervisors considered an increasing number of typologies of risks under the Risk Weighted Assets (RWAs) formula, so to contemplate the evolution of banking activity and avoid model obsolescence (BCBS, 1996, 1997, 1999). Also, over time they reviewed the modality of capital requirements calculation by different approaches, so to stimulate more sophisticated and relevant banks to invest in more advanced methodologies of risk evaluation (BCBS, 2005). These should eventually achieve sounder risk management together with lower capital absorption. Lastly, the Basel III framework aimed to improve the resilience of the banking sector by increasing the quality and quantity of the regulatory capital base, enhancing the risk coverage of the capital framework, proposing a new leverage ratio to protect against model risk and measurement error (Brei and Gambacorta, 2016), and finally introducing a number of macro-prudential elements to dampen the pro-cyclicality of prudential supervision (BCBS, 2011). A key potential pitfall of this regulatory framework is that banks might aggressively seek ways to reduce capital absorption. Especially the more significant banks – which generally relied most on funding sources other than common equity – might make discretionary use of regulation to upgrade their capitalization. For one, they could optimize their risk profile, e.g. moving from high capital absorbing (e.g. loans) to less capital consuming assets (e.g. state bonds and other financial assets). Also, they could improve the quality of portfolio assets, as well as choose risk measurement methods possibly lowering capital requirements. To this end, under Basel II and Basel III, using the Internal-Rating-Based (IRB) model can be decisive, because of both its significant differences vis-à-vis the alternative Standard model and the many portfolio assets considered in its calculation. Yet, a growing literature deems that sophisticated methodologies, such as all IRB − and chiefly A-IRB − models, embody large discretion. Specifically, via regulatory arbitrage banks might lower their capital commitments through lawful ways, alas not justified by sounder risk management. In this, large RWA dispersion may signal that, ceteris paribus, some banks engineered lower capital absorption by more leniently exploiting the regulatory framework (Fig. 1). The most fitting evidence of strategic risk-modelling via risk weights manipulation is by Mariathasan and Merrouche (2014) who study the relationship between banks’ IRB model approval and the ratio of RWAs to total assets across 115 banks from 21 OECD countries. Consistent with a risk-weight manipulation view, they find that RWA density drops after regulatory approval, and show that the decline in risk-weights is larger: i) at weakly capitalized banks; ii) in jurisdictions with weak legal supervisory framework, and iii) in countries with many supervised IRB banks. However, the dispersion among RWAs has become evident even across banks operating in the same region (e.g. Europe)

and with similar business specialization. So, supervisory worries about regulatory arbitrage taking place at banks via RWA calculations surfaced repeatedly. For instance, the European Banking Authority (EBA, 2013a) reviews RWA consistency via a top-down assessment of the banking book, EBA (2013b) performs an analogous exercise for low default portfolios, EBA (2013c) reports on the comparability of supervisory rules and practices, EBA (2013d) tells on the pro-cyclicality of capital requirements under the IRB Approach, EBA (2013e) reports on variability of RWAs for Market Risk Portfolios, and EBA (2014) testifies technical standards on supervisory benchmarking of internal approaches for calculating capital requirements.1 However, also other supervisory bodies have addressed the issue as the Basel Committee on Banking Supervision (BCBS, 2013a, 2013b, 2013c) or Argimón and Ruiz-Valenzuela (2010); ˜ or Ledo (2011), and Arroyo et al. (2012), at the Banco de Espana, Cannata et al. (2012) at the Banca d’Italia, or Gustin and Van Roy (2014) at the National Bank of Belgium, or Das and Sy (2012) and Le Leslé and Avramova (2012) at the IMF. In turn, Fratianni and Pattison (2015) show how the same Basel Accord can take significant deviations in national level implementations, suggesting that RWA dispersion might be easier where supervisors apply some form of benign neglect. All these contributions conclude that analogous amounts of RWAs may hide different levels of risk across countries/banks. However, these studies usually rely on few observations. For instance, Cannata et al. (2012) analyze 24 banks, Le Leslé and Avramova (2012) study 51 banks (18 Asian, 21 European, and 12 American). Three papers studying the sensitivity of RWAs to banks’ policies and macro circumstances are Beltratti and Paladino (2016), Vallascas and Hagendorff (2013), and Bruno et al. (2014). They all find significant indications of regulatory arbitrage. However, these studies either consider also non-European countries − B&P study 45 countries but only 22 of them are European with fewer than 150 banks, while V&H consider 41 countries but only 16 are European with only 61 banks − or rely only on the 50 largest European banking groups (B&A). Furthermore, V&H do not use information on EAD while B&P have EAD values only for a subsample of 86 banks (they don’t report how many of these are from Europe). Thus, to satisfy our perspective these studies should be improved. First, their results might depend on the variability entailed by comparing very different jurisdictions (B&P; V&H) and could thus have little bearing for regulatory arbitrage, when differences across countries are abated. Second, failing to consider the role of IRB models – as they don’t have EADs – V&H cannot really identify the type of regulatory arbitrage we have in mind; while B&P and B&A in their EAD analysis come close for Europe to the small sample of banks as Cannata et al. (2012). Third, B&P, and V&H stop their analysis in 2010, which does not allow them to tell whether European banks regulatory arbitrage intensified with the euro-crisis.2 Table 1 synthesizes the 13 papers most relevant for the issue of RWA dispersion. Although close to previous studies on the determinants of RWA density, our research question is relatively new. In fact, the literature only recently started to investigate how the use of IRB methods, introduced by Basel II, likely contributes itself to boost RWA dispersion. Thus, we compare those papers listing: their methodology, whether they consider the IRB model and/or the RollOut effect, the number of banks analyzed, the number of countries

1 The importance the EBA assigns to this problem is testified also by the fact that it organized a specific workshop on it hosting some relevant papers (e.g., Bruno et al., 2014). 2 In terms of span of the data we improve only by one year on Bruno et al. (2014) who reach up to 2012.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

Methodology

Consider IRB Model

Consider Roll-Out Effect

No. Banks

No. Countries

Time period

Main results

RWA/TA

Ledo (2011)

Top-down approach on public data

YES

NO

35

11 World

2000–2010

Arroyo et al. (2012)

YES

NO

16

8 EU

2011

Le Leslé and Avramova (2012)

1) Top-down approach on public data 2) Bottom-up approach on portfolio data Top-down approach on public data

YES

NO

51

19 World

1998–2011

Vallascas and Hagendorff (2013)

Top-down approach on public data

YES

NO

246

41 World

2000–2010

Mariathasan and Merrouche (2014)

Top-down approach on public data

YES

NO

115

21 OECD

2004–2010

Beltratti and Palladino (2016)

Top-down approach on public data

NO

NO

548

45 OECD

2007–2011

RWA drop due to risk profile (e.g. macro/institutional factors, business model), risk management, supervisory practices. RWA density depends on differences in type of business involved: higher RWA at more lending oriented banks. Move to Basel II and business mix changes caused RWA drop. Some A-IRB banks changed IRB methods to get a larger drop. Sensitivity of minimum regulatory capital to a market measure of the portfolio risk of banks is very weak. RWA density drops after IRB approval. Risk-weights drop acutely at weakly capitalized banks, and in countries with weak legal supervisory framework, or many IRB banks. No link of RWA density to cost of capital. EU banks engage less in RWA-saving in peripheral than in core countries.

Cannata et al. (2012)

Top-down approach on public data

YES

YES

24

8 EU

2010

BCBS (2013b)

1) Top-down approach on public data 2) Bottom-up approach on portfolio benchmarking exercise 3) Assessing bank/regulatory practices Top-down approach on public data

YES

NO

102

World

2012

YES

YES

89

16 EU

2011

RWA/EAD

EBA (2013a)

EBA (2013b)

1) Top-down approach on public data 2) Bottom-up approach on portfolio benchmarking exercise

YES

YES

35

13 EU

2012

EBA (2013c)

Top-down approach for IRB’s Retail and SME Corporate Portfolios

YES

NO

43

14 EU

2012

Bruno et al. (2014)

Top-down approach on public data

YES

YES

50

17 EU

2008–2012

Gustin and Van Roy (2014)

Bottom-up approach on portfolio benchmarking exercise

YES

NO

4

1

2013

Individual banks’ business mix, and different prudential approach (standard vs IRB) explain RWA dispersion. RWA dispersion increases from Basel I to Basel II, with an intensity depending on differences in the type of business involved, with higher values encountered when the weight of credit risk in a bank’s balance sheet is higher. RWA dispersion caused by fundamental effects, as IRB use; SA risk weights; IRB portfolio mix; share of IRB defaulted assets. RWA dispersion caused by fundamental effects, as IRB use; SA risk weights; IRB portfolio mix; share of IRB defaulted assets. Residual variance due to differences in inherent credit risk of banks’ IRB exposures and in supervisory and banks’ practices. RWA dispersion caused by share of IRB defaulted assets, portfolio mix of non-defaulted assets, different credit risk, use of credit risk mitigation, modelling and supervisory practices. RWA dispersion explained by banks’ size, business model and asset mix, and adoption of IRB approach, supervisory strength, together with market-based measures of bank risk. RWA dispersion driven not by PD estimates, but by differences in estimated LGDs, mostly depending on disparities in collateral valuation/management, ways to integrate collateral into internal models, and other banks’ modelling choices.

ARTICLE IN PRESS

Paper

G Model

JFS-491; No. of Pages 15

Dependent Variable

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

Table 1 Basics of key papers on the determinants of RWA dispersion.

3

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

4

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Fig. 1. The Regulatory Capital Framework Puzzle.

covered, the time period considered, the main results reached. A key distinction is between the six papers which try to identify regulatory arbitrage by estimating the determinants of RWA/TA and the seven papers that analyze RWA/EAD. It is important to stress that the RWA/TA measure is inappropriate to identify regulatory arbitrage. In fact, only focusing on RWA/EAD one can clean the risk weighted density from the Roll-Out effect (represented in the central column of Fig. 1) capturing the mechanic reduction of RWA density when a bank shifts portfolio shares from Standard to IRB. Since the Roll-Out effect is a fair use of regulatory options, the papers that fail to consider it will most likely find signals of regulatory arbitrage even where there was none. And even among the RWA/EAD papers, only four papers deal explicitly with the Roll-Out effect. Our aim in the rest of the paper is trying to improve on these four papers by extending the analysis to more banks/countries/years and to upgrade the empirical model of the determinants of RWA/EAD. 3. Our empirical approach and database 3.1. Main features of our empirical approach Our main contributions largely owe to the data we compiled. Namely, besides introducing other control variables, we augment BankScope data with information painstakingly gathered from individual banks’ statements and Pillar Three reports. This gives us for each bank: i) its Risk Weighted Assets (RWAs) and Exposures At Default (EADs), and ii) its percentage of EADs referred to, respectively, the Standard model, the Foundation IRB (F-IRB) model, and the Advanced-IRB (A-IRB) model. So, in the spirit of Mariathasan and Merrouche (2014), we can possibly identify regulatory arbitrage via Basel risk-weights manipulation to minimize capital requirements. Specifically, controlling for the progressive shift from Standard to F-IRB and to A-IRB models we check whether RWAs/EADs turn out to be systematically lower for the banks that were likely “capital constrained”. In practice, we expect that if capital constrained banks artificially reduce their risk-weights this should show up in a statistically significant positive link between Equity/Total Assets (E/TA) and RWA/EAD. Since the ability to manipulate risk-weights is nil for Standard banks, intermediate for F-IRB banks, and largest for A-IRB banks, we also expect the link between E/TA and RWA/EAD to strengthen moving from Standard to F-IRB to A-IRB banks. Our approach is more nuanced than the pioneering work of M&M in terms of measuring: i) RWA density, and ii) the extent of IRB methods. First, our RWA density is the ratio RWA/EAD, the appropriate measure rather than their RWA/TA. Second, while M&M have a 0–1 variable for F-IRB (and A-IRB) adoption by a bank,

we have a continuous measure of the percentage of the loan portfolio of each bank by the Standard, F-IRB, and A-IRB methods. That is a fundamental facet of our approach, and it is the only way to control for the Roll-Out effect, namely the effect of lawfully reducing RWA density thanks to the more parsimonious IRB methods. In other words, unlike M&M, controlling for the share of EAD portfolio measured by both the F-IRB and A-IRB methods we can investigate about potential other facets − able to reduce the RWA density − which can be rightly classified as regulatory arbitrage. This is allowed by the structure of our database, where a bank can be partly Standard, partly F-IRB and partly A-IRB. To exemplify, in year 2013 our 239 banks exhibit the following features. 131 banks are single method: 109 are only Standard, 11 only F-IRB, 11 only A-IRB. 101 banks apply two methods: 44 mix Standard and F-IRB, 2 mix F-IRB and A-IRB, 55 mix Standard and A-IRB. The remaining 7 banks apply all three methods together. A fundamental prerequisite for our approach is that we correctly measure a bank’s “true risk exposure” (Barakova and Palvia, 2014).3 As detailed in the next sub-section, we introduce a large list of controls partly at the individual bank’s level and partly at the level of the country a bank belongs to. In addition, given that during the observation period Eurozone countries underwent the asymmetric euro sovereign crisis (Pagano and Sedunov, 2016), we also need to control for the macroeconomic conditions of the country where a bank is established. Overall, our database has the following main features. First, we have a quite large number of individual banks (239) and of total bank-year observations (1128) from 29 European countries, covering on average above 80% of total assets of European banks.4 Second, our data covers a unique period. Specifically, we estimate effects from 2008 up to 2013, thus going well into the euro-crisis, which in various euro countries was much deeper than the sub-prime crisis. At the same time, our period is the one leading to the arrival of Basel III, when especially larger banks (generally relying more on IRB methods) should strive to save capital in achieving the new regulatory requirements, possibly engaging in regulatory arbitrage. Third, we managed to collect bank data in jurisdictions using different RWA calculation methods. In particular, having many more

3 We are particularly thankful to two anonymous referees for providing key insights in this respect. 4 The countries included are: Austria, Belgium, Bulgaria, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, Switzerland, United Kingdom.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

European banks than previous studies, we can analyze whether and to what extent there is regulatory arbitrage in Europe, an area where regulatory cross-country differences exist but are certainly smaller than when comparing Europe with other world areas (Cihak et al., 2013). Finally, having collected data for each bank for more than one year, allows us to use state-of-the-art methodologies for dynamic panel data, where the consideration of the autoregressive term of the dependent variable gives us the possibility to control for portfolio business mix composition in a more effective way (Köhler, 2016), despite the lack of detailed information about the portfolio’s composition under each method. 3.2. Dataset description To test whether and to which extent there was “regulatory arbitrage”, we focus on a set of key variables. These include potential predictors of the business specialization of a bank and measures of its risk level. Although we consider also the ratio RWA/TA (the ratio between RWA and Total Assets), as said our key dependent variable is the RWA density, defined as the ratio between the RWA measure from the credit portfolio and its corresponding EAD (RWA/EAD).5 We consider also the variables F-IRB and A-IRB, respectively the percentage coverage of the EAD credit portfolio by Foundation-IRB (F-IRB) and Advanced-IRB (A-IRB) approaches. By this manner, we aim to add to the existing literature, by distinguishing between the different determinants of RWA dispersion, namely a fair use of the regulatory options − like the different composition of Standard and IRB methods − vs. regulatory arbitrage via manipulation of Basel risk weights. Table 2 reports the complete list of variables considered in our regressions, their definitions and sources, while the basic descriptive statistics are presented in Table 3. The bank level independent variables we consider are: - F-IRB − EAD portfolio coverage by Foundation-IRB methodology (F-IRB), which we view as the most common regulatory option aiming to reduce RWA/EAD (namely “roll out” effect); ˆ − squared of F-IRB, which we consider to control for the - F-IRB SQ non-linearity of F-IRB effects on RWA density6 ; - A-IRB − EAD portfolio coverage by Advanced-IRB methodology (A-IRB), which we view as a further common regulatory option aiming to reduce RWA/EAD (namely “roll out” effect); ˆ − squared of A-IRB, which we consider to control for the - A-IRBSQ non-linearity of A-IRB effects on RWA density; - OFF/TA − ratio of Off-Balance Sheet Items to Total Assets, which we view as a variable potentially controlling for a bank’s “true risk exposure”. In this case, the variable can be considered like a regulatory option offered by regulation, even if we cannot exclude that it may be used as an instrument of regulatory arbitrage, especially by banks adopting the more sophisticated F-IRB and A-IRB methods;

5 Despite having collected data also on market and operational risks, we focus only on credit risks, still representing the most relevant component of European banks’ overall risk, or at least of the banks oriented to lending activity. 6 The key reason why one may envisage a non-linear relationship between the extent of F-IRB (and also of A-IRB) and RWA/EAD has to do with the behavior of supervisors and banks. In practice, supervisors might be lenient seeing reductions in RWA/EAD at a bank that is starting to shift its initial portfolio shares from Standard to F-IRB (or from F-IRB to A-IRB) but they might worry noticing analogous RWA/EAD reductions when that bank has already transferred a large part of its portfolio from Standard to F-IRB (or from F-IRB to A-IRB). Anticipating possible supervisors’ reactions, banks already using F-IRB (or A-IRB) to a large extent might limit the reduction of RWA/EAD when shifting additional portfolio shares from Standard to F-IRB (or from F-IRB to A-IRB).

5

- OTHER/TA − ratio of Other On-Balance Sheet Items considered in EAD portfolio to Total Assets, which we consider as a further variable potentially controlling for the bank’s “true risk exposure”. (We estimated this variable as the residual between EAD minus the OFF-Balance Sheet and Loans). Also here, the variable can be viewed as an option offered by regulation, even if we cannot exclude that it may be used as an instrument of regulatory arbitrage, especially by banks adopting the more sophisticated F-IRB and A-IRB approaches; - NPL − ratio of Not Performing Loans to Gross Loans, which we expect by its nature to increase RWA/EAD; - ASSETS GROWTH − increase in total assets, perhaps negatively related to RWA/EAD since faster growing banks can more easily adjust their portfolio composition; - NLOANS/LIABILITIES − ratio between net loans and total liabilities, viewed as a proxy of the leverage realized by each bank between borrowed funds and loans granted. We expect this variable by its nature to raise RWA/EAD; - SIZE − logarithm of total assets, to control for possible size related differences, though we have no a priori on its sign; - EQUITY − ratio of equity to total assets, where we expect a positive relationship since higher values of it should less likely lead to seek capital saving options. We define this ratio similarly to the leverage ratio of the Basel III capital framework, which is viewed as a more effective safeguard against model risk and measurement error than other ratios controlling for the level of bank capitalization − i.e. ratio between equity and EAD, or ratio between equity and RWA; ˆ − squared of EQUITY, introduced to control for the non- EQUITYSQ linearity of EQUITY effects on RWA density; - ROA − return on assets, measured as the ratio between net income and total assets, which we introduce to control for the profitability of each bank, though we have no a priori on its sign; - Z-SCORE − measure of the bank’s probability of insolvency (defined as in Hesse and Cihak, 2007), which we view as a variable potentially controlling for the bank’s “true risk exposure”; - LISTED − dummy variable − with value of 1 if the bank is listed and 0 otherwise − to control for the potential discipline exercised by capital markets. We expect it to raise RWA/EAD; - STATE AID − dummy variable which takes the value of 1 if the bank received any specific intervention during the period of our analysis. We expect it to raise RWA/EAD; - STRESS TEST − dummy variable which takes the value of 1 if the bank has been included in at least one of the EBA and ECB stress tests during the period of our analysis. While doubts on the efficacy of macro stress tests seem to emerge (Borio et al., 2014), the expected sign is unclear; - LESS CAPITAL − dummy variable which takes the value of 1 if the bank during the whole period we consider shows a mean value for the ratio Equity to RWA lower than 10%; - UNDER CAPITAL − dummy variable which takes the value of 1 if the bank during the whole period we consider shows a mean value for the ratio Equity to RWA lower than 8%; - LOWER CAPITAL − dummy variable which takes the value of 1 if the bank during the whole period we consider shows a mean value for the ratio Equity to RWA between 8% and 10%; - DUMMY F-IRB − dummy variable which is 1 from year t and on if the bank adopts the F-IRB method, no matter for the portion of EAD’s portfolio, from year t; - DUMMY A-IRB − dummy variable which is 1 from year t and on if the bank adopts the A-IRB method, no matter for the portion of EAD’s portfolio, from year t.

We also include the following macro level variables:

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

6

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Table 2 Variables, their definitions and sources. Name RWA/TA RWA/EAD F-IRB ˆ F-IRBSQ A-IRB ˆ A-IRBSQ DUMMY F-IRB DUMMY A-IRB OFF/TA OTHER/TA NPL ASSETS GROWTH NLOANS/LIABILITIES SIZE EQUITY ˆ EQUITYSQ Z-SCORE DUMMY LISTED STATE AID STRESS TEST LESS CAPITALIZED POOR CAPITALIZED UNDER CAPITALIZED NPL SYSTEM L.GDP GROWTH RESOLVING INSOLVENCY OVERALL STANDARD HIGH DEBT LOW DEBT HIGH SPREAD LOW SPREAD

Definition

Source

Bank level variables Ratio of RWA to Total Assets Ratio of RWA to Exposure at Default % of Foundation IRB methodology upon EAD % of Foundation IRB methodology upon EAD *squared % of Advanced IRB methodology upon EAD % of Advanced IRB methodology upon EAD *squared Dummy variable for use of F-IRB , irrespective of EAD portfolio share Dummy variable for use of A-IRB, irrespective of EAD portfolio share (Total Business Volume-Total Assets)/Total Assets * winsorized at 1% (EAD – Loans – OFF TA)/Total Assets * winsorized at 1% Ratio of Non-Performing Loans to Gross Loans * winsorized at 5% Increase in total assets Ratio of net loans to liabilities Logarithm of total assets Ratio of equity to total assets Ratio of equity to total assets *squared Z-score (distance to default) Dummy variable for Listed State Aided Dummy Variable EBA and ECB Stress Test Dummy Variable Dummy variable for less capitalized banks Dummy variable for poor capitalized banks Dummy variable for under capitalized banks Country level variables Country level ratio of non-performing loans to total loans Lagged increase in GDP Resolving Insolvency Rank (Time invariant) % of Standard Methodology upon Country EAD (Time variant) Dummy for countries having a Public Debt/GDP ratio >= 65.14% Dummy for countries having a Public Debt/GDP ratio < 65.14% Dummy for countries having a government bond spread >= 1.60 Dummy for countries having a government bond spread < 1.60

- NPL SYSTEM − country level ratio of non-performing loans to total loans, which, ceteris paribus, might increase RWA/EAD but could have the opposite effect if deteriorating macro conditions depress loan demand and/or amplify supervisory scrutiny; - L.GDP GROWTH − lagged increase in Country’s GDP. We have no a priori on its sign; - RESOLVING INSOLVENCY − Resolving Insolvency Rank as obtained from the World Bank’s dataset Doing Business. We consider this variable to control for the potential discipline exercised by the strength of the national legal system. We expect it to raise RWA/EAD; - OVERALL CREDIT STANDARD − national EAD portfolio coverage under Standard Methodology. We consider that variable to control for the potential discipline exercised by supervisors of the national legal system, as well as when the increase of banks utilizing IRB may lead to a relaxation of supervisory scrutiny. We expect it to raise RWA/EAD. - HIGH DEBT − Dummy variable valued 1 for countries having a Public Debt/GDP ratio > = 60.14%, and 0 otherwise. We consider that variable to control for the potential macroeconomic conditions influencing RWA dispersion, though we have no a priori on its sign; - LOW DEBT − Dummy variable valued 1 for countries having a Public Debt/GDP ratio <60.14% and 0 otherwise. We consider this variable as the opposite to the above one; - HIGH SPREAD − Dummy variable valued 1 for countries having a government bond spread > = 1.60 b.p. and 0 otherwise. We consider that variable to control for the potential macroeconomic conditions influencing the RWA dispersion, though we have no a priori on its sign; - LOW SPREAD − Dummy variable valued 1 for countries having a government bond spread <1.60 b.p. and 0 otherwise. We consider this variable as the opposite to the above one.

Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Banks’ Pillar 3 Bankscope & Pillar 3 Bankscope Bankscope Bankscope Bankscope Bankscope Bankscope Bankscope Bankscope Bankscope Bankscope ECB, EBA Bankscope Bankscope Bankscope World Bank World Bank World Bank ECB World Bank World Bank World Bank World Bank

4. Empirical analysis 4.1. Methodology of analysis The riskiness of a bank’s asset portfolio shows high persistence, we should say a memory, so that changes in RWA density from one period to the next tend to be small relative to the variable’s levels. This is a noteworthy property of our dataset we must consider to adopt an econometric approach able to address the issues arising from high persistence and autocorrelation of the series, with the potential endogeneity problems coming from reciprocal causality links among different variables. In these situations, the literature generally points to the dynamic regression model as the most effective approach, using a time lag of the dependent variable as an additional regressor. As stated by Arellano and Bond (1991); Arellano and Bover (1995), and Blundell and Bond (1998), our dataset is a “small T, large N” panel. Not only our dependent variable (RWA/EAD) depends on its own past realizations but also some independent variables are not strictly exogenous, meaning they could be correlated with past and possibly current realizations of the error. Then, we have fixed individual effects. Finally, we may have heteroskedasticity and autocorrelation within individuals. For those reasons, we framed our analysis through a “classic” approach (Holtz-Eakin et al., 1988; Arellano and Bond, 1991; Arellano and Bover, 1995; Blundell and Bond, 1998; Roodman, 2009) taking into consideration for our analysis the results achievable by the OLS estimator clustered by each individual, the Within Fixed Effects estimates, the Anderson and Hsiao (1982) estimates with IV in levels, the Dif-GMM Arellano and Bond (1991) model, the Sys-GMM Blundell and Bond (1998) model. Moving from the specific characteristics of our dataset, we consider the Sys-GMM Blundell and Bond (1998) model, like the most suitable of our porpoise. For all specifications we included time dummies and applied

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model

JFS-491; No. of Pages 15

RWA/TA

RWA/EAD

F-IRB

A-IRB

OFF/TA

OTHER/TA

NPL

EQUITY

ASSETS GROWTH

NLOANS/ LIABILITIES

Mean Max P90 P75 P50 P25 P10 Min SD N

47.15 134.15 78.18 63.17 45.45 30.00 19.14 0.03 22.94 1128

47.84 118.00 73.00 59.00 47.00 35.00 24.00 0.00 18.94 1128

16.17 100.00 76.00 0.00 0.00 0.00 0.00 0.00 31.44 1128

21.24 100.00 83.00 48.00 0.00 0.00 0.00 0.00 35.27 1128

12.93 40.50 28.50 18.96 11.02 4.38 0.58 -2.77 10.84 1128

29.59 536.20 68.56 44.03 25.28 13.15 -0.88 -84.97 49.23 1128

5.76 46.00 13.00 7.00 4.00 1.00 0.00 0.00 6.48 1128

6.60 25.00 11.00 8.00 6.00 4.00 2.00 1.00 3.99 1128

3.07 260.29 16.69 7.68 1.40 -5.35 -12.13 -79.89 19.61 1113

60.81 114.02 87.14 77.58 64.71 45.00 28.00 0.09 22.71 1128

Stats

ROA

DUMMY F-IRB

DUMMY A-IRB

LESS CAPITAL

UNDER CAPITAL

POOR CAPITAL

NPL SYSTEM

SIZE

Z-SCORE

GDP GROWTH

Mean Max P90 P75 P50 P25 P10 Min SD N

20.87 482.00 105.00 57.00 26.00 6.00 -62.00 -623.00 116.66 1128

0.25 1.00 1.00 0.00 0.00 0.00 0.00 0.00 0.43 1128

0.29 1.00 1.00 1.00 0.00 0.00 0.00 0.00 0.45 1128

0.25 1.00 1.00 1.00 0.00 0.00 0.00 0.00 0.44 1128

0.09 1.00 0.00 0.00 0.00 0.00 0.00 0.00 0.29 1128

0.16 1.00 1.00 0.00 0.00 0.00 0.00 0.00 0.37 1128

5.49 31.90 12.47 7.48 3.95 2.82 0.77 0.08 4.81 1147

17.61 21.51 20.13 19.01 17.43 16.38 15.23 12.47 1.82 1128

41.58 5883.14 66.85 40.42 20.77 7.63 2.23 0.00 249.70 1122

0.03 8.28 3.59 1.81 0.38 -1.59 -4.31 -14.74 2.90 1128

Stats

LISTED

RESOLVING INSOLVENCY

STATE AID

STRESS TEST

OVERALL STANDARD

Mean Max P90 P75 P50 P25 P10 Min SD N

0.42 1.00 1.00 1.00 0.00 0.00 0.00 0.00 0.49 1128

21.48 77.00 42.00 23.00 21.00 10.00 3.00 1.00 16.83 1128

0.19 1.00 1.00 0.00 0.00 0.00 0.00 0.00 0.39 1128

0.29 1.00 1.00 1.00 0.00 0.00 0.00 0.00 0.45 1128

51.86 97.89 87.86 76.83 51.52 28.59 14.81 8.67 25.34 1122

ARTICLE IN PRESS

Stats

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

Table 3 Descriptive statistics.

7

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

8

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

the Windmeijer correction to reported standard errors, showing the results for Sargan/Hansen tests of overidentifying restrictions and Arellano-Bond tests for autocorrelation of second-order.

Table 4 Determinants of RWA/EAD.

L.RWA/EAD

4.2. Results of the econometric analysis F-IRB

We first report a regression on the determinants of the RWA/EAD ratio. In particular, Table 4 presents the same general specification run on three different samples. Column 1 refers to the total sample while column 2 excludes the banks reporting no IRB adoption and column 3 excludes the banks reporting no AIRB adoption. In column 1, regulatory arbitrage via risk-weights manipulation might be partly confounded by the presence of purely Standard model banks in the regression. Thus, the most meaningful test lies in columns 2 and 3 where we can zero in on the differences of behavior between F-IRB and A-IRB banks. First, note how the coefficient of the lagged dependent variable, capturing persistence, is higher for the Total sample and notably lower for F-IRB and A-IRB banks. This is consistent with the characteristics of those methodologies, with Standard being least sensitive to portfolio evolution. In turn, the coefficient estimates for F-IRB suggest that RWA density is reduced by the migration from Standard to IRB methods, with a larger coefficient for F-IRB banks compared to the total sample. This is consistent with the socalled “Roll-Out” effect, which in this case we properly measured as the percentage of EAD portfolio under different methodologies. This effect remains in the first 2 columns despite the squared value of the regressor F-IRB is positive and significant for the total sample. When focusing on column 3, we see the coefficient of F-IRB losing its significance, whilst its squared value becomes negative and significant at 5%. This is not surprising, considering that A-IRB banks rely more on A-IRB methods, while progressively abandoning the IRB-Foundation approach. A-IRB shows an increasing negative and significant coefficient moving from column 1 to column 3, confirming the capability of this method to further reduce capital absorption. Also in this case, the squared term of the regressor shows always a positive value, which never dominates the magnitude of the linear term. To enhance our measures of a bank’s true risk exposure, we consider different variables suitable as potential determinants of RWA dispersion. An additional way to reduce capital absorption appears to be Assets Growth, which is negative and always significant at 1%, with a slightly higher coefficient for the Total sample. On the opposite, NPL is positive and significant at 5% only for A-IRB banks. NPL System exhibits a negative and significant coefficient for A-IRB banks, which might underscore a kind of disciplinary effect and/or drop in demand for credit for banks in high NPL contexts. Similarly, the value of Off Balance Sheet (OFF/TA) reduces RWA/EAD for A-IRB (Karim et al., 2013), whilst the value of Other Items on Balance Sheet (OTHER/TA) reduces RWA/EAD both for F-IRB and A-IRB banks. In particular, that variable exhibits an increasing coefficient moving from Standard to A-IRB banks, which becomes significant at 1% for F-IRB and A-IRB banks. In reality, that difference does not seem to be directly identifiable as potential manipulation, but be ascribed to the stronger capability of IRB-methodologies to capture the portfolio diversification, rather than to intense calibration activity. After controlling for all variables, we find that EQUITY has a positive and increasing value moving from Standard to A-IRB banks,7 which becomes significant at 1% for F-IRB and A-IRB banks, confirming the hypothesis that only for those banks the scarcity of capital can be a significant determinant for further reducing RWA/EAD via recali-

ˆ F-IRBSQ A-IRB ˆ A-IRBSQ NPL ASSETS GROWTH NLOANS/LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013 Constant N N(g) AR2-p J Hansen-df Hansen-p R-squared

Total

IRB

IRB ADV

0.8714*** 0.110 −0.1380*** 0.050 0.0010* 0.001 −0.1537** 0.062 0.0012* 0.001 −0.1180 0.076 −0.0687*** 0.023 0.0043 0.083 0.1002 0.090 0.0717 0.351 0.5625 0.674 −0.0199 0.019 −0.0009 0.001 −0.0165 0.042 −0.0331 0.026 0.0045 0.005 0.1775 0.179 −0.4140 0.658 0.6852 1.038 1.3516* 0.701 −0.0074 0.034 0.0087 0.015 1.1107 1.358 0.8502 0.740 −1.1454 0.864 −0.7087 1.295 3.7728 11.973

0.4869*** 0.145 −0.1862** 0.094 0.0011 0.001 −0.1770** 0.081 0.0010 0.001 0.1225 0.152 −0.0629*** 0.022 0.0126 0.087 −0.2528 0.187 0.3611 0.637 2.1066*** 0.799 −0.0803*** 0.031 −0.0238 0.018 −0.0713 0.054 −0.1156*** 0.035 −0.0098 0.007 −0.1726 0.186 1.4094 1.174 1.4487 1.116 −0.4330 1.085 −0.0563 0.049 0.0464 0.032 −1.2141 1.160 1.0572 0.970 −2.3105** 1.063 −3.8442*** 1.488 16.1799 16.048

0.4958*** 0.107 0.0251 0.102 -0.0029** 0.001 −0.2492*** 0.087 0.0012* 0.001 0.4346** 0.193 −0.1324*** 0.036 −0.1051 0.154 −0.6628** 0.287 −0.4472 1.254 2.7234*** 0.895 −0.1170*** 0.039 −0.0122 0.018 −0.1648* 0.095 −0.1960*** 0.054 −0.0024 0.007 −0.3153 0.263 0.3361 1.571 0.3219 1.260 1.3449 2.210 −0.0166 0.054 0.0311 0.044 −1.2361 1.360 2.1393* 1.217 −1.6623 1.103 −4.5154*** 1.686 42.3311 29.805

881 208 0.2550 50 23 0.5315 0.8889

505 117 0.3029 50 23 0.3530 0.7913

296 70 0.3426 50 23 0.5438 0.7740

The table represents different Sys-GMM model estimations for the dependent variable RWA/EAD upon the different groups of banks (Total, IRB, IRB ADV). In this case an approach 1-step is considered with lags (2 3) for instrumenting the endogenous variables L1.RWA/EAD and NLOANS/LIABILITIES. All regression include time dummies. N represents the number of observations available, N(g) represents the number of banks available, AR2-p represents the p-value of Arellano–Bond test for autocorrelation of second order; J represents the number of instruments; Hansendf represents the degree of freedom for the Hansen statistic; Hansen-p represents the p-value of the Hansen statistic. Windmeijer-corrected, bank-clustered standard errors are given in italics. ***, ** and * indicate statistical significance at the 1%, 5%, and 10% levels respectively.

7 The difference between the value of the coefficient in the total sample vs its value in the IRB and IRB ADV samples is statistically significant at 1% running a t-test for difference in means.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

bration of risk weights. Notice that in this case, the square of EQUITY has a positive and significant coefficient for both the F-IRB and AIRB banks, which never dominates the magnitude of the linear term. As already mentioned in Section 3.2, we define the variable EQUITY as the ratio of equity to total assets, similarly to the leverage ratio of Basel III, which is viewed as a more effective safeguard against model risk and measurement error than other ratios controlling for the level of bank capitalization.8 In Table 5 we introduce some additional regressors to better control for the different level of capitalization of each bank. Specifically, moving from the third column of Table 4 we present alternative specifications on A-IRB banks to assess the interaction between the level of capitalization and the adoption of IRBmethodologies. In column 2, we introduce the additional dummy LESS CAPITAL, which is 1 if the bank during the whole period we consider shows a mean value for the ratio Equity to RWA lower than 10%. We can interpret its positive coefficient as an effect which comes almost mechanically from its definition. Nevertheless, an interesting result comes from the interaction of that variable with the adoption of IRB-methodologies. The terms LESS CAPITAL * DUMMY F-IRB and LESS CAPITAL * DUMMY A-IRB aim to detect whether less capitalized banks try harder to reduce their RWA density via IRB methods. In this case, the adoption of F-IRB reveals no particular effects (remember we are focusing attention on A-IRB banks) whilst the adoption of A-IRB confirms our hypothesis of potential manipulation by less capitalized banks. In fact, the interaction term LESS CAPITAL * DUMMY A-IRB is negative and significant at 5%, while EQUITY still remains significant at 1%. On the opposite, A-IRB loses its significance. Also the other control terms maintain, on average, their magnitude and significance. Whilst the dummy LESS CAPITAL refers to banks with Equity/RWA ratio lower than 10%, in column 3 we introduce the alternative dummies UNDER CAPITAL and LOWER CAPITAL, so to distinguish between banks with an Equity/RWA ratio lower than 8% (UNDER CAPITAL) and the ones between 8% and 10% (LOWER CAPITAL). Also here, we interact those terms with the adoption of IRBmethodologies, aiming to identify if different level of capital can lead to different behavior across similar banks. Table 5 indeed confirms that hypothesis, because both the interaction terms POOR CAPITAL * DUMMY F-IRB and POOR CAPITAL * DUMMY A-IRB show a negative and significant value, whilst the interaction terms UNDER CAPITAL * DUMMY F-IRB and UNDER CAPITAL * DUMMY A-IRB are insignificant. This is likely explained by the fact that under-capitalized banks operate under stringent scrutiny by supervisors, whilst the poor-capitalized banks may have more leeway to optimize their RWA density. In columns from 4 to 7 we verify the robustness of our findings, by replicating the analysis considering separately the dummies relative to the level of capital – respectively LESS CAPITAL vs UNDER CAPITAL and LOWER CAPITAL in column 4 and 5–and the interaction terms with the DUMMY F-IRB and DUMMY A-IRB in column 6 and 7. This analysis confirms the previous results. 4.3. Robustness checks We perform several types of robustness checks. First, we consider another measure of RWA density, the RWA/TA ratio, the variable some authors focus on (among others, Mariathasan and Merrouche, 2014). Second, we consider an alternative method to control for IRB adoption by each bank, to better distinguish it from

8 Moreover, we consider this variable also to avoid the negative correlation which can be otherwise determined with the dependent variable (RWA/EAD), if adopting as regressor a different measure of capital, like the ratio between equity and RWA, the Tier-One Capital Ratio or the Total Capital Ratio.

9

the Roll-Out effect. Then, we study whether the indications of regulatory arbitrage intensified for the banks from the peripheral countries hit by the euro-crisis. Finally, we run our regressions by using rank-transformed variables. The first robustness check aims to verify our result throughout the use of an alternative measure of the RWA density. Though the RWA/EAD ratio is the appropriate measure for that phenomenon, in Table 6 we present the regression results on the determinants of RWA density, as commonly measured by the RWA/TA ratio. Again, we concentrate on columns 2 and 3. Here we find that persistence drops less moving from column 1 to column 2, confirming that RWA/TA behaves differently with respect to RWA/EAD. On the opposite, the coefficients estimated for F-IRB and A-IRB suggest almost similar reductions of RWA density in the migration from Standard to IRB methods. Again, this is consistent with the so-called Roll-Out effect. Also here linear term effects dominate squared term effects. Here as well, in column 3 the coefficient of F-IRB loses its significance, whilst its squared term becomes negative and significant suggesting that A-IRB Banks concentrate on A-IRB methods, while the Foundation method is progressively abandoned. Indeed, A-IRB shows an increasing negative and significant coefficient moving from left to right, confirming the ability of this method to further reduce capital absorption. As to the measures of a bank’s true risk exposure, ASSETS GROWTH reduces RWA/TA. NPL System still turns out negative, again possibly suggesting some disciplinary effect and/or drop in demand for credit for banks in high NPL contexts. Although the regressors OFF/TA and OTHER/TA seem to play a stronger role in changing RWA density than in the RWA/EAD regressions, we interpret that result as almost mechanical given their definition. Most importantly for us, the role of EQUITY as a determinant of RWA density is confirmed, even if with a positive and significant sign only for A-IRB banks.9 The second robustness check aims to better control for the adoption of IRB-methods. For that reason, in Table 7 we re-estimate the model of Table 4 on different samples of banks, after introducing the terms DUMMY F-IRB and DUMMY A-IRB and removing the squared terms of F-IRB and A-IRB. This specification is a robustness check for the Roll-Out effect. Here we notice that F-IRB shows always negative and significant, whilst the DUMMY F-IRB is insignificant. On the opposite, for the A-IRB method, the DUMMY A-IRB is always negative and significant, with the term A-IRB showing a positive and significant effect only for A-IRB banks. This is consistent with the hypothesis of gradual migration from Standard to F-IRB and A-IRB (both methods appear significant for all the subgroups of banks), together with the possibility to achieve a more relevant reduction of capital absorption with A-IRB against F-IRB methods. Finally, the other regressors – including the regressore maintain the significance already found in the previous estimates. The third robustness check assesses whether the indications of regulatory arbitrage intensified for the banks from the peripheral countries hit by the euro-crisis. This is tackled by the regressions in Table 8 where we identify the banks most affected by those severe macro conditions on the basis of two sample partitions. First, we dichotomize banks from countries with high public debt and banks from low public debt countries. Second, we separate the banks belonging to countries suffering a high interest rate spread – on their 10-year government bonds vis-à-vis Germany – from the banks belonging to countries enjoying a low interest rate spread. Despite the worry that the decrease in the number of observations might endanger the GMM approach, as highlighted by the Hansen statistic, the model appears to be satisfactorily stable. Both for the Total group and F-IRB banks, the main determinants of cap-

9 That result, in particular, is not disappointing, if one considers that both the regressors and the dependent variable are ratios to Total Assets.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

10

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Table 5 Interaction between lower level of capital and IRB approaches.

L.RWA/EAD F-IRB ˆ F-IRBSQ A-IRB ˆ A-IRBSQ NPL ASSETS GROWTH NLOANS/ LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013

IRB ADV Mod 01

IRB ADV Mod 02

IRB ADV Mod 03

IRB ADV Mod 04

IRB ADV Mod 05

IRB ADV Mod 06

IRB ADV Mod 07

0.4958*** 0.107 0.0251 0.102 −0.0029** 0.001 −0.2492*** 0.087 0.0012* 0.001 0.4346** 0.193 −0.1324*** 0.036 −0.1051 0.154 −0.6628** 0.287 −0.4472 1.254 2.7234*** 0.895 −0.1170*** 0.039 −0.0122 0.018 −0.1648* 0.095 −0.1960*** 0.054 −0.0024 0.007 −0.3153 0.263 0.3361 1.571 0.3219 1.260 1.3449 2.210 −0.0166 0.054 0.0311 0.044 −1.2361 1.360 2.1393* 1.217 −1.6623 1.103 −4.5154*** 1.686

0.3998*** 0.113 0.1079 0.172 −0.0036* 0.002 −0.1439 0.100 0.0004 0.001 0.4032** 0.198 −0.1332*** 0.035 −0.0907 0.139 −0.6123** 0.284 0.0301 1.115 3.3306*** 1.085 −0.1311*** 0.045 −0.0163 0.020 −0.1380 0.084 −0.1846*** 0.045 −0.0049 0.008 −0.2772 0.217 0.5772 1.515 −0.3273 1.298 0.9855 2.118 −0.0309 0.055 0.0107 0.047 −1.1805 1.212 1.7006 1.235 −2.3241* 1.252 −5.4617*** 1.885 13.6908*** 5.300

0.3242*** 0.115 0.0485 0.170 −0.0030 0.002 −0.1312 0.098 0.0002 0.001 0.5215*** 0.202 −0.1422*** 0.035 −0.0050 0.140 −0.7244*** 0.275 0.9471 1.181 3.0503*** 1.050 −0.1223*** 0.043 −0.0180 0.022 −0.1051 0.085 −0.1684*** 0.046 −0.0042 0.008 −0.2843 0.211 0.0380 1.751 −0.3540 1.369 −0.3512 2.231 −0.0270 0.056 0.0100 0.050 −1.3633 1.185 1.6521 1.282 −2.2958* 1.318 −5.5739*** 1.903

0.4854*** 0.110 0.0088 0.095 −0.0028*** 0.001 −0.2491*** 0.085 0.0012 0.001 0.3697* 0.203 −0.1312*** 0.036 −0.0833 0.155 −0.6106** 0.273 0.1268 1.197 3.0500*** 1.146 −0.1264*** 0.045 −0.0095 0.018 −0.1518* 0.090 −0.1879*** 0.051 −0.0028 0.008 −0.2737 0.246 0.0173 1.398 −0.7774 1.141 1.0085 2.166 −0.0125 0.052 0.0095 0.046 −1.1499 1.307 1.9304 1.263 −1.8446 1.247 −4.7141*** 1.775 3.8235* 2.064

0.4458*** 0.111 0.0252 0.098 −0.0031*** 0.001 −0.2458*** 0.082 0.0011 0.001 0.3998* 0.206 −0.1375*** 0.035 −0.0431 0.156 −0.6478** 0.271 0.4633 1.240 2.9821*** 1.095 −0.1251*** 0.043 −0.0122 0.020 −0.1371 0.090 −0.1821*** 0.051 −0.0038 0.008 −0.2661 0.241 −0.1183 1.559 −0.9107 1.189 0.4914 2.240 −0.0088 0.053 0.0055 0.047 −1.0905 1.270 1.9223 1.284 −1.8212 1.263 −4.8265*** 1.762

0.4887*** 0.109 0.0332 0.151 −0.0032* 0.002 −0.2610*** 0.084 0.0013* 0.001 0.4290** 0.189 −0.1362*** 0.036 −0.0453 0.148 −0.6794** 0.268 0.0599 1.293 2.4641** 0.986 −0.1082*** 0.040 −0.0152 0.019 −0.1437 0.090 −0.1839*** 0.052 −0.0020 0.007 −0.2813 0.257 0.2500 1.531 −0.2692 1.275 0.6713 2.174 −0.0045 0.052 0.0216 0.049 −1.0437 1.335 2.1728* 1.265 −1.4873 1.196 −4.2673** 1.746

0.4605*** 0.116 −0.0237 0.139 −0.0025* 0.001 −0.2584*** 0.086 0.0012 0.001 0.5051*** 0.196 −0.1409*** 0.036 −0.0020 0.144 −0.7465*** 0.264 0.3796 1.299 2.2728** 0.964 −0.1028*** 0.039 −0.0191 0.020 −0.1282 0.089 −0.1771*** 0.052 −0.0016 0.007 −0.2743 0.250 0.0557 1.723 −0.4584 1.237 0.0515 2.199 0.0001 0.052 0.0187 0.049 −1.0529 1.327 2.1497* 1.279 −1.4718 1.231 −4.1954** 1.772

LESS CAPITAL UNDER CAPITAL

8.4948* 5.007 22.4970*** 4.702

POOR CAPITAL LESS CAPITAL* DUMMY F-IRB

UNDER CAPITAL* DUMMY F-IRB

42.3311 29.805

31.3106 25.201

2.8832 3.192 −6.7282 4.696 −14.3922*** 4.760 −18.9473*** 3.872 14.1167 26.044

296 70 0.3426 50 23 0.5438 0.7740

296 70 0.2679 53 23 0.4900 0.7824

296 70 0.4109 56 23 0.5555 0.7839

UNDER CAPITAL* DUMMY A-IRB POOR CAPITAL* DUMMY F-IRB POOR CAPITAL* DUMMY A-IRB

N N(g) AR2-p J Hansen-df Hansen-p R-squared

−0.6650 3.670 0.7623 2.083

0.2598 3.852 −10.9086** 5.436

LESS CAPITAL* DUMMY A-IRB

Constant

3.2651 2.349 4.2693* 2.364

29.6658 27.950

23.5620 28.616

31.1517 29.702

296 70 0.319 51 23 0.4557 0.7884

296 70 0.304 52 23 0.4990 0.7892

296 70 0.3077 52 23 0.5116 0.7882

2.3153 2.667 0.3962 1.991 −12.0386** 5.305 0.6605 2.861 25.1700 29.415 296 70 0.3064 54 23 0.5137 0.7909

The table represents different Sys-GMM model estimations for the dependent variable RWA/EAD upon the A-IRB banks. In this case an approach 1-step is considered with lags (2 3) for instrumenting the endogenous variables L1.RWA/EAD and NLOANS/LIABILITIES. All regression include time dummies. N represents the number of observations available, N(g) represents the number of banks available, AR2-p represents the p-value of Arellano–Bond test for autocorrelation of second order; J represents the number of instruments; Hansen-df represents the degree of freedom for the Hansen statistic; Hansen-p represents the p-value of the Hansen statistic. Windmeijer-corrected, bank-clustered standard errors are given in italics. ***, ** and * indicate statistical significance at the 1%, 5%, and 10% levels respectively.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model

ARTICLE IN PRESS

JFS-491; No. of Pages 15

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx Table 6 Determinants of RWA/TA: Robustness checks on determinants of RWA density.

L.RWA/EAD F-IRB ˆ F-IRBSQ A-IRB ˆ A-IRBSQ NPL ASSETS GROWTH NLOANS/LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013 Constant

N N(g) AR2-p J Hansen-df Hansen-p R-squared

Total

IRB

IRB ADV

0.9003*** 0.083 −0.1080** 0.043 0.0008* 0.000 −0.1304** 0.056 0.0010* 0.001 −0.1147 0.082 −0.1437*** 0.048 0.0425 0.126 0.0462 0.100 0.1362 0.303 −0.3152 0.667 0.0174 0.034 0.0000 0.001 0.0909 0.073 0.0189 0.037 0.0021 0.003 0.0825 0.137 0.3154 0.652 −0.0422 0.656 0.1403 0.688 0.0139 0.024 0.0018 0.017 −0.4821 0.907 −0.6036 0.738 −1.6675* 1.004 −1.3122 0.996 0.6585 11.914

0.7360*** 0.096 −0.1793*** 0.065 0.0012* 0.001 −0.1728*** 0.066 0.0011* 0.001 −0.0501 0.111 −0.1106*** 0.022 0.2066** 0.098 −0.1238 0.129 0.8098 0.530 0.7197 0.535 -0.0516*** 0.018 −0.0190 0.014 0.1651*** 0.051 0.0880** 0.035 −0.0027 0.005 −0.0615 0.157 -0.6746 0.854 −0.0313 0.766 −0.2528 0.934 0.0332 0.030 0.0136 0.022 −0.8918 0.995 0.0308 0.764 −1.2288 1.042 −1.5329 1.246 −17.2604 14.235

0.6602*** 0.112 0.0176 0.083 −0.0026*** 0.001 −0.2285*** 0.073 0.0011** 0.001 0.2128 0.200 −0.1221*** 0.030 0.0770 0.117 −0.5500** 0.240 −0.1651 0.896 1.5061* 0.876 -0.0805** 0.035 −0.0107 0.013 0.1058 0.072 0.0262 0.036 0.0008 0.008 −0.2222 0.227 -0.5301 1.089 −0.3168 1.004 0.9601 1.652 0.0202 0.041 −0.0089 0.022 −2.0765* 1.130 0.2788 0.956 −1.4053 0.986 −2.8834** 1.393 15.9030 22.824

881 208 0.2639 50 23 0.2421 0.4631

505 117 0.2993 50 23 0.3333 0.3051

296 70 0.2802 50 23 0.5783 0.5160

The table represents different Sys-GMM model estimations for the dependent variable RWA/TA upon the different groups of banks (Total, IRB, IRB ADV). In this case an approach 1-step is considered with lags (2 3) for instrumenting the endogenous variables L1.RWA/TA and NLOANS/LIABILITIES. All regression include time dummies. N represents the number of observations available, N(g) represents the number of banks available, AR2-p represents the p-value of Arellano–Bond test for autocorrelation of second order; J represents the number of instruments; Hansen-df represents the degree of freedom for the Hansen statistic; Hansen-p represents the p-value of the Hansen statistic. Windmeijer-corrected, bank-clustered standard errors are given in italics. ***, ** and * indicate statistical significance at the 1%, 5%, and 10% levels respectively.

11

Table 7 Determinants of RWA/EAD: Robustness checks on IRB adoption.

L.RWA/EAD F-IRB A-IRB DUMMY F-IRB DUMMY A-IRB NPL ASSETS GROWTH NLOANS/LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013 Constant

N N(g) AR2-p J Hansen-df Hansen-p R-squared

Total

IRB

IRB ADV

0.8744*** 0.107 −0.0435* 0.023 0.0193 0.030 −0.2483 1.329 −4.8239** 2.289 −0.1356* 0.075 −0.0689*** 0.023 −0.0063 0.082 0.1284 0.086 −0.0594 0.315 0.6258 0.646 −0.0219 0.019 −0.0009 0.001 −0.0178 0.042 −0.0353 0.025 0.0046 0.005 0.1591 0.180 −0.4594 0.663 0.7748 1.030 1.0643 0.661 −0.0019 0.034 0.0096 0.014 1.0173 1.365 0.8117 0.730 −1.1734 0.838 −0.8408 1.223 5.9242 11.219

0.5050*** 0.138 −0.0961*** 0.030 −0.0174 0.036 1.2099 1.731 −4.8125** 2.443 0.0812 0.147 −0.0630*** 0.022 0.0214 0.089 −0.2104 0.173 0.1799 0.614 2.1261*** 0.814 -0.0853*** 0.030 −0.0265 0.018 -0.0568 0.054 −0.1137*** 0.035 -0.0087 0.006 −0.2059 0.187 1.4769 1.116 1.5050 1.111 −0.7028 1.149 −0.0340 0.051 0.0517 0.032 −1.4239 1.188 1.0050 0.944 −2.2330** 1.042 −3.9190*** 1.461 16.3552 16.623

0.5105*** 0.103 −0.2158*** 0.057 −0.0808* 0.042 2.1043 1.962 −4.6678* 2.423 0.3759* 0.204 −0.1327*** 0.036 -0.1100 0.167 −0.6017** 0.284 −0.5540 1.279 2.6311*** 0.917 -0.1135*** 0.039 −0.0111 0.018 -0.1600 0.098 −0.1895*** 0.056 −0.0023 0.007 −0.3064 0.272 0.0053 1.493 0.7044 1.304 1.2243 2.401 −0.0169 0.055 0.0421 0.047 −1.2020 1.423 2.0944* 1.222 −1.6128 1.062 −4.4664*** 1.594 43.1457 30.222

881 208 0.247 50 23 0.4872 0.8886

505 117 0.3182 50 23 0.4226 0.7950

296 70 0.3621 50 23 0.6179 0.7759

The table represents different Sys-GMM model estimations for the dependent variable RWA/EAD upon the different groups of banks (Total, IRB, IRB ADV). In this case an approach 1-step is considered with lags (2 3) for instrumenting the endogenous variables L1.RWA/TA and NLOANS/LIABILITIES. All regression include time dummies. N represents the number of observations available, N(g) represents the number of banks available, AR2-p represents the p-value of Arellano–Bond test for autocorrelation of second order; J represents the number of instruments; Hansendf represents the degree of freedom for the Hansen statistic; Hansen-p represents the p-value of the Hansen statistic. Windmeijer-corrected, bank-clustered standard errors are given in italics. ***, ** and * indicate statistical significance at the 1%, 5%, and 10% levels respectively.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

12

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

Table 8 Determinants of RWA/EAD: robustness checks for sensitivity to macro conditions. Variable L.RWA/EAD F-IRB ˆ F-IRBSQ A-IRB ˆ A-IRBSQ NPL ASSETS GROWTH NLOANS/ LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013 Constant

N N(g) AR2-p J Hansen-df Hansen-p R-squared

Total

Total Country High Debt

Total Country Low Debt

Total Country High Spread

Total Country Low Spread

IRB

IRB Country High Debt

IRB Country Low Debt

IRB Country High Spread

IRB Country Low Spread

0.8714*** 0.110 −0.1380*** 0.050 0.0010* 0.001 −0.1537** 0.062 0.0012* 0.001 −0.1180 0.076 −0.0687*** 0.023 0.0043 0.083 0.1002 0.090 0.0717 0.351 0.5625 0.674 −0.0199 0.019 −0.0009 0.001 −0.0165 0.042 −0.0331 0.026 0.0045 0.005 0.1775 0.179 −0.4140 0.658 0.6852 1.038 1.3516* 0.701 −0.0074 0.034 0.0087 0.015 1.1107 1.358 0.8502 0.740 −1.1454 0.864 −0.7087 1.295 3.7728 11.973

0.7115*** 0.110 −0.1922*** 0.074 0.0017** 0.001 −0.2137*** 0.066 0.0016** 0.001 0.0371 0.122 −0.0933*** 0.028 0.0508 0.074 0.0228 0.121 0.3736 0.402 1.2434 0.809 −0.0373* 0.022 −0.0051 0.011 −0.0010 0.052 −0.0316 0.021 0.0000 0.006 0.0145 0.211 0.9154 0.996 0.6347 1.119 1.4706 1.063 −0.0779 0.053 0.0117 0.028 −0.7008 1.521 0.6286 0.857 −2.2832** 1.149 −3.3259** 1.627 1.8464 10.989

0.9642*** 0.163 −0.1163 0.095 0.0012 0.001 0.0628 0.112 -0.0005 0.001 0.0012 0.107 −0.0373 0.039 0.0132 0.132 −0.1224 0.431 −0.5658 0.974 0.5380 0.820 −0.0412 0.032 −0.0013 0.001 −0.0357 0.072 −0.0730 0.084 0.0094* 0.005 0.2929 0.244 −0.5434 1.696 −1.3574 3.100 −0.7468 1.209 0.0411 0.037 0.0105 0.020 4.2428* 2.195 0.8994 1.207 1.2203 1.762 2.1798 1.952 7.9495 29.276

0.8561*** 0.128 −0.0172 0.134 −0.0002 0.002 −0.0491 0.120 0.0015 0.002 −0.0156 0.095 −0.0330 0.043 0.0290 0.156 0.1433 0.222 −0.9734 0.654 0.6825 0.782 −0.0277 0.022 0.0001 0.014 -0.1512** 0.074 −0.0754 0.059 0.0067 0.005 0.2346 0.248 0.8865 1.051 1.2473 1.898 0.2737 1.326 −0.0832 0.073 0.0158 0.072 1.7543 2.123 0.9789 1.250 −1.1937 2.280 −0.3299 2.248 21.7835 21.563

0.6777*** 0.092 −0.1676** 0.066 0.0010 0.001 −0.2242*** 0.062 0.0017*** 0.001 0.0219 0.146 −0.0721*** 0.026 0.0797 0.073 −0.0232 0.142 0.6751 0.434 1.0358 0.655 −0.0257 0.024 −0.0010 0.001 0.0419 0.049 −0.0258 0.024 −0.0059 0.008 0.0314 0.195 −0.0334 0.991 1.0103 1.148 0.8403 1.068 −0.0176 0.037 0.0084 0.021 −0.0275 1.446 0.4291 0.779 −1.5431* 0.898 −2.3098* 1.223 −3.9200 11.529

0.4869*** 0.145 −0.1862** 0.094 0.0011 0.001 −0.1770** 0.081 0.0010 0.001 0.1225 0.152 −0.0629*** 0.022 0.0126 0.087 −0.2528 0.187 0.3611 0.637 2.1066*** 0.799 −0.0803*** 0.031 −0.0238 0.018 −0.0713 0.054 −0.1156*** 0.035 −0.0098 0.007 −0.1726 0.186 1.4094 1.174 1.4487 1.116 −0.4330 1.085 −0.0563 0.049 0.0464 0.032 −1.2141 1.160 1.0572 0.970 −2.3105** 1.063 −3.8442*** 1.488 16.1799 16.048

0.3920** 0.172 −0.2085 0.135 0.0014 0.001 −0.1929** 0.097 0.0009 0.001 0.2182 0.189 −0.0635** 0.028 −0.0353 0.098 −0.2276 0.240 0.0027 0.923 3.2315*** 1.184 −0.1284** 0.051 −0.0347 0.027 −0.0864 0.069 −0.1467*** 0.036 −0.0131 0.009 −0.0920 0.240 3.4622** 1.684 0.3877 1.236 −0.1605 1.429 −0.2167** 0.091 0.0052 0.043 −1.3279 1.538 −0.2967 1.004 −4.1174*** 1.364 −6.8662*** 2.179 32.1265 22.721

0.5319*** 0.127 −0.1466 0.111 0.0002 0.001 −0.0974 0.135 -0.0007 0.001 −0.1027 0.434 −0.0130 0.044 −0.1946* 0.102 1.1997*** 0.420 −1.1302 1.046 1.5809** 0.795 −0.0347 0.040 0.0060 0.013 −0.1318** 0.066 −0.1824*** 0.070 −0.0107 0.008 −0.6254* 0.321 −0.6302 2.250 3.4611 5.382 0.6319 2.133 0.0580 0.057 −0.0288 0.042 −4.4995*** 1.552 1.1586 1.538 −2.5102* 1.501 −2.5529 2.146 56.5196* 29.757

0.1757 0.111 0.1899 0.174 −0.0012 0.002 0.1572 0.149 -0.0005 0.002 0.4311 0.279 −0.0586 0.050 −0.0534 0.176 −0.3574 0.381 0.2466 0.868 1.6994* 0.939 −0.0444 0.029 −0.0140 0.011 0.0472 0.094 −0.1508*** 0.038 −0.0120* 0.007 −0.4695* 0.252 4.1076* 2.334 −2.1876 2.468 3.5957 2.344 −0.3870*** 0.099 0.2842* 0.161 −3.6398 2.426 -0.2564 3.123 −5.2823 3.362 −7.6618* 4.403 16.4461 30.711

0.5477*** 0.098 −0.2385** 0.097 0.0013 0.001 −0.2581*** 0.077 0.0016** 0.001 0.1676 0.213 −0.0557** 0.022 0.0403 0.087 −0.3440 0.234 0.6604 0.710 1.8498*** 0.625 −0.0755** 0.031 −0.0297 0.020 −0.0686 0.065 −0.1006** 0.045 −0.0136 0.011 0.0134 0.206 0.7878 1.208 0.8904 1.162 −1.1253 1.200 0.0005 0.032 0.0337 0.032 0.0943 1.207 0.6761 0.887 −1.8173* 1.068 −2.9713** 1.240 9.3378 17.936

881 208 0.255 50 23 0.5315 0.8889

607 145 0.5105 50 23 0.3754 0.8807

274 63 0.1797 50 23 0.1057 0.7913

249 62 0.1815 50 23 0.6853 0.7794

632 146 0.505 50 23 0.1827 0.9071

505 117 0.3029 50 23 0.3530 0.7980

384 88 0.2987 50 23 0.6635 0.8670

121 29 0.6618 50 23 1.0000 0.7992

88 21 0.164 49 22 1.0000 0.8610

417 96 0.2272 50 23 0.2536 0.8001

The table represents different Sys-GMM model estimations for the dependent variable RWA/EAD upon the different groups of banks, determined by the category of banks (Total, IRB) and the country macroeconomic condition. In this case an approach 1-step is considered with lags (2 3) for instrumenting the endogenous variables L1.RWA/EAD and NLOANS/LIABILITIES. All regression include time dummies. N represents the number of observations available, N(g) represents the number of banks available, AR2-p represents the p-value of Arellano–Bond test for autocorrelation of second order; J represents the number of instruments; Hansen-df represents the degree of freedom for the Hansen statistic; Hansen-p represents the p-value of the Hansen statistic. Windmeijer-corrected, bank-clustered standard errors are given in italics. ***, ** and * indicate statistical significance at the 1%, 5%, and 10% levels respectively.

ital absorption are confirmed.10 We refer both to the capabilities of IRB methodologies to reduce capital absorption, as well as the other instruments already identified in Table 4 as potential determinants for reducing RWA density, like OTHER/TA, ASSETS GROWTH and OFF/TA. In particular, we confirm the role of the variable EQUITY which always acts as a positive and significant determinant of RWA/EAD. Overall, we do not obtain clear signals that regulatory arbitrage intensified more at the banks from crisis-hit peripheral countries. In fact, focusing on the sensitivity of the RWA/EAD to

10 In this case, the limited number of A-IRB banks hampers us to consider those banks in this part of the analysis.

EQUITY we do find that the coefficient for crisis country banks is higher than for their homologues from non-crisis countries if we refer to the partition based on the level of public debt, but we obtain the opposite result if we consider the partition based on the level of the spread. The rationale behind our fourth and last robustness check stems from aiming to limit the potential effects of measurement errors. Since some of our variables might suffer from measurement errors, we employ a quantile regression approach, so to obtain results that are more robust to outliers than other least squared regressions (Arellano and Bonhomme, 2015). Moreover, by considering different quantile levels, in Table 9 we control for the effect of our variables of interest upon different quantile levels of the depen-

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model

ARTICLE IN PRESS

JFS-491; No. of Pages 15

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

13

Table 9 Determinants of RWA/EAD: Robustness checks on IRB Sample via Quantile Analysis. Variable L.RWA/EAD F-IRB ˆ F-IRB SQ A-IRB ˆ A-IRB SQ NPL ASSETS GROWTH NLOANS/ LIABILITIES NPL SYSTEM SIZE EQUITY ˆ EQUITYSQ Z-SCORE OFF/TA OTHER/TA ROA L.GDP GROWTH LISTED STATE AID STRESS TEST RESOLVING INSOLVENCY OVERALL STANDARD Tau2010 Tau2011 Tau2012 Tau2013 Constant N N(g) AR2-p J Hansen-df Hansen-p R-squared

IRB Sys-GMM

IRB Q(0.10)

IRB Q(0.25)

IRB Q(0.50)

IRB Q(0.75)

IRB Q(0.90)

0.4869*** 0.145 −0.1862** 0.094 0.0011 0.001 −0.1770** 0.081 0.0010 0.001 0.1225 0.152 −0.0629*** 0.022 0.0126 0.087 −0.2528 0.187 0.3611 0.637 2.1066*** 0.799 −0.0803*** 0.031 −0.0238 0.018 −0.0713 0.054 -0.1156*** 0.035 −0.0098 0.007 −0.1726 0.186 1.4094 1.174 1.4487 1.116 −0.4330 1.085 −0.0563 0.049 0.0464 0.032 −1.2141 1.160 1.0572 0.970 −2.3105** 1.063 −3.8442*** 1.488 16.1799 16.048 505 117 0.3029 50 23 0.3530 0.7913

0.3777*** 0.047 −0.0569 0.082 0.0006 0.001 0.0152 0.068 -0.0003 0.001 0.2489** 0.112 −0.0502 0.033 −0.0661** 0.034 −0.0607 0.134 1.8645*** 0.566 4.3053*** 0.777 −0.1780*** 0.042 0.0108 0.010 -0.0285 0.040 -0.0851*** 0.014 −0.0037 0.008 -0.5730* 0.335 0.0117 1.051 4.0978*** 1.097 1.0638 1.057 −0.1220*** 0.039 0.0527* 0.030 -3.3311 2.065 2.4466 1.587 −1.2491 1.570 −5.3359*** 1.501 −28.4421** 11.513 618

0.5438*** 0.069 −0.0663 0.073 0.0004 0.001 −0.0515 0.063 0.0002 0.001 0.0234 0.101 −0.0581* 0.035 −0.0520* 0.027 −0.0177 0.128 1.0997*** 0.398 3.8024*** 0.901 −0.1703*** 0.047 −0.0008 0.011 −0.0468 0.036 −0.0777*** 0.016 −0.0057 0.007 −0.1167 0.277 0.9563 1.019 3.2632*** 0.986 −0.2150 0.969 −0.0927** 0.038 0.0489** 0.024 −0.8719 1.878 1.7215 1.333 −1.7261 1.303 −3.5464** 1.549 −13.1023 8.077 618

0.6691*** 0.046 −0.0323 0.058 0.0001 0.001 −0.0284 0.056 0.0000 0.001 0.0472 0.108 −0.0484** 0.020 −0.0185 0.021 −0.1432 0.143 0.6661** 0.300 2.1525*** 0.546 −0.0961*** 0.029 −0.0077 0.008 −0.0532 0.036 -0.0818*** 0.018 −0.0065 0.006 −0.1385 0.209 1.0596 0.823 1.7559** 0.787 −1.4900* 0.832 −0.0248 0.032 0.0566*** 0.021 −3.2334** 1.525 −1.1749 1.043 -4.3522*** 1.087 −4.2264*** 1.258 −0.7749 6.376 618

0.5078*** 0.063 −0.0812 0.094 0.0003 0.001 −0.0367 0.090 −0.0001 0.001 0.2409 0.170 −0.0633*** 0.021 −0.0031 0.031 −0.3148 0.203 0.0790 0.395 2.6716*** 0.520 -0.0962*** 0.025 -0.0213** 0.009 −0.0254 0.050 -0.1092*** 0.032 −0.0161* 0.008 -0.1285 0.288 1.9765 1.206 1.2227 0.946 −2.6928*** 0.970 −0.0388 0.031 0.0555* 0.029 −5.7487*** 2.076 −4.2889*** 1.356 -8.2416*** 1.353 −8.9687*** 1.627 23.0483** 10.583 618

0.3499*** 0.060 −0.2052* 0.124 0.0003 0.001 −0.2519** 0.101 0.0011 0.001 0.6489 0.430 −0.0238 0.035 −0.0554 0.057 −0.8113** 0.370 −1.8557** 0.770 3.5451*** 0.750 -0.1315*** 0.040 −0.0388*** 0.012 −0.0124 0.084 −0.1660*** 0.048 −0.0247** 0.012 −0.4124 0.429 3.0238 1.965 2.0111 2.151 −2.4842* 1.376 −0.1351*** 0.047 0.0062 0.040 −8.2693*** 2.351 −6.1580*** 2.027 −8.5536*** 2.300 −10.3921*** 2.427 86.6509*** 17.597 618

0.5187

0.5608

0.5618

0.5887

0.5186

The table represents different model estimations for the dependent variable RWA/EAD upon the IRB group of banks. More in particular, the first column represents the result obtained by the Sys-GMM model tested in the earlier table, whilst the other columns represents the estimation obtained by different Panel Quantile Regression performed upon the quantile 0.10, 0.25, 0.50, 0.75, 0.90.

dent variable. In this way, F-IRB and A-IRB regressors are confirmed significant especially for the largest quantile, similarly to Assets Growth, significant above the median. Noteworthy, SIZE exhibits a positive and significant coefficient for the lower quantiles, whilst in the higher quantiles it shows a negative sign. EQUITY appears to be almost stable, with a positive and significant coefficient, which is higher for both bottom and top quantiles. OTHER/TA exhibits a stable value over the alternative estimates. Z-SCORE in this case

gains significance and, as expected, it negatively influences the RWA density in case of higher quantiles. STATE AID shows a positive and significant value for quantiles below the median. STRESS TEST exhibits a negative and significant value for quantiles above the median. RESOLVING INSOLVENCY highlights a negative and significant value for quantiles below the median and top quantile. Finally, the relevance of the Standard methodology over the whole national banking system (OVERALL STANDARD), which we consider like a

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS

14

G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

proxy of the possible effort by national supervisors to perform their control, exerts a positive impact.

measurement errors, we run our regressions on rank-transformed variables and obtained confirmation of the previous results.

4.4. Summary of the main contributions and results

5. Conclusions

In this paper we investigated the extent of regulatory arbitrage via RWA dispersion at European banks providing various novel contributions. They can be summarized as follows. As a first step, we innovated with respect to most previous studies improving from considering RWA/TA (Total Assets) to focusing on RWA/EAD (Exposure At Default). We argued, in fact, that studying RWA/TA will most likely lead to overestimating regulatory arbitrage. Indeed, IRB models allow lawful capital-saving Roll-Out effects which RWA/TA analyses disregard and likely misidentify as regulatory arbitrage. Instead, encapsulating Roll-Out effects, RWA/EAD avoids false positive identification. Accordingly, to study the phenomenon we had to undertake a long and complex compilation of data that are currently unavailable in the usual databases employed by researchers (e.g., BankScope), which we derived from banks Basel Pillar Three statements. Next, employing state of the art econometric methods to deal with endogeneity, we ran regressions to test whether RWA/EAD is sensitive to the bank employing either a Standard or F-IRB or A-IRB approach. It turns out that, controlling for other possible determinants, F-IRB (A-IRB) banks have a systematically lower value of RWA/EAD with respect to Standard (F-IRB) banks. Overall, two different channels of capital requirements attenuation seem to be at work. The first channel describes how formerly Standard banks save on capital requirements by migrating assets to the IRB approach. Instead, the second channel kicks in once the asset migration from Standard to IRB is (almost) completed, and is particularly visible for A-IRB banks. Here equity-short banks might be tempted to engineer reductions in capital requirements by re-calibrating their internal asset risk weights. While the first channel of “capital saving” might include some regulatory arbitrage along with some appropriate asset risk re-classification, the concern is there that the second channel might include a larger component of regulatory arbitrage. We found strong evidence of those two channels. Several additional specifications beyond the baseline regression helped us to check the robustness of our results as well as to more precisely delineate under which circumstances F-IRB and/or A-IRB adoption leads to regulatory arbitrage. First, we zeroed in on poorly capitalized (Equity/RWA between 8 and 10%) and on under-capitalized (Equity/RWA below 8%) banks. We found that poorly capitalized banks manage to achieve significantly lower RWA/EAD when they adopt F-IRB or A-IRB, thus substantiating the suspicion of regulatory arbitrage via risk weights manipulation. Such signals of regulatory arbitrage were instead missing at undercapitalized banks, supposedly under closest supervisory scrutiny. Second, though the right measure of RWA dispersion is RWA/EAD, we estimated also the determinants of RWA/TA given that many studies focus on it. While some differences emerged, our results on the determinants of RWA/TA were qualitatively consistent with those found for RWA/EAD. Third, identifying the year of adoption of F-IRB and/or A-IRB methods and controlling for the portfolio share conferred to each IRB method (and so cleaning for the RollOut effects), we showed that new A-IRB users managed to obtain a significant reduction of RWA/EAD over and above the reduction achieved by the banks that were already A-IRB since previous years. The fact that this reduction shows up for new A-IRB users but not for new F-IRB users is consistent with the idea that advanced IRB methods offer banks more leeway to mitigate their risk weights. Fourth, studying whether the indications of regulatory arbitrage intensified for the banks from the peripheral countries hit by the euro-crisis we found no clear evidence of that, while validating the qualitative results of the previous estimates. Finally, addressing potential

A popular adage says: ‘Every law has its loophole’. And, indeed, though the Basel Committee on Banking Supervision worked for decades to come up with sophisticated rules to measure banking risks in objective ways, something is still left to banks’ subjective judgment. So, even with the sturdy armor of Basel II and Basel III, supervisors fret that banks could underreport their true risks. And the cruelty is that those suspected to underreport are the most sophisticated banks, not those using primitive methods. The main culprit is the mechanism by which the banks endowed with Internal Rating Based (IRB) – especially A-IRB – Models may manage to assuage their Risk Weighted Assets (RWAs) via risk weight manipulation, and thus lower their capital requirements. This would be regulatory arbitrage, something unfair and dangerous. In other words, those banks would be engaging in practices that, while being formally legitimate, end up in reducing (eluding a rise of) regulatory capital while risk doesn’t decrease (increases). This would lead to an artificial reduction of their RWA ratios. In this paper we investigated the extent of regulatory arbitrage via RWA dispersion at European banks providing various novel contributions as detailed in last section. We found evidence that regulatory arbitrage: i) was present; ii) likely materialized via risk weights manipulation with IRB models; iii) was stronger at Advanced-IRB vs Foundation-IRB banks. A caveat is in order. Our results are trustworthy to the extent that we succeeded in identifying the “true risk” underlying banks’ intermediation. To that end we deployed a large set of control variables aimed to capture the various multiple dimensions of on-balance and off-balance risks at each bank. Bearing that caveat in mind, our results suggest that regulatory arbitrage may be present for the more sophisticated banks adopting the IRB − especially the A-IRB − approach. The policy implication we can sketch is twofold. Firstly, refinements of the Basel technical standards may be needed to make sure that the same risks get the same treatment irrespective of a bank’s regulatory approach. Moreover, supervisors should probably improve the procedures to grant validation to a bank’s IRB model. Else, they should consider opting for simpler regulation (Haldane and Madouros, 2012). References Anderson, T.W., Hsiao, C., 1982. Formulation and estimation of dynamic models using panel data. J. Econometrics 18, 47–82. Arellano, M., Bond, S., 1991. Some tests of specification for panel data: monte Carlo evidence and an application to employment equations. Rev. Econ. Stud. 58, 277–297. Arellano, M., Bonhomme, S., 2015. Nonlinear panel data estimation via quantile regressions. CEMFI Working Paper No. 1505, July. Arellano, M., Bover, O., 1995. Another look at the instrumental variable estimation of error-components models. J. Econometrics 68, 29–51. Argimón, I., Ruiz-Valenzuela, J., 2010. The Effects of National Discretions on Banks. Banco de Espana Working Paper No. 1029, September. Arroyo, J.M., Colomer, I., García-Baena, R., González-Mosquera, L., 2012. Comparing Risk-Weighted Assets: The Importance of Supervisory Processes. Fin. Stabil. J., ˜ May. Banco de Espana, Ayadi, R., Naceur, S.B., Casu, B., Quinn, B., 2016. Does Basel compliance matter for bank performance? J. Fin. Stab. 23, 15–32. BCBS, 1988. International Convergence of Capital Measurement and Capital Standards. July. BCBS, 1996. Amendment to the Capital Accord to Incorporate Market Risks. January. BCBS, 1997. Principles for the Management of Interest Rate Risk. September. BCBS, 1999. A New Capital Adequacy Framework. June. BCBS, 2005. Studies on the Validation of Internal Rating Systems. May. BCBS, 2011. Basel III: A Global Regulatory Framework for More Resilient Banks and Banking Systems (revised version). June. BCBS, 2012. Core Principles for Effective Banking Supervision (revised version). September.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006

G Model JFS-491; No. of Pages 15

ARTICLE IN PRESS G. Ferri, V. Pesic / Journal of Financial Stability xxx (2016) xxx–xxx

BCBS, 2013. Regulatory Consistency Assessment Programme (RCAP) −Analysis of risk-weighted assets for market risk. January. BCBS, 2013. Regulatory Consistency Assessment Programme (RCAP) −Analysis of risk-weighted assets for credit risk in the banking book. July. BCBS, 2013. The regulatory framework: balancing risk sensitivity, simplicity and comparability. discussion paper, July. Barakova, I., Palvia, A., 2014. Do banks’ internal Basel risk estimates reflect risk? J. Fin. Stab. 13, 167–179. Beltratti, A., Paladino, G., 2016. Banks risk-weights and cost of equity. J. Empir. Fin., forthcoming. Blundell, R., Bond, S., 1998. Initial conditions and moment restrictions in dynamic panel data models. J. Econometrics 87, 115–143. Borio, C., Drehmann, M., Tsatsaronis, K., 2014. Stress-testing macro stress testing: does it live up to expectations? J. Fin. Stab. 12, 3–15. Brei, M., Gambacorta, L., 2016. Are bank capital ratios pro-cyclical? New evidence and perspectives. Econ. Policy 31, 357–403. Bruno, B., Nocera, G., Resti, A., 2014. The credibility of European banks’ risk-weighted capital: structural differences or national segmentations? 3rd EBA Policy Research Workshop How to Measure the Riskiness of Banks. Cannata F., Casellina, S., Guidi, G., 2012. Inside the labyrinth of RWAs: how not to get lost, Bank of Italy Occas. Paper No. 132, September. Cihak, M., Demirgüc-Kunt, A., Martinez Peria, M.S., Mohseni-Cheraghlou, A., 2013. Bank regulation and supervision in the context of the global crisis. J. Fin. Stab. 9, 733–746. Das, S., Sy, A.N.R., 2012. How Risky Are Banks’ Risk-Weighted Assets? Evidence from the Financial Crisis. IMF WP 12/36. EBA, 2013. Interim results of the EBA review of the consistency of risk-weighted assets. Top-down assessment of the banking book. February. EBA, 2013. Interim results update of the EBA review of the consistency of risk-weighted assets −Low default portfolio analysis. August. EBA, 2013. Report on the comparability of supervisory rules and practices. December.

15

EBA, 2013. Report on the pro-cyclicality of capital requirements under the Internal Ratings Based Approach. December. EBA, 2013. Report on variability of Risk Weighted Assets for Market Risk Portfolios. December. EBA, 2014. Consults on technical standards on supervisory benchmarking of internal approaches for calculating capital requirements. May. Fratianni, M., Pattison, J.C., 2015. Basel III in reality. J. Econ. Integr. 30 (1), 1–28. Gustin, E., Van Roy, P., 2014. The role of international models in regulatory capital requirements: a comparison of Belgian bank’s credit risk parameters. NBB Fin. Stab. Rev. Haldane, A.G., Madouros, V., 2012. The dog and the Frisbee. Fed. Res. B. Kansas City’s 366th economic policy symposium, 31 August. Hesse, H., Cihak, M., 2007. Cooperative banks and financial stability, IMF WP 07/02. Holtz-Eakin, D., Newey, W., Rosen, H.S., 1988. Estimating vector autoregressions with panel data. Econometrica 56, 1371–1395. Karim, D., Liadze, I., Barrell, R., Davis, E.P., 2013. Off-balance sheet exposures and banking crises in OECD countries. J. Fin. Stab. 9, 673–681. Köhler, M., 2016. Which banks are more risky? The impact of business models on bank stability. J. Fin. Stab. 16, 195–212. Ledo, M., 2011. Towards more consistent, albeit diverse, risk-weighted assets ˜ across banks. Riv. Estabilidad Fin. 21, Banco de Espana. Le Leslé, V., Avramova, S. 2012. Revisiting Risk-Weighted Assets. IMF WP 12/90. Mariathasan, M., Merrouche, O., 2014. The manipulation of Basel risk-weights. J. Fin. Intermed. 23, 300–321. Pagano, M.S., Sedunov, J., 2016. A comprehensive approach to measuring the relation between systemic risk exposure and sovereign debt. J. Fin. Stab. 23, 62–78. Roodman, D., 2009. A note on the theme of too many instruments, oxford bull. Econ. Stat. 71 (1), 135–158. Vallascas, F., Hagendorff, J., 2013. The risk sensitivity of capital requirements: evidence from an international sample of large banks. Rev. Fin., 1–42.

Please cite this article in press as: Ferri, G., Pesic, V., Bank regulatory arbitrage via risk weighted assets dispersion. J. Financial Stability (2016), http://dx.doi.org/10.1016/j.jfs.2016.10.006