Monitoring and Forecasting Cyclical Dynamics in Bank Credits: Evidence from Turkish Banking Sector

Transcription

1. Monitoring and Forecasting Cyclical Dynamics in Bank Credits : Evidence from Turkish Banking Sector Mehmet Selman ÇOLAK İbrahim Ethem GÜNEY Ahmet ŞENOL Muhammed Hasan YILMAZ September 2019 Working Paper No: 19/29
2. © Central Bank of the Republic of Turkey 2019 Address: Central Bank of the Republic of Turkey Head Office Structural Economic Research Department Hacı Bayram Mh. İstiklal Caddesi No: 10 Ulus, 06050 Ankara, Turkey Phone: +90 312 507 80 04 Facsimile: +90 312 507 78 96 The views expressed in this working paper are those of the author(s) and do not necessarily represent the official views of the Central Bank of the Republic of Turkey. The Working Paper Series are externally refereed.
3. Monitoring and Forecasting Cyclical Dynamics in Bank Credits : Evidence from Turkish Banking Sector Mehmet Selman Çolak İbrahim Ethem Güney Ahmet Şenol Muhammed Hasan Yılmaz Abstract Credit growth rate deviating from its long-run trend or equilibrium value holds importance for policymakers given the implications on economic activity and macro-financial interactions. In the first part of this study, the main aim is to construct indicators for determining the episodes of moderate-to-excessive credit slowdown and expansion by utilizing time-series filtering methods such as Hodrick-Prescott filter, Butterworth filter, Christiano-Fitzgerald filter and Hamilton filter over the time period 2007-2019. In addition to filtering choices, four different credit ratios (which are credit-to-GDP ratio, real credit growth, logarithm of real credit, credit impulse ratio) are included in the methodology to ensure the robustness. This framework enables one to generate monitoring tools for not only total loans, but also for financial intermediation activities with different loan breakdowns regarding type, sector and currency denomination. Moreover, industry-based dynamics of commercial loans are examined by using micro-level Credit Registry data set. In the following part, the credit cycle implied by macroeconomic dynamics are investigated by using factor-augmented predictive regression models. In this context, factors representing the global economic developments, banking sector outlook, local financial conditions and economic growth tendencies are created from large data set of 107 time series by utilizing principal component analysis. Analysis conducted for January 2009-April 2019 interval seems to be in line with exogenous shocks affecting the credit market in the corresponding period. To gain more knowledge about the predictive power of factor-augmented regression models, out-of-sample forecasting exercises are performed. It is found that global forces and economic activity provide substantial improvement in terms of predictive power over simple autoregressive benchmark models given low level of relative forecast errors. Özet İktisadi faaliyet ve makro-finansal etkileşimler dikkate alındığında, kredi büyüme oranlarının uzun vadeli eğiliminden ya da denge değerinden sapması politika yapıcılar açısından önem arz etmektedir. Bu çalışmanın ilk aşamasında Hodrick-Prescott, Butterworth, Christiano-Fitzgerald ve Hamilton filtrelerinin kullanıldığı zaman serisi analizleri yoluyla, 2007-2019 dönemi için aşırı kredi daralması ve genişlemesine işaret eden dönemleri takip eden göstergeler oluşturulmuştur. Filtreleme yöntemlerinin dışında sağlamlığı desteklemek için dört farklı kredi rasyosu (kredi/GSYİH oranı, reel kredi büyümesi, reel kredilerin logaritması, kredi etki oranı) metodolojiye dahil edilmiştir. Kullanılan yöntem çerçevesi toplam kredilere ek olarak, kredi türü, para birimi ve niteliği kırılımlarında finansal aracılık faaliyetlerinin sayısal takibine imkan sunmaktadır. Ek olarak, sektör bazlı ticari kredi dinamikleri Risk Merkezi mikro verisi kullanılarak analiz edilmektedir. Çalışmanın takip eden aşamasında, faktör eklenmiş regresyon modelleri kullanılarak makroekonomik ve finansal dinamiklerin ima ettiği kredi çevrimleri elde edilerek gerçekleşmelerle karşılaştırılmıştır. Faktörlerin oluşturulması adına küresel ekonomik gelişmeler, bankacılık sektörü görünümü, yurtiçi finansal koşullar ve iktisadi faaliyet dinamiklerini içeren 107 değişkenlik geniş bir veri setine temel bileşenler analizi uygulanmıştır. Ocak 2009-Nisan 2019 dönemine ilişkin bulgular modellerin ima ettiği kredi çevrimi ile gerçekleşmelerin seyrinin tarihsel olarak dışsal şokları yansıttığına işaret etmektedir. Faktör eklenmiş regresyon modellerinin açıklayıcılık gücü hakkında ek bilgiler elde etmek amacıyla tahmin egzersizleri yapılmıştır. Göreli tahmin hataları incelendiğinde, özellikle küresel gelişmeler ve yurtiçi iktisadi faaliyete ilişkin faktörlerin kredi çevriminin tahmininde baz modellere göre ek bilgi sağladığı gözlenmektedir. JEL Classification: G21, E51, C38, C53 Keywords: Credit Cycle, Macroeconomic Dynamics, Filtering, Factor Models, Forecasting 1
4. Non-Technical Summary Cyclical dynamics in financial intermediation activities holds importance for the course of economic activity in emerging markets . Particularly, in countries with banking sectororiented financial structure, episodes characterized with drastic slowdown in credit allocation could hurt firm prospects, investment tendencies and economic growth. On the other hand, excessive credit growth periods are equivalently hazardous as it could lead to accumulation of financial risks, occurrence of asset price bubbles and emergence of current account imbalances. In fact, in the post-Global Financial Crisis era, constructed banking sector regulations on the global scale pay attention to some methods of quantifying the cyclical credit behavior. In assessing the credit cycle dynamics, multiple parameters are crucial, as it is seen in the literature and practice. The statistical filtering method and credit ratio definition are two of such significant parameters. Our study aims to contribute to the existing literature by undertaking a comprehensive empirical analysis embodying four different time-series filters and four different credit ratios. In terms of aggregate banking sector analysis with nine credit sub-groups, we are able to monitor the moderate-to-excessive credit slowdown and expansion episodes in Turkey with monthly frequency for recent time period. Moreover, similar methodological framework is applied on sector-specific commercial loans retrieved from Credit Registry data to characterize the differentiation across sectors. On the top of monitoring tasks, later parts of this study deal with predicting and forecasting credit cycle behavior in Turkey. To define the credit cycle level implied by macroeconomic and financial outlook, we utilize factor-augmented predictive regressions which summarize the informative content of a large data set. In particular, data set of 107 variables representing global economic developments, banking sector outlook, local financial conditions and economic growth tendencies are added to predictive models via principal component analysis. The relationship between in-sample fitted values and credit cycle realizations reflect the impact of exogenous shocks faced by our economic structure from historical perspective. Furthermore, out-of-sample forecasting exercises have yielded the conclusion that global and growth-related forces improve the forecasting accuracy in predicting credit cycle dynamics. 2
5. 1 . Introduction Credit growth, as one of the prominent indicators to be monitored by the regulatory authorities, plays a critical role in macroeconomic dynamics of emerging markets. In the era of credit shrinkage, consumption and investment growth rates are expected to decline, and ultimately, leads to deceleration in economic activity and worsening in employment. On the other hand, excessive credit growth is expected to cause financial instabilities and macroeconomic imbalances such as bubbles in asset prices, deterioration in current account balance and inflationary pressures. The earlier studies in the literature of financial crisis also manifest that moderate-to-excessive credit expansion is associated with banking crisis episodes (Demirguç-Kunt and Detragiache, 1998; Borio and Lowe, 2002; Kaminsky and Reinhart, 1999; and Eichengreen and Arteta, 2002). In this context, regulatory authorities have been closely monitoring the abnormal credit developments. Particularly, in the aftermath of global financial crisis, excessive rise in credit use in emerging markets due to accommodative monetary policies in major central banks and drastic increase in capital inflows necessitated authorities to implement macro-prudential measures against credit booms. For all these reasons, economists and policymakers have given particular attention to the determination of extreme movements in credits as well as equilibrium level of credit growth implied by macroeconomic fundamentals (Buncic and Melecky, 2013; Kiss et al., 2006; Jakubik and Moinescu, 2015; Drehmann et al., 2010). This is crucial in the sense that most policymakers are concerned with achieving credit growth rate being consistent with macro dynamics. For instance, for the countries that are in early phase of financial development, higher credit growth rates might not be sufficient given the fact that financial deepening process fundamentally requires even higher loan growth. And similarly, for others which have already had well-functioning financial markets, relatively low level of credit growth might be excessive. The aim of this study is to create tools to determine the moderate-to-excessive credit growth or crunch periods for total loans and their breakdowns, e.g. commercial/consumer1 and FX/TRY disaggregation, through implementation of econometric techniques. As a first 1 Types of consumer loans used in this study are personal finance, vehicle and housing loans. 3
6. step , we are de-trending credit series by using statistical filters and obtain credit cycle realizations. Then, the boom/bust periods are determined by looking at the position of the cycle value with respect to the assigned certain threshold value. In the following step, we predict fair values for the cycle series of aggregate credits (and different types of credits) with macroeconomic variables via factor-augmented regressions. This will enable us to interpret the degree of compatibility between credit cycle realizations and what macroeconomic/financial fundamentals imply. As a last step of analysis, credit cycle is being forecasted with macroeconomic and financial factors. The outline of the paper is as follows: In Section 2, we cover related literature by focusing on cycle measures, filtering techniques and widely used econometric methods. In Section 3, the methodology about cycle extraction, data compression technique (principle component), fair value estimation and out-of-sample forecasting framework are all explained. Section 4 presents the data concerning the credit indicators, several breakdowns of aggregate credits and macroeconomic variables used in the fair value estimations. In the following section, we provide and discuss the empirical results on de-trending, factoraugmented fair value regression models and forecasting exercises. In the final section, we will conclude the paper by summarizing the overall findings and policy implications derived from the paper. 2. Literature Review Defining and quantifying financial cycle as a tool to monitor financial stability and direct macroprudential policies have been intensively focused in the finance and banking literature. In this section, we focus on the existing literature on credit indicators, filtering methodologies and macroeconomic determinants of credit indicators. There exist several credit measures proposed in the literature so as to separate cyclical component from its trend. Calza et al. (2006); Eller et al. (2010), and IMF2 (2004) utilize the level of real credit as an indicator for credit measure in their studies. While Utari et al. (2014) and Guo and Stepanyan (2011) prefer nominal credit growth rate in their boom-bust analyses, Elekdag and Wu (2011) use real credit growth rate as an indicator in their study 2 Terrones, M., Mendoza E., Sutton Bennet.(2004). Are Credit Booms In Emergıng Markets A Concern?. IMF World Economic Outlook, April 2004: Advancing Structural Reforms, 147-166 4
7. on excess credit movements . Considering small sample size and the data with structural breaks, utilization of credit level and growth ratios have some drawbacks. To exemplify, if initial level of credit is small, the growth rate of credit level for the subsequent period may unusually high notwithstanding the credit measure returns its historical level in the following period. From the demand side perspective, the use of credit growth measures might be misleading if they do not taking into account the income levels of countries. In order to eliminate the abovementioned problems, several empirical studies propose aggregate credit level to be scaled by some macro level aggregates. In this case, the most preferred indicator is the ratio of credit level to gross domestic product (GDP) (Gourinchas et. al., 2001; Cottarelli et. al, 2005; Barajas, Dell’Ariccia, Levchenko, 2007; Dell’Ariccia et. al., 2016; Dell’Ariccia et. al., 2012; Ottens, Lambregts and Poelhekke, 2005; Castro and Martins, 2018; Hosszú et al., 2015; Kocsis and Sallay, 2018). Furthermore, Mendoza and Terrones (2008) focus on real credit per capita by scaling aggregate real credit by population, while Arena et al. (2015) suggest the logarithm of real credit per capita. In addition to these macro level aggregates, Kara and Tiryaki (2013, 2014) developed credit impulse ratio in order to comprise proper credit growth rate by using historical private loan growth paths for emerging economies. They propose the ratio of change in credit stock to GDP measure and exhibit it for Turkey in their research. With this calculation, stock credit data can be transformed into flow structure. For seperating the cycle series from the trend series of the credit indicators mentioned above, there exist a variety of filtering methods in the literature. The most prominent methodology is Hodrick-Prescott (HP) filter developed by Hodrick and Prescott (1997) and further utilized by Hilbers et al. (2005), Drehman et al. (2010), Elekdağ and Wu (2011). Due to some drawbacks observed in the practice of HP filter such as sensitivity to smoothing parameter and end-point bias, some works in the literature choose to utilize other ones (Coşar et al., 2012). Hosszú et al. (2015), for instance, implement variety of univariate filtering methods such as HP, Christiano-Fitzgerald (CF), Beveridge-Nelson and complement their empirical approach with multivariate filtering methods such as multivariate HP filter in their studies on credit cycles. Furthermore, Kocsis and Sallay (2018) calculate credit-toGDP gap by using multivariate HP method in terms of corporate, household and aggregate 5
8. credits . In addition to these filtering methods, Schüler (2018) prefer Hamilton filter in his work for determining the de-trended credit-to-GDP ratio. Another strand of the literature, related to abovementioned credit boom investigations, is estimating binary outcome models on a cross-country setting to analyze the impact of macroeconomic, global and banking sector-related variables on the probability of experiencing credit boom episodes. In this context, with logit models, Barajas et al. (2007) find that high inflation and bad quality banking supervision are coincided with credit boom episodes. Dell’Ariccia et al. (2016) estimate a probit model and show that stronger economic activity, surge in capital flows, financial reforms (liberalization) contribute to the probability of credit boom; while embracing flexible exchange rate regime and having lessoriented banking system are found to decrease the probability of booms. Arena et al. (2015) benefit from panel logit estimations to argue that capital flows, financial development and GDP growth increase the probability of credit booms, whereas global funding conditions (US Fed rate) and trade openness have negative impacts. Third group of studies approaches the issue from equilibrium credit perspective. Kiss et al. (2006) analyze the long-run relationship between credit and GDP in Eastern European countries by using a system of equations and pooled mean group estimator. Guo and Stepanyan (2011) focus on a cross-country setup involving the estimation of credit growth rate with fixed effects panel regressions. Buncic and Melecky (2013) proceed with an errorcorrection form of ARDL model to determine the equilibrium credit level which is constructed with the inputs such as GDP, real interest rate, lending-deposit spread, inflation and cost of borrowing for banks by using pooled mean group estimator. Gersl and Seidler (2010) identify a similar framework by using financial development, GDP per capita and consumption to compose long-run equilibrium credit level, again by using pooled mean group estimator. There is an extensive literature dealing with the determination of extreme and optimal credit cycles for advanced and emerging markets, whereas fewer studies exists for Turkish economy. As an early attempt to quantify credit gap developments, Binici and Köksal (2012) identify the phases of excessive credit movements in Turkish banking sector for the period 6
9. covering December 2002-April 2012 . Total credit extended in the sector is decomposed into cyclical part by utilizing three different indicators including credit/GDP ratio, nominal credit and real credit. Without resorting to statistical filters, Binici and Köksal (2012) employ 12months moving average to isolate the deviations of amount of extended loans from the trend. Their results indicate that years like 2006, 2008 and 2011 are characterized by excessive credit movements. To complement their empirical investigation, they design binary variables taking value of 1 for identified episodes (in the first stage), and 0 otherwise. In the following step, they assess the impact of pre-selected global and local financial variables on the probability that credit dynamics display excessive movements. It is found that capital flows, US interest rates, real exchange rate and slope of the yield curve have statistically significant effects on credit cycle. More recently, as an insightful study, Aydın and Yılmaz (2019) examine credit gap indicators for Turkish banking sector. In contrast to previous works for Turkey, this study implements univariate time-series filtering technique to calculate the credit gap indicators. Given the drawbacks of Hodrick-Prescott (HP) filter, they highlight the advantages of Hamilton filter, in terms of being independent of smoothing parameter determination and the ability to allow volatile trend structure, in the analysis. Their results show the existence of excess credit growth in the post-crisis period. The literature on the interaction between credit indicator and macroeconomic variables has evolved into three main categories. One group of studies in the literature is directly focusing on the credit boom episodes and provides descriptive measures about how macroeconomic variables behave around those accelerated credit dynamics. Hilbers et al. (2005) examine the average realizations of GDP, trade balance, inflation and current account around boom phases. Similarly, Dell’Ariccia et al. (2016) distinguish between boom and non-boom episodes of credit cycle by providing descriptive averaged economic performances in these two cases. Arena et al. (2015) apply event study analysis to examine the behavior of macroeconomic variables around the peak of credit boom episodes. In our paper, we are choosing to use rather a flexible approach employing factoraugmented models. Instead of restricting our methodology to certain macroeconomic and financial variables, a broad data set incorporating 107 variables are pre-determined and the 7
10. respective information is extracted through the application of factor models. Then, what this information implies is compared with credit cycle realizations through factoraugmented predictive regressions. This method is similar to equilibrium credit calculations (which we use rather loose term of fair value estimations) and it allows to identify credit cycle movements diverging from what local and global economic fundamentals imply. 3. Methodology In this section, firstly, filtering methods for the extraction of credit cycle from its long-run trend are reviewed. Data compression technique applied on large macroeconomic/financial data set to form the factor-augmented regressions will also be discussed. In the last subsection, forecasting framework utilized to assess predictive performance of macroeconomic/financial factors for credit cycle dynamics is explained. 3.1. Filtering Methods We choose to proceed with four different statistical filters, namely HP filter, CF filter, Butterworth (BW) filter and lastly Hamilton filter. Use of multiple univariate time series filters is preferred to enhance the robustness of credit cycle calculation. HP filter, developed by Hodrick and Prescott (1997), has been widely used as a method in macroeconomic analysis to isolate smooth estimate of long-term trend, particularly with respect to business cycles. On the other hand, large number of empirical works use it to identify financial cycles determined by credit growth and asset prices3. Technically, two-sided HP filter technique is a frequency-pass filter taking the historic and future information on the time series into consideration. It assumes that a data series