Interest rate convergence across maturities: Evidence from bank data in an emerging market economy

Accepted Manuscript Interest rate convergence across maturities: Evidence from bank data in an emerging market economy Mark J. Holmes, Ana María Iregui, Jesús Otero PII: DOI: Reference:

S1062-9408(18)30471-6 https://doi.org/10.1016/j.najef.2019.03.008 ECOFIN 943

To appear in:

North American Journal of Economics & Finance

Received Date: Revised Date: Accepted Date:

4 September 2018 12 March 2019 12 March 2019

Please cite this article as: M.J. Holmes, A.M. Iregui, J. Otero, Interest rate convergence across maturities: Evidence from bank data in an emerging market economy, North American Journal of Economics & Finance (2019), doi: https://doi.org/10.1016/j.najef.2019.03.008

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

Interest rate convergence across maturities: Evidence from bank data in an emerging market economy∗ Mark J. Holmes Waikato Management School University of Waikato New Zealand

Ana Mar´ıa Iregui Unidad de Investigaciones Banco de la Rep´ ublica Colombia

Jes´ us Otero Facultad de Econom´ıa Universidad del Rosario Colombia March 2019

Abstract Against a background of financial liberalisation reforms, we assess the extent of market integration and competition in Colombian retail deposits and loans markets. We employ a dataset comprising banklevel interest rate data for different financial products across a range of banks. We utilise and further develop the Phillips and Sul convergence club approach by estimating the drivers of club formation. We find integration of the deposits market, but not loans where portfolio riskiness and loan maturity explain why there is not a fully integrated market. Also, the degree of loan market convergence responds asymmetrically to changes in monetary policy. JEL Classification: C33, E43, G21 Keywords: Convergence clubs, interest rates, maturity, Colombia. We are grateful to two anonymous reviewers, Luis Eduardo Arango, and participants at the 2018 meetings of the New Zealand Association of Economists and the Annual Conference of the International Association for Applied Econometrics (Montreal, Canada) for their comments and suggestions. Research assistance by Sergio Montoya is also acknowledged. The views expressed in the paper are the responsibility of the authors, and should not be interpreted as reflecting those of the Board of Directors of the Banco de la Rep´ ublica, or other members of its staff. The usual disclaimer applies. E-mail addresses: [email protected] (M.J. Holmes; corresponding author); [email protected] (A.M. Iregui); [email protected] (J. Otero). ∗

1

Introduction

Financial liberalisation reforms have been made by many emerging economies since the 1990s, where a key component has been the drive towards increased competition in retail banking. It is of interest to consider whether or not the new banking regimes that have emerged are characterised by integrated domestic markets for retail deposits and loans. Indeed, measuring the extent of integration or convergence across interest rates set by banks can provide valuable insights into how fragmented retail markets might still be despite significant changes in regulations. A further consideration is that the extent of convergence across maturities has implications for our understanding of the term structure of interest rates as based on expectations theory or liquidity preference considerations. This paper investigates the extent of interest rate convergence in the retail banking sector of an emerging market economy, namely Colombia, which embarked upon a financial liberalisation process in the early 1990s. Prior to these reforms, Colombia, like many other developing countries, was characterised by “financial repression”, a term used to describe an economy with a banking sector that is subjected to important restrictions in the form of high reserve requirements and liquidity ratios, interest rate ceilings, and subsidised credit to specific sectors of the economy, among others. Uribe and Vargas (2003) note that the reforms aimed at creating an independent central bank; redefining of the structure of the financial sector; relaxing the requirements for entry and exit of intermediaries; regulating mergers, acquisitions and liquidations; reducing reserve requirements; and liberalising interest rates although, despite the reforms, lending rate caps are still in place today through the operation of usury laws.1 1 The Colombian Superintendency of Financial Institutions has two categories of interest rate caps: the usury rate for conventional (i.e. ordinary and consumption) loans and the usury rate for microcredit (i.e. loans to very small enterprises). Supervised institutions must report their rates by category on a weekly basis, and these data are used by the

1

Tests of the different aspects of convergence started with the works of Baumol (1986), Barro (1991) and Barro and Sala-i-Martin (1992) that typically focussed on output growth. During the last decade or so, there has been an increasing number of applications to financial markets. Indeed, authors such as Fung (2009) and Narayan et al. (2011) have investigated the notions of absolute and conditional convergence in the banking sectors and stock markets of several low, middle and high income countries, respectively. In turn, variations of time-series and cointegration-based methodologies have been employed by Brada et al. (2005), Kim et al. (2005, 2006), Arghyrou et al. (2009), Mylonidis and Kollias (2010), Herwartz and Roestel (2011) and Baharumshah et al. (2013) to analyse convergence of key financial variables in European and US markets, by Kaul and Mehrotra (2007) to study Toronto-listed Canadian stocks that are also traded in stock exchanges in the United States, and by Yu et al. (2010) to examine financial markets in Asia. More recently, the policy initiatives undertaken by the European Union (EU) to integrate national markets, through the introduction of the Single Market Programme, have also provided a fertile area of research to test the hypothesis of price convergence in retail banking. Some examples are provided in the analyses of convergence patterns among retail interest rates in the European banking sector by Rughoo and Sarantis (2012, 2014), banking efficiency scores by Matousek et al. (2015), and sovereign bond yield spreads by Antonakakis et al. (2017). The distinctive aspect of this latter line of research is that it is based on the idea of convergence club formation put forward by Phillips and Sul (2007, 2009), according to which convergence may not hold for all individual series under consideration but for sub-groups of them. This paper analyses convergence club formation in Colombian retail banking using the Phillips and Sul (2007, 2009) methodology. These authors put Superintendency to fix the level of the usury rates. In both cases, the usury rate is set as 1.5 times the corresponding average rate for the system as a whole.

2

forward the notion of “relative” long-run equilibrium or convergence, according to which two series, let us say yit and yjt , are said to exhibit relative convergence when the ratio between them converges to unity as time passes, that is when limt→∞ (log yit / log yjt ) = 1, and suggest testing the null hypothesis of relative convergence using a regression that involves log-t as a regressor. Here, an algorithm based on this log-t regression approach clusters interest rates with a common unobserved factor in their variance. In this respect, sigma-convergence as opposed to beta-convergence deals with the reduction in the variance of the distribution of retail interest rates over time. In addition to detecting panel convergence, if present, a key benefit attached to the Phillips and Sul clustering algorithm test is that it can also reveal whether club formation is present. There do exist other time-varying approaches when it comes to convergence analysis. In our paper, we opt for the Phillips and Sul view of convergence which exploits all the information contained in the time-series dimension of the underlying data. Unlike methods based on unit root testing or cointegration, the Phillips and Sul (2007, 2009) approach does not necessitate any specific assumptions regarding the order of integration of the variables and allows for cases where the individual time series may be transitionally divergent. Indeed, the method by Phillips and Sul (2007, 2009) enables the detection of convergence where other methods such as stationarity tests fail insofar as stationary time series methods are unable to detect the asymptotic co-movement of two time series and therefore erroneously reject the convergence. Further to this, the concept of relative convergence advocated by Phillips and Sul (2007, 2009) is different from the concept of level convergence considered by Bernard and Durlauf (1995) and Evans and Karras (1996), which is defined as limt→∞ (log yit − log yjt ) = 0. Chatterji and Dewhurst (1996) offer an alternative view of club convergence formation within the context of the cross-section regression model advocated by Bau-

3

mol (1986).2 However, Evans (1996) explains that tests of convergence based on regressions that relate the growth rate of a variable to its initial value (and possibly other characteristics) are only valid under strong conditions. We provide three main distinctive features that add significant value to the existing literature. First, we are able to exploit a database of retail deposit and lending interest rates at bank-level over a period of more than 15 years. The information at our disposal permits us to analyse convergence club formation across different maturities in deposits and loans, which is a step forward in comparison to Rughoo and Sarantis (2012, 2014) where the analysis is carried out at country-level for EU economies. Second, we develop further the club convergence analyses by Rughoo and Sarantis (2012, 2014), Matousek et al. (2015) and Antonakakis et al. (2017), by looking at the factors driving convergence club membership. Third, the data span offers the possibility of implementing the Phillips and Sul algorithm in a recursive fashion, allowing us to gain insights into the potential effects of monetary policy on the number of (or membership of) convergence clubs resulting from monetary tightening or loosening. Although it can be argued that Colombia is not representative of the rest of the emerging economies, Colombia nonetheless shares two key common features with this group of countries. First, there is a heavily concentrated banking sector and, against this background, access to bank-level data offers the opportunity to examine whether financial liberalisation has had similar effects on the deposits and loans sides of the market. Second, like in many other developing (and for that matter also developed) countries, in Colombia banks operate in the presence of interest rate caps in the form of usury laws. Apart from the fact that the effect of such caps has typically been overlooked by the existing literature, what makes the Colombian experience interesting 2 See, for example, Degl’Innocenti et al. (2018) for a recent application and further elaboration of this approach to the analysis of financial centres’ competitiveness and economic convergence in European Union regions.

4

is that the effects of the Global Financial Crisis (GFC) have not been regarded as serious as has been the case for other economies, mainly because of favourable commodity price movements related to Colombian exports and the country’s strong economic policy framework. This suggests that a zero lower bound on interest rates has been much less of an issue when it comes to lax monetary conditions. Instead, the existence of usury laws gives rise to an upper boundary on lending rates. In assessing interest rate co-movements in retail banking, this opens up the interesting possibility of an upper bound, which is in sharp contrast to most other countries which have had concerns with a lower zero bound during the GFC era. The paper is organised as follows. Section 2 offers a brief review of the existing literature on interest rate convergence clubs. Section 3 outlines the econometric methodology. Section 4 describes the data. Section 5 presents the results of the club convergence analysis in deposit and lending interest rates. Section 6 examines the factors that drive the formation of convergence clubs. Section 7 relates the existence of convergence clubs to the stance of monetary policy. Section 8 concludes.

2

Literature on interest rate convergence clubs

The studies that apply the Phillips and Sul methodology to interest rate club convergence have focused on European interest rates. Rughoo and Sarantis (2014) consider the European Union retail banking sector by analysing deposit and lending rates to the household sector during the period 2003–2011. The presence of a single convergence club driving all interest rate series is taken as being indicative of an integrated market. Based on finding a single convergence club in each case, their empirical results point to the presence of convergence in all deposit and lending rates up to 2007. In sharp contrast, the null of convergence is rejected in all deposit and credit markets after the onset of the 2008 financial crisis. It is argued that the crisis has 5

brought an abrupt halt to the integration process in both the deposit and credit markets. In the consumer credit and mortgage markets, the crisis has most likely further compounded the existing structural and legal barriers prevalent in this market. In an earlier paper, Rughoo and Sarantis (2012) draw similar conclusions when they analyse deposit and lending rates to nonfinancial corporations. More recently, Antonakakis et al. (2017) examine the convergence patterns of Euro Area (EA) 17 countries’ sovereign bond yield spreads (relative to the German bund) over the period of March 2002 to December 2015. They reject full convergence across the EA17 bond yields spreads. In particular, three subgroup convergence clubs emerge. Despite short-run divergences, it is argued that EA17 sovereign bond yield spreads tend to converge in the long-run in most cases. Studies of Colombian interest rate setting across banks include Iregui and Otero (2013) over a 2002-10 study period. Using a pairwise unit root resting approach to test the law of one price for deposit and lending rates, they find that when banks are of different sizes, deposit rates adjust quickly, suggesting a competitive environment. By contrast, lending rates adjust rapidly when banks are of similar sizes, supporting market segmentation. Holmes et al. (2015) examine the impact of monetary policy changes on Colombian retail bank rates. They find evidence that some banks are more willing to reduce lending and deposit rates than increase them in response to monetary policy changes. Other studies of interest rate convergence in retail banking include Mart´ın-Oliver et al. (2007) who investigate the level and the determinants of retail banking interest rate differences among Spanish banks in the period 1989–2003. They find that the interest rates of twenty five different bank loan and deposit products adjust rather rapidly to their long-term values in response to external shocks as the relative version of the Law of One Price predicts, but the evidence runs contrary to the absolute version of the Law. In this respect, the credit risk premium is an important source of interest rate dispersion across banks and loan products. 6

3

Econometric methodology: A brief review

Previous support of the stochastic convergence hypothesis in the sense of Bernard and Durlauf (1995) where series share a common stochastic trend, have employed variations of individual unit root testing and cointegrationbased methodologies.3 By way of contrast, the Phillips and Sul clustering algorithm that we employ here provides an empirical modelling of long-run equilibria within a heterogeneous panel, outside of the cointegration setup. Following the detailed presentation of the methodology by Phillips and Sul (2007, 2009), we begin by considering the following time varying factor representation of a panel data Xit : Xit = git + ait ,

(1)

where Xit consists of common permanent components, git , which give rise to cross section dependence, and transitory components, ait , where i = 1, ..., N is the number of individuals in the panel, and t = 1, ..., T is the number of time observations. In this representation, Xit is assumed to contain both common and idiosyncratic terms, so that in order to separate one from the other, equation (1) can be re-written as: Xit =



git + ait µt



µt = δit µt ,

(2)

for all i and t, where δit is a time varying idiosyncratic element, while µt is a time varying component that is common to all individuals in the panel Xit . In equation (2), Phillips and Sul assume that the term δit can be modelled in a semiparametric form as: δit = δi + σi ξit L (t)−1 t−α , 3

(3)

See, for example, Iregui et al. (2002), Iregui and Otero (2013) and Holmes et al. (2015) in earlier studies of Colombian retail interest rates.

7

where δi is fixed, ξit ∼ iid (0, 1) across i but weakly dependent over t, and L (t) is a slowly varying function such that L (t) → ∞ as t → ∞; an illustration of the function L (t) is log(t). Phillips and Sul indicate that the formulation in (3) ensures that for all α ≥ 0, δit converges to δi , and so these conditions become the null hypothesis of interest. Within this framework, the Phillips and Sul develop a test of the null hypothesis of convergence based on: H0 : δ i = δ

and α ≥ 0,

(4)

against the alternative hypothesis of divergence: HA : δi 6= δ

∀i or α < 0.

(5)

The implementation of the test of convergence involves three stages. In the first stage, one needs to construct the cross sectional variance ratio H1 /Ht given by: N 1 X (hit − 1)2 , Ht = N i=1

(6)

where hit is referred to as the relative transition parameter, which traces out the transition path of each individual time series i in relation to the average of the panel, that is: hit =

N −1

Xit PN

i=1

Xit

=

N −1

δit PN

i=1 δit

.

(7)

In equations (6) and (7), the notion of relative convergence adopted by Phillips and Sul implies that, for a given number of time series in the panel N , hit → 1 and therefore Ht → 0 as t → ∞. In the second stage one runs the so-called log t regression: log



H1 Ht



− 2 log L (t) = a ˆ + ˆb log t + uˆt ,

8

(8)

for t = [rT ] , [rT ] + 1, ..., T, with r > 0. In this regression, Phillips and Sul use L (t) = log (t + 1) and further observe that the estimate of the slope coefficient ˆb = 2ˆ α, where α ˆ is the estimate of α in the null hypothesis. Here, interest resides not only in the sign of the coefficient associated to b = 2α, but also in its magnitude since it provides information regarding the speed of convergence of δit . For estimation purposes, the regression model uses a fraction rT of the observations, where the trimming parameter r is recommended to be set equal to 0.30 (in our empirical application we check the robustness of our findings when this parameter is modified taking on the values of 0.15, 0.20, 0.25 and 0.30). Finally, in the third stage the notion of relative convergence is tested through the use of a one-sided t test of the inequality part of the null hypothesis α ≥ 0, where the test statistic is constructed using a heteroskedasticity and autocorrelation consistent standard error (HAC). Using a 5% significance level, the null hypothesis is rejected when tˆb < −1.65. Phillips and Sul (2007, 2009) indicate that rejection of the null hypothesis of convergence does not rule out the possibility of convergence in subgroups of the individuals that conform the whole panel. Accordingly, to allow for this possibility, the authors develop a clustering algorithm with the purpose of determining the number of potential convergence clubs as well as their conforming members (also permitting for the existence of a category of nonconvergent time series). The original version of the Phillips and Sul algorithm was implemented in the software GAUSS, although versions of this code in R and Stata have also been produced respectively by Schnurbus et al. (2017) and Du (2017). For the purposes of our empirical analysis, we rely on the latter version of the computing code. The clustering algorithm consists of four steps, which can be sketched as follows (the interested reader is referred to the above mentioned references for a more detailed exposition): • Step 1: Remove the trend component of Xit by means of the Hodrick 9

and Prescott (1997) HP filter, and order the resulting de-trended version of the series according to the value of the last observation or, in case of highly volatile time series, the average value of the last half of observations. • Step 2: Create a core primary group of k ∗ individuals using the maximum tk , subject to tk > −1.65, from the sequential log(t) regression based on the k highest individuals in the panel, where 2 ≤ k < N . • Step 3: Add one individual at a time to the core primary group with k ∗ members, estimate the log(t) regression again, and add the new individual to the core primary group if the resulting t-statistic is greater than the criterion c∗ . Phillips and Sul refer to this step as the sieve condition, and recommend setting c∗ = 0 in small T samples, while for large T the asymptotic 5% critical value -1.65 is the recommended value. • Step 4: Form a second group of individuals for which the sieve condition stated in step 3 fails. Estimate the log(t) regression on this group and verify if the resulting t statistic is greater than -1.65. If this condition is satisfied then conclude that there are two convergence clubs, namely the core primary group and the second group. Otherwise, repeat steps 1 through 3 to see if the second group can be subdivided into smaller convergence clubs. If in step 2 there is no k for which tk > −1.65, conclude that the remaining individuals in the panel do not contain a convergence subgroup, and so the remaining individuals exhibit divergent behaviour.

4

Data

As a result of the financial reforms mentioned above, from the late 1980s to 2016 the share of government-owned banks in the assets of the financial 10

sector decreased from 43% to 4%, while the share of banks with foreign participation increased from 3% to 25%. Regarding the evolution of the number of banks through time, the total increased from 36 in 1985 to 41 banks in 1997; see Appendix 1 in Arango (2006). In the late 1990s, a period during which Colombia experienced the deepest recession ever recorded, the number of commercial banks decreased as a result of liquidations, acquisitions and mergers. By 2002, the number of banks was reduced to 29, and by 2016 this figure had reduced even further to 25 institutions. In terms of the state of competitive conditions prevailing in the banking sector, market concentration has increased in recent years. That is, while in 2001 five banks accounted for as much as 43% of the total assets, in 2016 this percentage had risen to approximately 68% of the banking sector. These developments can be put in the context of our empirical analysis. Clearly, if the above mentioned reforms succeeded in making the banking sector more competitive, then we would expect stronger support for the hypothesis of convergence. Associated with this, there is also the question of whether support for this hypothesis is generalised throughout the banking sector or specific to some segments of the market. The dataset comprises individual deposit and lending rates for 13 banking institutions from May 2002 to August 2017 thereby providing a total of 184 monthly observations for each series. In June 2017, these banks collectively accounted for almost the totality of the assets of the banking system in the country; that is, approximately 97%. The choice of banks included in the sample is dictated by the need to assemble a consistent database over the largest possible study period given data availability. The data are taken from information on all new deposits and (consumption and ordinary) loans reported by the banks to the Colombian Superintendency of Financial Institutions. More specifically, for each bank, maturity and time period deposit interest rates are consolidated into a single interest rate using a depositweighted average, and similarly for lending interest rates. The resulting 11

consolidated deposit rate can then be thought of as being about the same as the return of all original separate deposits; in the case of loans, the consolidated rate refers to the cost of only a subset of loan categories, namely consumption and ordinary loans.4 For the empirical analysis, we denote di,j,t as the deposit-weighted average rate paid by bank i on time deposit certificates (CDTs) at maturities j = 1 (short-term, i.e. less than 120 days), 2 (medium term, i.e. between 120 and 179 days), and 3 (long-term, i.e. 180 to more days), at time t. In turn, we denote li,j,t as the loan-weighted average rate charged by bank i on consumption and ordinary loans at maturities j = 1 (short-term, i.e. less than 365 days), 2 (medium term, i.e. between 366 and 1095 days), and 3 (long-term, i.e. 1096 to more days), at time t. The index i = 1, ..., 13 denotes the banks under consideration where, to facilitate the interpretation of the results, hereafter banks are reported in descending order according to the size of their assets as of June 2017. Thus, d1,j,t (l1,j,t ) denotes the deposit (lending) rate of the bank with the largest assets at maturity j and time t, while d13,j,t (l13,j,t ) denotes the corresponding rate of the bank with the smallest assets in the sample, also for maturity j and time t. For both deposits and loans the data are assembled as two separate 3D-panels, each consisting of i = 13 banks and j = 3 maturities.

5

Club convergence analysis

We begin our empirical analysis by reporting the results when the log(t) test is applied to the 3D-panels of deposit and lending interest rates. The 4

Consumption loans are given by the banks to customers in order to meet one time needs, while ordinary loans are given for business purposes. The aggregation of these two categories into a single one is a necessary step to have consistent time series of lending interest rates for each bank and maturity over the study period. Although the database has information on lending rates for construction firms, mortgages, small-scale enterprises, preferential customers, treasury operations, credit cards, and overdrafts, limited data availability across banks makes the inclusion of all these categories into our analysis problematic.

12

interest rate series are expressed in percentage terms, and after removing their underlying cyclical component using the HP filter with a smoothing parameter equal to 14400 (as recommended for monthly time series). In addition, we discard the first 25% of the sample period, that is the first 46 time observations for each series; qualitatively similar results are obtained when this trimming parameter is varied by taking on the values of 30%, as recommended by Phillips and Sul, 20% and 15%. Ordinary least squares (OLS) estimation over the 2002-17 study period of the log(t) regression for deposit rates yields ˆb = −0.061, with an associated HAC standard error of 0.181, and a t-statistic of -0.336. This indicates that the slope coefficient is statistically equal to zero, thereby supporting the view of convergence across all deposit interest rates. In the European Union, Rughoo and Sarantis (2014) uncovered support for the null hypothesis of convergence in deposits rates based on monthly data for the pre-GFC years 2003-7; for example, the slowest rate of convergence they report is ˆb = 0.836, with a t-statistic of 11.182. Our finding of a much smaller ˆb thus suggests that the speed of adjustment is much slower for Colombia. Post-GFC however, Rughoo and Sarantis find no evidence of deposits rates convergence, so the intra-country convergence that we find in Colombia can more latterly be compared with the lack of inter-country convergence across European Union countries. Rughoo and Sarantis (2014) point to a retrenchment in the European Union deposit market due to the crisis. In doing so, it is noted that banking activities were affected more by the GFC than other markets due to the nature of the crisis which dealt a serious blow to the confidence in banks. In contrast to this, Colombia was much less directly affected by the GFC in this way. To have an idea of the time path of all deposits rates, Figure 1 presents time plots of their range of variation, median rate, and usury interest rate. Although the usury rate applies only to loans, it is included here because it is indirectly linked to deposit rates through the margin on lending rates. 13

As can be seen, the range of variation of deposit interest rates has remained relatively constant over the sample period, except between 2005 and 2008 when it is noticeably wider.5 Turning to lending rates, the results unequivocally reject the null hypothesis of convergence among all series: more specifically, the slope coefficient is equal to -1.179, with an associated HAC standard error of 0.023, and a resulting t-statistic of -50.650. Developing further this latter result, we examine the possibility of smaller clubs of convergence for lending rates through the application of the Phillips and Sul clustering algorithm outlined before. Table 1 presents the initial classification of convergence clubs and the statistical tests of potential club merging. The left-hand side of the table shows that all the lending interest rates are initially classified into five convergence clubs and a non-convergent group. Tests of club merging were subsequently performed for convergence clubs [1] and [2], [2] and [3], [3] and [4], [4] and [5], and [5] and the non-convergent group [6], and in all of the cases the merging hypothesis was clearly rejected. The final classification for lending rates is therefore one in which there are 5 convergence clubs and a non-convergent group. Further examination of Table 1 reveals that the speed of adjustment across the 5 clubs is varied. For three out of the five clubs, ˆb is statistically different from zero (i.e. 0.125, 0.346 and 1.411 in clubs 1, 2 and 5). This points to faster speeds of adjustment compared to the single converge club that accommodates all Colombian deposits rates. For clubs 3 and 4 which collectively accommodate three quarters of the loan rates, the speed of adjustment is slow since ˆb is not significant. Table 2 summarises the composition of the convergence clubs. The table 5

Important changes in reserve requirements might be behind the widening range of variation of deposit rates. Indeed, as indicated by authors such as Vargas and Cardozo (2012) and Mora-Arbel´ aez et al. (2015), during the first half of the 2000s reserve requirements remain largely unaltered. Then, in the mid 2000s they were increased to control capital inflows and the amount of credit to the private sector. By the end of the 2000s, most of the changes in reserve requirements were reversed because of liquidity restrictions in international markets due to the GFC.

14

shows that the composition of the first two convergence clubs is somewhat mixed, in the sense that they include institutions of large, medium and small sizes along with interest rates on medium- to long-term maturities. Club 3, which comprises the largest number of interest rates (that is, 22), includes banks of almost all sizes and interest rate maturities. Clubs 4 and 5 include from mid- to small-size banks, and interest rates on short- to medium-term maturities. Lastly, regarding the interest rate of the non-convergent group, it is given by the interest rate charged by the smallest bank in the sample at medium-term maturity, which has moved very closely with the usury rate. Figure 2 shows the time plot of lending rates by convergence club, also including in each case the corresponding median rate, and usury rate. The plots show that, except for convergence Club 1, the absolute magnitude of the range of variation of lending interest rates has been decreasing within each convergence club as time passes. In sum, our results thus far are consistent with all deposit rates being part of an integrated market, but this is not the case with respect to lending rates. Indeed, this might point towards limited success, or even some degree of failure, attached to the financial reforms which have not delivered in terms of a competitive loans market.6 6

To assess the robustness of these findings to the choice of the filtering procedure, we experimented with the application of the Hamilton (2018) filter. This is where the cyclical component of a series is removed by running an OLS regression of the variable of interest against an intercept and several of its lagged values, and using the resulting fitted values as an estimate of the underlying trend component. Usage of the Hamilton filter provided results that are qualitatively the same in terms of deposit rates being more converged than loan rates. More specifically, the application of this filter points to the presence of two convergence clubs in deposit rates, and four convergence clubs (with two interest rates in a non-convergent group) in lending rates. Given that existing results for the EU in earlier work are based on employment of the HP filter, in what follows we shall report the results based on this filter in order to facilitate comparisons.

15

6

Drivers of club convergence

Our results so far indicate the presence of a single convergence club for deposit rates, but multiple convergence clubs for lending rates. This could be down to more competition in the market for deposits operating as an integrated market compared with the market for loans. With this in mind, we might note that Rughoo and Sarantis (2014) also find that convergence is more evident for the EU deposit market whereas the consumer credit and mortgage markets look to the more heterogeneous markets. In the case of deposit rates, the aggregation of broad classes of deposit rates into a single convergence club is perhaps understandable, as banks usually have uniform deposit rates across depositors. However, the aggregation of loan rates into a single convergence club is perhaps less expected insofar as loans are granted to borrowers with different risk characteristics. Moreover, one might expect a lower degree of integration in the loan market compared to the deposit market, as banks serve different clienteles. Given the presence of multiple convergence clubs in the case of Colombian loan rates, we now turn to examine the potential drivers of club formation which can throw light on understanding why the market for loans may not be acting as a fully integrated market. To this end, we begin by estimating a binary (probit) model to identify those variables that affect the probability that a pair of loan rates belongs to the same convergence club; the lending rate classified in the non-convergent group is therefore omitted from the analysis. Here, the main economic variable that we take into account is introduced with the intention of capturing the riskiness part of a bank’s portfolio, for which we consider two alternative measures; the first is given by the ratio between non-performing loans and total outstanding loans, denoted by npf(t) , and the second is the ratio between non-performing loans and total outstanding consumption loans, denoted by npf(c) .7 Using npf(t) and npf(c) in turn we then compute the corresponding 7

When we applied the club convergence algorithm, the individual interest rates were

16

absolute differential for each bank combination. All other things being equal, we might expect a negative sign on the riskiness variable if convergence club formation draws on banks of more similar degrees of riskiness. Additionally, we also consider as potential determinants dummy variables to denote any pair of lending rates that are of the same maturity, and also those rates that are set by the same bank. Finally, indicator variables are employed to allow for the potential effect of foreign and government ownership of banks on the setting of lending interest rates.8 Table 3 summarises the estimation results from the probit model for lending rates. Both the estimated coefficients on the absolute differential on npf(t) and npf(c) are negative and statistically different from zero, thereby supporting the view that the likelihood of any pair of lending rates belong to the same convergence club decreases for banks with dissimilar risk components.9 Similarly, the likelihood of any pair of lending rates being in the same convergence club falls when both banks are foreign-owned. Previous related research by Iregui and Otero (2013) indicates that the speed of adjustment of differentials in lending interest rates is slow when banks are foreign-owned, which may be viewed as further evidence of loan market fragmentation or, to put it another way, banks serving different market segments. The other sorted according to the value observed in the final period; thus, to construct the ratios npf(t) and npf(c) we use the balance sheet information of each bank as of August 2017. 8 When investigating the factors that can explain the probability that loan rates cluster into a convergence club, one might wonder whether there are potentially more variables than non-performing loans and ownership that could affect market integration. For example, one might think that the geographical coverage of banks (rural versus urban) or their mandate (commercial versus cooperative) are a determinant of different pricing policies. In a sense, geographical coverage is already being considered through the indicator variable for government ownership, since the operations of the only bank in the sample that is owned by the government (that is, Banco Agrario) are primarily focussed on the rural sector. As for the mandate of the banks, cooperative banks account for a very small share of the total assets of the banking sector, and for this reason they are not included among the thirteen banks that are used in our econometric analysis. 9 Although the results of the probit analysis are not intended to be causal, but rather correlational, exogeneity tests using the share of outstanding loans to total assets as an instrument for non-performing loans were supportive of our specification. These results are not reported here but are available upon request.

17

drivers that we consider are found to be insignificant. With three maturity categories spread across five convergence clubs, the estimated coefficient on same maturity dummy variable does not turn out to be statistically different from zero. We find that some banks straddle multiple convergence clubs. The same bank lending pair dummy is insignificant. Despite the possibility that loan pricing policy within a given bank might be similar or the same across products, our results suggest that those products may not necessarily be in the same convergence club. Finally, the estimated coefficient on the government ownership dummy is not significant either. There is significant change in the probability of being in the same convergence club if both interest series are from the same government-owned bank. With these latter considerations in mind, it is possible for us to explore the drivers or factors that determine the probability that any given lending rate belongs to a specific rather than the same convergence club. For this, a suitable econometric tool of analysis is the ordered probit model, where the dependent variable, say y, takes on five possible integer values {1, 2, 3, 4, 5}, associated to each and every one of the five convergence clubs determined by the Phillips and Sul algorithm. The right-hand-side variables we employ then aim to capture the potential role of maturity and riskiness. Since this is no longer a pairwise analysis, all variables are expressed in levels form rather than as absolute differentials. Table 4 reports the estimation results from the ordered probit model for lending rates. Two sets of results are presented, depending on whether one is using npf(t) (top panel) or npf(c) (bottom panel) as a riskiness regressor. The estimated coefficients are presented along with the resulting marginal effects. Using a 10% significance level, the results for the ordered probit indicate that the medium and long maturities (respectively denoted as mt and lt) are very much driving the membership of Clubs 2 and 3. Moreover, Club 3 membership appears to be driven by the medium maturities, while the Club 4 result indicates a significant exit as maturities go up. Comparing the coefficients, 18

it seems that a given rise in medium (long) maturities will most likely lead to a shift of interest rates out of Club 4 into Club 3 (2). Maturity duration is therefore important in driving the Colombian convergence results. For a theoretical perspective on this result, we may draw on the term structure of interest rates in terms of the expectations and liquidity preference theories. The presence of multiple convergence clubs implies that loan rates are not operating as a single market, and maturity is found to play a significant role in driving the creation of convergence clubs. This evidence suggests that the Expectations Theory of the Term Structure does not hold. Indeed, the role played by the longer maturities can be explained by the liquidity preference theory insofar as longer term interest rates not only reflect market expectations, but also include a risk or liquidity premium to factor in the higher level of risk for the lending bank concerned. The results reported in Table 4 also suggest that with loans as the riskier part of a bank’s portfolio, an increase in the relative importance of loans is associated with more risk exposure and so higher interest rates which impact on convergence club membership. We find that loan rates set by banks with the riskier balance sheets, means moving towards a convergence club characterised by higher interest rates. This is highlighted by a movement out of Club 4 and into Club 2. In terms of the other drivers that we consider, the estimated coefficient on government ownership is not estimated with a great deal of precision because there is only one publicly-owned institution in the sample of 13 banks. Likewise, foreign ownership does not appear to affect the probability of belonging to a specific convergence club.

7

Club convergence and monetary policy

The earlier cointegration-based study of Holmes et al. (2015) points to changes in Colombian monetary policy being associated with asymmetries where banks are subsequently more partial to raising rather than reducing lend19

ing rates. Given that other studies have also found evidence that monetary policy changes affects interest rate setting behaviour in retail banking markets, it is also interesting to consider whether the number of (or membership of) lending rate convergence clubs changes under monetary tightening or loosening. Following changes in monetary policy, a smaller or larger number of convergence clubs can arguably reflect a greater or lesser degree of market integration and so predictability in assessing the impact on lending rates. To examine the effect of monetary policy changes over time, we implement the Phillips and Sul algorithm in a recursive way, so that we begin by performing the estimation for the sub-sample May 2002 to January 2007; then May 2002 to February 2007; and so on until reaching the last observation available, that is over the full sample period May 2002 to August 2017. The choice of the first sub-sample is intended to shed some light on club convergence just prior to the start of the GFC. Figure 3 summarises the results of the recursive estimation. The pictures that emerge from the two plots are markedly different. Indeed, after an initial period of (about) three years of relative stability in the number of convergence clubs, in the case of deposit rates the number of clubs has been decreasing over time, while for lending rates it has been increasing. This finding can be compared with Rughoo and Sarantis (2014) on European retail banking, for instance, who find that deposit rates are more convergent than lending rates. They find that the GFC stalled the integration process. As mentioned above, the GFC is not regarded as having been as big an issue for Colombia as compared to other economies. This is reflected in Figure 3 where there is no marked change in the number of convergence clubs until several years after the GFC. Indeed, 2013-4 and onward is marked by a noticeable decrease (increase) in the heterogeneity of deposit (lending) rates in terms of a decreased (an increased) number of convergence clubs. This finding is consistent with a consequence of the Colombian financial reforms which led to a system in which banks compete for the same “pool” of deposits and so offer similar deposit rates. 20

But when it comes to lending, banks tend to specialise in specific segments of the market (for example, short- versus long-term credit, or industry, agriculture, microcredit, etc.) which explains the increased fragmentation in loan rates. Furthermore, a key event that has been taking place since 2010 is the entrance of new intermediaries, from 15 to currently 25. Some of these new banks are financial institutions that became banks, while some of the other new banks are foreign. This added participation in Colombian banking towards the end of the study period served to reinforce increased lending rate heterogeneity. As for the relationship between the number of convergence clubs and the interbank rate (ibt ), Table 5 shows that an increase (decrease) in the bank rate leads to a fall (increase) in the number of deposits convergence clubs.10 In the case of lending rates, the opposite applies. Moreover, a tight monetary policy appears to be associated with increased bank heterogeneity (more convergence clubs) in loan rates. In terms of theory, the collusive pricing hypothesis points to a downward rigidity of lending rates that can be attributed to the reluctance of banks to decrease lending rates in fear of disrupting collusive arrangements and/or the hesitation by consumers to change lenders due to switching costs. Instead, our results give support to the consumer behaviour hypothesis in that the reaction from customers to lending rate increases and/or the adverse selection problem faced by lenders in an increasing interest rate environment may translate into a degree of upward rigidity in lending rates. This perhaps makes the predictions of monetary effectiveness more problematic in contractionary monetary regimes. These results can be compared with the past study by Holmes et al. (2015) which focuses on the response in interest rate levels across banks rather than interest rate club formation to monetary policy changes. In 10 The estimation period for Table 5 (and Table 6 below) starts from 2010m1. The period 2007m1-2009m12 is not included because the dependent variable exhibits little variation over this time, and so several parameter estimates were not found to be statistically significant.

21

that study, lending rates were found to be more rapidly adjusted downwards than upwards in response to monetary policy changes. This earlier empirical finding adds support to the view that lending rates are relatively more rigid in terms of upward than downward movements. As with our new findings regarding club formation, there is earlier support for the consumer behaviour hypothesis when it comes to loan rate setting. Further insight into this issue can be obtained from taking the first difference in the interbank rate, and then splitting this into its positive (∆ib+ t ) and negative components (∆ib− t ); see Table 6. With our focus on the impact of policy changes on club formation, we find key differences based on monetary expansions versus contractions. When we introduce these asymmetric terms, both positive and negative components lose significance in the case of deposit rates. However, there is a noticeable difference in the case of lending rates. Indeed, we can identify an asymmetry here insofar it is a policy rate increase rather than decrease that has a significant effect on loan rate club formation. In the case of a monetary tightening where policy rates are increased, the number of loan rate convergence clubs is likely to increase. This suggests that the extent of loan market integration will decrease. A corresponding decrease in the number of convergence clubs or increase in loan market integration is likely to be absent in the case where policy rates are decreased. In addition to these insights into retail bank behaviour, there is a further implication with respect to the term structure of interest rates. The Expectations Hypothesis of the Term Structure is more likely to hold where there exists a single convergence club for loan rates. According to the results reported in Table 5, this is most likely to happen during regimes of monetary easing and low interest rates. This is in contrast to the divergence experienced by GFC-affected countries insofar as policy rates being at or near the zero lower bound were not matched by similar reductions in longer rates. Lastly, an interesting question that arises here is whether increased banking sector competition for deposits has been beneficial for the median in22

vestor. To answer this, Figure 4 plots the evolution through time of the interbank rate, the range of variation of deposit rates, and the median deposit rate.11 As can be seen, both the interbank and median rates are very close to each other, and so it appears that banking sector competition for deposits is leading to a case where deposit rates are converging to a single level, and this level turns out to be close to the market (or policy) rate. The issue here is that this level is perhaps too low to be attractive for the median investor. In fact, for a given average annual inflation rate of 4.6% over the sample period, there are several instances of negative real interest rates for deposits.

8

Concluding remarks

The application of the Phillips and Sul convergence club methodology to a bank-level data set provides the opportunity to assess the extent of retail banking market integration in an emerging economy. In doing so, we further develop the Phillips and Sul approach by conducting a formal analysis of the drivers behind convergence club formation. Against a background of financial liberalisation of the Colombian banking sector, and the fact that the top five banks account for a higher market share in 2016 (68%) compared to 2001 (43%), one would have initially expected higher integration with more deposit and loan rates converging in the retail banking industry. However, this investigation shows that one obtains very different stories regarding deposits and loans. Thus, even if concentration is an issue for deposits convergence, matters are much different for loans. Indeed, competition is very strong in 11

The interbank rate is the price of the operations in domestic currency that are undertaken between financial intermediaries to solve liquidity problems overnight; although the maturity of the rate typically refers to one day, it can vary during weekends or when there are holidays during weekdays. Loans between intermediaries do not require collateral, and so the interbank rate is considered to reflect the credit risk associated to the parts involved in the operations. Additionally, the level of the interbank rate reflects liquidity conditions in the money market, and so it can be regarded as a suitable indicator of the stance of monetary policy in Colombia; see, e.g. the discussion in www.banrep.gov.co/es/tib.

23

the market for deposits where interest rates are driven by a single convergence club. This is not the case for interest rates on loans which are driven by multiple convergence clubs. The factors that drive the lack of market integration for loans include variations in bank riskiness and loan maturities. The study offers further insights through recursive convergence club estimation which indicates the presence of asymmetric effects from changes in monetary policy on loan rate convergence club formation. Moreover, tighter (looser) monetary regimes are likely to be met with more (less) convergence clubs and so less (more) competition in the case of interest rates on loans. A number of modelling and policy implications arise from this study when looking at the findings concerning loan rates. First, it seems that bank lending behaviour is more consistent with the consumer hypothesis rather than the collusion hypothesis. Second, despite the financial reforms that have taken place, evidence in support of the law of one price in the Colombian loans market is lacking. There does not appear to be an integrated loan market and the role played by maturity in convergence club creation suggests that the expectations hypothesis of the term structure is not applicable in the case of Colombia. Instead, the evidence appears to be more supportive for hypotheses of liquidity preference. In terms of policy, two main implications might be noted. Although there is an upper bound present on loan rates, it seems unlikely to have been binding. The analysis suggests that there is still some distance from achieving an integrated market for loans, and so this should be at the centre of future reforms of the banking sector. The ongoing focus needs to be in terms of additional measures aimed at enhancing competition in loans. At the practical level, these measures could include efforts to make information on lending rates more readily available to the public in general and not only to specialised audiences; e.g., on weekdays El Tiempo and La Rep´ ublica, two of the most influential newspapers in the country, publish information (albeit aggregate) on deposit but not lending rates. Such measure could well be 24

complemented through educational campaigns that remind individuals about the benefits of searching across banks and comparing their rates. Perhaps part of the reason for the limited competition in loans has to do with banks viewing their customers as having restricted financing options. Thus, at the institutional level there may still be room for removing further regulatory restrictions imposed on the entrance of new banks. A second policy implication concerns monetary policy. The findings here suggest that there is a relative difficulty in predicting the effectiveness of a monetary contraction (being accompanied by an increase in loan rates convergence clubs) as opposed to monetary easing. From the point of view of monetary policymakers, there is the need to play closer attention to the riskiness and asset maturity profile of particular banks and how these factor interplay with loan rate setting.

25

References Antonakakis, N., C. Christou, J. Cunado, and R. Gupta (2017).

Con-

vergence patterns in sovereign bond yield spreads: Evidence from the Euro Area. Journal of International Financial Markets, Institutions and Money 49 (C), 129–139. Arango, M. (2006). Evoluci´on y crisis del sistema financiero colombiano. Serie Estudios y Perspectivas 11. Oficina de la CEPAL en Bogot´a. Arghyrou, M. G., A. Gregoriou, and A. Kontonikas (2009). Do real interest rates converge? evidence from the European union. Journal of International Financial Markets, Institutions & Money 19 (3), 447–460. Baharumshah, A. Z., S.-V. Soon, and D. Borˇsiˇc (2013). Real interest parity in Central and Eastern European countries: Evidence on integration into EU and the US markets. Journal of International Financial Markets, Institutions and Money 25 (C), 163–180. Barro, R. J. (1991). Economic growth in a cross section of countries. Quarterly Journal of Economics 106 (2), 407–443. Barro, R. J. and X. Sala-i-Martin (1992). Convergence. Journal of Political Economy 100 (2), 223–251. Baumol, W. J. (1986). Productivity, growth, convergence, and welfare: what the long run data show. American Economic Review 76 (5), 1072–1085. Bernard, A. B. and S. N. Durlauf (1995). Convergence in international output. Journal of Applied Econometrics 10 (2), 97–108. Brada, J. C., A. M. Kutan, and S. Zhou (2005). Real and monetary convergence between the European Union’s core and recent member countries: A rolling cointegration approach. Journal of Banking and Finance 29 (1), 249–270. 26

Chatterji, M. and J. H. L. Dewhurst (1996). Convergence clubs and relative economic performance in Great Britain: 1977-1991. Regional Studies 30 (1), 31–40. Degl’Innocenti, M., R. Matousek, and N. G. Tzeremes (2018). Financial centres’ competitiveness and economic convergence: Evidence from the European Union regions. Environment and Planning A: Economy and Space 50 (1), 133–156. Du, K. (2017). Econometric convergence test and club clustering using Stata. The Stata Journal 17 (4), 882–900. Evans, P. (1996). Using cross-country variances to evaluate growth theories. Journal of Economic Dynamics and Control 20 (6-7), 1027–1049. Evans, P. and G. Karras (1996). Convergence revisited. Journal of Monetary Economics 37 (2-3), 249–265. Fung, M. K. (2009). Financial development and economic growth: Convergence or divergence? Journal of International Money and Finance 28 (1), 56–67. Hamilton, J. D. (2018). Why you should never use the Hodrick-Prescott filter. The Review of Economics and Statistics 100 (5), 831–843. Herwartz, H. and J. Roestel (2011). Convergence of real capital market interest rates - evidence from inflation indexed bonds. Journal of Money, Credit and Banking 43 (7), 1523–1541. Hodrick, R. J. and E. C. Prescott (1997). Postwar U.S. business cycles: An empirical investigation. Journal of Money, Credit and Banking 29 (1), 1–16.

27

Holmes, M. J., A. M. Iregui, and J. Otero (2015). Interest rate pass through and asymmetries in retail deposit and lending rates: An analysis using data from Colombian banks. Economic Modelling 49 (C), 270–277. Iregui, A. M., C. Milas, and J. Otero (2002). On the dynamics of lending and deposit interest rates in emerging markets: A non-linear approach. Studies in Nonlinear Dynamics and Econometrics 6, Article 4. Iregui, A. M. and J. Otero (2013). Testing the law of one price in retail banking: An analysis for Colombia using a pair-wise approach. Economics Letters 118 (1), 29–32. Kaul, A. and V. Mehrotra (2007). The role of trades in price convergence: A study of dual-listed Canadian stocks. Journal of Empirical Finance 14 (2), 196–219. Kim, S. J., F. Moshirian, and E. Wu (2005). Dynamic stock market integration driven by the European Monetary Union: An empirical analysis. Journal of Banking and Finance 29 (10), 2475–2502. Kim, S.-J., F. Moshirian, and E. Wu (2006). Evolution of international stock and bond market integration: Influence of the European Monetary Union. Journal of Banking and Finance 30 (5), 1507–1534. Mart´ın-Oliver, A., V. Salas-Fum´as, and J. Saurina (2007). A test of the law of one price in retail banking. Journal of Money, Credit and Banking 39 (8), 2021–2040. Matousek, R., A. Rughoo, N. Sarantis, and A. G. Assaf (2015). Bank performance and convergence during the financial crisis: Evidence from the ‘old’ European union and Eurozone. Journal of Banking and Finance 52 (C), 208–216.

28

Mora-Arbel´aez, T. A., A. J. Garc´ıa-Bernal, J. E. G´omez-Gonz´alez, and M. Villamizar-Villegas (2015). Una historia exhaustiva de la regulaci´on financiera en Colombia. Technical Report Borradores de Investigaci´on 887, Banco de la Rep´ ublica. Mylonidis, N. and C. Kollias (2010). Dynamic European stock market convergence: Evidence from rolling cointegration analysis in the first euro-decade. Journal of Banking and Finance 34 (9), 2056–2064. Narayan, P. K., S. Mishra, and S. Narayan (2011). Do market capitalization and stocks traded converge? new global evidence. Journal of Banking and Finance 35 (10), 2771–2781. Newey, W. K. and K. D. West (1987). A simple, positive semi-definite, heteroskedastic and autocorrelation consistent covariance matrix. Econometrica 55 (3), 703–708. Phillips, P. C. B. and D. Sul (2007). Transition modeling and econometric convergence tests. Econometrica 75 (6), 1771–1855. Phillips, P. C. B. and D. Sul (2009). Economic transition and growth. Journal of Applied Econometrics 24 (7), 1153–1185. Rughoo, A. and N. Sarantis (2012). Integration in European retail banking: Evidence from savings and lending rates to non-financial corporations. Journal of International Financial Markets, Institutions and Money 22 (5), 1307–1327. Rughoo, A. and N. Sarantis (2014). The global financial crisis and integration in European retail banking. Journal of Banking and Finance 40 (C), 28–41. Schnurbus, J., H. Haupt, and V. Meier (2017). Economic transition and growth: A replication. Journal of Applied Econometrics 32 (5), 1039–1042.

29

Uribe, J. D. and H. Vargas (2003). Financial reform, crisis and consolidation in Colombia. In P. L. V. der Haegen and J. Vi˜ nals (Eds.), Regional Integration in Europe and Latin America. Monetary and Financial Aspects, pp. 419–446. Aldershot: Ashgate. Vargas, H. and P. Cardozo (2012). El uso de encajes en un marco de pol´ıtica monetaria o´ptima. Technical Report Borradores de Investigaci´on 716, Banco de la Rep´ ublica. Yu, I.-W., K.-P. Fung, and C.-S. Tam (2010). Assessing financial market integration in Asia - equity markets. Journal of Banking and Finance 34 (12), 2874–2885.

30

Table 1: Classification of convergence clubs for lending rates Initial classification Club [1] [2] [3] [4] [5] Group [6]

Tests of club merging

β coeff.

t-stat.

p-value

Club

0.125 0.346 0.027 -0.111 1.411

3.217 2.581 0.736 -1.239 2.038

(0.999) (0.995) (0.769) (0.108) (0.979)

[1+2] [2+3] [3+4] [4+5] [5] + Group [6]

β coeff.

t-stat.

p-value

-0.115 -0.375 -0.274 -0.501 -2.438

-6.819 -7.091 -7.322 -5.628 -167.157

(0.000) (0.000) (0.000) (0.000) (0.000)

Note: The numbers in brackets denote the number of convergence clubs. t-statistics are based on Newey and West (1987) heteroskedasticity and autocorrelation (HAC) standard errors. Lower tail probability values in parentheses. Group [6] is a non-convergent club consisting of one lending interest rate (see next table).

31

Table 2: Composition of convergence clubs for lending rates Club

Bank

Maturity

[1] [1] [1]

2 6 13

mt: 366 to 1095 days lt: 1096 to more days lt: 1096 to more days

[2] [2] [2] [2] [2]

1 8 8 11 11

lt: 1096 to more days mt: 366 to 1095 days lt: 1096 to more days mt: 366 to 1095 days lt: 1096 to more days

[3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3] [3]

1 1 2 2 3 3 3 4 4 4 5 5 6 6 7 7 9 9 10 11 12 12

st: 1 to 365 days mt: 366 to 1095 days st: 1 to 365 days lt: 1096 to more days st: 1 to 365 days mt: 366 to 1095 days lt: 1096 to more days st: 1 to 365 days mt: 366 to 1095 days lt: 1096 to more days mt: 366 to 1095 days lt: 1096 to more days st: 1 to 365 days mt: 366 to 1095 days st: 1 to 365 days lt: 1096 to more days mt: 366 to 1095 days lt: 1096 to more days lt: 1096 to more days st: 1 to 365 days mt: 366 to 1095 days lt: 1096 to more days

Club

Bank

Maturity

[4] [4] [4] [4] [4] [4]

5 7 8 9 10 12

st: 1 to 365 days mt: 366 to 1095 days st: 1 to 365 days st: 1 to 365 days mt: 366 to 1095 days st: 1 to 365 days

[5] [5]

10 13

st: 1 to 365 days st: 1 to 365 days

Group [6]

13

mt: 366 to 1095 days

Note: Banks are identified according to the size of their assets as of June 2017, from the largest (Bank 1) to the smallest (Bank 13). Group [6] is a non-convergent club consisting of one lending interest rate. Short-, medium- and long-term maturity are respectively denoted as st, mt and lt.

32

Table 3: Estimation results from probit model for lending rates Model 1 Regressors Abs. Dif. npf(t)

β coeff. −8.928‡ (4.129)

Model 2

Marginal effects

β coeff.

Marginal effects

−3.304‡ (1.512)

Abs. Dif. npf(c)

−6.966† (3.885)

−2.584† (1.431)

Same maturity

0.025 (0.106)

0.009 (0.039)

0.025 (0.106)

0.009 (0.039)

Same bank

0.118 (0.243)

0.045 (0.093)

0.135 (0.244)

0.051 (0.093)

Both foreign

−0.515‡ (0.158)

−0.176‡ (0.048)

−0.504‡ (0.158)

−0.173‡ (0.048)

Both government

−0.400 (0.783)

−0.136 (0.236)

−0.396 (0.783)

−0.135 (0.237)

Obs. Pseudo-R2 χ2 p-value

703 0.019 16.434 [0.006]

0.016 14.201 [0.014]

Notes: The dependent variable takes the value of one when a pair of lending interest rates belong to the same convergence club. Both models include intercept (not reported). Standard errors (in parentheses) are heteroskedasticity consistent. Rejection probabilities in the coefficients are denoted † p < 0.10 and ‡ p < 0.05.

33

Table 4: Estimation results from ordered probit model for lending rates β coeff.

Marginal effects ∂ Pr(y=2) ∂x

∂ Pr(y=1) ∂x

Regressors

∂ Pr(y=3) ∂x

∂ Pr(y=4) ∂x

∂ Pr(y=5) ∂x

mt

−1.713‡ (0.546)

0.024 (0.034)

0.096 (0.065)

0.344‡ (0.157)

−0.393‡ (0.135)

−0.072 (0.058)

lt

−2.535‡ (0.537)

0.126 (0.104)

0.239‡ (0.095)

0.139 (0.187)

−0.431‡ (0.138)

−0.073 (0.058)

Foreign ownership

0.646 (0.457)

−0.017 (0.022)

−0.062 (0.042)

−0.033 (0.075)

0.106 (0.099)

0.006 (0.009)

Government ownership

1.343 (0.857)

−0.015 (0.020)

−0.069† (0.040)

−0.290 (0.314)

0.323 (0.244)

0.050 (0.091)

−67.108‡ (18.454)

2.029 (2.180)

7.118‡ (3.412)

1.141 (5.734)

−9.845‡ (4.228)

−0.442 (0.510)

npf(t)

Obs. Pseudo-R2 χ2 p-value

38 0.307 29.105 [0.000]

mt

−1.857‡ (0.540)

0.021 (0.028)

0.101 (0.069)

0.350‡ (0.158)

−0.392‡ (0.134)

−0.080 (0.077)

lt

−2.639‡ (0.561)

0.106 (0.079)

0.246‡ (0.101)

0.148 (0.182)

−0.420‡ (0.139)

−0.080 (0.078)

Foreign ownership

0.529 (0.447)

−0.011 (0.014)

−0.049 (0.040)

−0.019 (0.058)

0.075 (0.083)

0.004 (0.008)

Government ownership

0.137 (0.761)

−0.003 (0.014)

−0.013 (0.066)

−0.004 (0.042)

0.019 (0.114)

0.001 (0.007)

−55.526‡ (12.634)

1.254 (1.329)

5.650† (2.966)

0.525 (4.449)

−7.091‡ (3.268)

−0.339 (0.474)

npf(c)

Obs. Pseudo-R2 χ2 p-value

38 0.321 36.249 [0.000]

Notes: The estimates of the threshold parameters are not reported. Standard errors (in parentheses) are heteroskedasticity consistent. Rejection probabilities in the coefficients are denoted † p < 0.10 and ‡ p < 0.05.

34

Table 5: Convergence clubs and interbank rate Number of convergence clubs in: Regressors:

Deposits rates

Lending rates

−0.643‡ (0.068)

0.404‡ (0.089)

Constant

6.084‡ (0.397)

1.632‡ (0.375)

R2 F stat. p-value

0.373 89.775 [0.000]

0.218 20.709 [0.000]

ib

Notes: The estimation period runs from 2010m1 to 2017m8, for a total of 92 observations. Standard errors (in parentheses) are heteroskedasticity consistent. Rejection probabilities in the coefficients are denoted † p < 0.10 and ‡ p < 0.05.

35

Table 6: Convergence clubs and variations in the interbank rate Number of convergence clubs in: Regressors:

Deposits rates

Lending rates

∆ib+

−0.931 (0.844)

2.166‡ (0.849)

∆ib−

1.082 (1.055)

−0.746 (0.807)

Constant

3.245‡ (0.211)

3.279‡ (0.140)

R2 F stat. p-value

0.015 0.885 [0.416]

0.066 3.417 [0.037]

Notes: The estimation period runs from 2010m1 to 2017m8, for a total of 92 observations. Standard errors (in parentheses) are heteroskedasticity consistent. Rejection probabilities in the coefficients are denoted † p < 0.10 and ‡ p < 0.05.

36

0

5

10

Percentage 15 20

25

30

35

Figure 1: Time plot of deposit rates (one convergence club)

2001

2003

2005

2007

2009

Range of variation

2011

2013

Median rate

37

2015

2017

Usury rate

35 30 25 Percentage 15 20 10 5 0

0

5

10

Percentage 15 20

25

30

35

Figure 2: Time plot of lending rates by convergence club

2001

2003

2005

2007

2009

Range of variation

2011

2013

Median rate

2015

2017

2001

Usury rate

2003

2005

2007

2009

Range of variation

2013

2015

2017

Usury rate

30 25 Percentage 15 20 10 5 0

0

5

10

Percentage 15 20

25

30

35

(b) Convergence club 2

35

(a) Convergence club 1

2011

Median rate

2001

2003

2005

2007

2009

Range of variation

2011

2013

Median rate

2015

2017

2001

Usury rate

2003

2005

2007

2009

Range of variation

2013

2015

2017

Usury rate

30 25 Percentage 15 20 10 5 0

0

5

10

Percentage 15 20

25

30

35

(d) Convergence club 4

35

(c) Convergence club 3

2011

Median rate

2001

2003

2005

2007

Range of variation

2009

2011

2013

Median rate

2015

2017

2001

Usury rate

2003

2005

2007

2009

Non−convergent rate

(e) Convergence club 5

2011

2013

(f) Non-convergent group

38

2015

Usury rate

2017

0

1

2

3

4

5

6

7

Figure 3: Number of convergence clubs over time

2007

2008

2009

2010

2011

2012

2013

2014

2015

2016

2017

2014

2015

2016

2017

0

1

2

3

4

5

6

7

(a) Deposit rates

2007

2008

2009

2010

2011

2012

2013

(b) Lending rates

39

0

5

Percentage 10

15

20

Figure 4: Interbank rate and range of variation of deposit rates

2001

2003

2005

2007

2009

Range of variation

2011

Median rate

40

2013

2015

2017

Interbank rate