Monetary news shocks and credit spread predictability : evidence of the European Central Bank - Draghi era

(1)

Monetary News Shocks and Credit Spread Predictability:

Evidence of the European Central Bank – Draghi era

Master ’s Thesis Max de Lorijn, 10445838

-

Amsterdam School of Economics

University of Amsterdam

-

Master’s Program:

MSc Economics

International Economics and Globalisation

-

Supervisor: Dr. P. Foldvari

Second reader: Dr. D.J.M. Veestraeten

-

(2)

2

Statement of Originality

This document is written by Student Max de Lorijn who declares to take full responsibility for the contents of this document.

I declare that the text and the work presented in this document are original and that no sources other than those mentioned in the text and its references have been used in creating it.

The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

(3)

3

Abstract

This paper aims to shed light on the impact of ECB communication on Eurozone sovereign bond spreads through a communication index, which is constructed so as to capture the variety of ECB communication. In addition, this paper evaluates whether the impact differs for triple-A credit spreads and BBB rated credit spreads. This is done by estimating a VEC model for the period November 2011 – April 2015. The results of this paper indicate that ECB communication tightens the Eurozone sovereign bond spreads rated AAA and BBB in the long-run. No evidence is found that supports a significant impact of ECB communication in the short-run for either of the two asset classes.

Keywords: communication, credit spreads, European Central Bank, monetary policy JEL-codes: E44, E52, E58, G12 , G15

(4)

4

Central bank transparency has been a long lasting topic for debate. Montagu Norman, a former Governor of the Bank of England, believed in the sharp sentence “Never explain, never excuse” (Boyle, 1967). His view encompasses the consensus of many monetary economists of his time. They believed that more central bank communication results in inefficient policy. In recent decades, however, disclosure of information has been promoted by many governors, such as Mario Draghi from the European Central Bank (ECB) and Ben S. Bernanke from the United States (U.S.) Federal Reserve Bank (FED). They argue that more communication enhances the predictability of central banks, and hence increases the efficiency of their policy. This was especially the case during the 2007-2009 sovereign debt crisis, in which central bankers were forced to implement unconventional monetary policies (e.g. communication), and financial markets could only count on verbal references as the traditional monetary tool was not effective in stabilising the interbank market and reducing the cost of credit. In all, the thesis of news-driven monetary policy has become a prevailing topic in the literature of monetary policy.

Central bank communication could serve as a signalling channel to economic agents. The ECB has two main windows for the latter, namely its press conferences and its monthly economic bulletins, in which they could indicate their keenness to act when the markets experience economic downturns. An example of this is the speech given by Mario Draghi in the summer of 2012 when he announced the OMT program1 with the well-known statement “we will do whatever it takes”. Moreover, Ben Zeev et al. (2016) argue that economic agents react upon announcements made by central banks, which could have significant effects on the real economy. This could for instance justify why communication intensity could increase to calm down the markets in times of economic downturns. These effects induced by central bank announcements are captured by the phenomenon of “Monetary news shocks”.

Several contributions have been made to identify real economic effects of monetary news shocks by analysing the impact on central bank rate decisions. Moreover, Gerlach (2007) and Ben Zeev et al. (2016) all find evidence that monetary news shocks are involved with macroeconomic fundamentals, which is in line with the thinking of many other monetary economists. For instance, Rosa and Verga (2007) state that the ECB’s

1

The Outright Monetary Transactions (OMT) program was introduced to support the secondary sovereign bond market by trying to remove the unwarranted and self-reinforcing fears of a Eurozone break up.

(6)

6

communication contributes to the forecast of short-term interest rate paths. It might therefore be plausible that there is a link between central bank communication and credit spreads, as Cochrane and Piazzesi (2005) and Krishnan et al. (2010) all write that the predictability of bond spreads can be improved upon by including macroeconomic fundamentals. To the best of my knowledge, a direct link between monetary news shocks and credit spreads has yet to be made. This will be a significant contribution to the strand of literature assessing determinants of credit spread alternations. Furthermore, the ability to forecast credit spreads would be valuable for both corporate finance agents and fixed income investors as they can obtain significant trading gains. For example, hedge funds regularly take leveraged positions in bonds while hedging alterations in interest rates by taking a short position in treasuries. This, however, makes their portfolios more exposed to bond volatility than changes in yields. Thus, if communication can further explain credit spreads movements, major losses accompanied with this exposure will be mitigated. Therefore, this paper studies the relationship between credit spreads and central bank communication and evaluates the following research question: “Does more intensive central

bank communication aid credit spread predictability?” For the purpose of this analysis, this

study will focus on the credit spreads of European (EU) government bonds over commensurate risk-free benchmarks.

Although there are some applied econometric papers that asses to what extent ECB communication provides significant information for economic agents, Rosa and Verga (2007) argue that the existing literature disregards the rules of hermeneutic theory. Hence, neglecting the changing meaning of words according to the situation. To control for this, a continuous variable will be employed that better reflects the variety of ECB communication by allowing the hawkish-ness indicator to fluctuate more over time. Moreover, previous papers as regards ECB communication either used ordinary least squares (OLS), ordered-probit, or generalized autoregressive conditional heteroscedasticity (GARCH) models as an econometric technique to analyse central bank communication. Therefore, the contribution of this paper is threefold as a vector error correction model (VECM) will be estimated so as to evaluate the effect of central bank communication on credit spreads.

In order to answer the research question, monthly economic and financial data will be used. Furthermore, this paper employs 2 dependent variables, namely Credit spread BBB, 10-year and Credit spread AAA, 10-year. This is done, since AAA credit spreads have

(7)

7

lower forecast errors compared to BBB credit spreads. To that end, including different asset classes enables one to distinguish between their possible improvement in predictability after the start of unconventional monetary policy. The main explanatory variable is the communication index. The latter will be accompanied with several macroeconomic fundamentals that embody the interest rate environment, investor risk preferences and the stage of the economic cycle.

The remainder of this paper is organized as follows: Section 2 functions as literature review. Herein several essential papers regarding the thesis will be addressed. With this knowledge, section 3 elaborates on the data sources and variables used in this paper. Particular weight is put on the construction of the communication index. In section 4, the hypotheses are presented. After this, the empirical analysis is provided so as to answer the hypotheses. Section 5 assesses the results of the methodology. Section 6 will provide a discussion of the results. Finally, Section 7 presents the main conclusions and directions for forthcoming research.

(8)

8

2. Literature review

In order to answer the research question, several essential articles will be presented in the literature review. First, research on monetary news shocks will be elucidated. Second, literature on credit spread predictability will be addressed. The latter is accompanied with a prognosis on the link between credit spreads and monetary news shocks.

2.1. Monetary news shocks

Central bankers long lived by the maxim that a certain unpredictability linked to their activities is preferable. They argued that making monetary policy should be left solely to the initiates and that communicating, hence letting economic agents get involved, would degrade the effectiveness of policy (Bernanke, 2004). In line with the latter, Morris and Shin (2002) argue that more public information does not necessarily lead to welfare gains, but that it often causes more harm than good. In contrast to this view, nowadays many central bankers follow the opposite of this propensity. This can be illustrated by the 2007-2009 sovereign debt crisis, in which the ECB was confronted with the complex task of establishing monetary transmission to aid the economy. Nevertheless, the traditional monetary tool2 of the ECB was not able to establish a stable interbank market (Szczerbowicz, 2015). Accordingly, the ECB introduced various unconventional monetary policies to accomplish its targets (e.g. stabilizing the malfunctioning interbank market). This paper attempts to examine one of these unconventional monetary policies. Namely, evaluating the effectiveness of more ECB communication. In the following, the most essential papers will be discussed as regards this.

Blattner et al. (2008) highlight the main elements related to the central bank’s predictability and the involvement of transparency as one of its predominant determinants. They emphasise that besides the nature of the information, the channel of communication is important for two reasons. First, if the channel does not transmit the information properly, economic agents perceive a continuous mismatch between realised policy decisions and public announcements by central bankers, and in turn, give market participants no reason to believe that central bankers will maintain the announced strategy in the future. Second, they write that the channel enabling a central banker to communicate clearly and unambiguously

(9)

9

will result in more predictable monetary policy decisions, which allow financial market participants to be able to anticipate accurately the broad path of monetary policy. The ECB for instance, communicates through two predominant channels: The President's introductory statement at the monthly press conference, which discloses the monetary policy decisions taken simultaneously by the Governing Council on the monetary stance in real time, and the Economic monthly Bulletins explaining in more detail the basis for the Governing Council’s policy decisions and a comprehensive analysis of the staff macroeconomic and financial indicators (Papademos, 2008). Gerlach (2007) examines the ECB’s communication using the monthly economic bulletins. He finds evidence that these provide valuable information for market participants with regards to acquiring better understanding of the conduct of ECB monetary policy. Furthermore, Rosa and Verga (2007) study the communication of the ECB by confining their approach to the introductory statement at the monthly press conference. After controlling for the monetary policy shock, they find evidence that market expectations are responsive to the unexpected constituent of the information released by the ECB. Moreover, they argue that announcements of the ECB provide complementary information concerning macroeconomic variables such as short-term interest rate paths. Ben Zeev et al. (2016) findings are somewhat in line with those of Gerlach (2007) and Rosa and Verga (2007). They restrict their analyses to the U.S. economy and find evidence that monetary news shocks are involved with real economic effects. They write that real GDP declines in an incessant hump-shaped manner after a positive monetary shock and that this is accompanied by an expeditious increase in the nominal interest rate and a fall in inflation.

The empirical studies above suggest that central bank communication (i.e. monetary news shocks) can amend expectations and with it market outcomes, and hence, have impact on the economy. However, De Haan and Jansen (2010), who aim their attention at the monthly bulletins, draw different conclusions than Rosa and Verga (2007), Gerlach (2007) and Ben Zeev et al. (2016). They argue that despite the statistical significance of the ECB’s communication, the economic impact of the latter is merely small. Moreover, they write that the financial markets are only responsive regarding the ECB’s communication when the latter is in line with the perception of upcoming alterations in the ECB’s monetary policy stance.

(10)

10

2.2. Credit spread predictability

Fama and Bliss (1987), who examine the U.S. bond market, suggest that the current yield curve is significantly involved with future yield spreads. They write that 1-year forward rates contain predominant information with regard to riskless interest rates 2- to 4-years ahead. Furthermore, Campbell and Shiller (1991), who study post-war U.S. term structure data, argue that for almost any combination of maturities between one month and ten years the current yield curve is able to predict subsequent deviations in the short-term. However, they write that it fails to correctly forecast long-term bond spreads. Contrary to the latter, Diebold and Li (2006), who use variations of the Nelson-Siegel framework3 so as to be able to capture the entire yield curve, find evidence that bond spread forecasts seem to be significantly involved with both the short and long horizons.

More recently, contributions have investigated whether control variables can be used in combination with yield term structure information in order to enhance credit spread predictability. For instance, Cochrane and Piazzesi (2005) state that the model used by Fama and Bliss (1987) can be significantly improved by adding information about several yields obtained from the riskless yield curve. Krishnan et al. (2010) concur with Cochrane and Piazzesi (2005). They state that credit spread forecasts can be improved upon by adding auxiliary variables that include macroeconomic indicators. Moreover, they suggest that yield spreads linked to different credit ratings might experience a different impact as they differ in their likelihood of default and risk premia. Figure 1 on the next page, illustrates Eurozone sovereign 10-year bond yield spreads over German benchmark from which one can infer that bonds with different credit classes differ in their spreads.

As pointed out above, the literature suggests that there are two main factors that predict credit spreads: information about the riskless yield curve and control variables that include macroeconomic indicators. The most commonly used framework to identify these determinants are the so-called “structural models of default risk”, which will also be employed in this paper.4 This framework enables one to apply the standard option pricing theory, equivalent to the Black-Scholes formula, so as to price a default-risky bond (Collin-Dufresne et al., 2001). Boss and Scheicher (2002), employ this framework to evaluate changes in corporate credit spreads by using explanatory variables that gauge three different

3

For a review on the Nelson-Siegel framework, refer to Diebold and Rudebusch (2013).

(11)

11

types of risk: interest rate risk, liquidity risk and credit risk. Although the first two apply to all bond spreads, sovereign bonds are assumed to be free of credit risk (Boss and Scheicher 2002). Hence, this risk type is omitted in the empirical framework of this paper.

As mentioned in section 2.1, Rosa and Verga (2007) argue that monetary news shocks are involved with macroeconomic variables such as short-term interest rate paths. Furthermore, Ben Zeev et al. (2016) write that economic agents’ expectations about the stance of monetary policy are significantly influenced by news about future monetary policy. These findings in conjunction with the findings of Cochrane and Piazzesi (2005) and Krishnan et al. (2010), who write that adding macroeconomic indicators enhance credit spread predictability, suggests that it is plausible that the predictability of credit spreads is involved with central bank communication. Mink and De Haan (2013) provide some support for this prognosis, as they suggest that Greek bailout news resulted in anomalous returns on sovereign bonds in the Eurozone. Nonetheless, a direct link between credit spreads and communication has yet to be proven.

Figure 1: Eurozone Sovereign 10-Year Bond Yield Spreads over German Benchmark

Source: Bloomberg (2014).

Note: credit ratings for sovereign bonds Spain, Italy, Belgium, France and Austria are respectively BBB+, BBB-, AA, AA and AA+.

(12)

12

3. Data description

This section provides a description of the data employed. First, the main elements and the construction of the communication index will be provided. The latter will embody hawkish-ness indicators in order to control for the fluctuating tune of ECB communication as described by Rosa and Verga (2007). Second, a description of the dependent and explanatory variables is presented.

3.1. The communication index

Private markets are regularly faced with large amounts of new information. Therefore, it is plausible that corporate finance agents and fixed income investors make use of filtering instruments to digest the incoming data. Concurrently, central banks follow a strategy of employing key sentences and keywords in their communication.

Since the ECB has been established, it has focused particular attention on its communication strategy. The ECB uses several communication channels. The President's introductory statement at the monthly press conference, which is the main communication window of the ECB, outlines the monetary policy decisions taken simultaneously by the Governing Council on the monetary stance in real time. Each sentence and word in the introductory statement is chosen precisely and guided by the “single voice” principle, reflecting the position of the Governing Council as a whole. Moreover, organizing this press conference enables the ECB to reach a wider audience than by promulgating a press release. A second important instrument is the ECB’s Monthly Bulletin, which is issued one week after the introductory statement, offering a more detailed explanation about the monetary policy decisions and staff macroeconomic projections for the Eurozone. Finally, the interactions between the Governing Council members and several different audiences during the inter-meeting term are also considered to be an important window on the ECB’s communication strategy. This channel provides additional opportunities for the Governing Council members, as it enables them to disclose the Councils views in their own languages, overcoming any difficulties as regards linguistic and national barriers of communication. This reduces the possible noise of the ECB’s communication strategy, which has become a more significant factor as the Eurozone grows. For the empirical framework in this paper, the President's monthly introductory statements will be employed starting from Draghi’s press conference

(13)

13

in November 2011 until March 2015, resulting in 41 data points. The statements will be retrieved from the ECB’s website which contains an archive for every single press conference given by the ECB’s President. Furthermore, the President's introductory statement is chosen due to the following underlying motivations. This paper joins a strand of the research agenda that assesses the significance of ECB communication windows. The introductory statement appears to be the natural candidate, as it is considered to be the most important communication channel of the ECB as stated by Papademos (2008). Besides the fact that the introductory statement is straightforward and systematic in terms of its recurrence, format and employment of keywords, there is also a more apt reason for one to use the monthly press conference. Namely, the introductory statement of the monthly press conference releases equivalent information regarding the ECB monetary policy stance as the editorial section of the monthly bulletin. Nonetheless, the latter is disclosed to the public after a time lag in comparison to the introductory statement.

Several contributions have been made to investigating the response of financial markets to introductory statements. The majority of these papers employ a numerical index in order to make ECB communication suitable for statistical computation. These communication indexes are then employed to complement macroeconomic fundamentals typically included in the Taylor-rule, so as to estimate the ECB reaction function. For instance, Rosa and Verga (2007) use an index in which each statement can take five values depending on the language employed, namely +2 (very hawkish), +1 (hawkish), 0 (neutral), -1 (dovish) or -2 (very dovish). De Haan and Jansen (20-10), who confine their approach to the ECB Monthly Bulletin, rely on an analogous strategy. They distinguish between comments made on the main refinancing rate, the outlook for Eurozone economic growth and the outlook for Eurozone inflation. Per topic, they assign to each comment a value of -1 (dovish), 0 (neutral) or +1 (hawkish). The approach of Gerlach (2007) is somewhat in line with that of De Haan and Jansen (2010), as he employs the Monthly Bulletin to establish indicators that embody the Governing Council’s view on developments in real economic activity, M3 growth and inflation pressures. He classifies each comment on a ternary scale: -2 (very dovish), -1 (dovish), 0 (neutral), +1 (hawkish) and +2 (very hawkish). These methodologies seem to be justified, since in the course of time the ECB has used reoccurring Keywords and Key phrases

(14)

14

when communicating. Examples of these standardized expressions and their corresponding5 hawkish-ness values are illustrated in Table 1.

Table 1: Examples of ECB Keywords/Key phrases and their assigned hawkish-ness values

ECB’s significant sentences Index

Imperative that upward pressure to be contained – Risks [to price stability] are upward (upside) – Vigilant (vigilance) [with regard to upside risks to price stability]– Close monitored (or: continuous close attention) [upside risks] ₊₂ Both confident and vigilant [upside risks] – Upward pressure remains contained – A number of (or: Some) upside risks need to be carefully monitored – Alert to emerging of upward risks – Vigilance with regard to the materialisation of upside risks

+1

Appropriate – Favourable – Compatible – Consistent – In line – Balanced –Absence of significant pressures either upwards or downwards – The downside risks have

disappeared ₀

Favourable, but there are some [downside] risks – Appropriate but remain downside risks – Downside risks are not vanished – Some of the downward risks had

materialised _-1

Consistent, but carefully monitor all [downside] risks to economic growth – Downside risks are still relevant – Economic slowdown is still cause for concern –

[Strong] downside risks for economic activity _-2

Source: Rosa and Verga (2007)

Note: The table shown is a modified version of the one constructed by Rosa and Verga (2007). For the full version, please refer to their paper which can be found in the bibliography.

Nevertheless, coding introductory statements on a ternary scale with discrete values ranging from -2 to +2 is accompanied with the risk that one does not fully capture the effect of ECB, as this coding does not fully control for the variety in ECB communication and therefore neglects parts of the information disclosed. This can be illustrated by analysing the passage from the introductory statement below:

5

Note: The assigned hawkish-ness values are based on the authors opinion. The classifications shown in the table are obtained from the paper of Rosa and Verga (2007).

(15)

15

The economic outlook continues to be subject to particularly high uncertainty and intensified downside risks. Some of these risks have been materialising, which makes a significant downward revision to forecasts and projections for average real GDP growth in 2012 very likely (ECB, Press Conference, November 2011).

If one would apply the method of the existing literature, the above passage will be coded with a value of -2 due to the use of intensified downside risks and significant downward revision. However, one would neglect parts of the information released as the keyword downward/downside is mentioned twice. The latter as the repetition of a certain keyword is done with a purpose by the Governing Council so as to emphasize the importance of that notion. Also, the frequent use of an equivalent sentence will most likely lead to information being more easily transmitted to the public. The communication index employed in this paper will therefore assign a hawkish-ness value to each sentence in a statement so as to allow for more variation. The use of intensified downside risks and significant downward revision will therefore lead to a value of -4 (-2x2) for the above passage. Besides reflecting the level of hawkish-ness or dovish-ness, the index in this paper will therefore also control for the tendency that information is transmitted to the public.

As mentioned above, this paper will only restrict its analysis to the President’s introductory statement. The format of the introductory statement has remained roughly the same since January 1999. It begins with an introductory paragraph that examines the ECB’s perception of price risk to price stability, economic outlook and monetary policy stance. Then follows a detailed economic analysis, which includes a part for price developments and the real sector in the Eurozone. This is followed up by a paragraph that provides a monetary analysis. Subsequently, a summing-up paragraph is given as regards the price developments, the real sector and the monetary policy stance. Then a paragraph is dedicated to draw attention to structural reforms and fiscal policies (e.g. labour market and pension reforms). Finally, the statement ends with a “Question and Answers” session which enables the ECB to further elaborate on the meaning of some keywords used or to reiterate its view on the economic outlook in general. Rosa and Verga (2007) restricted their attention to the two main parts of the introductory statement, namely the introductory part and the summing-up paragraph. These two assessments usually contain information regarding the ECB’s

(16)

16

perception of price risk to price stability, monetary policy and economic outlook. This approach will also be employed in this paper. However, the structure of the introductory statement has slightly changed during the sample period used in this paper. Namely the summing-up paragraph has decreased substantially in its elaboration averaging only 41.45 words over the period February 2012 and March 2015, and 158.67 over the period November 2011 and January 2012. This cutback in the number of words used can be explained by the fact that the summing-up part solely conveyed information on risk to price stability after February 2012. For this reason, only the first main part of the introductory statement will be codified so as to determine the hawkish-ness value of each statement6. In deciding them, however, the parts regarding the structural reforms and fiscal policy are not taken into account, as they do not contain information regarding the future monetary policy stance of the ECB.

The communication index is based on the assumption that new information is distributed to the public for each new press conference that takes place. Thus, the new information is valid until a new press conference is held. Hence, the number of days in between each press conference are taken into account. This is especially needed, as the interval between each press conference fluctuates over time. For instance, the number of days between the press conference on the 4th of December 2014 and the 22nd of January 2015 was 48. In another case, the interval was only 27 days, in 2012 from the 12th of January until the 9th of February. Although, the President normally conducts one introductory statement per month, this is not always true. In October 2014 a press conference was held on the 2nd and 26th. On one occasion, however, no introductory statement for that month was released at all, namely February 2015. A different weight is put on each press conference, in order to control for the differences described above. The more days in between each press conference, the higher the weight accredited to the first one. In this way, each month receives a weighted average hawkish-ness score. The latter results in the following equation:

6_{This is also an underlying motivation for the sample period employed in this paper (November 2011 – March}

2015). As the bias regarding the difference in the communication of ECB presidents is mitigated. For instance, Draghi repeated the introductory paragraph only 3 times (during the sample period) in the summing-up part while Jean-Claude Trichet did it in all of his introductory statements over the period October 2003 – October 2011.

(17)

17 𝑆𝑐𝑜𝑟𝑒_𝑡 = 𝛾_1,𝑡((∅_𝑡− 𝜃_1,𝑡+ 1) − ∑(𝛾_𝑖,𝑡𝜃_𝑖,𝑡)) ∞ 𝑡=1 𝐻_1,𝑡+ ∑(𝛾_𝑖,𝑡∅_𝑡− 𝜃_𝑖,𝑡+ 1)𝐻_𝑖,𝑡 ∞ 𝑡=1 +𝜆𝑡𝑆𝑐𝑜𝑟𝑒𝑡−1 (1) 𝑖 ∈ {2,3, … , 𝑁}, 𝑡 ∈ {1,2, … , 𝑇}

With 𝑆𝑐𝑜𝑟𝑒𝑡 being the weighted average hawkish-ness score in month 𝑡; 𝑆𝑐𝑜𝑟𝑒𝑡−1 being the weighted average hawkish-ness score in month 𝑡 − 1; 𝐻1,𝑡 being the hawkish-ness value of the first press conference in month 𝑡 and 𝐻𝑖,𝑡 being the hawkish-ness value of press conference 𝑖 in month 𝑡. 𝛾1,𝑡 is a dummy variable that takes the value of 1 to indicate if one or more press conferences take place in month 𝑡, and 0 to indicate the absence of a press conference. 𝛾𝑖,𝑡 is a dummy variable that takes the value of 1 to indicate if two or more press conferences take place in month 𝑡, and 0 to indicate the absence of two or more press conferences in month 𝑡. ∅𝑡 represent the total number of days in month 𝑡, and 𝜃1,𝑡 the day number of the first press conference in month 𝑡. 𝜃𝑖,𝑡 is the day number of press conference 𝑖 in month 𝑡. Finally, 𝜆𝑡 is a dummy variable that takes the value of 1 if no press

conferences are held in month 𝑡, an 0 if not.

However, formulating the communication index in such a way neglects one aspect of ECB communication. Namely, second round effects. In the majority of the introductory statements a reference is made to the previous press conference. Such a constellation suggests that ECB key interest rate decisions might be partially based on preceding statements. The aforementioned can be illustrated below:

Second, following today’s rate cut, the Governing Council reviewed the forward guidance provided in July and confirmed that it continues to expect the key ECB interest rates to remain at present or lower levels for an extended period of time (ECB, Press Conference, November 2013).

A communication index allowing for second round effects appears then more fitting. Thus, an extra term is needed to capture all relevant information since the last ECB press conference. 𝑆𝑐𝑜𝑟𝑒𝑡 will therefore also depend on the hawkish-ness value of month 𝑡 − 1 (𝐻𝑡−1). Hence, one can now compute the hawkish-ness score for month 𝑡 through the following equation:

(18)

18 𝑆𝑐𝑜𝑟𝑒_𝑡= 𝛾_1,𝑡((∅_𝑡− 𝜃_1,𝑡+ 1) − ∑(𝛾_𝑖,𝑡𝜃_𝑖,𝑡)) ∞ 𝑡=1 𝐻_1,𝑡+ ∑(𝛾_𝑖,𝑡∅_𝑡− 𝜃_𝑖,𝑡 + 1)𝐻_𝑖,𝑡 ∞ 𝑡=1 +𝜆_𝑡𝑆𝑐𝑜𝑟𝑒_𝑡−1+ 𝛾_1,𝑡(𝜃_1,𝑡− 1)𝐻_𝑡−1 (2) 𝑖 ∈ {2,3, … , 𝑁}, 𝑡 ∈ {1,2, … , 𝑇}

Figure 2, plots the weighted average hawkish-ness score of the ECB communication index. Furthermore, Appendix A provides a few examples of introductory statements along with their hawkish-ness coding. Moreover, the dataset will be completed with credit spreads and its determinants which will be elucidated in the next section (3.2).

Figure 2: Plot weighted average hawkish-ness score (𝑺𝒄𝒐𝒓𝒆𝒕).

Source: Author’s ECB communication index. November 2011 – March 2015. 3.2 Dependent and explanatory variables

As mentioned, this study uses monthly economic and financial data from November 2011 until March 2015. For the purpose of this analysis, this paper will solely focus on the credit spreads of Eurozone sovereign bonds. Table 2.1B in Appendix B provides an overview of all Eurozone government bonds, whether they were employed and the reason for their exclusion. Two dependent variables will be employed, namely credit spread AAA, 10-year and credit spread BBB, 10-year. Distinguishing between different asset classes has two advantages. First, it controls for the fact that for instance BBB credit spreads have higher

-300 -250 -200 -150 -100 -50 0 Oct-11 De c-1 1 Fe b -12 Ap r-12 Ju n -12 Au g-12 Oct-12 De c-1 2 Fe b -13 Ap r-13 Ju n -13 Au g-13 Oct-13 De c-1 3 Fe b -14 Ap r-14 Ju n -14 Au g-14 Oct-14 De c-1 4 Fe b -15

Score

(19)

19

forecast errors than AAA credit spreads. Second, it allows one to distinguish between the possible improvement in predictability of each asset class after the start of unconventional monetary policy. The determinants employed as explanatory variables can be divided into two categories: equity related variables and interest rate risk sensitivity variables. These are factors that embody the stage of the economic cycle, the interest rate environment and investment risk preferences. These determinants are partially motivated by Collin-Dufresne et al. (2001); Boss and Scheicher (2002) and Oliveira et al. (2012), who state that the factors are based on the “structural models of default risk”. The latter allows one to link credit risk to for instance macroeconomic fundamentals.

In Table 3.1C in Appendix C, an overview of all variables is provided along with their mean, standard deviation and number of observations. All variables are measured in percentage changes. The next section will further elaborate on the variables used in this paper.

3.2.1 Dependent variables

Sovereign bond data are obtained from the St. Louis FED FRED database. The ratings of each asset class is retrieved from Standards and Poor (S&P) Global7. As outlined above, this paper examines the following asset classes: credit spread AAA, year and credit spread BBB, 10-year. AAA bonds are deemed to be of the highest quality, while BBB bonds are considered to be more risky and hence carry more risk of default. This index includes all asset class ratings. The credit spread AAA is reflected by the value-weighted index on 10-year Eurozone government bonds rated AAA minus the yield on a 10-year German government bond. The last dependent variable, the credit spread BBB, is embodied by a value-weighted index on 10-year Eurozone government bonds rated BBB minus the yield on a 10-year German government bond. This index includes all bonds rated BBB and below. Bonds were omitted from the value-weighted indices if no monthly data could be obtained or if no data value is provided at all.

(20)

20

3.2.2. Explanatory variables 3.2.2.1. Equity related variables

In structural models of default risk the return of equity markets is a key variable in determining credit spreads. The return of equity markets could be a relevant factor to fixed income investors for several reasons. For instance, it could be considered as an alternate investment. Credit spreads are therefore expected to be a decreasing function of the return one would receive on the equity market. To proxy the return of Eurozone equity markets, this paper will use the Dow Jones EURO STOXX 50 index which measures the returns of 50 large European companies operating in the Eurozone. In order to account for the risk level in the equity market, the Dow Jones EURO STOXX 50 Volatility index is employed. The correlation is expected to be positive. The underlying motivation for this choice is very straightforward: an increase in volatility entails an increase in the market risk and, hence, in the probability of default. Data on the variables outlined above are retrieved from Datastream.

3.2.2.2. Interest rate sensitive variables

As pointed out by Cochrane and Piazzesi (2005), predictability of credit spreads can be enhanced by adding information about the riskless yield curve. Even more so, Boss and Scheicher (2002), write that information about the risk-free term structure is the pre-eminent benchmark for sovereign bond spreads with different maturities. Furthermore, in line with Litterman and Scheinkman (1991); Collin-Dufresne et al. (2001) and Longstaff and Schwartz (1995), Oliveira et al. (2012) write that the slope, the curvature, and the level of the term structure are indicators to gauge the interest rate environment. However, Clerc et al. (2002) suggest that if one wants to compare yield spreads with different credit risks at any given maturity, the slope of the yield curve emerges as the clear individual choice over the other indicators. As mentioned above, this paper will use two dependent variables that embody different credit risks so as to distinguish between different asset classes. Hence, the empirical framework will only incorporate the slope of the yield curve. Data on this variable is obtained from the St. Louis FED FRED database.

In the previous literature, the slope of the yield curve is broadly measured as the difference between a 10-year sovereign bond’s rate and a short-term rate. For example,

(21)

21

Hamilton and Kim (2002) compute the slope of the yield curve as the difference between the 10-year Treasury bond and the three-month Treasury bill. This paper, however, will focus its attention on the Eurozone. Hence, the slope is defined as the difference between the 10-year and 2-10-year German government bond rates. According to Longstaff and Schwartz (1995), an increase in the slope of the risk-free yield will lead to an increase of the expected future spot rates, which will most likely result in a reduction of the credit spread. Furthermore, one could also argue that a negative yield curve implies a weak economy, which in turn leads to an increase in the sovereign bond spread. In conclusion, the slope of the riskless yield curve is most likely to have an inverse effect on credit spreads.

(22)

22

4. Empirical analysis

In this section the empirical analysis is presented. The analysis starts by presenting the hypotheses. The empirical framework is outlined in the second part of this chapter, so as to test the hypotheses.

4.1. Hypotheses

After taking the variable statistics, and previous literature into consideration, this section will formulate the hypotheses. The first hypothesis focusses on the relation between credit spreads and ECB communication.

1st hypothesis: Central bank communication has a significant negative effect on credit

spreads.

The underlying motivation for this hypothesis goes as follows. More intensive central bank communication results in more transparency in the markets (i.e. less asymmetric information between financial market participants and central banks), and hence less risk. As the risk premia will decrease, credit spreads should tighten.

The second hypothesis concerns, whether the effects differ for different asset classes. As mentioned, this paper will evaluate AAA and BBB credit ratings.

2nd hypothesis: The effect of ECB communication on credits spreads rated AAA and BBB is

presumed to be heterogeneous.

This is due to the fact that credit spread volatility rises with inferior credit ratings. Hence, the effect is assumed to be stronger for higher credit ratings.

To test these two hypothesis this study will provide an empirical framework, which will be exemplified in the succeeding section.

4.2. Empirical analysis

To empirically determine the effectiveness of ECB communication on credit spreads, I model monthly changes to explain the impact on the two credit spreads8 for the Eurozone as a linear function of the communication index described in section 3.1. The econometric

(23)

23

modelling software that will be employed is EViews. As this study considers time series data, several points need to be addressed for statistical purposes. First, there is a possibility that the variables contain a stochastic trend, which in turn, makes the conventional hypothesis tests, confidence intervals, and forecasts unreliable (Stock and Watson, 2012). In order to avoid these spurious results, one must construct a linear combination of variables that eliminates the stochastic trend and produces stationary residuals. This is better known as cointegrated time series variables (i.e. there is some long‐run equilibrium relation binding the set of variables together). Figure 3 below, that plots the two credit spreads, suggests that the data contains a stochastic trend as there is a persistent downward movement over time. Therefore, the augmented Dickey-Fuller (ADF) test9 will be employed to test for a stochastic trend. The ADF test is preferred over the standard Dickey-Fuller test, as the former is able to control for all autocorrelation in the dependent variable. In this case, a higher-order autoregression is more fitting (Stock and Watson, 2012). The lag length will be automatically computed by EViews, which employs automatic lag length selection. Under the null hypothesis in the ADF test given the present analysis, the credit spread contains a unit root (i.e. data contains a stochastic trend); under the alternative hypothesis the credit spread is stationary.

Figure 3: Plots credit spreads (levels).

Source: The St. Louis FED FRED database (2017).

9

There are various other test available so as to detect stochastic trends, nonetheless the ADF test is the most frequently employed test in empirical work and is considered the most trustworthy (Stock and Watson, 2012).

0,0000 2,0000 4,0000 6,0000 8,0000 10,0000 12,0000

(24)

24

Performing the ADF test does not result in a rejection of the null hypothesis of unit root (i.e. data contains a stochastic trend) for neither of the two credit spreads employed.10 As mentioned before, central banks try to steer expectations about future policy by communicating about the central bank’s view on the economic outlook and signalling through the decision rates. Hence, there could be a common stochastic trend between the credit spread series and the communication index series. In order to determine whether the communication index and the credit spreads have to be modelled as cointegrated, the Johansen test for cointegration will be used. EViews follows VAR-based cointegration tests employing the methodology described in Johansen (1991, 1995) using a computed vector autoregression (VAR) object (EViews, 2017). Applying this methodology to the present analysis, the procedure is carried out as follows: Consider the VAR of credit spread 𝑙 (𝐶𝑆𝑡𝑙) of order 𝑝:

𝐶𝑆_𝑡𝑙 _{= 𝐴}

1𝐶𝑆𝑡−1𝑙 + ⋯ + 𝐴𝑝𝐶𝑆𝑡−𝑝𝑙 + 𝐵𝑆𝑐𝑜𝑟𝑒𝑡+ 𝜀𝑡 (3)

With 𝐶𝑆𝑡𝑙 being a 𝑘-vector of non-stationary variables; 𝑆𝑐𝑜𝑟𝑒𝑡 being a 𝑑-vector of explanatory variables (the weighted average hawkish-ness score) and 𝜀𝑡 being a vector of innovations. This can be rewritten to the following:

∆𝐶𝑆𝑡𝑙= ∏ 𝐶𝑆𝑡−1𝑙 + ∑ 𝛤𝑖 𝑝−1 𝑖=1 ∆𝐶𝑆_𝑡−𝑖𝑙 + 𝐵𝑆𝑐𝑜𝑟𝑒𝑡+ 𝜀𝑡 (4) With: ∏ = ∑ 𝐴_𝑖 𝑝 𝑖=1 − 𝐼 , (5) 𝛤_𝑖 = − ∑ 𝐴_𝑗 𝑝 𝑗=𝑖+1 (6)

If the coefficient matrix ∏ has reduced rank 𝑟 < 𝑘, then we have 𝑘 x 𝑟 matrices 𝑧 and 𝑞 with rank 𝑟 such that ∏ = 𝑧𝑞′ and 𝑞′𝐶𝑆𝑡𝑙 is stationary ~ 𝐼(0) according to Granger’s representation theorem (EViews, 2017). 𝑟 is the cointegrating rank and each column of 𝑞 is the cointegrating vector. Moreover, 𝑧 and 𝑞 are also known as the adjustment elements in

(25)

25

the VECM. Johansen (1991, 1995) follows a method that derives the ∏ from an unrestricted VAR and evaluates whether one can reject the restrictions indicated by a reduced rank of ∏. Running this procedure in EViews and using the Akaike Information Criterion (AIC), one can conclude that credit spread AAA is integrated with 𝑆𝑐𝑜𝑟𝑒𝑡 of order 1 ~ 𝐼(1) with some linear trend, and credit spread BBB is integrated with 𝑆𝑐𝑜𝑟𝑒𝑡 of order 1 ~ 𝐼(1) with some linear trend.11

Lütkepohl (1993) suggests that in case of cointegration, model order selection is an impotant element. They write that adding to much lags increases the mean square forecast errors of your model and that underfitting the lag length results in too little predictive power. The lag selection procedure is commonly estimated by applying an information criteria. According to Stock and Watson (2012), one must construct the model suggested by AIC rather than the Schwartz information criterion (SIC)12 when determining the order 𝑝 of an autoregression. They state that employing the SIC is accompanied with the risk of neglecting potential significant information. The AIC, however, is not unblemished either as it might overestimate 𝑝 (Stock and Watson, 2012). Therefore, for robustness, a second approach will be used as well. Namely, performing hypothesis tests on the final lag. This will be done by starting with a second-order autoregressive model and test whether the parameter of the final lag is significant. In case the coefficient is not significant, it will be omitted and a first-order autoregressive model will be computed, and so on.

Lütkepohl and Krätzig (2004) state that if some of the variables are cointegrated, a VECM is the preferred modelling technique as it allows for various cointegrating relationships, and long-run dynamics of the parameters. Moreover, it treats all variables as endogenous. Therefore, the VECM technique will be employed in this paper. Furthermore, the VECM is calculated following a two-step procedure. First, employing the Johansen method, the cointegrating relations are computed. These long-run equilibrium relations are then formulated as follows:

𝐶𝑆_𝑡𝐴𝐴𝐴 _{= 𝛽} 1𝐴𝐴𝐴𝑆𝑐𝑜𝑟𝑒𝑡+ 𝛽2𝐴𝐴𝐴𝑇𝑅𝑡+ ∁ (7) 𝐶𝑆𝑡𝐵𝐵𝐵 = 𝛽1𝐵𝐵𝐵𝑆𝑐𝑜𝑟𝑒𝑡+ 𝛽2𝐵𝐵𝐵𝑇𝑅𝑡+ ∁ (8) 𝑡 ∈ {1,2, … , 𝑇} 11

Refer to Table 5.1E and 5.2E in Appendix E. These results will be elucidated in section 5.1.

(26)

26

With 𝐶𝑆𝑡𝐴𝐴𝐴 being the credit spread AAA in month 𝑡; 𝐶𝑆𝑡𝐵𝐵𝐵 being the credit spread BBB in month 𝑡; 𝑇𝑅𝑡 being the trend and ∁ being the constant. 𝛽1𝐴𝐴𝐴 is the coefficient that represents the long-term impact of ECB communication on credit spread AAA. 𝛽1𝐵𝐵𝐵 represents the coefficient regarding the effect of ECB communication on credit spread BBB. The values of these coefficients will be provided in section 5. In the second step, the error correction terms are derived from the cointegrating relations of step one and a first-difference VAR model is estimated containing the error correction terms as parameters. This step represents the short-run relations, from which one can for instance evaluate the following question: How does 𝐶𝑆𝑡𝐴𝐴𝐴 react when 𝑆𝑐𝑜𝑟𝑒𝑡 deviates from its long-run

equilibrium?.

Finally, Stock and Watson (2012) argue that one must be prudent as regards serial-correlation. Hence, a test for serial-correlation is conducted in order to check whether a sufficient amount of lags are included. The Durbin-Watson (DW) test and the Breusch-Godfrey LM test are the most commonly used tests to evaluate the null hypothesis of no serial correlation. Asteriou and Hall (2007), however, state that the DW test has various drawbacks. For instance, they write that it is not applicable when a lagged dependent variable is employed, and it does not capture higher orders of serial correlation. The Breusch-Godfrey LM test for serial correlation would then be more appropriate. The test results for the latter do not provide evidence to reject the null hypothesis of no serial correlation for both credit spread AAA and credit spread BBB (refer to Table 7.1G and 7.2G in

Appendix G for the results).13

13

These results will be elaborated in chapter 5. Note: three has been chosen as the highest order to be tested for serial correlation.

(27)

27

5. Results

This section provides the results of the ADF tests, Johansen’s cointegration tests, final lag hypothesis tests, the VECM estimations and the serial correlation tests. First, the results of the credit spread AAA series are deliberated. Subsequently, the results of the credit spread BBB series are discussed. Lastly, the results of the two different series are compared.

5.1. Results: credit spread AAA

Table 4.1D in Appendix D presents the results of the ADF test for the credit spread AAA series in levels. The ADF t-statistic is -1.875093, which indicates that the null hypothesis (credit spread AAA in levels has a unit root) cannot be rejected given a 1%, 5% and 10% significance level (𝑝 − 𝑣𝑎𝑙𝑢𝑒 = 0.6491). In Table 4.2D in Appendix D, the results of the ADF test for the credit spread AAA series in first differences are provided. The results for the latter, indicate that the null hypothesis (credit spread AAA in first difference contains a unit root) cannot be rejected given a 1%, 5% and 10% significance level (𝑝 − 𝑣𝑎𝑙𝑢𝑒 = 0.1657).

The Johansen cointegration test results are presented in Table 5.1E in Appendix E. Given a 5% significance level, the AIC indicates that we have a single cointegration vector with a linear trend. Furthermore, the final lag hypothesis test coincides with the AIC as three out of the four first lags are significant given a 5% significance level, and just one out of the four second lags (see Table 6.1F in Appendix F).

The VECM baseline regression results are summarized in Table 2 on the next page. We can immediately notice that there is one restriction in the cointegration vector, namely that 𝐶𝑆𝑡𝐴𝐴𝐴 is normalized to one. The results for the long-run equilibrium relations, provide evidence that more ECB communication tightens credit spreads rated AAA in the long-run as the sign of the coefficient (0.014876) is negative and the t-statistic (5.10807) exceeds its critical value (1.96) at a 5% significance level. Furthermore, if we evaluate the error correction terms we can see that the long-run disequilibrium of ECB communication is corrected downwards, as the coefficient is -59.17281 and the t-statistic (-4.82052) is lower than its critical value (-1.96). Nonetheless, no evidence is provided for a correction regarding the disequilibrium of the credit spread AAA given a 5% significance level as the t-statistic (-0.42789) is not lower than its critical value (-1.96). Moreover, the first lag of 𝐶𝑆𝑡𝐴𝐴𝐴

(28)

28

and 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference are not able to explain short-run deviations of both series (credit spread AAA and the ECB communication index) given a 5% significance level.

Table 2: VECM baseline regression credit spread AAA. Included observations: 39 after adjustments

Standard errors in ( ) & t-statistics in [ ] Cointegrating Eq: CointEq1

CSAAA(-1) 1.000000 SCORE(-1) 0.014876 (0.00291) [ 5.10807] @TREND(1) 0.027794 (0.00754) [ 3.68615] C 0.927902

Error Correction: D(CSAAA) D(SCORE)

CointEq1 -0.010360 -59.17281 (0.02421) (12.2752) [-0.42789] [-4.82052] D(CSAAA(-1)) 0.377074 199.5176 (0.22544) (114.296) [ 1.67265] [ 1.74562] D(SCORE(-1)) 0.000294 0.070774 (0.00025) (0.12772) [ 1.16725] [ 0.55414] C -0.006292 4.833578 (0.01019) (5.16530) [-0.61756] [ 0.93578] R-squared 0.164415 0.462795 Adj. R-squared 0.092794 0.416749 Sum sq. resids 0.131788 33875.99 S.E. equation 0.061363 31.11086 F-statistic 2.295612 10.05069 Log likelihood 55.61882 -187.2932 Akaike AIC -2.647119 9.809905 Schwarz SC -2.476497 9.980527 Mean dependent -0.009615 2.820513 S.D. dependent 0.064424 40.73656 Number of coefficients 11

Note: CSAAA = 𝐶𝑆𝑡𝐴𝐴𝐴, SCORE = 𝑆𝑐𝑜𝑟𝑒𝑡, @TREND = Trend (𝑇𝑅𝑡), C = constant (𝐶), D(CSAAA) = 𝐶𝑆𝑡𝐴𝐴𝐴 in

(29)

29

Figure 4 illustrates the response of credit spread AAA after an impulse of ECB communication. First, a positive shock can be noted. However, in the long-run after month four it becomes negative. Hence, the credit spread tightens. The VECM augmented regression results are presented in Table 3 below. In accordance with the VECM baseline regression results, the cointegration vector has one restriction. Namely, that 𝐶𝑆𝑡𝐴𝐴𝐴 is normalized to one. Moreover, the coefficient of the ECB communication, at 0.10352, has not changed much. Its effect remained significant, at a 5% significance level, as the t-statistic (4.95226) still exceeds its critical value (1.96). Hence, the long-run equilibrium relations coincide with the baseline regression as they suggest that more ECB communication tightens credit spreads rated AAA in the run. The error correction terms indicate that the long-run disequilibrium of ECB communication is still corrected downwards, as the sign of the coefficient is negative (-8.555785) given a 5% significance level as the t-statistic (-4.37654) is lower than its critical value (-1.96). The results for the error correction term of the credit spread AAA, however, do not provide evidence for an adjustment as regards long-run disequilibrium given a 5% significance level as the t-statistic (-0.09472) is not lower than its critical value (-1.96). Furthermore, the short-run relations of the first lag of 𝐶𝑆𝑡𝐴𝐴𝐴 and 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference are still not significant for both series at a 5% significance level. Lastly, the added control variables (the Dow Jones EURO STOXX 50 index, the Dow Jones EURO STOXX 50 volatility index, and the slope of the term structure) are not able to explain short-run alternations for both series given a 5% significance level.

Figure 4: Impulse-response function

IMPULSE: SCORE RESPONSE: CSAAA

-.002 .000 .002 .004 .006 .008 5 10 15 20 25 30 35 40 MONTH

(30)

30

Table 3: VECM augmented regression credit spread AAA. Included observations: 39 after adjustments

CSAAA(-1) 1.000000 SCORE(-1) 0.103052 (0.02081) [ 4.95226] @TREND(1) 0.251558 (0.17470) [ 1.43998] C 6.947989

Error Correction: D(CSAAA) D(SCORE)

CointEq1 -0.000348 -8.555785 (0.00368) (1.95492) [-0.09472] [-4.37654] D(CSAAA(-1)) 0.317931 163.4481 (0.22507) (119.674) [ 1.41261] [ 1.36578] D(SCORE(-1)) 0.000311 0.111199 (0.00026) (0.13909) [ 1.18722] [ 0.79948] C 0.124533 72.55848 (0.21660) (115.172) [ 0.57494] [ 0.63000] DWJ50 index -2.91E-05 -0.001569 (4.7E-05) (0.02498) [-0.61950] [-0.06280] V2TX -0.003268 -1.816968 (0.00312) (1.65707) [-1.04860] [-1.09649] SLTS 0.015650 -20.15887 (0.03845) (20.4424) [ 0.40708] [-0.98613] R-squared 0.214491 0.444530 Adj. R-squared 0.067207 0.340379 Sum sq. resids 0.123890 35027.79 S.E. equation 0.062222 33.08502 F-statistic 1.456315 4.268146 Log likelihood 56.82390 -187.9451 Akaike AIC -2.555072 9.997187 Schwarz SC -2.256484 10.29577

(31)

31

Mean dependent -0.009615 2.820513

S.D. dependent 0.064424 40.73656

Number of coefficients 17

Note: CSAAA = 𝐶𝑆𝑡𝐴𝐴𝐴, SCORE = 𝑆𝑐𝑜𝑟𝑒𝑡, @TREND = Trend (𝑇𝑅𝑡), C = constant (𝐶), D(CSAAA) = 𝐶𝑆𝑡𝐴𝐴𝐴 in

first difference, D(SCORE) = 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference, DWJ50 index = the Dow Jones EURO STOXX 50 index,

V2TX = Dow Jones EURO STOXX 50 volatility index, and SLTS = the slope of the term structure.

Concerning the Breusch-Godfrey LM serial correlation test, no evidence is found for serial correlation, at a 5% significance level. This is due to the fact that the p-values of lag one, two and three are 0.5600, 0.2870 and 0.5284 respectively (refer to Table 7.1G in

Appendix G).

5.2. Results: credit spread BBB

For credit spread series BBB, the results of the ADF test are provided in Table 4.3D in

Appendix D. The ADF t-statistic is -0.380854, which suggests that the null hypothesis (credit

spread BBB in levels has a unit root) cannot be rejected at a 1%, 5% and 10% significance level (𝑝 = 0.9840). With respect to the ADF t-statistic for the credit spread BBB series in first difference, a significant result is found given a 1% significance level (𝑝 − 𝑣𝑎𝑙𝑢𝑒 = 0.0010) (see Table 4.4D in Appendix D). Hence, we can reject the null hypothesis (credit spread BBB in first difference has a unit root).

The results for the Johansen cointegration test are summarized in Table 5.2E in

Appendix E. The AIC suggests that we have a single cointegration vector with a linear trend

at a 5% significance level. However, the final lag hypothesis test differs in its conclusion, as it indicates that the second lag of the credit spread BBB in first difference is significant given a 5% significance level (refer to Table 6.2F in Appendix F). Adding the second lag of 𝐶𝑆𝑡𝐵𝐵𝐵 and the 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference seems then more appropriate.

The credit spread BBB VECM baseline results are provided in Table 4 on the next page. First, we can see that there is one restriction in the cointegration vector, to be specific 𝐶𝑆_𝑡𝐵𝐵𝐵_{is normalized to one. The long-run equilibrium results suggest that more ECB} communication tightens credit spreads rated BBB in the long-run as the sign of the coefficient (0.165684) is negative and the t-statistic (5.32119) is above its critical value (1.96) given a 5% significance level. Moreover, if we examine the error correction term no evidence is found for an adjustment regarding disequilibrium of the credit spread BBB, at a 5% significance level, as the t-statistic (1.51118) does not cross the boundary (1.96) of the

(32)

32

critical region. The opposite is true for the ECB communication index, as a disequilibrium is significantly adjusted downwards given a 5% significance level as the t-statistic (-4.23427) is lower than its critical value (-1.96). Furthermore, evidence is found that the second lag of 𝐶𝑆_𝑡𝐵𝐵𝐵_{in first difference is able to explain short-run deviations of the credit spread BBB} series, at a 5% significance level. The positive sign of the coefficient indicates that a unit increase of the second lag results in an increase of the current value. The first lag, however, is insignificant given a 5% significance level. The second and first lag of 𝑆𝑐𝑜𝑟𝑒𝑡 are also insignificant for both series given a 5% significance level.

Table 4: VECM baseline regression credit spread BBB. Included observations: 38 after adjustments

CSBBB (-1) 1.000000 SCORE(-1) 0.165684 (0.03114) [ 5.32119] @TREND(1) 0.581947 (0.09636) [ 6.03950] C 1.479243

Error Correction: D(CSBBB) D(SCORE)

CointEq1 0.036772 -5.037572 (0.02433) (1.18972) [ 1.51118] [-4.23427] D(CSBBB (-1)) 0.119177 -0.685632 (0.15887) (7.76725) [ 0.75017] [-0.08827] D(CSBBB (-2)) -0.311637 -1.137703 (0.15046) (7.35624) [-2.07124] [-0.15466] D(SCORE(-1)) -0.001188 0.184284 (0.00320) (0.15632) [-0.37162] [ 1.17887] D(SCORE(-2)) -0.002993 0.057343 (0.00291) (0.14215) [-1.02950] [ 0.40341]

(33)

33 C -0.215329 0.535739 (0.11617) (5.68001) [-1.85349] [ 0.09432] R-squared 0.210777 0.398421 Adj. R-squared 0.087461 0.304425 Sum sq. resids 14.99552 35845.55 S.E. equation 0.684551 33.46899 F-statistic 1.709239 4.238676 Log likelihood -36.25281 -184.0580 Akaike AIC 2.223832 10.00306 Schwarz SC 2.482398 10.26162 Mean dependent -0.201053 1.289474 S.D. dependent 0.716605 40.13013 Number of coefficients 15

Note: CSBBB = 𝐶𝑆𝑡𝐵𝐵𝐵, SCORE = 𝑆𝑐𝑜𝑟𝑒𝑡, @TREND = Trend (𝑇𝑅𝑡), C = constant (𝐶), D(CSBBB) = 𝐶𝑆𝑡𝐵𝐵𝐵 in

first difference and D(SCORE) = 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference.

The response of credit spread BBB after an impulse of ECB communication is illustrated in Figure 5. In this figure, the credit spread tightens after a communication shock. In Table 5 below, the VECM augmented regression results are provided. Immediately we can see that the cointegration vector has one restriction (𝐶𝑆_𝑡𝐵𝐵𝐵 is normalized to one), which is in line with the baseline regression results. Considering the cointegration equation, the sign of the coefficient (0.108840) of the ECB communication index remains unchanged and is still significant given a 5% significance level as the t-statistic (4.45526) does not exceed its critical value (1.96). Therefore, we can conclude that more ECB communication tightens credit spreads rated BBB in the long-run. Furthermore, the error correction terms suggest that for the ECB communication index a disequilibrium is corrected downwards, at a 5% significance level, as the t-statistic (-4.33543) is lower than the critical value (-1.96). Nonetheless, no evidence is found for an adjustment regarding disequilibrium of the credit spread BBB series. Furthermore, the second lag of 𝐶𝑆𝑡𝐵𝐵𝐵 in first difference remains significant given a 5% significant level. Nonetheless, the first lag is still not able to explain short-run deviations of the credit spread BBB series, at a 5% significance level. The first and second lag of 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference remain insignificant for both series given a 5% significance level. Lastly, none of the added control variables (the Dow Jones EURO STOXX 50 index, the Dow Jones EURO STOXX 50 volatility index, and the slope of the term structure) are able to explain deviations of the long-run equilibrium for credit spread BBB, at a 5% significance level. However, the Dow Jones EURO STOXX 50 index and the slope of the term structure are both significant as

(34)

34

regards disequilibrium of the ECB communication index given a 5% significance level. The former indicates a decrease and the latter suggests an increase of ECB communication.

Figure 5: Impulse-response function

IMPULSE: SCORE RESPONSE: CSBBB

-.35 -.30 -.25 -.20 -.15 -.10 -.05 .00 5 10 15 20 25 30 35 40 MONTH

Note: the impulse response function employs Cholesky (d.f. adjusted) Factors

Table 5: VECM augmented regression credit spread BBB. Included observations: 38 after adjustments

CSBBB (-1) 1.000000 SCORE(-1) 0.108840 (0.02443) [ 4.45526] @TREND(1) 0.914617 (0.24668) [ 3.70777] C -12.93626

Error Correction: D(CSBBB) D(SCORE)

CointEq1 0.045364 -7.006177 (0.03159) (1.61603) [ 1.43597] [-4.33543] D(CSBBB (-1)) 0.006232 0.354490 (0.16438) (8.40858) [ 0.03791] [ 0.04216]

(35)

35 D(CSBBB (-2)) -0.330095 -1.248905 (0.14686) (7.51239) [-2.24773] [-0.16625] D(SCORE(-1)) -0.000531 0.101995 (0.00296) (0.15132) [-0.17936] [ 0.67402] D(SCORE(-2)) -0.001378 -0.035867 (0.00280) (0.14334) [-0.49189] [-0.25021] C 0.198452 -119.5243 (2.61968) (134.008) [ 0.07575] [-0.89192] DWJ50 index -0.000301 0.067851 (0.00063) (0.03231) [-0.47601] [ 2.09993] V2TX 0.036734 -0.341068 (0.03726) (1.90595) [ 0.98591] [-0.17895] SLTS -0.278968 -52.09002 (0.45065) (23.0527) [-0.61904] [-2.25960] R-squared 0.327452 0.438811 Adj. R-squared 0.141921 0.284001 Sum sq. resids 12.77866 33438.88 S.E. equation 0.663810 33.95681 F-statistic 1.764949 2.834503 Log likelihood -33.21328 -182.7375 Akaike AIC 2.221751 10.09145 Schwarz SC 2.609601 10.47930 Mean dependent -0.201053 1.289474 S.D. dependent 0.716605 40.13013 Number of coefficients 21

Note: CSBBB = 𝐶𝑆𝑡𝐵𝐵𝐵, SCORE = 𝑆𝑐𝑜𝑟𝑒𝑡, @TREND = Trend (𝑇𝑅𝑡), C = constant (𝐶), D(CSBBB) = 𝐶𝑆𝑡𝐵𝐵𝐵 in

first difference, D(SCORE) = 𝑆𝑐𝑜𝑟𝑒𝑡 in first difference, DWJ50 index = the Dow Jones EURO STOXX 50 index,

V2TX = Dow Jones EURO STOXX 50 volatility index, and SLTS = the slope of the term structure.

Table 7.2G in Appendix G summarizes the results of the Breusch-Godfrey LM serial correlation tests with respect to the credit spread BBB series. The null hypothesis of no serial correlation cannot be rejected, at a 5% significance level, as the p-values of lag one, two and three are 0.4568, 0.1191 and 0.6078 respectively.

(36)

36

5.3. Comparing series

Focussing on the impulse response functions, both credit spread series tighten in the long-run after a communication shock. This prognosis is supported by the results of the VECM results, as they suggest that for credit spreads rated AAA and BBB more ECB communication has a significant negative effect in the long-run. Contrary to the BBB rated credit spread, the triple-A response function shows a positive shock in the short-run. Nonetheless, no supportive evidence is found for the latter. One more point is worth highlighting. Namely, that neither of the two credit spread series can be explained by their first lag or by the first two lags of the ECB communication index. Contrary to the credit spread AAA, however, the credit spread BBB series is significantly involved with its second lag.

Table 6, provides a summary of the results regarding the effect of ECB communication on the credit spread series rated AAA and BBB.

Table 6: Effect from ECB communication on AAA and BBB-rated credit spreads.

IMPACT ECB COMMUNICATION AAA BBB

SHORT-RUN No evidence found No evidence found

Monetary news shocks and credit spread predictability : evidence of the European Central Bank - Draghi era