Are CBRT’s monetary policy statements affected by ECB and FED statements?
Necmettin Alpay KOÇAK1
Abstract
This paper aims to examine whether CBRT’s statements are affected by ECB and FED statements. The sentiment analysis is performed to build the balance sentiment indicators (BSI) for CBRT, ECB, and FED from 2008 January to 2020 August using by their statements. Individual and group properties of BSI’s are examined by unit root tests and the Bounds cointegration tests. The short and long-run effects of ECB and FED sentiments on CBRT’s sentiments are investigated by the ARDL and UECM models. The analysis applied over the full sample period and two sub-samples which they represent the periods before and after 2013 June. The results imply that the sentiment indicators of CBRT, ECB, and FED have statistically significant cointegration relationship over the full sample period and the period after 2013 June. Particularly after June 2013, the CBRT’s statements are positively related with the ECB’s and FED’s statements in the long run however CBRT’s statements are negatively affected by FED’s statements in the short run.
Keywords: Communication, Central bank, Monetary policy statements, Sentiment analysis, Cointegration
JEL classification: C32, C53, E31, E37
TCMB para politikası açıklamaları ECB ve FED açıklamalarından etkileniyor mu?
Öz
Bu çalışma, TCMB, ECB ve FED’in para politikası açıklamalarındaki duyguların ampirik bir analizini sunmaktadır. Merkez Bankası açıklamalarının ECB ve FED açıklamalarından etkilenip etkilenmediğini incelemeyi amaçlamaktadır. Üç merkez bankası için 2008 Ocak’tan 2020 Ağustos’a kadar göstergelerin oluşturulmasına yönelik bir duyarlılık analizi içerir Duygu göstergeleri arasındaki kısa ve uzun vadeli ilişkiyi ve TCMB’nin duygularının belirleyicileri olan ECB ve FED’in duyguları, sınırsız hata düzeltmesi model ve Sınır testi analizi tüm örneklem periyodu boyunca uygulanmıştır ve iki alt örnek 2013 Haziran öncesi ve sonrasını temsil etmektedir. Sonuçlar, TCMB, ECB ve FED’in duyarlılık göstergelerinin tüm örneklem dönemi ve 2013 Haziran sonrası dönemde istatistiksel olarak anlamlı bir eş bütünleşme ilişkisine sahip olduğunu göstermektedir. TCMB açıklamaları uzun vadede ECB açıklamalarıyla uyumludur, ancak TCMB kısa dönemde FED açıklamalarını tersine değerlendirmektedir.
Anahtar Sözcükler: İletişim, Merkez bankası, Para politikası beyanları, Duyarlılık analizi, Eşbütünleşme.
JEL kodları: C32, C53, E31, E37
Introduction
The communication strategy of a central bank has crucial importance to be understood its policy actions and assessments by the economic agents (Bernanke and Reinhart 2004). Especially, the monetary policy statements (hereafter statements) have a role of “forward guidance” to manage the economic agents expectations and to provide a transparent monetary policy (Hansen and McMahon 2016).
In the literature, several studies have explored the statements’ role in the communications strategy of central banks using sentiment analysis. Recently, Bholat (2015), Hansen, McMahon, and Prat (2018), and Stegmann (2019) analyzed the Federal Reserve (FED) statements while Berger, De Haan, and Sturm (2011) and Coenen et al. (2017) examined the European Central Bank (ECB) statements. Kahveci and Odabaş (2016) and Iglesias, Ortiz, and Rodrigo (2017) studied the Central Bank of the Republic of Turkey (CBRT) statements. The related studies are focused on time series properties of the statements’ sentiment indicators, and a possible relationship between them and economic indicators.
There exists a consensus in the literature that the monetary policies of central banks are related to each other symmetrically or asymmetrically (Throop (1994), Siklos and Wohar (1997), Bec, Salem, and Collard (2002)). For instance, Bec, Salem, and Collard (2002) examines non-linear taylor-type monetary reaction functions for US, France and Germany. The authors suggest a model framework which allows one country’s interest rate may have a role in other country’s reaction function which data covers the years between 1982 and 1997. They use Generalized Method of Moments estimation method to estimate threshold models. The study concludes that the inter-related monetary reaction functions for the US, Germany and France can be represented by non-linear models.
Therefore, it can be argued whether there is a relationship among the communication strategies of central banks. There are very few studies which examines the relationship. Kahveci and Odabaş (2016) is the only study that visually analyzes the sentiments in ECB, CBRT and FED statements by comparing their behaviors over pre & post-global economic crisis period. Recently, Armelius et al. (2020) suggest that there is a co-movement in the sentiments among the central banks. They analyzed the governor’s speeches data set for 22 central banks, not including the CBRT. Therefore, it seems that there is a gap in the literature regarding the examination of the dynamic relationships among the CBRT, ECB and FED statements.
This paper aims to explain the dynamics of the sentiments in the CBRT statements by using the sentiments in the ECB and the FED statements for the period between 2008 January and 2020 August. Individual and group properties of the sentiment indicators are examined by unit root tests and the Bounds testing approach suggested by Pesaran, Shin, and Smith (2001). The short and long-run dynamics of this relationship are investigated by the Autoregressive Distributed Lag (ARDL) and the Restricted Error Correction (RECM) models.
This paper is structured as follows. The research model, the data and the methodology are described in the Section @ref(sec:model). The empirical findings are reported in the Section @ref(sec:result). Several implications and concluding remarks are given in the conclusion.
Model, data and method
A model can be suggested which includes CBRT statements as dependent variable and ECB and FED statements as independent variables. However, it would be wrong to assume that the CBRT’s statements is only be affected by ECB and FED statements. Therefore, I also added the annual consumer inflation variable to the model. Annual consumer inflation variable may be related with CBRT’s statements since the primary target of CBRT is to achieve price stability. Therefore, their actions and discourses may be shaped according to the annual consumer inflation. Hence, I use the following model structure given in Eq. @ref(eq:n1) to explain CBRT’s sentiments.
\[\begin{equation} BSI_{CBRT,t} = \alpha_{0} + \alpha_{1} CPI_{TR,t} + \alpha_{2} BSI_{ECB,t} + \alpha_{3} BSI_{FED,t} + \epsilon_{t} (\#eq:n1) \end{equation}\]
\(CPI_{TR}\) is the annual consumer price inflation by months which covers the period between January 2008 and August 2020. CPI data is provided by TURKSTAT. \(BSI\) represents the balance sentiment indicators for CBRT, ECB and FED statements. I applied the sentiment analysis to the CBRT, ECB, and FED statements from 2008 January to 2020 August. I consider two types of sentiments, namely positivity and negativity, by following the method suggested by Rinker (2019). The analysis is performed to obtain positive and negative sentiments using the dictionaries defined in Hu and Liu (2004), Henry (2008), Loughran and McDonald (2011), Young and Soroka (2012), Mohammad and Turney (2013). The positivity (negativity) score is the ratio of the number of positive (negative) words to the total number of words in a statement. I assume the scores to show an exponential decay process running from the date of a speech until the next one. Afterward, I perform the aggregation by taking monthly averages on each score. The monthly score is calculated by the average of daily sentiments. Each \(BSI\) is calculated by subtracting the negativity score from the positivity score for CBRT, ECB, and FED statements. Finally, I rescale the monthly BSI. The BSI indicator lies between 0 and 100 with 50 mean. If BS>50, the positivity is dominant on the statements, otherwise the negativity. If BS=50, the statements have a neutral sentiment. The data used in this paper is available at https://bit.ly/33mzXKd .
Unit root tests
A time series is considered as stationary in which level the mean and variance of the time series do not change depending on time and the relationship between the two periods depends only on the distance (Gujarati 2009). Identification of stationary level is important to selection of the methods for further empirical analysis. Five different tests, which are frequently used in the literature, are used in this paper. These are Augmented Dickey-Fuller (hereafter ADF) (Dickey and Fuller 1981), Philips-Perron (hereafter PP) (Phillips and Perron 1988), Elliott, Rothenberg and Stock (hereafter ERS) (Elliott, Rothenberg, and Stock 1996), Kwiatkowski et al. (hereafter KPSS) (Kwiatkowski et al. 1992), and Zivot-Andrews (hereafter ZA) (Zivot and Andrews 1992) tests. The ADF performs a test which the series is not stationary in the null hypothesis. The PP also performs a test which the series is not stationary in the null hypothesis. The possible autocorrelation problem in the test equation is solved by determining the appropriate lag length in the ADF test, while the autocorrelation and variance problem in the test equation is solved with a non-parametric approach (HAC) in the PP test. On the other hand, the most important disadvantage of the PP test is that it is based on asymptotic theory and the suggested HAC correction is invalid in short time series. ADF and PP tests are sensitive to structural breaks and have poor test strengths in small samples. In the ERS test, the ADF test equation is used, but it clears the series from the linear trend as “local.” Thus, the ERS test is superior to the ADF test in terms of test strength. Unlike other tests, the KPSS test tests that the series is stationary in the null hypothesis, and in this way it is a more advanced version of the ADF test in terms of test power. The ZA test is a stronger stationary test compared to other tests in terms of sensitivity to structural breaks. In this test, a single structural break is allowed in the series and the minimum observation of the t-statistics in the ADF test is the observation that the break is experienced. Zivot-Andrews test is the same as ADF, PP and ERS in terms of hypothesis structure. The optimal lag lengths for the test equations are selected according to the AIC (Akaike Information Criterion) criterion.
Cointegration test
According to the studies of Engle and Granger (1987), Søren Johansen (1988) and S. Johansen and Juselius (1990), all series must have same integration order and the order should not be zero. The cointegration relationship with these tests cannot be investigated if one or more series are I(O). The Bounds test approach suggested by Pesaran, Shin, and Smith (2001) eliminates this problem. The Bounds test is applied regardless of whether the variables used in the model are I(0) or I(1) and the null hypothesis is there is no cointegration relationship among the variables. To do this, a preliminary ARDL model with the orders ARDL\((p,q_{1},q_{2},q_{3})\) is estimated to determine the optimal orders, then UECM are estimated with the orders minus one, i.e. \((p-1,q_{1}-1,q_{2}-1,q_{3}-1)\). The adaptation of UECM to this paper is as given in Eq. @ref(eq:n2):
\[\begin{equation} \begin{aligned} \Delta BSI_{CBRT,t} = & \nu + \phi_{1}BSI_{CBRT,t-1} + \phi_{2}CPI_{TR,t-1} + \phi_{3}BSI_{ECB,t-1} + \phi_{4}BSI_{FED,t-1}\\ & + \sum_{i=1}^{p-1}{\beta_{i}\Delta BSI_{CBRT,t-i}} + \sum_{j=0}^{q_{1}-1}{\gamma_{j} \Delta CPI_{TR,t-j}}\\ & + \sum_{k=0}^{q_{2}-1}{\rho_{k} \Delta BSI_{ECB,t-k}} + \sum_{l=0}^{q_{3}-1}{\pi_{l} \Delta BSI_{FED,t-l}}\\ & + \xi_{t};\qquad t = Jan 2008,\ldots,Aug 2020\\ \end{aligned} (\#eq:n2) \end{equation}\]
The optimal values for the orders \((p,q_{1},q_{2},q_{3})\) are selected according to the \(Adjusted \ R^2\) criterion up to maximum orders equal to \(3\) for each. The absence of auto-correlation problem also taken into cosideration in the selection of lag length selection procedure. Long-run properties of the Eq. @ref(eq:n2) are investigated by the Wald bounds-test for no cointegration proposed by Pesaran, Shin, and Smith (2001). According to Pesaran, Shin, and Smith (2001), the long-run relationship distinguishes among 5 different cases. These differ in terms of whether the ‘intercept’ and/or the ‘trend’ are restricted to participate in the long-run relationship, or they are unrestricted and so they participate in the short-run relationship. In this paper, I use the Case 3 which the cointegration model contains unrestricted intercept and no trend. The bounds test is a Wald test on the parameters of Eq. @ref(eq:n2) expressed as an F-statistic. Since the distribution of the F test statistics used is not standard in the test, Pesaran, Shin, and Smith (2001) suggests to use the upper and lower critical value bounds in a certain significant level. A Wald test is applied on the parameters of Eq. @ref(eq:n2) for the following hypotheses given in Eq. @ref(eq:n3):
\[\begin{equation} \begin{aligned} \mathbf{H_{0}:} \phi_{1} = \phi_{2} = \phi_{3} = \phi_{4} = 0\\ \mathbf{H_{1}:} \phi_{1} \neq \phi_{2} \neq \phi_{3} \neq \phi_{4} \neq 0\\ \end{aligned} (\#eq:n3) \end{equation}\]
If the calculated F statistic is less than the Pesaran lower critical value, there is no cointegration relationship between the series. If the calculated F statistic is between the lower and upper critical value, a definite interpretation cannot be made and other cointegration test approaches should be applied. If the calculated F statistic is above the upper critical value, there is a cointegration relationship between the series.
Pesaran, Shin, and Smith (2001) suggests the t-bounds test for no cointegration. It is a t-test on the parameters of Eq. @ref(eq:n2). The t-bounds test can be applied for cases 1,3 and 5. Pesaran, Shin, and Smith (2001) also suggests to use the upper and lower critical value bounds for t-test in a certain significant level. A t-test is applied on the parameter of Eq. @ref(eq:n2) for the following hypotheses given in Eq. @ref(eq:n4):
\[\begin{equation} \begin{aligned} \mathbf{H_{0}:} \phi_{1} = 0\\ \mathbf{H_{1}:} \phi_{1} \neq 0\\ \end{aligned} (\#eq:n4) \end{equation}\]
The evaluation approach described above for the F-test is equally valid for the t-test.
Examination of long-run and short-run dynamics
ARDL model is used to measure the long-run effects of the \(BSI_{ECB}\) and the \(BSI_{FED}\) on the \(BSI_{CBRT}\) after the cointegration relationship between \(BSI\) indicators has been determined. The ARDL model with the orders \((p,q_{1},q_{2},q_{3})\) specification is given in Eq. @ref(eq:n5):
\[\begin{equation} \begin{aligned} BSI_{CBRT,t} = & c_{0} + \sum_{i=1}^{p}{\delta_{1,i}BSI_{CBRT,t-i}} + \sum_{j=0}^{q_{1}}{\delta_{2,j}CPI_{TR,t-i}}\\ & + \sum_{k=0}^{q_{2}}{\delta_{3,k}BSI_{ECB,t-j}} + \sum_{l=0}^{q_{3}}{\delta_{4,l}BSI_{FED,t-l}}\\ & + \ \epsilon_{t};\qquad t = Jan \ 2008,\ldots,Aug \ 2020\\ \end{aligned} (\#eq:n5) \end{equation}\]
The estimates of the coefficients in Eq. @ref(eq:n5) are robust under the assumption of lack of auto-correlation, heteroskedasticity, non-linearity and structural change in \(\hat\epsilon_{t}\). The fitted residuals are tested by the BG(p) test (Breusch 1978) against auto-correlation for \(p\) order, the GQ test (Goldfeld and Quandt 1965) against heteroskedasticity, the HC test (Harvey and Collier 1977) against non-linearity and the M-test (Zeileis 2006) against structural change.
The long-run coefficients are calculated from Eq. @ref(eq:n5) following the method suggested by Bårdsen (1989). The delta method is used for approximating the standard errors (and thus the t-statistics) of the estimated coefficients.
\[\begin{equation} \begin{aligned} \mu = \frac{c_{0}}{1-∑_{i=1}^{p}} \\ \alpha_{j} = \frac{∑_{l=0}^{q_{j}}\delta_{j,l}}{1-∑_{i=1}^{p}{\delta_{1,i}}} \;\;\;\;\; \forall j=CPI_{TR},BSI_{ECB},BSI_{FED} \end{aligned} (\#eq:n6) \end{equation}\]
\(\mu\) is the constant term in the long-run relationship, \(\alpha\) is the long-run coefficients of the variables in question.
By Eq. @ref(eq:n7), the RECM has been established as follows to investigate the relationship between variables in the short-run.
\[\begin{equation} \begin{aligned} \Delta BSI_{CBRT,t} = & \nu + \sigma_{1}ECT_{t-1}+ \sum_{i=1}^{p-1}{\beta_{i}\Delta BSI_{CBRT,t-i}} + \sum_{j=0}^{q_{1}-1}{\gamma_{j} \Delta CPI_{TR,t-j}}\\ & + \sum_{k=0}^{q_{2}-1}{\rho_{k} \Delta BSI_{ECB,t-k}} + \sum_{l=0}^{q_{3}-1}{\pi_{l} \Delta BSI_{FED,t-l}}\\ & + \ \xi_{t};\qquad t = Jan\ 2008,\ldots,Aug\ 2020\\ \end{aligned} (\#eq:n7) \end{equation}\]
The \(ECT_{t-1}\) is the lagged value of \(\hat\epsilon_{t}\) derived from the Eq. @ref(eq:n5). \(\sigma_{1}\) indicates how much of the imbalance in the short run will be corrected in the long run. The estimates of the coefficients in Eq. @ref(eq:n7) are robust under the assumption of lack of auto-correlation, heteroskedasticity, non-linearity and structural change in \(\hat\xi_{t}\). Same diagnostic tests applied to Eq. @ref(eq:n5) are also applied to the \(\hat\xi_{t}\).
As mentioned before, the analysis period covers the date between January 2008 and August 2020. I called this coverage as “Full sample.” On the other hand, the relationship between the \(BSI_{CBRT}\), the \(BSI_{ECB}\) and the \(BSI_{FED}\) can vary by time. Especially, the financial relationship between Turkey and International Monetary Fund (IMF) can affect the central banks actions and decisions, hence statements. On the other hand, the announcement of FED about the roadmap to end of stimulus program may influence this relationship. To show the effect of time on this relationship, I divide the sample into two sub-samples. The first sample covers the period from January 2008 and May 2013. The date of May 2013 is the important since it is the date when the financial relationship between Turkey and IMF officially ended and FED announced a roadmap to end stimulus program. The second sample covers the period after June 2013 until August 2020. Therefore, the cointegration and long/short-run analysis are performed to full and two sub-samples.
Empirical results
The descriptive statistics for BSI indicators of CBRT, ECB, and FED are reported in Table @ref(tab:desc). The \(BSI_{FED}\) implies positivity, \(BSI_{ECB}\) implies neutrality and \(BSI_{CBRT}\) implies negativity on average. The \(BSI_{CBRT}\) has the highest standard deviation which means it shows big fluctuations around the mean. Beside, \(CPI_{TR}\) has 3.99% and 25.24% values of minimum and maximum, respectively.
| Statistics | \(BSI_{CBRT,t}\) | \(BSI_{ECB,t}\) | \(BSI_{FED,t}\) | \(CPI_{TR}\) |
|---|---|---|---|---|
| N.obs. | 152.00 | 152.00 | 152.00 | 152.00 |
| Mean | 41.54 | 50.21 | 56.83 | 9.73 |
| Std.Dev. | 10.67 | 4.19 | 7.12 | 3.78 |
| Max. | 67.13 | 61.32 | 78.23 | 25.24 |
| Min. | 9.23 | 36.91 | 36.31 | 3.99 |
\(BSI_{CBRT}\), \(BSI_{ECB}\), and \(BSI_{FED}\) as well as \(CPI_{TR}\) are presented in Figure @ref(fig:gr1). It is seen that the \(BSI_{ECB}\) show stable at the neutrality level after the year 2013. By contrast, it can be claimed that \(BS_{FED}\) shows upward trend clearly after the election Donald Trump as president in 2016. Even though \(BS_{CBRT}\) shows fluctuations by time, it is seen that it moves around a stable mean which is lower than neutrality after 2012.
Balance sentiment indicators of CBRT, ECB, FED and Turkey Consumer Price Inflation Rate
Unit root results
Table @ref(tab:tab0) shows the results of five different unit root tests for \(BSI\) and \(CPI_{TR}\) variables. The first column represents the name of the series. The second column provides the null and alterative hypothesis of the tests in terms of integration order. While I(0) expression expresses stationarity in its level, I(1) expresses stationarity at first differences. The fourth column provides the information about model specification of the test. From Table @ref(tab:tab0), it may be claimed that \(BS_{CBRT,t}\) and \(BS_{ECB,t}\) are stationary at the I(0) order, however the \(BS_{FED,t}\) and \(CPI_{TR,t}\) series are stationary at I(1) order.
| Series | Null vs. Alternative Hypothesis | Test name | Test type | Lag length | Test Statistics | Critical Values |
|---|---|---|---|---|---|---|
| \(BSI_{CBRT,t}\) | I(1) vs I(0) | Augmented Dickey-Fuller Test | drift | 1 | -4.3097*** | -3.46; -2.88; -2.57 |
| \(BSI_{CBRT,t}\) | I(1) vs I(0) | Phillips-Perron | with intercept | 13 | -5.7387*** | -3.47; -2.88; -2.58 |
| \(BSI_{CBRT,t}\) | I(1) vs I(0) | Elliot, Rothenberg and Stock | with intercept | 12 | -2.7262*** | -2.58; -1.94; -1.62 |
| \(BSI_{CBRT,t}\) | I(0) vs I(1) | KPSS | mu | 13 | 0.3451 | 0.74; 0.46; 0.35 |
| \(BSI_{CBRT,t}\) | I(1) vs I(0) | Zivot-Andrews | intercept | 12 | -4.3899 | -5.34; -4.8; -4.58 |
| \(BSI_{CBRT,t}\) | I(2) vs I(1) | Augmented Dickey-Fuller Test | drift | 1 | -11.8947*** | -3.46; -2.88; -2.57 |
| \(BSI_{CBRT,t}\) | I(2) vs I(1) | Phillips-Perron | with intercept | 13 | -19.3705*** | -3.47; -2.88; -2.58 |
| \(BSI_{CBRT,t}\) | I(2) vs I(1) | Elliot, Rothenberg and Stock | with intercept | 11 | -4.7458*** | -2.58; -1.94; -1.62 |
| \(BSI_{CBRT,t}\) | I(1) vs I(2) | KPSS | mu | 13 | 0.0864 | 0.74; 0.46; 0.35 |
| \(BSI_{CBRT,t}\) | I(2) vs I(1) | Zivot-Andrews | intercept | 12 | -4.7285** | -5.34; -4.8; -4.58 |
| \(BSI_{ECB,t}\) | I(1) vs I(0) | Augmented Dickey-Fuller Test | drift | 1 | -4.8314*** | -3.46; -2.88; -2.57 |
| \(BSI_{ECB,t}\) | I(1) vs I(0) | Phillips-Perron | with intercept | 13 | -7.6225*** | -3.47; -2.88; -2.58 |
| \(BSI_{ECB,t}\) | I(1) vs I(0) | Elliot, Rothenberg and Stock | with intercept | 12 | -1.1014 | -2.58; -1.94; -1.62 |
| \(BSI_{ECB,t}\) | I(0) vs I(1) | KPSS | mu | 13 | 0.2447 | 0.74; 0.46; 0.35 |
| \(BSI_{ECB,t}\) | I(1) vs I(0) | Zivot-Andrews | intercept | 12 | -3.8197 | -5.34; -4.8; -4.58 |
| \(BSI_{ECB,t}\) | I(2) vs I(1) | Augmented Dickey-Fuller Test | drift | 1 | -12.9712*** | -3.46; -2.88; -2.57 |
| \(BSI_{ECB,t}\) | I(2) vs I(1) | Phillips-Perron | with intercept | 13 | -28.8055*** | -3.47; -2.88; -2.58 |
| \(BSI_{ECB,t}\) | I(2) vs I(1) | Elliot, Rothenberg and Stock | with intercept | 11 | -0.5501 | -2.58; -1.94; -1.62 |
| \(BSI_{ECB,t}\) | I(1) vs I(2) | KPSS | mu | 13 | 0.0766 | 0.74; 0.46; 0.35 |
| \(BSI_{ECB,t}\) | I(2) vs I(1) | Zivot-Andrews | intercept | 12 | -5.3801*** | -5.34; -4.8; -4.58 |
| \(BSI_{FED,t}\) | I(1) vs I(0) | Augmented Dickey-Fuller Test | drift | 1 | -3.9107*** | -3.46; -2.88; -2.57 |
| \(BSI_{FED,t}\) | I(1) vs I(0) | Phillips-Perron | with intercept | 13 | -5.5372*** | -3.47; -2.88; -2.58 |
| \(BSI_{FED,t}\) | I(1) vs I(0) | Elliot, Rothenberg and Stock | with intercept | 12 | -0.8976 | -2.58; -1.94; -1.62 |
| \(BSI_{FED,t}\) | I(0) vs I(1) | KPSS | mu | 13 | 0.4722** | 0.74; 0.46; 0.35 |
| \(BSI_{FED,t}\) | I(1) vs I(0) | Zivot-Andrews | intercept | 12 | -4.7347* | -5.34; -4.8; -4.58 |
| \(BSI_{FED,t}\) | I(2) vs I(1) | Augmented Dickey-Fuller Test | drift | 1 | -13.9488*** | -3.46; -2.88; -2.57 |
| \(BSI_{FED,t}\) | I(2) vs I(1) | Phillips-Perron | with intercept | 13 | -22.4172*** | -3.47; -2.88; -2.58 |
| \(BSI_{FED,t}\) | I(2) vs I(1) | Elliot, Rothenberg and Stock | with intercept | 11 | -0.8655 | -2.58; -1.94; -1.62 |
| \(BSI_{FED,t}\) | I(1) vs I(2) | KPSS | mu | 13 | 0.0579 | 0.74; 0.46; 0.35 |
| \(BSI_{FED,t}\) | I(2) vs I(1) | Zivot-Andrews | intercept | 12 | -4.4809 | -5.34; -4.8; -4.58 |
| \(CPI_{TR}\) | I(1) vs I(0) | Augmented Dickey-Fuller Test | drift | 1 | -2.6608* | -3.46; -2.88; -2.57 |
| \(CPI_{TR}\) | I(1) vs I(0) | Phillips-Perron | with intercept | 4 | -2.3773 | -3.47; -2.88; -2.58 |
| \(CPI_{TR}\) | I(1) vs I(0) | Elliot, Rothenberg and Stock | with intercept | 1 | -2.5334** | -2.58; -1.94; -1.62 |
| \(CPI_{TR}\) | I(0) vs I(1) | KPSS | mu | 4 | 1.1963*** | 0.74; 0.46; 0.35 |
| \(CPI_{TR}\) | I(1) vs I(0) | Zivot-Andrews | intercept | 1 | -4.1409 | -5.34; -4.8; -4.58 |
| \(CPI_{TR}\) | I(2) vs I(1) | Augmented Dickey-Fuller Test | drift | 1 | -8.8246*** | -3.46; -2.88; -2.57 |
| \(CPI_{TR}\) | I(2) vs I(1) | Phillips-Perron | with intercept | 4 | -9.205*** | -3.47; -2.88; -2.58 |
| \(CPI_{TR}\) | I(2) vs I(1) | Elliot, Rothenberg and Stock | with intercept | 1 | -6.8258*** | -2.58; -1.94; -1.62 |
| \(CPI_{TR}\) | I(1) vs I(2) | KPSS | mu | 4 | 0.0368 | 0.74; 0.46; 0.35 |
| \(CPI_{TR}\) | I(2) vs I(1) | Zivot-Andrews | intercept | 12 | -6.3853*** | -5.34; -4.8; -4.58 |
Note: 1 Critical values for 1, 5 and 10 pct. are given respectively. p<0.1; p<0.05; p<0.01
Cointegration results
The appropriate lag length is determined for the ARDL specification by the dicrimation of the sample coverage, then the UECM model is estimated by different samples using these appropriate lag lengths minus one due to the differenced variables are used. Three different samples are used, i.e. Full sample, Jan 2008-May 2013 and June 2013-Aug 2020. Table @ref(tab:lagselection) presents the results of the lag length determination by sample coverage.
| Sample | \(BSI_{CBRT,t}\) | \(CPI_TR_{t}\) | \(BSI_{ECB,t}\) | \(BSI_{FED,t}\) | \(Adj.R^{2}\) | \(\chi^{2}\) | AIC |
|---|---|---|---|---|---|---|---|
| Full Sample | 1 | 1 | 1 | 1 | 0.494 | 3.59** | 620.29 |
| Full Sample | 1 | 2 | 1 | 1 | 0.494 | 2.43 | 617.36 |
| Full Sample | 1 | 2 | 2 | 1 | 0.503 | 2.7 | 615.73 |
| Full Sample | 1 | 2 | 2 | 2 | 0.502 | 1.51 | 616.9 |
| Full Sample | 1 | 2 | 3 | 1 | 0.502 | 2.43 | 613.82 |
| Full Sample | 1 | 3 | 2 | 1 | 0.502 | 3.35** | 613.94 |
| Full Sample | 2 | 2 | 2 | 2 | 0.507 | 0.33 | 616.51 |
| Full Sample | 2 | 2 | 2 | 3 | 0.531 | 0.49 | 606.77* |
| Full Sample | 2 | 2 | 3 | 2 | 0.506 | 1.02 | 614.54 |
| Full Sample | 2 | 2 | 3 | 3 | 0.533* | 1.09 | 606.82 |
| Full Sample | 2 | 3 | 2 | 2 | 0.505 | 0 | 614.8 |
| Full Sample | 2 | 3 | 2 | 3 | 0.53 | 0.07 | 607.85 |
| Full Sample | 2 | 3 | 3 | 3 | 0.532 | 0.47 | 608.21 |
| Full Sample | 3 | 3 | 3 | 3 | 0.529 | 1.04 | 609.94 |
| 2008 Jan-2013 May | 1 | 0 | 1 | 0 | 0.573 | 0.65 | 632.5 |
| 2008 Jan-2013 May | 1 | 0 | 1 | 1 | 0.573 | 1.94 | 621.12 |
| 2008 Jan-2013 May | 1 | 0 | 2 | 0 | 0.58 | 0.91 | 627.11 |
| 2008 Jan-2013 May | 1 | 0 | 2 | 1 | 0.578 | 2.31 | 617.29 |
| 2008 Jan-2013 May | 1 | 1 | 1 | 1 | 0.567 | 3.59** | 620.29 |
| 2008 Jan-2013 May | 2 | 2 | 2 | 2 | 0.583 | 0.33 | 616.51 |
| 2008 Jan-2013 May | 2 | 3 | 2 | 2 | 0.585 | 0 | 614.8 |
| 2008 Jan-2013 May | 2 | 3 | 2 | 3 | 0.622 | 0.07 | 607.85* |
| 2008 Jan-2013 May | 2 | 3 | 3 | 2 | 0.576 | 0.23 | 616 |
| 2008 Jan-2013 May | 2 | 3 | 3 | 3 | 0.615 | 0.47 | 608.21 |
| 2008 Jan-2013 May | 3 | 3 | 3 | 3 | 0.626* | 1.04 | 609.94 |
| 2013 Jun-2020 Aug | 1 | 1 | 1 | 1 | 0.436 | 3.59** | 620.29 |
| 2013 Jun-2020 Aug | 1 | 2 | 0 | 1 | 0.438 | 2.25 | 615.53 |
| 2013 Jun-2020 Aug | 1 | 2 | 0 | 2 | 0.449 | 0.98 | 616.26 |
| 2013 Jun-2020 Aug | 1 | 2 | 1 | 1 | 0.443 | 2.43 | 617.36 |
| 2013 Jun-2020 Aug | 1 | 2 | 1 | 2 | 0.453* | 1.09 | 618.06 |
| 2013 Jun-2020 Aug | 1 | 2 | 1 | 3 | 0.452 | 0.88 | 608.8 |
| 2013 Jun-2020 Aug | 1 | 3 | 1 | 2 | 0.452 | 1.59 | 616.39 |
| 2013 Jun-2020 Aug | 2 | 2 | 0 | 2 | 0.443 | 2.34 | 615.87 |
| 2013 Jun-2020 Aug | 2 | 2 | 1 | 2 | 0.449 | 1.86 | 617.82 |
| 2013 Jun-2020 Aug | 2 | 2 | 1 | 3 | 0.448 | 1.5 | 608.27* |
| 2013 Jun-2020 Aug | 2 | 2 | 2 | 2 | 0.442 | 0.33 | 616.51 |
| 2013 Jun-2020 Aug | 2 | 3 | 1 | 2 | 0.448 | 0.51 | 616.11 |
| 2013 Jun-2020 Aug | 2 | 3 | 2 | 2 | 0.44 | 0 | 614.8 |
| 2013 Jun-2020 Aug | 3 | 3 | 3 | 3 | 0.439 | 1.04 | 609.94 |
Note: 1 * shows highest adj. \(R^{2}\) in relevant column, * represents significant auto-correlation problem in BG test column, and * shows minimum AIC value in the relevant column
Table @ref(tab:lagselection) presents Adj.\(R^{2}\), BG test value for auto-correlation, and AIC values for each ARDL estimation with the orders for \(BSI_{CBRT}\), \(BSI_{ECB}\), \(BSI_{FED}\) and \(CPI_{TR}\). Table implies that ARDL(2,2,3,3) model is found appropriate for the full sample, ARDL(3,3,3,3) model is found appropriate for the sample Jan 2008-May 2013, and ARDL(1,2,1,2) model is found appropriate for the sample June 2013-August 2020 considering the Adj. \(R^2\) criterion.
The Bounds tests are performed after proper lag lengths are determined as respectively, UECM(1,1,2,2), UECM(2,2,2,2) and UECM(0,1,0,1) over the full sample, the sample Jan 2008-May 2013,and the sample June 2013-August 2020 using by Eq. @ref(eq:n2) to investigate the cointegration relationship among the BSI indicators and \(CPI_{TR}\). The F-test and t-test results are presented in Table @ref(tab:bounds). k is the number of independent variables in Eq. @ref(eq:n2). The critical values are taken from Table CI (V) in Pesaran et al. (2001: 300).
| Model | Test type | Test statistics | Lower bound 5 pct. | Upper bound 5 pct. |
|---|---|---|---|---|
| Full Sample | F-test | 4.708 | 3.230 | 4.350 |
| Full Sample | t-test | -4.160 | -2.860 | -3.780 |
| 2008 Jan-2013 May | F-test | 1.239 | 3.230 | 4.350 |
| 2008 Jan-2013 May | t-test | -1.609 | -2.860 | -3.780 |
| 2013 Jun-2020 Aug | F-test | 8.282 | 3.078 | 4.155 |
| 2013 Jun-2020 Aug | t-test | -5.382 | -2.860 | -3.780 |
Note: 1 k = 3, Critical bound’s values (three regressors, and no trends in the model), Pesaran et al. (2001), p. T.2, Table Cl. iii: Case III.
The bounds \(F\)-statistics are found statistically significant for the full and the sample 2013 Jun-2020 Aug since the test statistics are higher than the upper critical values at 5% level. Moreover, the bounds t-tests are also found significant for those samples. This implies that \(BS_{CBRT,t}\) has co-movement with \(BS_{ECB,t}\), \(BS_{FED,t}\) and \(CPI_{TR}\) in the long-run for the full sample and after 2013 June.
On the other hand, it is an interesting finding that the long-run relationship does not exist before 2013 June. It can be argued that CBRT sentiments has become closer to ECB and FED statements after FED declared a road map for ending financial stimulus which started after 2009 global financial crisis, and the end of the relationship between IMF and Turkey, which these two events are happened at June 2013.
Long and short run properties
The long-run and short-run properties of the relationships can be examined since a cointegration relationship is determined between the series. The selected orders for ARDL models are already given in Table @ref(tab:lagselection) for the three samples. The estimation results of ARDL models (Eq. @ref(eq:n5)) are presented in Table @ref(tab:tab1) for two samples. Full sample and the sample after 2013 June are used to estimation of ARDL since the cointegration relationships are valid for these two samples.
| Model | Test type | Test statistics | Lower bound 5 pct. | Upper bound 5 pct. |
|---|---|---|---|---|
| Full Sample | F-test | 4.708 | 3.230 | 4.350 |
| Full Sample | t-test | -4.160 | -2.860 | -3.780 |
| 2008 Jan-2013 May | F-test | 1.239 | 3.230 | 4.350 |
| 2008 Jan-2013 May | t-test | -1.609 | -2.860 | -3.780 |
| 2013 Jun-2020 Aug | F-test | 8.282 | 3.078 | 4.155 |
| 2013 Jun-2020 Aug | t-test | -5.382 | -2.860 | -3.780 |
The diagnostics are reported at the lower part of Table @ref(tab:tab1). It can be claimed that the estimated models are robust in terms of the auto-correlation, heteroscedasticity, non-linearity and structural breaks. The estimated coefficients in Table @ref(tab:tab1) are used to estimate long-run coefficients in Eq. @ref(eq:n1). Long-run coefficients are presented in Table @ref(tab:lrelas). Table shows that \(CPI_{TR}\), \(BSI_{ECB}\) and \(BSI_{FED}\) have significant effect on \(BSI_{CBRT}\) in the long-run. However, this finding is valid for the estimation for the sample 2013 June- 2020 Aug, not for the full sample over which the bounds tests shows the cointegration relationship is significant at %5 level.
The coefficients of \(BSI_{ECB}\) and \(BSI_{FED}\) variables are positive and statistically significant at 5% level for the sample June 2013 and Aug 2020. The size of \(BSI_{ECB}\) and \(BSI_{FED}\) coefficients are equal and over unit size. This finding suggests that \(BSI_{CBRT}\) is affected by \(BSI_{ECB}\) and \(BSI_{FED}\) with same direction. COnsidering the geographical proximity of Euro area countires and heavity of European countries in foreign trade of Turkey, the closeness of sentiments can be expected. Beside, the closeness of sentiments between CBRT and FED are also expected since FED has great importance in the world moetary balance and also Turkish external debt composition is mainly based on US Dollars.
Additionally, the \(CPI_ {TR}\)’s coefficient is found statistically significant at 10% level and negative. This finding indicates that the \(CPI_{TR}\) and \(BSI_{CBRT}\) have an inverse relationship in the long-run. This results are also expected because CBRT applies inflation targeting and if annual CPI rise, the sentiments of the monetary policy committee turns out to be negative and vice versa.
| Sample | Full Sample | 2008 Jan-2013 May | 2013 Jun-2020 Aug |
|---|---|---|---|
| Constant | 68.252 (45.623) | 176.103 (122.607) | -78.509* (42.359) |
| CPI_TR | -0.248 (0.729) | -3.196 (3.963) | -1.24* (0.64) |
| ECB | -0.671 (0.722) | -1.89 (1.654) | 1.549** (0.634) |
| FED | 0.162 (0.436) | -0.256 (1.582) | 0.999** (0.424) |
UECM model given in Eq. @ref(eq:n7) are estimated to examine how \(BSI_{ECB}\), \(BSI_{FED}\) and \(CPI_{TR}\) affect the \(BSI_{CBRT}\) in the short-run for three sample coverage. Estimation of UECM for the sample before 2013 June is a valid option for \(BSI\) variables if there is no cointegration relationship among them. The estimation results are given in Table @ref(tab:tab1). The diagnostics are also reported at the lower part of Table @ref(tab:tab1). It can be claimed that the estimated models are robust in terms of the auto-correlation, heteroscedasticity, non-linearity and structural breaks.
| Model | Test type | Test statistics | Lower bound 5 pct. | Upper bound 5 pct. |
|---|---|---|---|---|
| Full Sample | F-test | 4.708 | 3.230 | 4.350 |
| Full Sample | t-test | -4.160 | -2.860 | -3.780 |
| 2008 Jan-2013 May | F-test | 1.239 | 3.230 | 4.350 |
| 2008 Jan-2013 May | t-test | -1.609 | -2.860 | -3.780 |
| 2013 Jun-2020 Aug | F-test | 8.282 | 3.078 | 4.155 |
| 2013 Jun-2020 Aug | t-test | -5.382 | -2.860 | -3.780 |
The coefficients of the error correction variable (\(ECT_{t-1}\)) are negative and between 0 and -1. The fact that the coefficient is negative and less than one means that the system is stabilized by fluctuating, and this fluctuation will gradually decrease each time and return to balance in the long run. As a result, it is negative and statistically significant as expected.
The effect of \(BSI_{ECB}\) on \(BSI_{CBRT}\) is positive and noticeably for the full sample and the sample before 2013 June, however the effect has become weak after 2013 June. The effect of \(BSI_{FED}\) on \(BSI_{CBRT}\) has mixed properties. For the full sample, it has negative effect at \(t\), but it has positive effect at \(t-2\). So, the whole effect is become neutral. For the sample 2008 Jan - 2013 June, it has positive effect on \(BSI_{CBRT}\). On the other hand, it has sronge negative effect on \(BSI_{CBRT}\) after the period 2013 June. Beside, the effect of \(CPI_{TR}\) on \(BSI_{CBRT}\) is positive and noticeably for the full sample and the sample before 2013 June, however the effect has become weak after 2013 June.
To sum up, it may be concluded that CBRT sentiments are positively affected by the ECB and FED sentiments in the long run, and only negatively affected by the FED sentiments in the short run The common movement of statements in the long run is seneible because monetary policies are also related in the long run. On the other hand, the management of capital movements and tightening measures for price stability taken by CBRT may cause an inverse relationship with the FED in the short-run.
Conclusion
The monetary policy statements are important communication tools for central banks. However, there are very few studies in the literature regarding the examination of the relationships among the central banks’ statements. Therefore, the main concern of this paper is to investigate how the CBRT sentiments are affected by the ECB and the FED sentiments. The analysis covers the period between January 2008-August 2020. Two sub-samples, i.e. before and after 2013 June, are analysed in addition to the full sample. In this paper, the monthly sentiment indicators are calculated from the statements of CBRT, ECB, and FED. These indicators show the balance between the positivity and the negativity scores of the statements. Several unit root test and the Bounds test are performed to investigate the individual and group time series characteristics. Then, ARDL and UECM models are estimated to examine the short and long-run relationship between the sentiments of CBRT, ECB, and FED statements.
The bounds test results suggest that the balance sentiment indicators of CBRT, ECB, and FED have statistically significant cointegration relationship for the full sample and after June 2013. This implies that sentiments of CBRT statements are in line with ECB and FED statements’ after June 2013 when the relationship between IMF and Turkey is officially ended. Long-run and short-run dynamics shows that ECB statements have strong effect on the CBRT statements in the long run but weak in the short run. On the other hand, FED statements have strong effect on the CBRT statements in the long and short-run. The sign of the effect is positive in the long-run however it is negative in the short-run.
Several implications can be derived from the findings. The estimated long-run elasticity of ECB indicates that the CBRT statements are in line with the ECB statements in the long-run. This behavior can be explained by the geographic proximity and high-level foreign trade partnership. The long-run relationship between CBRT and ECB statements is positive after 2013 June when this date coincides with two events. June 2013 is the date when CBRT made a change in monetary policy due to the FED outlined an end to its stimulus to the USA economy, and announced an official end the relationship between IMF and Turkey.
By contrast, the FED statements inversely affect the CBRT statements in the short-run. The negative relationship between them can be explained by the increase in the needs for the short-term funds in Turkey due to the FED continuously hikes the Federal fund rate after the stimulus period.
Conflict of Interest The author declares that he has no conflict of interest.
Statement of Human and Animal Rights This article does not contain any studies with human participants or animals performed by the author.
References
Dr., Hacettepe Üniversitesi, Ankara, Türkiye, ORCID: 0000-0002-4232-9985, alpaykocak@hacettepe.edu.tr↩︎