Effectiveness of monetary and macroprudential shocks on consumer credit growth and volatility in Turkey

Chadwick, Meltem Gulenay Article Effectiveness of monetary and macroprudential shocks on consumer credit growth and volatility in Turkey Central Bank Review (CBR) Provided in Cooperation with: Central Bank of The Republic of Turkey, Ankara Suggested Citation: Chadwick, Meltem Gulenay (2018) : Effectiveness of monetary and macroprudential shocks on consumer credit growth and volatility in Turkey, Central Bank Review (CBR), ISSN 1303-0701, Elsevier, Amsterdam, Vol. 18, Iss. 2, pp. 69-83, https://doi.org/10.1016/j.cbrev.2018.03.001 This Version is available at: https://hdl.handle.net/10419/217318 Standard-Nutzungsbedingungen: Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden. Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen. Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in EconStor may be saved and copied for your personal and scholarly purposes. You are not to copy documents for public or commercial purposes, to exhibit the documents publicly, to make them publicly available on the internet, or to distribute or otherwise use the documents in public. If the documents have been made available under an Open Content Licence (especially Creative Commons Licences), you may exercise further usage rights as specified in the indicated licence. https://creativecommons.org/licenses/by-nc-nd/4.0/ Effectiveness of monetary and macroprudential shocks on consumer credit growth and volatility in Turkey Meltem Gulenay Chadwick Central Bank of the Republic of Turkey, Istiklal Cad. No:10, Ulus, Ankara, 06050, Turkey article info Article history: Received 12 February 2018 Received in revised form 22 March 2018 Accepted 22 March 2018 Available online 30 March 2018 JEL classification: C54 E44 E52 Keywords: Consumer loans Monetary policy Macroprudential policy Stochastic volatility models Credit growth volatility IV probit model Panel VAR model abstract This paper proposes a panel VAR model to uncover the effect of monetary policy and macroprudential tightening probability on general purpose loans, housing loans, vehicle loans, credit cards and their respective volatilities in Turkey. To conduct our analysis, first, we compare a number of stochastic volatility models using our loan and credit card series in a formal Bayesian model comparison exercise, in order to determine the best volatility model for our series. Second we disclose the latent probability of macroprudential tightening from the binary information of policy episodes, using an instrumental variable probit model estimated by conditional maximum likelihood with heteroscedasticity robust standard errors. Lastly we estimate the dynamic impact of monetary policy and macroprudential measures using a panel VAR, incorporating the latent probability of tightening episodes, credit growth, industrial production growth, loan rates, inflation and credit growth volatilities into the endogenous system of equations. We conclude that macroprudential tightening is effective in dampening credit growth, credit growth volatility and reducing consumer price inflation. Besides, this effect is more prominent when macroprudential tools are administered in coordination with monetary policy. ©2018 Central Bank of The Republic of Turkey. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/). 1. Introduction After the global financial crisis macroprudential tools are seen as useful policies to reduce financial imbalances. The global crisis reminded us that financial stability has a macroprudential or systemic dimension that should not be ignored. 1 In addition to that, global crisis helps us understand the importance of global build-up of systemic-risk and financial imbalances whose sudden unfurling turned out to have severe global macroeconomic consequences. Recent global crisis also highlighted the need to go beyond microprudential approach to macro based financial regulation and supervision. The policy stance is concentrating notably on the usage, implementation and effectiveness of macroprudential tools as well as their impact on macroeconomic outcomes and their relationship with monetary policy. The implementation of macroprudential policies for financial stability raises a number of challenges. One important challenge is that little is known about their effects as it is difficult to quantify the effectiveness of these measures, especially when macroprudential actions involve multitude of instruments. These instruments are taken at infrequent intervals and they are in use for a very short time span only making traditional regression analysis difficult. Accordingly, in the wake of the financial crisis, macroprudential policy has attracted considerable attention among researchers and policy makers and the literature on the usage, implementation and the effectiveness of macroprudential policies now is growing very fast. Turkey faced rapid credit growth after the 2001 crisis, which was a local crisis, with recovering economic fundamentals afterwards. Institutions respond to this crisis with several structural reforms agenda that has fiscal, monetary and prudential dimensions. Rapid credit growth after 2001 crisis is accompanied with tight regulations and supervision within the banking system. In this respect, most of the prudential policies in Turkey are enforced through the banking system till 2011. For example, banks E-mail address: [email protected].tr. Peer review under responsibility of the Central Bank of the Republic of Turkey. 1 See Galati and Moessner (2017),Cerutti et al. (2017) and Cakir (2017) for a conceptual framework of macroprudential design. Contents lists available at ScienceDirect Central Bank Review journal homepage: http://www.journals.elsevier.com/central-bank-review/ https://doi.org/10.1016/j.cbrev.2018.03.001 1303-0701/©2018 Central Bank of The Republic of Turkey. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http:// creativecommons.org/licenses/by-nc-nd/4.0/). Central Bank Review 18 (2018) 69e83 were not allowed to have currency mismatches, foreign currency loans to consumers were prohibited, and there were restrictions on foreign currency lending to non-financial firms. Tight restrictions were introduced on distributing bank dividends, new bank entry and branch openings. During this regulatory period the Banking Regulation and Supervision Agency (BRSA) in Turkey enforced significantly higher minimum capital adequacy and liquidity coverage ratios than required by international standards. Against the high volatility in capital flows during the post-global crisis period, which materialized with the quantitative easing policies of advanced economies, Turkey has taken more steps towards implementing explicit macroprudential policies after 2011. Accordingly, the Central Bank of the Republic of Turkey (CBRT) reshaped the inflation targeting framework by incorporating financial stability as an additional objective. 2 With this paper we aim to offer a mixture of methodologies to correctly measure the effects of monetary and macroprudential policies in Turkey especially on consumer credit market, i.e. credit growth and credit growth volatility. In this framework, first we start with estimating the volatility for credit data employing various stochastic volatility models. Second, we use binary macroprudential policy indicators, acknowledging their endogenous nature, i.e. a macroprudential tightening is an endogenous response to a previous heating in credit markets and we employ instrumental variable probit model to uncover the latent propensity to macroprudential tightening from the observed binary policy data. We estimate the instrumental variable probit model by conditional maximum likelihood with heteroscedasticity robust standard errors. Lastly, we apply panel VAR of Love and Zicchino (2006) to uncover the effect of monetary and macroprudential polices effectively used in Turkey. The literature on the effectiveness of macroprudential policy tools is still in its infancy. In recent years, however, increasing efforts have been made to fill this gap. This paper complements other studies on the effectiveness of macroprudential policies. Different from the existing literature on the effects of macroprudential tools, our main contribution is analyzing this effect by exploiting the endogenous nature of these tools and analyzing the transmission from a possible shock to these polices to credit growth and credit growth volatility. This is the first study in the literature that discusses the importance of the effect of such policies on the second moment of credits. The empirical literature on macroprudential policies has broadly followed two approaches in assessing the effects of macroprudential tools: reduced-form regression analysis conducted using cross country panel regressions and reduced-form regression analysis based on microdata. 3 Cross-country panel data studies can do a relatively good job of controlling for global and local factors, by including a host of global variables in the regressions as well as fixed effects to capture unobserved heterogeneity. Such control variables often include global variables, such as the VIX, and macroeconomic variables to control for local factors. The literature also makes use of information on various policy actions as an independent variable to explain asset price movements and credit growth in a time-series or dynamic panel regression framework. 4 Each method, i.e. reduced form regression analysis using crosscountry panel regression or microdata has both advantages and disadvantages. However, it is important to address the main problem, that is the endogenous nature of the macroprudential policies. A key issue in both the academic literature and the policy debate is how the macroprudential policy interacts with monetary policy, i.e. should monetary policy be regarded as a complement or even a substitute for macroprudential policy for restraining a potential credit boom? Both the theoretical literature and the empirical literature gives different answers to this question. Yet, most of the papers offer an optimal cooperation between two sets of instruments. Collard et al. (2017) and Svensson (2017) claim that the optimal monetary policy alone is not efficient enough for financial instability and illustrates the optimal conditions for the complementarity conditions of monetary and macroprudential policies together to serve as the first line of defense against financial instability. Brunnermeier and Sannikov (2016) show that welfare is significantly improved by a combination of macroprudential policy and monetary policy. Within this framework our results highlight that - using information for consumer loans over the period of 2006e2017 monetary policy and macroprudential policy are complements and the existence of macroprudential policies besides monetary policy increases the effectiveness of these tools. We manifest these results for the consumer credit market, i.e. credit cards, general purpose loans, housing loans and vehicle loans. The results should be interpreted with the following interpretations in mind. First limitation is related to the tightening periods related to the macroprudential policies. The policy measure used for the estimations reflect the direction of the policy action, but not the strength of the action. When we estimate the macroprudential tightening probability, we use a binary variable for the tightening periods and for certain periods more than one policy action is taken or for some other periods the intensity of the action taken is higher than the other periods in question, which might cause some measurement errors. Measurement error related to the intensity of the macroprudential policy actions are commonly mentioned in the literature, therefore one must be wary of the binding effect of this measurement error which is likely to weaken the estimated effect of macroprudential policies. 5 A second limitation is related to the difficulty in completely encountering the potential endogeneity of macroprudential policies. To alleviate such concerns we use a probit model with instrumental variables to extract the macroprudential tightening probability and a GMM estimation within panel VAR, which employs additional dynamic instruments in the empirical framework. Our paper is related to a large body of literature on the effectiveness of macroprudential tools, but there are few papers that are particularly related to this study. Tillmann (2015) proposes a VAR augmented by qualitative variables (Qual VAR) to estimate the effects macroprudential tightening on the housing market of Korea and conclude that macroprudential tightening is effective in dampening credit growth and reducing the appreciation of house prices. Tovar Mora et al. (2012) examine the role of reserve requirements and other macroprudential instruments with crosscountry evidence on how they influence real private bank credit growth and their results show that these instruments have a moderate and transitory effect and play a complementary role to monetary policy. Greenwood-Nimmo and Tarassow (2016) examine the implications of monetary shocks and macroprudential shocks for aggregate financial fragility using a sign restricted VAR with US data and they suggest that combined 2 Kara (2016a) and Kara (2016b) for detailed information about Turkey's experience with macroprudential polices. 3 See Aiyar et al. (2014),Akinci and Olmstead-Rumsey (2018),Erdem et al. (2017), Fendo glu (2017),Bruno et al. (2017),Jim enez et al. (2017),Dell’Ariccia et al. (2017) and Altunbas et al. (2018). 4 See Lim, Costa, Columba, Kongsamut, Otani, Saiyid, Wezel, and Wu (Lim et al.), Galati and Moessner (2013),Claessens (2015),Cerutti et al. (2017) and Kahou and Lehar (2017) for an overview and the use of macroprudential policies. 5 See Fendo glu (2017) for a similar argument. M.G. Chadwick / Central Bank Review 18 (2018) 69e8370 monetary and macroprudential approach is more effective for financial stability. Lastly Gambacorta and Murcia (2017) evaluate the effectiveness of macroprudential tools and their interaction with monetary policy for five Latin American countries and they propose that macroprudential tools have a greater effect on credit growth when reinforced by the use of monetary policy to push in the same direction. Similar to all Tillmann (2015),Tovar Mora et al. (2012), Greenwood-Nimmo and Tarassow (2016) and Gambacorta and Murcia (2017), this paper also examines whether macroprudential policy tools are effective besides monetary policy for restraining consumer credit growth yet our paper differs from the literature in two fundamental ways. First, we try to measure the effectiveness of both the monetary policy shocks and macroprudential tightening shocks on not only credit growth but also on credit growth volatility. It is common to measure the impact of macroprudential tools on credit growth and house prices but not on credit growth volatility, which is very important for financial stability and in this respect, it is one of the few papers to measure this impact. Using capital requirements as a macroprudential policy tool Aguirre and Blanco (2015) find that macroprudential policy smooths output, price, interest rate and credit volatility over the business cycle. 6 Second, we combine the “Qual VAR”and the “Panel VAR”methodology to measure the latent probability of macroprudential tightening as the first step via instrumental variable probit model, and then we use this probability within the Panel VAR in the second step. In this way, we try to encounter the endogeneity problem related to the macroprudential policy shocks. 7 The remainder of this paper is organized as follows. Section 2 provides information on the data employed in this study. This section also discusses various methodologies employed in this paper, i.e. stochastic volatility models to extract the volatility of loan data, instrumental variable probit method to extract the latent probability of macroprudential tightening and lastly panel VAR to measure the impact of macroprudential shocks. Section 3provides the empirical results of the panel VAR, policy implications and some robustness exercises. Section 4concludes. 2. Data and methodology Our loan-level data, i.e. loans to households by banks and the interest rates related to those loans come from the CBRT Electronic Data Delivery System. 8 We use three different types of loans, i.e. general purpose loans, housing loans, vehicle loans and credit cards. This data set has weekly frequency. We use weekly frequency data to extract the credit growth volatilities and then convert weekly volatilities to monthly frequency to use them in IV probit and Panel VAR. Weighted average funding rate (WAFR) is used as the monetary policy rate. However, since this data started to be reported in 2011, the Istanbul Stock Exchange (BIST) interbank overnight borrowing rate was used as a proxy for previous periods. WAFR come from the CBRT Electronic Data Delivery System and we use Bloomberg for BIST interbank overnight borrowing rate. Seasonally adjusted industrial production index and consumer price index comes from the Turkstat. We use TRAMO-SEATS to seasonally adjust consumer price index. The source of VIX is Bloomberg. To define macroprudential tightening we employ a dummy variable, which are designed to take the value of one for periods when strict macroprudential policies are implemented and zero otherwise. We decide on the macroprudential tightening periods using Cerutti et al. (2015),Fendo glu (2017) and Ero glu (2018). 9 It is important to state that we exclude all the macroprudential policy actions that are related to commercial credits, as those types of credits are beyond the scope of this study. We use only those instruments that directly address the consumer loans of the households. Table (1) illustrates the descriptive statistics of the data used in this paper. 2.1. Modelling time-varying volatility The effect of monetary policy and macroprudential policy on credit growth volatility is a largely unexplored area. Mishkin and Schmidt-Hebbel (2007) state that, when the monetary policy is suboptimal, the economy will exhibit large output and inflation volatility and will be located at a significant distance from the frontier. A similar result will be observed given a suboptimal macroprudential policy which will create credit growth volatility that will result in output volatility, especially in a developing emerging market country which depends on credits to grow. Indeed, Gould et al. (2016) employ credit volatility as a financial stability indicator and find that countries with higher volatility of credit tend to have lower growth rates. Bennani et al. (2017) state that the reduction in credit volatility is viewed as a proxy for reducing the welfare cost of fluctuations within a dynamic stochastic general equilibrium (DSGE) model. Accordingly, credit volatility enters DSGE as an objective function (ad hoc credit volatility function) that needs to be minimized to smooth financial cycles. 10 We assess the credit growth volatility using stochastic volatility models, where the volatility is a latent variable that follows a stochastic process. 11 We choose the best stochastic volatility model for out credit data out of seven different models, performing a formal Bayesian model comparison exercise given the data. For each seven volatility model, we compute the marginal data density, which evaluates how likely it is for the observed data to have occurred given the model. Using this measure we can obtain the posterior probabilities of the models. 12 We compare these seven timevarying volatility models against each other and choose the one that is better modelled as a latent stochastic process. The first model is the standard stochastic volatility (SV) model: yt¼ m þ s t; s tN0;eh t ;(1) ht¼ m hþfhðht1 m hÞþεt;εtN0; u 2 h:(2) where the log-volatility h t follows a stationary ARð1Þprocess with jf h j<1 and unconditional mean m h . The second stochastic volatility (SV-2) model has the same observation equation as in Eq. (1), but the log-volatility h t follows a 6 See Grydaki and Bezemer (2013) on the importance of lower credit volatility. 7 Qual VAR methodology comes with its own problems. El-Shagi and von Schweinitz (2016) discuss the identification problem that is related to the procedure. 8 https://evds2.tcmb.gov.tr/index.php?/evds/serieMarket. 9 See Table (A1) in the Appendix, which provides a summary of macroprudential policy actions (tightenings) and how they are defined. 10 In Bennani et al. (2017), macroprudential authority to minimize the loss function given by: f v ¼arg minf s 2 creditgrowth þ l y s 2 GDPgrowth þ l v s 2 v g, where s 2 x denotes the unconditional variance of variable xand the parameters l y ; l v ,reflect the policymaker's priorities when trading off a reduction in credit and GDP volatility and a variation in the instrument of an acceptable magnitude. f v is calibrated like a Basel III-type capital buffer rule. 11 The results are generated by MATLAB codes provided by Chan and Grant (2016). 12 See Koop (2003) for a detailed discussion on Bayesian model comparison. M.G. Chadwick / Central Bank Review 18 (2018) 69e83 71 stationary ARð2Þprocess: ht¼ m hþfhðht1 m hÞþ r hðht2 m hÞεt;εtN0; u 2 h: SV-2 model reduces to the standard SV model when r h ¼0. The third stochastic volatility model allows for the possibility of infrequent jumps, which can accommodate drastic changes in credit growth. Under the stochastic volatility model with jumps (SV-J), the observation equation becomes: yt¼ m þktqtþ s t; s tN0;eh t ; where the log-volatility h t follows a stationary ARð1Þprocess. q t 20;1 is a jump variable with success probability Pðq t ¼1Þ¼ k . Hence, if q t ¼1, a jump occurs at time tand its size is determined by if k t , which is modelled as k t Nð m k ; n 2 k Þ. Next we examine the stochastic volatility in mean (SV-M) model of Koopman and Hol Uspensky (2002), under which the stochastic volatility enters the observation equation as a covariate: yt¼ m þ l eh t þ s t; s tN0;eh t ; As before, the log volatility follows a stationary ARð1Þprocess as in Eq. (2). The parameter l captures the extent of volatility feedback. SV-M model reduces to the standard SV model when l ¼0. The fifth model is a version of the stochastic volatility models with moving average innovations (SV-MA) in Chan (2013). The first order SV-MA model will be: yt¼ m þ s t; s t¼utþ j ut1;utN0;eh t : where u 0 ¼0 and j j j<1 and the log volatility follows a stationary ARð1Þprocess as in Eq. (2). The sixth model is the stochastic volatility model with t innovations (SV-t): yt¼ m þ s t; s tt n 0;eh t : where the log-volatility follows a stationary ARð1Þprocess as in Eq. (2). The last model is the stochastic volatility model with leverage (SV-L), which allows a leverage effect. Specifically, the innovations in the observation and state equations can potentially be correlated: yt¼ m þ s t; s tN0;eh t ; htþ1¼ m hþfhðht m hÞþεt; where the innovations s t and ε t jointly follow a bivariate normal distribution:  s t εtN0 B B B B @0; 0 B B B B @ eh t r e 1 2 h t u h r e 1 2 h t u h u 2 h 1 C C C C A 1 C C C C A where, if r <0, given a negative shock to y t at time t, the volatility at time tþ1 tends to be larger. It is also clear that when r ¼0, this model reduces to the standard SV. All the seven stochastic volatility models are estimated using Bayesian techniques; Markov Chain Monte Carlo (MCMC) methods. 13 We sample from the posterior distributions of the models by constructing Markov samplers and use the posterior draws obtained to compute various quantities of interest. For the stochastic volatility models, log-volatilities are sampled jointly Table 1 Panel summary statistics (2005-December to 2017-December). Variable D IP D PR Pr MaP D LR D CRD D CPI CRD VOL D VIX Credit cards mean 0.003 0.009 0.339 0.239 0.011 0.007 1.208 0.000 median 0.003 0.000 0.299 0.041 0.011 0.007 1.207 0.016 s.d. 0.021 0.678 0.269 1.940 0.017 0.005 0.038 0.211 max. 0.085 2.860 1.000 12.507 0.065 0.025 1.337 0.853 min. 0.075 2.340 0.000 8.397 0.036 0.007 1.145 0.486 General purpose mean 0.003 0.009 0.384 0.016 0.021 0.007 0.495 0.000 median 0.003 0.000 0.365 0.095 0.017 0.007 0.429 0.016 s.d. 0.021 0.678 0.240 0.924 0.018 0.005 0.167 0.211 max. 0.085 2.860 0.999 5.490 0.119 0.025 1.088 0.853 min. 0.075 2.340 0.005 2.790 0.012 0.007 0.302 0.486 Housing mean 0.003 0.009 0.274 0.015 0.018 0.007 0.269 0.000 median 0.003 0.000 0.209 0.135 0.015 0.007 0.218 0.016 s.d. 0.021 0.678 0.254 0.786 0.018 0.005 0.194 0.211 max. 0.085 2.860 0.999 3.890 0.119 0.025 1.334 0.853 min. 0.075 2.340 0.000 3.330 0.014 0.007 0.147 0.486 Vehicle mean 0.003 0.009 0.291 0.021 0.000 0.007 0.290 0.000 median 0.003 0.000 0.254 0.115 0.002 0.007 0.288 0.016 s.d. 0.021 0.678 0.221 0.932 0.021 0.005 0.043 0.211 max. 0.085 2.860 0.999 5.350 0.113 0.025 0.450 0.853 min. 0.075 2.340 0.000 1.920 0.048 0.007 0.171 0.486 Definitions: D IP is the monthly logarithmic difference of seasonally adjusted industrial production index. D PR is the monthly difference of policy rate. Pr MaP is the macroprudential tightening probability. D LR is the monthly difference of consumer loan rates. D CRD is the monthly logarithmic difference of consumer loans. D CPI is the monthly logarithmic difference of seasonally adjusted consumer price index. CRD VOL is the volatility of consumer loans. D VIX is the monthly logarithmic difference of VIX. 13 See the Appendix for prior Definitions and hyperparameters. M.G. Chadwick / Central Bank Review 18 (2018) 69e8372 using acceptance-rejection Metropolis-Hastings described in Chan (2017). 14 The marginal likelihoods for the stochastic volatility models are computed using the adaptive importance sampling approach in Chan and Eisenstat (2015). To choose the best volatility model for the credit growth we use Bayesian model comparison via the Bayes factor and the Bayes factor in favour of M i against M j is defined as: BFij ¼pðyjMiÞ pyMj; where pðyjMkÞ¼Zpðyj q k;MkÞpð q kjMkÞd q k is the marginal likelihood under model M k ;k¼i;j,pðyj q k ;M k Þbeing the likelihood function that depends on the model specific parameter vector q k of dimension p k and a prior density pð q k jM k Þ. This marginal likelihood can be interpreted as a density forecast of the data under model M k evaluated at the actual observed data y. Since the Bayes factor BF ij is simply a ratio of two marginal likelihoods, researchers often only report the marginal likelihoods of the set of competing models and we choose to do so in this paper. 15 The results of the marginal likelihoods for seven SV models estimated for credit cards, general purpose loans, housing loans and vehicle loans are reported in Table (2). The data frequency is weekly and the sample period is from December 2005 to December 2017. The data are transformed into rates of change by taking the first difference of the logs. We estimate the GARCH counterparts of these SV models but for all series they perform considerably worse than these models therefore we choose not to report them in the paper. Under GARCH models the conditional variance is a deterministic function of the parameters and past data in contrast to SV models, in which the log-volatility is a random variable. Therefore, SV models are more robust to misspecification and to drastic changes in the time series. This helps explain why SV models perform much better then their counterpart GARCH models. Table (2) shows that the best model is either SV-MA or SV-M for credit growth series and credit cards. The Bayes factors for the four series clearly favour of either SV-MA or SV-M models for these series. 16 Figure (1) exhibits the volatility series of credit cards, general purpose, housing and vehicle loans. We will use these volatilities later within a panel VAR to observe whether monetary policy and macroprudential shocks are able to affect them significantly. 2.2. Modelling latent macroprudential policy probability There is no continuous indicator of macroprudential policy actions. What do we have at hand are a few tightening or easing episodes of macroprudential policy. Including a binary endogenous variable into a panel VAR, i.e. a linear probability model will unfortunately produce results that are difficult to interpret. Therefore, we uncover the latent probability of macroprudential tightening from the binary information of policy episodes, using an instrumental variable probit model estimated by conditional maximum likelihood with heteroscedasticity robust standard errors. In order to account for the endogeneity of macroprudential shocks, the probit function for the tightening outcome is simultaneously estimated with the dynamic interaction with other variables that are assumed to be endogenous to policy actions. 17 We thus specify the following model: y t¼Zt b þX0t g þεt where y  t is a latent variable for unobservable propensity for macroprudential tightening. We only observe a binary dependent variable y t 20;1, which is driven by a continuous latent variable y  . g is defined so that covðε t ;X t Þ¼0. Z t is potentially endogenous and thus correlated with ε t .Wedefine an instrument W t which does not influence y  t directly but is correlated with Z t . In our case Table (3) presents the results from the IV probit regressions for each of the four series. To save some space we only list the results related to two variables, i.e., credit growth D CRD and credit growth volatility CRD VOL. We expect a negative sign for credit growth and credit growth volatility, and as can be observed from Table (3), for most of the regression results we get correct signs. To begin with, the Wald test of exogeneity provides evidence that Table 2 Log marginal likelihoods of the SV models. Credit cards General purpose Housing Vehicle SV 1052.7 490.1 107.4 481.7 (0.00) (0.01) (0.03) (0.01) SV-2 1047.1 490.5 110.1 480.9 (0.00) (0.14) (0.06) (0.02) SV-J 1055.9 495.1 112.9 485.3 (0.09) (0.10) (0.06) (0.34) SV-M 1056.6 456.2 61.7 357.0 (0.02) (0.03) (0.02) (0.04) SV-MA 1013.0 493.1 39.6 434.5 (0.00) (0.01) (0.02) (0.01) SV-t1054.4 491.3 108.7 482.0 (0.01) (0.02) (0.02) (0.02) SV-L 1052.7 491.0 107.6 482.1 (0.01) (0.02) (0.07) (0.01) Note: The numerical standard errors are in parantheses. Zts¼ D IPts; D PRts;Pr Mapts; D LRts; D CRDts; D CPIts; D LRts;CRD VOLtsfor s¼0;1 and Wts¼ D IPts; D PRts;Pr Mapts; D LRts; D CRDts; D CPIts; D LRts;CRD VOLtsfor s>1: 14 See Chan and Grant (2016) for more details. 15 The computation of the marginal likelihood is non-trivial as it is often high dimensional and therefore cannot be obtained analytically. That is why the marginal likelihoods for the stochastic volatility models are computed using the adaptive importance sampling approach in Chan and Eisenstat (2015). 16 The posterior estimates of the model parameters for the SV models are illustrated by Tables (A2), (A3), (A4) and (A5) in the Appendix. 17 See Wooldridge (2010). The model is estimated using ivprobit in Stata. M.G. Chadwick / Central Bank Review 18 (2018) 69e83 73 Fig. 1. Graphs of stochastic volatility models. Table 3 IV Probit results. Credit cards General purpose Housing Vehicle Coefficient of Credit Growth 3.903 5.039 1.569 81.569*** Marginal Effects (SE) (5.520) (11.861) 4.645027 (31.639) Coefficient of Credit Growth Volatility 2.237 0.286 0.283 34.559*** Marginal Effects (SE) (2.603) (1.008) (0.816) (13.219) Instrument Significance ♯ c 2 ½df  164.26 [10] 220.05 [10] 12.30 [10] 5.54 [10] (p-value) (0.000) (0.000) (0.031) (0.853) Over-identification test c 2 ½df 5.45 [10] 2.31 [10] 0.74 [10] 6.34 [10] (p-value) (0.859) (0.804) (0.946) (0.786) % of Correct Classification 88.19 95.35 97.67 90.48 Note: ♯ Amemiya-Lee-Newey minimum c 2 statistic. Marginal effects are estimated holding model covariates at their means. Standard errors (SE) of marginal effects are listed in parentheses. The Chi-square test statistics ( c 2 ) and their degrees of freedom [df] for testing the joint significance of the instruments and for testing the appropriateness of the over-identification restrictions of the instruments are reported. *** refers to 1% significance level. M.G. Chadwick / Central Bank Review 18 (2018) 69e8374 macroprudential tightening is, indeed an endogenous variable. 18 The validity of the instruments was tested using the AmemiyaLee-Newey over identification test. 19 One measure of how well probit models fit the data is to consider the number of correct predictions or classifications that they generate. The approach here is to predict individual's scoring on 1 on the dependent variable “y”, i.e. in our case it is macroprudential tightening”, based on his or her predicted probability. A common cutoff for the predicted probability is 0.5 so that: by¼1if b p >0:5 by¼0if b p 0:5 where byis the predicted score on the response variable. Whenever there is agreement between byand y, there is correct prediction. The higher the percentage of correctly classified observations, the better the model fit is. We can observe from the last row of Table (3) that the correct classification rate is very high for all the loan categories which illustrates that the IV-probit gives us good predictions with respect to the probability of macroprudential tightening. 2.3. Panel data VAR The panel data VAR methodology combines the traditional VAR approach, which treats all the variables in the system as endogenous, and employs panel data with a VAR that also allows for unobserved individual heterogeneity. In its general form, a panel data VAR model can be written as follows: Yit ¼ G 0þ G 1Yit1þftþdtþeit (3) where Y it is a vector of seven key variables: D IP that is the industrial production growth (log-difference of seasonally adjusted industrial production index), D PR that is the log-difference of policy rate, D LR that is the log-difference of general purpose, housing, vehicle loan and credit card rates, D CRD that is the credit growth rate (log-difference of general purpose, housing, vehicle loans and credit cards), D CPI that is the log-difference of consumer price index, CRD VOL that is the volatility of general purpose, housing, vehicle loans and credit card growth rates and Pr MaP, i.e. probability of a macroprudential tightening for general purpose, housing, vehicle loans and credit cards. 20 The advantage of the panel VAR is the same as the advantage of any panel approach, i.e., in allowing for explicit inclusion of a fixed effects in the model, denoted f i in Eq. (3), which captures all unobservable time-invariant factors at a loan-type level. This is important for our purposes as inclusion of these fixed effects allows each loan category to have a loan specific level of each of these factors in the model. However, inclusion of these fixed effects presents an estimation challenge that arises in any dynamic model including lags of the dependent variable: the fixed effects are correlated with the regressors and, therefore, the meandifferencing procedure commonly used to eliminate fixed effects would create biased coefficients. To avoid this problem we use forward mean-differencing, which is a very common procedure in panel VAR literature, also referred to as the “Helmert procedure”. 21 This procedure removes only the forward-mean, i.e., the mean of all the future observations available for each loan-month. This transformation preserves the orthogonality between transformed variables and lagged regressors, which allows us to use lagged regressors as instruments and estimate the coefficients by system GMM. 22 To deal with the time effects, we time difference all the variables prior to inclusion in the model, which is equivalent to putting time dummies in the system. Model represented by Eq. (3) is commonly referred to as reduced form, which contains lagged values of all other variables in the system. The prime benefit of the VAR system is in allowing one to evaluate the impact of the orthogonal shocks, which is accomplished with the impulse response functions. Since, the actual variance-covariance matrix of the errors is unlikely to be diagonal, to isolate shocks to one of the variables in the system it is necessary to decompose the residuals in such a way that they become orthogonal. The usual convention is to adopt a particular ordering and allocate any correlation between the residuals of any two elements to the variable that comes first in the ordering. This procedure is known as Cholesky decomposition of the variancecovariance matrix of residuals and is equivalent to transforming the system in a recursive VAR for identification purposes. 23 The identification assumption via recursive ordering assumes that the variables which come earlier in the ordering affect all the following variables contemporaneously, while the variables that come later affect the previous variables only with a lag. Accordingly, we employ the following ordering: D IP/ D PR/Pr MaP/ D LR/ D CRD/ D CPI/CRD VOL. D IP is placed at the very beginning of the ordering because the state of the real economy is expected to affect many, if not all, other variables contemporaneously, but be affected by other variables in the model with a lag. Indeed, the phase of the business cycle in any given period affects all the other variables, and D IP serves a s a proxy for the state of the business cycle. Given the state of the economy, which is represented by D IP, the authorities can change the policy rate D PR and the macroprudential policies Pr MaP. These policy variables can respond to changes in D IP quickly, but will only be expected to affect the real state of the economy with a lag as it takes time for the monetary policy response. It is expected that the policy rate is transmitted to loan rates which in turn will affect the credit growth and volatility, therefore we put these variables last. For the analysis of the impulse response functions, we need an estimate of their confidence intervals and we generate the confidence intervals for the impulse responses using Monte Carlo simulations. In addition, we evaluate variance decompositions, which show the percent of variation in one variable that is explained by the shock to another variable, accumulated over time. We report the total effect accumulated over the 10 months, but longer time horizons produced equivalent results. 3. Empirical results This section discusses the simulations of impulse responses, which are represented by Figures (2)e(4). 24 In these figures, we plot the responses of variables to a one-standard deviation shock to 18 The Wald test of exogeneity tests whether the correlation between the errors in the full probit equation and reduced-form equation for the endogenous regressor, macroprudential tightening, is equal to zero. Accepting the null hypothesis would have meant that the suspected endogenous variable is in fact exogenous and, therefore, a normal probit could be used. 19 It tests the joint null hypothesis that the excluded instruments are uncorrelated with the error term (and therefore are valid instruments). See Baum et al. (2016). 20 A global risk aversion variable proxied by the VIX index is included as an exogenous variable. 21 See Arellano and Bover (1995) for details. 22 Our panel VAR estimation routine follows Love and Zicchino (2006). 23 See Hamilton (1994) for details. 24 The coefficient estimates and standard errors are given by Table (A6) and Table (A7) in the Appendix. M.G. Chadwick / Central Bank Review 18 (2018) 69e83 75 the monetary policy and macroprudential policy variables. All graphs show responses for the first 10 months, and nearly all of the responses converge to zero in this time frame. The two lines on either side of the impulse response are 5th and 95th percentile bounds constructed using Monte Carlo simulations with 1000 repetitions. Thus, for the periods where the zero line is outside of the error bands we can be 95 percent certain that there is indeed a non-zero effect on the variable under consideration. Figure (2) is based on our panel VAR model (model 1) that does not include the macroprudential tightening probability variable Pr MaP: D IP/ D PR/ D LR/ D CRD/ D CPI/CRD VOL. D IP. First, we will discuss the responses of credit growth and credit growth volatility to monetary policy shocks. It is clear from Figure (2) that monetary policy shock has a negative and significant impact on both credit growth and volatility. The sign of both of the responses is expected, i.e. an increase in policy rate will result in a decreasing credit growth rate and volatility. Second, the response of industrial production growth is significant, negative and very close to the response of credit growth, which is an expected falling industrial outcome of the contractionary monetary policy. Inflation also falls in response to production growth. Finally, it is not surprising to observe a significant hike in the loan rates after a contractionary Fig. 2. Impulse responses for 1 lag VAR of monetary policy shock (Model 1/no macroprudential policy). M.G. Chadwick / Central Bank Review 18 (2018) 69e8376 Economic Research. Technical report. Bruno, V., Shim, I., Shin, H.S., 2017. Comparative assessment of macroprudential policies. J. Financ. Stabil. 28, 183e202. Cakir, M., 2017. A conceptual design of “what and how should a proper macroprudential policy framework be?”A globalistic approach to systemic risk and procuring the data needed. IFC Bulletins chapters 45. Cerutti, E., Claessens, S., Laeven, L., 2015. The Use And Effectiveness Of Macroprudential Policies: New Evidence. IMF Working Paper 15(61). Cerutti, E., Claessens, S., Laeven, L., 2017. The use and effectiveness of macroprudential policies: new evidence. J. Financ. Stabil. 28, 203e224. Chan, J.C., 2013. Moving average stochastic volatility models with application to inflation forecast. J. Econom. 176 (2), 162e172. Chan, J.C., 2017. The stochastic volatility in mean model with time-varying parameters: an application to inflation modeling. J. Bus. Econ. Stat. 35 (1), 17e28. Chan, J.C., Eisenstat, E., 2015. Marginal likelihood estimation with the cross-entropy method. Econom. Rev. 34 (3), 256e285. Chan, J.C., Grant, A.L., 2016. Modeling energy price dynamics: garch versus stochastic volatility. Energy Econ. 54, 182e189. Chen, H., Chow, K., Tillmann, P., 2017. The effectiveness of monetary policy in China: evidence from a qual var. China Econ. Rev. 43, 216e231. Claessens, S., 2015. An overview of macroprudential policy tools. Ann. Rev. Financ. Econ. 7 (1), 397e422. Collard, F., Dellas, H., Diba, B., Loisel, O., 2017. Optimal monetary and prudential policies. Am. Econ. J. Macroecon. 9 (1), 40e87. Dell'Ariccia, G., Laeven, L., Suarez, G.A., 2017. Bank leverage and monetary policy's risk-taking channel: evidence from the United States. J. Finance 72 (2), 613e654. El-Shagi, M., von Schweinitz, G., 2016. Qual var revisited: good forecast, bad story. J. Appl. Econ. 19 (2), 293e321. Erdem, F., € Ozen, E., Unalmis, I., 2017. Are Macroprudential Policies Effective Tools to Reduce Credit Growth in Emerging Markets? Research and Monetary Policy Department, Central Bank of the Republic of Turkey. Technical report. Ero glu, E., 2018. Makroihtiyati Politikalar: Turkiye Uygulamasi Ve Uygulamanin Hanehalki Borcluluguna Etkileri. Mimeo. Central Bank of the Republic of Turkey. Fendo glu, S., 2017. Credit cycles and capital flows: effectiveness of the macroprudential policy framework in emerging market economies. J. Bank. Finance 79, 110e128. Galati, G., Moessner, R., 2013. Macroprudential policyea literature review. J. Econ. Surv. 27 (5), 846e878. Galati, G., Moessner, R., 2017. What do we know about the effects of macroprudential policy? Economica. Gambacorta, L., Murcia, A., 2017. The Impact Of Macroprudential Policies And Their Interaction With Monetary Policy: An Empirical Analysis Using Credit Registry Data. BIS Working Papers No 636. Gould, D.M., Melecky, M., Panterov, G., 2016. Finance, growth and shared prosperity: beyond credit deepening. J. Pol. Model. 38 (4), 737e758. Greenwood-Nimmo, M., Tarassow, A., 2016. Monetary shocks, macroprudential shocks and financial stability. Econ. Modell. 56, 11e24. Grydaki, M., Bezemer, D., 2013. The role of credit in the great moderation: a multivariate garch approach. J. Bank. Finance 37 (11), 4615e4626. Hamilton, J.D., 1994. Time Series Analysis, vol. 2. Princeton university press Princeton. Jim enez, G., Ongena, S., Peydr o, J.-L., Saurina, J., 2017. Macroprudential policy, countercyclical bank capital buffers, and credit supply: evidence from the Spanish dynamic provisioning experiments. J. Polit. Econ. 125 (6), 2126e2177. Kahou, M.E., Lehar, A., 2017. Macroprudential policy: a review. J. Financ. Stabil. 29, 92e105. Kara, H., 2016a. A brief assessment of Turkey's macroprudential policy approach: 2011e2015. Cent. Bank Rev. 16 (3), 85e92. Kara, H., 2016b. Turkey's Experience with Macroprudential Policy. BIS Papers (No 86). Kim, K., Lee, J.Y., 2017. Estimating the effects of fx-related macroprudential policies in korea. Int. Rev. Econ. Finance 50, 23e48. Koop, G., 2003. Bayesian Econometrics. Wiley &Sons, New York. Koopman, S.J., Hol Uspensky, E., 2002. The stochastic volatility in mean model: empirical evidence from international stock markets. J. Appl. Econom. 17 (6), 667e689. Lambertini, L., Mendicino, C., Punzi, M.T., 2013. Leaning against boomebust cycles in credit and housing prices. J. Econ. Dynam. Contr. 37 (8), 1500e1522. Lim, C. H., A. Costa, F. Columba, P. Kongsamut, A. Otani, M. Saiyid, T. Wezel, and X. Wu. Macroprudential Policy: what Instruments and How to Use Them? Lessons from Country Experiences. IMF Working Paper. Love, I., Zicchino, L., 2006. Financial development and dynamic investment behavior: evidence from panel var. Q. Rev. Econ. Finance 46 (2), 190e210. Mishkin, F.S., Schmidt-Hebbel, K., 2007. Does Inflation Targeting Make a Difference? NBER Working Paper No. 12876. Tovar Mora, C.E., Garcia-Escribano, M., Vera Martin, M., 2012. Credit Growth and the Effectiveness of Reserve RequiTovar Morarements and Other Macroprudential Instruments in Latin America. IMF Working Paper WP/12/142. Svensson, L.E., 2017. Cost-benefit analysis of leaning against the wind. J. Monetary Econ. 90 (C), 193e213. Tillmann, P., 2015. Estimating the effects of macroprudential policy shocks: a qual var approach. Econ. Lett. 135, 1e4. Wooldridge, J.M., 2010. Econometric Analysis of Cross Section and Panel Data. MIT press. M.G. Chadwick / Central Bank Review 18 (2018) 69e83 83