The effectiveness of the Fed’s quantitative easing policy: New evidence based on international interest rate differentials

Journal of International Money and Finance 73 (2017) 335–349

Contents lists available at ScienceDirect

Journal of International Money and Finance journal homepage: www.elsevier.com/locate/jimf

The effectiveness of the Fed’s quantitative easing policy: New evidence based on international interest rate differentials q Angar Belke a,b,c,⇑, Daniel Gros c, Thomas Osowski a a

University of Duisburg-Essen, Berliner Platz 6-8, 45127 Essen, Germany Institute for the Study of Labor (IZA), Schaumburg-Lippe-Straße 5-9, 53113 Bonn, Germany c Centre for European Policy Studies (CEPS), Place du Congrès 1, 1000 Brussels, Belgium b

a r t i c l e

i n f o

Article history: Available online 20 February 2017 JEL classification: E58 F42 G15 Keywords: Quantitative easing Unconventional monetary policies Time series econometrics Cointegrated VAR Recursive methods

a b s t r a c t This paper explores the effects of non-standard monetary policies on international yield relationships. We first document that long-term rates followed a common global downward trend that had already manifested itself prior to the financial crisis. The bondbuying operations (commonly dubbed Quantitative Easing (QE)) of the US Federal Reserve did not disturb this global co-movement – i.e. the global downward trend in interest rates. We model the relationship between USD and euro (riskless) long-term interest rates using a Cointegrated Vector Autoregressive Model (CVAR) employing recursive estimation methods. We find no evidence that QE1 (or the QE episodes) destabilized the transatlantic interest-rate relationship, nor the relationship between interest rates and the US dollar exchange rate. A robustness test using a Vector Autoregressive Model (VAR) with interest rates, inflation rates and output differentials for 11 countries (relative to US) yielded the same result. There is thus little evidence that central bank bond-buying in the US had an independent, distinct impact on US interest rates. Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction Huge adverse shocks generated by the financial crisis caused a deep recession and turmoil in the financial markets in 2008–09. Even reducing policy (short-term) interest rates to zero proved insufficient to stabilize the economy. Central banks around the globe thus resorted to different, so-called ‘non-standard measures’. Regarding the size of the measures undertaken, the Fed was, initially, the most active central bank in implementing several non-standard measures – most notably several rounds of large purchase programs of public sector bonds, which are usually called quantitative easing (QE). The first round of QE (QE1) was announced in November of 2008, mainly with the aim of calming the turmoil on financial markets and thus stabilizing the US economy.1 After the termination

Abbreviations: BoE, Bank of England; BoJ, Bank of Japan; CVAR, Cointegrated Vector Autoregression Model; VIX, CBOE volatility index; ECB, European Central Bank; FOMC, Federal Open Market Committee; GSE, government sponsored enterprise; MBSs, mortgage backed securities; OT, Operation Twist; QE, quantitative easing; VAR, Vector Autoregression Model. q

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

⇑ Corresponding author at: University of Duisburg-Essen, Berliner Platz 6-8, 45127 Essen, Germany.

E-mail addresses: [email protected] (A. Belke), [email protected] (D. Gros), [email protected] (T. Osowski). The start date of QE1 is subject to different interpretations. It is frequently denoted in the literature as having been launched on March 18, 2009, the actual date on which government paper was first purchased. But we feel justified in starting our analysis on the date of its formal announcement in November 2008, in accordance with the advice of an anonymous referee. 1

http://dx.doi.org/10.1016/j.jimonfin.2017.02.011 0261-5606/Ó 2017 Elsevier Ltd. All rights reserved.

336

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

of QE1 in March 2010, QE2 started in November 2010, followed by Operation Twist in September 2012 and an additional round of QE (QE3) in September 2012. The central banks of many other large industrialized countries also implemented similar asset purchase programs: the Bank of England (since 2009), the Bank of Japan (since 2010) and finally the European Central Bank (since 2015). The common aim of these variations of QE has been to put pressure on long-term yields, which in turn are expected to stimulate demand.2 The basic mechanism by which central bank asset purchases lower interest rates is usually supposed to be some portfolio balance effect: as central banks purchase riskless assets, i.e. government bonds, investors who previously held these bonds are forced out of their normal ‘habitat’ and have to buy other, more risky assets (Borio and Zabai, 2016; Deutsche Bundesbank, 2016). These portfolio balance effects have rarely been modelled explicitly, however, and it is thus difficult to determine the exact channels through which QE should be expected to stimulate demand. As the outgoing Chairman of the Fed, Ben Bernanke, quipped on his last days in office: ‘‘The problem with QE is that it works in practice, but it doesn’t work in theory” (Financial Times, 2014). Another channel through which QE might be effective is the exchange rate. The exact channel linking purchases of domestic assets to the exchange rate has rarely been made clear. The implicit assumption is usually that lower interest rates at home should lead to a weaker exchange rate. For the purpose of our discussion later on, we note that QE can be expected to influence the exchange rate only if the impact of the QE operation on domestic interest rates is larger than any reaction of foreign interest rates. In other words, the exchange rate channel should work only if QE affects international interest rate differentials. A key problem in estimating the impact of QE is that asset markets tend to anticipate future policy actions. Current bond prices (and thereby long-term interest rates) as well as exchange rates are often said to be more affected by expectations about the future than by current economic conditions. The conjecture by Deutsche Bundesbank (2016) therefore, is that the announcement of a program can have a stronger impact than its actual implementation. The announcement effects seem to be large, but if they are not permanent, there is no effect. Measuring the longer-term impact of QE on interest rates and exchange rates is thus an inherently difficult exercise. One approach used in a number of cases has been to study the behavior of interest rates around the announcement of several QE episodes. For example, Borio and Zabai (2016), Thornton (2013) and Deutsche Bundesbank (2016) provide surveys of these ‘event studies’. These studies have generally concluded that QE did have a significant impact on interest rates in the US in the sense that they find that long-term US interest rates tended to fall by a substantial amount at or around the same time as the announcement dates. Another, less-often used approach is to construct a macroeconomic counterfactual. But in this case one has to make many assumptions about how asset prices such as the exchange rate and the interest rate would have evolved in the absence of QE.3 We propose a different approach to test the hypothesis that large-scale asset purchases had a separate, identifiable impact on long-term interest rates in the US. We estimate the cointegrated relationship between US and euro-area interest rates, and then test whether one finds a structural break in this relationship around the time that QE was undertaken in the US. To our knowledge, this paper is one of the first that tries to test empirically whether QE has changed economic relations in international financial markets (Taylor, 2016). Another rare example is Thornton (2014a), who did this by looking at the effect of QE on the difference between the US 10-year Treasury yields and the 10-year sovereign yields for Germany, France and the UK. Thornton argued that if QE affected US long-term rates, the spread between the sovereign yields of countries that did not engage in QE and those in the US would have increased significantly and persistently following the Fed’s first QE announcement in November 2008. He showed that the spreads actually declined, that is, foreign yields fell relative to the US yield. Thornton tested for a structural break in the relationship using the Bai-Perron test. He found no statistically significant break for either Germany or the UK, but did find a statistically significant break for France, which occurred later and coincided with the European financial crisis. He noted that the results could be affected by differentials in these economies’ inflation and output performance, so he repeated the test using real rate differentials and found qualitatively similar results. He also showed that these countries did not have significantly poorer economic growth and concluded that QE had no effect on US long-term rates. Our paper uses a cointegration approach to analyze whether the Fed’s QE1 has caused a structural change in the USEuropean interest-rate relationship.4 However, we accept that the stochastic properties of interest rates are always an issue (as the discussion of it in our paper illustrates). So some academics may argue that interest rates are I(0), no matter what formal tests may show. This could possibly be the case in the low-interest rate environment that we have faced for several years now and which is part of our estimation period. Hence, strictly speaking, to test whether QE affected interest rate differentials in a particular way does not require the cointegration exercise conducted in this paper. Indeed, when we perform the test on other interest rates in Section 4.2.4, we do not use the cointegration approach, but rather follow the approach of Thornton (2014b). We estimate bilateral VAR models for a large amount of countries, including cross-country differences in real growth, inflation 2

There is an inherent contradiction here. If QE strongly increased expectations about future inflation and growth, nominal yields should actually increase. For a counterfactual analysis in macro-econometrics with an empirical application to QE, see Pesaran and Smith (2012). 4 Assessing QE effects in a cointegration or error-correction framework is not far-fetched. See, for instance, Chen et al. (2015), Cloyne et al. (2015), Troug and Murray (2015), and Saeidinezhad (2015). 3

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

337

and interest rates. Before, we test for structural breaks in the interest rate equation around the time of QE1’s announcement and implementation. The main advantage of the cointegration test methodology is that it allows us to test the hypothesis in several different ways and more rigorously. Moreover, it allows us to carry out some counterfactual tests, which we do later in Section 4.2.3. Even more important, if interest rates are I(1), the difference between the US and the euro-area interest rate is not necessarily stationary, except if both variables are cointegrated. Hence, we must test for cointegration as a preliminary step. At this point, we feel justified to proceed with a cointegration exercise. As an empirical approach, we use the Johansen CVAR procedure in order to estimate the long-run relationship between US and European interest rates, also taking developments of the nominal exchange rate into account. After estimating the potential long-run relationship, we employ recursive methods proposed by Johansen and Juselius (2006) in order to check for structural changes with respect to parameter constancy. Our focus on QE1 is motivated by two main arguments. First, the general impression and empirical evidence of QE in the US indicate that QE1 has been the program with the highest impact on financial variables. Therefore, if one aims to set out an independent effect on US interest rates relative to euro interest rates, QE1 might appear to be the natural choice. The second reason is related to the statistical approach of this paper. As recursive tests lose the power to detect structural breaks at the end of the data sample, it would be difficult to reject the assumption of parameter constancy/structural constancy for Operation Twist, QE3 and partly QE2, even if a structural break were present. This paper is organized as follows. The next section provides a description of the large asset purchase programs of the Federal Reserve and documents the continuing co-movement of interest rates across the Atlantic. The data used and the estimation approach are described in Section 3. The estimation cycle and the results are presented in Section 4, including some robustness checks. Section 5 sums up our results and provides an outlook for further research.

2. Quantitative easing and global financial markets 2.1. Bond purchase programs in the US In November 2008, the Federal Reserve announced the first round of asset purchases (QE1). These purchases were to include government-sponsored enterprise (GSE) debt and agency mortgage-backed securities (MBSs) of up to $600 billion.5 The motivation given was that the spread on agency bonds had increased, thus making house purchases more expensive.6 After announcing the intention to extend the program in January 2009, the Federal Open Market Committee (FOMC) decided to purchase an additional $750 billion in (agency) MBSs, $100 billion in agency debt, and also started to purchase long-term Treasury securities worth $300 billion in March 2009. These three different announcements explain why it is not possible to assign a unique date to the start of QE1. We keep to the widely accepted practice of dating the announcement of QE1 to November 2008. In total, the Fed purchased assets worth $1.75 trillion between November 2008 and March 2010. Its balance sheet more than doubled over this period (see also Borio and Zabai, 2016). However, a key, often overlooked aspect of this so-called QE1 operation was that the balance sheet of the Federal Reserve was supposed to remain unchanged. During 2008 the Federal Reserve had expanded its balance sheet by engaging in foreign currency swaps, providing money market funds and banks with liquidity. As can be seen in Fig. 1, these assets were reduced to almost zero over the period of QE1 and substituted with bonds, mostly mortgage-backed securities backed by government-sponsored enterprises (Stroebel and Taylor, 2012). In October 2010, the FOMC announced the second round of QE (QE2). It contained purchases of $600 billion worth of treasuries and was finished in June 2011. A few months later, the implementation of a maturity extension program, the so-called Operation Twist (OT, was launched (Borio and Zabai, 2016). By purchasing $400 billion worth of Treasury bonds with maturities of 6–30 years and selling bonds with maturities of less than 3 years, the FOMC intended to extend the average maturity of the Fed’s portfolio. Eventually, the third round of QE (QE3) started in September 2012. It targeted a monthly purchase of $85 billion through the purchase of mortgage-backed securities ($40 billion) and longer-term Treasury securities ($45 billion). In contrast to the other programs, the continuation of QE3 was tied to the improvement in the labour market. Overall, the Fed balance sheet increased by about $3.5 trillion (roughly 20% of GDP). As shown in Fig. 1, the balance sheet of the Federal Reserve has now reached over $4 trillion, or close to 25% of GDP. Two assets dominate the asset side: Treasury securities (about $2.5 trillion) and federal agency securities ($22.4 billion).7 The latter are all guaranteed by the federal government of the United States. It is thus formally true that the Federal Reserve has intervened in the market for securitized mortgages, but it has bought only securities guaranteed by the government. In terms of the evolution of the balance sheet, one can clearly see the impact of QE 1, 2 and 3. 5 6 7

Analogous to our paper, Stroebel and Taylor (2012) investigate the effectiveness of the MBS part of QE. See https://www.federalreserve.gov/newsevents/press/monetary/20081125b.htm. MBS, however, are about $1.8 trillion.

338

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

QE2

QE1

50,00,000

QE3

45,00,000 40,00,000 35,00,000

11% of GDP

30,00,000 25,00,000 20,00,000 15,00,000

14% of GDP

10,00,000 5,00,000

U.S. Treasury Securities

Federal Agency Securities

06/15

12/14

06/14

12/13

12/12

06/13

06/12

12/11

06/11

12/10

06/10

12/09

06/09

12/08

06/08

12/07

06/07

12/06

0

Other assets

Fig. 1. Federal Reserve balance sheet ($ mil). Notes: Percentage refers to 2014 GDP. Federal Agency Securities represent debt securities and mortgagebacked securities backed by government-sponsored enterprises. Source: Federal Reserve.

2.2. Literature review While numerous empirical papers focus on the domestic effects of QE, the empirical evidence on international effects is still limited. In this section, we provide a survey of the literature that has attempted to quantify the effects of the Fed’s QE programs. As our paper focuses on the effects of QE on interest rate relationships, the following literature review primarily focuses on the impacts on financial markets – namely interest rates and exchange rates. The purpose of QE is to influence long-term interest rates. But this implies that QE should have a strong impact on bond prices. In order to be effective QE must have a persistent if not permanent effect on interest rates.8 Most studies focus on its effect on long-term yields (especially Treasury bond yields).9 Gagnon et al. (2011) investigate the effects of QE1 by using event study as well as time series methods. They find that the cumulative effect of LSAP (large-scale asset purchase) announcements on yields of US Treasury bonds as well as US agency was a reduction of up to 150 basis points. By scaling the Fed purchases to ‘10-year equivalents’, the authors measure the duration that the Fed removed from the market. Across the three asset classes that were purchased during QE1, the purchases account for more than 20% of the total outstanding 10-year equivalents. Gagnon et al. (2011) argue that by reducing the net supply of assets with long duration, the program was successful in reducing the term premium by 30 to 100 basis points. In accordance with their results, the authors highlight the importance of the portfolio balance channel relative to the signaling channel.10 Thornton (2014b) shows that their time series results from a common trend in the data and that when the trend is accounted for there is no time series evidence of a portfolio balance effect. While Christensen and Rudebusch (2012) find similar cumulative reductions using an event study, their results stress the importance of the signaling channel.11 Wright (2012) generates interesting insights using a structural VAR with daily data to identify monetary policy shocks. While he finds significant effects on long-term yields, these effects evaporate quite fast, with an estimated half-life of two months. These results seem to be more consistent with ours. Further evidence is presented by Hamilton and Wu (2012) who use a term-structure model to predict the effect of a change in the central bank’s asset structure (short-for-long-term debt swap) and also indirectly the effect of buying $400 billion in long-term Treasuries.12 Their results are much lower compared to the event studies mentioned, as they find that such a policy would cause a reduction of the 10-year rate of (only) 13 basis points. This seems to be one of the lowest values achieved in the literature. Results of similar magnitude have been obtained by Neely (2015) and Meyer and Bomfim (2010).13 Chung et al. (2012) find effects that are not negligible. Based on counterfactual model simulations, they find that the past and projected expansion of the Federal Reserve’s securities holdings since late 2008 are roughly equivalent to a 300 basis point reduction in policy interest rates (from 2009 until 2012).

8

We are grateful to an anonymous referee to provide this insight to us. See Doh (2010), Gagnon et al. (2011), Krishnamurthy and Vissing-Jorgensen (2011), Meaning and Zhu (2011), D’Amico et al. (2012), D’Amico and King (2013) and Li and Wei (2013). For surveys of studies assessing the impact of QE on domestic long-term US yields see Borio and Zabai (2016), De Santis (2016) and Deutsche Bundesbank (2016). Few contributions analyze the impact of QE on corporate credit. See Gilchrist and Zakrajsek (2012). 10 For a coherent description of the portfolio balance versus the signaling channel, see Borio and Zabai (2016) and Deutsche Bundesbank (2016). 11 Further studies employing the event study methodology are Krishnamurthy and Vissing-Jorgensen (2011) and D’Amico and King (2013). 12 The purchase amount roughly corresponds to the amount of Treasury bonds bought during QE1. 13 For further evidence, see Krishnamurthy and Vissing-Jorgensen (2011). 9

339

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

14 12 10 8 6 4 2 0

United Kingdom

United States

2016

2015

2014

2012

2013

2011

2010

2009

2008

2007

2006

2005

2004

2002

2003

2001

2000

1999

1998

1997

1996

1994

1995

1992

1993

1991

1990

-2

Germany

Fig. 2. Long-term interest rates in major currency areas since 1990. Data source: OECD.

Liu et al. (2014) find smaller effects. By using a change-point VAR model, they estimated that the Fed’s asset purchase program reduced 10-year spreads by an average of 90 basis points over the crisis period.14 Regarding the effects on international financial markets, most papers find cross-border effects, as well as effects on exchange rates. Fratzscher et al. (2013) examine the international effects of QE1 and QE2. They find that QE1 was effective in lowering sovereign yields and raising equity markets in the US and abroad. According to their results, QE1 might have generated a safe haven effect causing a strong global rebalancing of portfolios out of emerging markets and into US equity and funds, thereby putting upward pressure on the US dollar (USD). However, regarding the effects of QE2, the authors find that this program has generally been ineffective in lowering yields worldwide and has caused sizeable capital outflows, mainly into emerging economies, and thereby marked a USD depreciation. Neely (2010, 2015a) puts more weight on the effects of QE on treasury yields of developed countries.15 Using an event study as well as a portfolio-balance model, Neely (2010, 2015a) finds substantial evidence that QE1 announcements have reduced sovereign yields in the US and abroad. Furthermore, Neely (2010, 2015a) comes up with significant evidence that QE has generated a general depreciation of the USD. Bauer and Neely (2014) use dynamic term structure models to uncover whether international yields have declined as a result of signaling or portfolio-balance effects. They find that the relative importance of the signaling channel increases with an economy’s sensitivity to signals from conventional US monetary policy. Consistent with the notion that Canada is highly sensitive to US monetary policy, the authors find large signaling effects for Canadian treasury yields. For Australian and German treasury bonds, the authors find especially large portfolio balance effects. 2.3. Global financial markets and national QE Our contribution is to go beyond the usual approach of looking for a link between asset purchases and financial market variables at the national level. In reality, however, financial markets in advanced countries are very open and highly integrated. This implies that one should not just look at US financial variables when trying to measure the impact of LSAPs by the Fed. However, disentangling the impact of QE in globally integrated financial markets is much more difficult, as one needs to adopt a comparative approach. The first key observation is that in reality (long-term) interest rates have followed a common long-term trend across major currency areas. Global financial markets are highly integrated and (long-term) rates have been highly correlated across advanced economies, not only along a downward trend, but also during cyclical ups and downs, as illustrated in Fig. 2. The correlation is too tight and has lasted too long to be just a coincidence (Gros et al., 2015). The most obvious interpretation is that there is a global capital market that is integrated across currency boundaries as emphasized also by Belke et al. (2010) and Rey (2013). The effectiveness of large-scale asset purchases by the Federal Reserve should thus not be measured simply by the associated fall in US interest rates, but by a fall in the interest rate differential between the US and other major markets. This contribution focuses on the euro area as the best comparator since it is of a similar size to the US. Fig. 3 displays the evolution of the long-term (10-year) interest-rate differential between the US and the euro area (proxied by the main riskless rate, i.e. the German rate) between 2007 and 2014. 14 For the definition of a change-point structural VAR in the context of assessing the economy-wide effects of the BoE’s quantitative easing, see Kapetanios et al. (2012). 15 Neely uses data for the US, Australia, Germany, Japan and the UK.

340

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

1.0

QE1

QE2

Twist

QE3

0.5 0.0 -0.5 -1.0 -1.5

07/15

01/15

07/14

01/14

07/13

01/13

07/12

01/12

07/11

01/11

07/10

01/10

07/09

01/09

07/08

01/08

07/07

01/07

-2.0

Fig. 3. Transatlantic long-term interest rate differential from 2007. Notes: The difference between the German and US long-term interest rates is calculated by deducting the US rate from the German rate. Long-term interest rates refer to the monthly average government bonds maturing in about 10 years. The vertical lines indicate the announcements by the Fed of the different quantitative-easing measures. Source: Own elaboration based on data from OECD.

Since the ECB only undertook QE at the very end of this period, one would expect that the repeated round of large-scale asset purchases by the Federal Reserve should have resulted in a lowering of long-term US rates relative to euro area rates; i.e. the differential should have gone up after each round of QE. The opposite has been the case, however: US rates increased relative to euro area rates if one compares the period just before QE1 (say May 2008) to January 2014 (i.e. long before it could have been anticipated that the ECB would also eventually engage in large-scale purchases of government bonds). Over this period, the Federal Reserve bought bonds worth over 20% of US GDP in total, but US interest rates actually increased (slightly) relative to euro area rates. The event study methodology would concentrate on the changes immediately following the announcement, i.e. any jumps just after the vertical lines indicating the announcement dates (Deutsche Bundesbank, 2016; Thornton, 2013). It is apparent that one finds a jump only around the announcement of QE1, not for all the other episodes (QE2, Twist and QE3). After the jump, the interest rate differential regresses again. However, as the analysis so far has solely focused on descriptive methods, we will now analyze the relationship between US interest rates, European interest rates and the nominal exchange rate in a more sophisticated econometric approach in Sections 3 and 4. 3. Data and the empirical approach To implement our research strategy we had to choose, interest rate measures for the US as well as for the euro area. The decision for the US is straightforward, as treasury bond yields are the most common choice. For the euro area, German treasury bond yields are used. We argue that German treasury bonds are considered to be the least risky bonds in the euro area. By choosing German bonds we can avoid the distortions of the cointegration relationships that might be generated by rising risk-premia on the bonds of other countries in the euro area at the time of the European debt crisis. Furthermore, the use of Treasury bond yields is also motivated by the fact that they are often regarded as a benchmark for domestic interest rates. As measures of long-term interest rates, we use 10-year bond yields. We further include the (natural logarithm of the) nominal exchange rate (USD/euro) in our estimations. We employ end-of-month data between 2002:01 and 2014:12. The econometric framework applied is a cointegrated vector autoregressive (CVAR) model, which allows us to model the impact of domestic interest rate shocks on foreign interest rates and the exchange rate, while taking care of the feedback between the variables. Our choice is also based on the CVAR’s feature to avoid an a priori division of variables into exogenous and endogenous. As we include interest rate measures for the US and the euro area, as well as the nominal exchange rate, any ex-ante causality classification would be arbitrary.

DX t ¼ PX t1 þ C1 DX t1 þ    þ Ck1 DX tk1 þ uDt þ et ;

t ¼ 1; . . . ; T

ð1Þ

where X t is a vector containing the p variables of interest; Dt is a vector of deterministic variables and et is a vector of Gaussian errors. Eq. (1) presents the vector error correction (VECM) representation of the VAR model. The VECM form of the model gives an intuitive explanation of the data, separating long and short-run effects. While Ci contains the short-run information, P contains the long-run relationships. Based on the insights of Stock and Watson (1988) that a cointegrating relationship represents that two or more time series share a common stochastic trend and assuming that our variables are I(1), the rank (r) of matrix P has to be reduced (r < p). The reduced rank matrix can be factorized into two r x p matrices a and b (P ¼ ab0 ). The factorization provides r stationary linear combinations of the variables (cointegrating vectors) and p  r common stochastic trends of the system. In accordance with our theoretical approach, if QE1 was effective in reducing long-term interest rates, it should have an effect on US interest rates for which the global downward trend in interest rates represented by the German yield development does not account. Therefore, QE might have had a separate, identifiable and persistent impact on long-term interest

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

341

rates in the US, which should show up as a break in the long-run relationship. We thus check for a potential structural break around the time of the Fed’s announcements of QE1. We use a large set of recursive techniques as proposed by Johansen and Juselius (2006) to check for structural changes in the relationship. The fundamental idea of recursive testing is to start with a baseline model estimated for a subsample and to gradually extend the endpoint of the sample until the full sample is covered. After every extension of the sample, the test statistics are re-estimated. The recursive methods used are: (1) The log likelihood test, which is a broad test of the general appropriateness of the model and is similar to the recursive Chow tests used in single equation models; (2) Recursive tests based on the eigenvalues, which allow us to obtain detailed information about the constancy of the individual cointegration relations, and (3) Recursive tests of the cointegration space. Because the eigenvalues are a quadratic function of a and b, we use additional tests that focus on spotting non-constancy in the b structure. 4. Empirical results 4.1. Unit root tests The first step of our analysis is to determine the order of integration of the time series. Augmented Dickey Fuller (ADF) and Phillips Peron (PP) tests are used. In order to generate robust results, we perform different specifications regarding deterministic components. The results indicate that the levels are at least integrated of order one as the time series possess at least one stochastic trend. With regard to the results of the first differences, the tests propose that they are (trend-) stationary. Therefore, the results suggest that the time series in levels are I(1).16 As described by Perron (1989), structural breaks might have strong effects on the results of unit root tests, sometimes leading to wrong implications generated by tests. In order to strengthen the robustness of our unit root tests, Zivot-Andrews tests are used, which allow for a single break in the intercept, the trend or both (Zivot and Andrews, 1992). The results confirm the outcome of the ADF and PP tests. Despite our test results, the I(1) property of interest rates is controversially discussed in the literature. However, nonstationarity is regarded as a sample property by quite a few applied econometricians. It is found by many authors such as Campbell and Shiller (1987). Stochastic trends in interest rates are often driven by business-cycle shocks. An important drawback of I(0) models is that they imply long-term rates that are not volatile enough. Empirical studies analyzing exchange rate behavior and monetary policy shocks usually adopt interest rates as I(1) variables, at least in subsamples. Prominent examples that we would like to follow in this context are Dickey et al. (1991), Dees et al. (2005), Stock and Watson (1988) and a lot more. 4.2. Cointegrated VAR estimations 4.2.1. Estimation of the long-run relationship We focus on the relationship between long-term (10-year) bond yields and neglect further real variables such as real GDP, since we are mainly interested in the impact of QE on financial markets. Furthermore, if a cointegrating relationship can be detected using a sub-system, the long-run relationship should also be present in a larger model that includes additional variables. In this regard, our model M 10Y contains the following variables: t

M10Y ¼ ðUS10Y; Ger10Y; LEXÞ0t t

ð2Þ

Regarding the specification of the model, we include an intercept in the unrestricted VAR and the cointegrating space. Because the model resembles the Uncovered Interest Parity (UIP), there appears to be no theoretical reason for a linear trend in the cointegrating space.17 We include three lags in the unrestricted VAR. Although the information criteria suggest two lags, the results of the residual analysis improve considerably if an additional lag is included. For the following dates, dummy variables are included into the VAR-equations: 2004:04, 2008:10 and 2008:12. However, no dummy variable enters the cointegrating vector(s). In order to avoid bias of the trace test, the model has to be well specified, especially regarding residual autocorrelation and the normality of the residuals. Diagnostic tests of the residuals show no evidence of non-normality or ARCH-effects and only weak evidence of autocorrelation.18 Overall, we conclude that the model is well-specified. The results of the trace tests are presented in Table 1.19 As indicated by our tests, the hypothesis of no cointegrating relationship (r ¼ 0Þ is rejected at the 5%-level. However, the hypothesis r 6 1 cannot be rejected. Therefore, the results clearly indicate the presence of one single cointegrating relationship. Therefore, the rank of the P-matrix is restricted to one. 16

The results of our unit root tests are available on request. While the results of the LR-Test of Exclusion do not recommend excluding a deterministic trend from the cointegrating space, its inclusion does not change the results of our analysis. After imposing over-identifying restrictions, the deterministic trend becomes insignificant and is therefore left out. 18 The results of the residual analysis are available on request. 19 We simulate the asymptotic distribution of the trace test. The following settings are used: length of random walk: 400, number of replications: 2500. 17

342

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

After the rank is determined, restrictions placed on the b-vector can be used to test for specific relationships suggested by economic theory. We restrict the coefficients of the interest rates to have the same magnitude, but different signs. Therefore, we check for a relationship of the interest rate differential between German and US yields and the exchange rate.20 The model is accepted (p-value ¼ 0:785) at the 10% significance level. The results are presented in Table 2. According to our estimations, we obtain the following long-run relationship:

0:400 þ 0:128  ðGer10Y  US10YÞ ¼ EXR

ð3Þ

The estimated relationship indicates that a reduction of the US (German) yield should lead to a depreciation (appreciation) of the US dollar vis-à-vis the euro, which is in line with economic theory. Movements of the exchange rate and the German yields thus reduce deviations from the long-run relationships. Based on our results, the estimated alpha coefficient of the US interest rate is not significantly different from zero. Therefore, the US interest rate affects the relationship without receiving significant feedback. In this respect, the US can be regarded as Stackelberg leader of the interest-rate relationship. 4.2.2. Did QE1 cause a structural break? Even though the trace tests recommend one cointegrating relationship, this does not exclude the possibility that the model suffers from parameter non-constancy. The purpose of recursive tests is to identify whether structural breaks are present. Regarding the research question of the paper, the time period between 2008:11 and 2010:03 is of importance in this respect. In particular, two dates might be of particular interest regarding the possibility of a structural break caused by QE1: firstly, the announcement of QE1 in 2008:11 and secondly, the FOMC’s decision to buy Treasury bonds and the quantitative expansion of the QE program in 2009:03. For the recursive tests, the baseline sample contains data from 2002:04 to 2005:12.21 Additionally, we performed backward recursive estimations for which the base sample contains data from 2011:12 to 2014:12.22 The log-likelihood test presented in Fig. 4 indicates a limited amount of instability starting around the beginning of the subprime mortgage crisis in 2007. The test statistic slightly exceeds the black line, which represents the critical value at the 5% significance level, and only for a brief period of time. After mid-2008, the test statistic stays below the critical value. In this regard, we find no evidence of parameter non-constancy around or after the introduction of QE1 in 2008:11. Further evidence of a structural break starts to appear in of 2011, but fails to reject the null hypothesis of parameter constancy. Fig. 5 depicts the recursively estimated trace test statistics explicitly checking the stability of the cointegrating relationship (and thereby the eigenvalues). In this regard, the upper graph displays the development of the statistical significance of the estimated long-run relationship over time. We observe some signs of instability in the estimated cointegration. Around the announcement and implementation of QE1, the cointegration relationship shows an almost linear development, indicating no signs of structural changes caused by QE1. The largest amount of instability is found around mid-2010, which corresponds to the beginning of the European debt crisis, as the cointegration relationship gradually declines and even becomes insignificant at the 5% level. Eventually, in mid-2011, the relationship once again starts to increase until the end of the sample. We also tested the fluctuations of the eigenvalues but find no sign of instabilities. As a next step, we exclusively focus on the components of our long-run relationship. The main idea of the ‘test of known ~ based on a chosen reference period and to test its stability after gradually extending the data sample. beta’ is to estimate b

We chose the reference sample as 2002:04 to 2006:12.23 The results are presented in Fig. 6.24 A first period of instability can be established around the beginning of 2008, around the meltdown of Bear Stearns. However, stability is not rejected. Surprisingly, we once again find no evidence of a structural meltdown around the peak of the financial crisis in late 2008. Correspondingly, we also find no evidence of any QE1 impact. The most striking result is the large amount of instability that begins to develop around the beginning of the European debt crisis in mid-2010. From this data point onwards, the test results indicate that the long-run relationship has changed in comparison to the pre-crisis era, as indicated by the rejection of the null hypothesis. With respect to the timing of the detected structural break, it appears most likely to have been associated with the start of the European debt crisis. To sum up our results, we detect some empirical evidence of structural changes in the model. First of all, we come up with limited evidence that there is some degree of instability prior to the outbreak of the global financial crisis (between mid2007 and mid-2008). Most importantly with regard to our research question, we do not find evidence that the announcement or the implementation of QE1 generated a structural change in the relationship of ‘risk-free’ US and European bonds. We find the highest amount of instability only in mid to late 2010. But the beginning of the European debt crisis might explain this result better than the asset purchases of the Federal Reserve. Although QE2 was announced in November 2010, it appears to be unlikely that the structural break is generated by the Fed’s announcement of QE2, because it is 20 We test for proportionality between the interest-rate measures by restricting the coefficient of the exchange rate variable to zero. The test of the restricted model rejects the restriction (p-value ¼ 0:002). Therefore, the exchange rate appears to be an important component of the long-run relationship. 21 In order to check for robustness, we also used the following baseline samples: 2002:04 to 2004:12 and 2002:04 to 2006:12. We obtained similar results, which are available on request. 22 The results of backward recursive tests do not reveal contradictory evidence for the following tests. In order to keep the amount of figures manageable, the results are available on request. 23 Because the results can be sensitive to the chosen reference sample, we performed several tests using the following reference sample periods: 2002:04 to 2004:12, 2002:04 to 2005:12, 2002:04 to 2007:06 and 2002:04 to 2007:12. However, the test does not appear to be sensitive to the reference sample. 24 While the XðtÞ model contains short-run and long-run information of the data, the R1 ðtÞ only contains long-run information.

343

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349 Table 1 LR trace test for the unrestricted VAR (3)-model, M 10Y . t pr

r

Eigenvalue

Trace

95% Crit. value

p-value

3 2 1

0 1 2

0.148 0.084 0.016

40.359 15.773 2.455

34.565 19.932 9.219

0.011 0.177 0.682

Notes: p -values for testing the null hypotheses of H0 : r ¼ 0 and H0 : r 6 i þ 1, i ¼ 0; 1; 2, respectively, for different ranks r.

Table 2 The over-identified long-run cointegration relations for r ¼ 1, M 10Y . t

^1 b

US10Y

Ger10Y

EXR

Constant

0.128 [2.041]

0.128 [2.041]

1 [NA]

0.400 [9.691]

a^ 1

DUS10Y

DGer10Y

DEXR

0.145 [1.238]

0.193 [2.002]

0.052 [4.021]

Test of restricted model: CHISQR(1) = 0.074 [0.785]

R1(t)

12/14

06/14

12/13

06/13

12/12

06/12

12/11

06/11

12/10

06/10

06/09

12/09

12/08

06/08

12/07

06/07

12/06

06/06

2.0 1.8 1.6 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0

12/05

Notes: The first column reports long-run coefficients b. The second column shows the adjustment coefficients a. t-values in brackets.

5% Critical Value

Fig. 4. Test for constancy of the log-likelihood, M10Y . Source: Own calculations. t

1.50 1.25 1.00 0.75 0.50

5% Crical Value

H(0)|H(3)

Fig. 5. Recursively estimated trace test statistics,

H(0)|H(3)

M10Y . t

12/14

06/14

12/13

06/13

12/12

06/12

12/11

06/11

12/10

06/10

12/09

06/09

12/08

06/08

12/07

06/07

12/06

06/06

0.00

12/05

0.25

H(0)|H(3)

Source: Own calculations.

considered to be the least effective QE program conducted by the Fed. Therefore, if our tests generate no significant evidence of a structural break generated by QE1, it appears unlikely that the QE2 announcement can be held accountable for the structural break in 2010. Uncertainty about the future of the euro and the probability of a breakup of the euro area might have destabilized the relationship between the nominal exchange rate and the trans-Atlantic interest rate differential. The euro crisis is widely

344

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

2.5 2.0 1.5 1.0 0.5

X(t)

12/14

06/14

12/13

06/13

12/12

06/12

12/11

06/11

06/10

5% Crival Value

12/10

12/09

06/09

12/08

06/08

12/07

Fig. 6. Test of known beta,

M10Y . t

06/07

12/06

06/06

12/05

0.0

R1(t)

Source: Own calculations.

believed to have generated a flight of capital from the euro area periphery to Germany, the so-called safe haven effect or flight to safety. This might have depressed German yields and could thus have led to the impression that QE in the US had an equal impact on yields in the US and the risk-free part of the euro area. However, the euro crisis did not coincide with the QE operated by the Fed. One could thus argue that if the safe haven effect had really been so important, it should have created a structural break in the co- movements of US and euro area or German interest rates. Moreover, it would a priori not make much sense to look at the impact of QE1 on the interest rate differentials among US bonds and bonds of an only double-A rated country such as Italy (for which risk differences would also play a role). In order to verify the robustness of the results presented, additional estimations were conducted using five- and sevenyear yields. The estimations generated almost identical results. Therefore, our results do not appear to be sensitive regarding the specific choice of the yields. Furthermore, we checked the robustness by including the VIX as an exogenous variable in order to correct for the risk perception of US financial markets. Again, the trace test recommends the presence of one cointegrating relationship (r ¼ 1Þ. While the VIX significantly enters the short-term dynamics of the model, the variable is not significant in the long-run relationship. Once again, we find no evidence that QE generated a structural change in the transatlantic interest-rate relationship.25 4.2.3. Robustness I: power of the test A key issue is of course the power of our approach. We find that it is only by end 2010, that one would have been forced to reject the null hypothesis that QE did not change the transatlantic interest rate relationship. The question that remains is what kind of impact QE would have to have had in order to show a statistically significant break. A priori this seems difficult. The daily standard deviation (of changes) in the US 10-year rate is about 8 basis points. Assuming changes are uncorrelated (as they should be in efficient markets, which is the reason why announcement dates are usually used), it follows that the two standard deviation intervals after one month would be approximately 70 basis points.26 The (differential) impact of US QE on US interest rates would thus seemingly have been greater than this order of magnitude. However, our approach tests for a deviation from the long-run relationship, and could thus potentially signal a break even for smaller magnitudes. In order to test for the power of our approach, we thus performed a simple counterfactual experiment. We re-computed our test statistics with a slightly modified data set: while the actual data for euro interest rates and the exchange rate are used, the US interest rate data are modified: the US 10-year rate is reduced by a fixed amount starting in November of 2008 (the conventional announcement date of QE1). This would mimic the idea that (the announcement of) QE1 reduced US interest rates (relative to euro interest rates) permanently by 10–50 basis points (see Fig. 7) as implicitly assumed by event studies. We preferred not to phase-in the interest rate reduction gradually since the literature universally assumes that interest rates should jump on the announcement date(s) and not be much affected by the actual implementation (Deutsche Bundesbank, 2016). We concentrate on the test statistics with a known beta. The results are shown in Fig. 7 below. We show five different lines in Fig. 7. One line (the lowest one) uses the original data and is thus identical to the one in Fig. 6 labelled R1. The other lines, corresponding to the use of modified US interest rate paths, are all higher. The fact that all the other lines are higher shows that any counterfactual permanent reduction in US interest rates increases the probability of finding a structural break. 25

Results of the robustness checks are available on request. The predictive confidence interval is given by 2 ⁄ sqrt(N) ⁄ 8 basis points for day N. With N = 20 this works out to about 70 bps. We are grateful to an anonymous referee for pointing this out to us. 26

345

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

25

20

15

10

5

5% Critival Value

10% Critival Value

US 10y Yield

US 10y Yield - 10 BP

US 10y Yield - 20 BP

US 10y Yield - 25 BP

12/14

06/14

12/13

06/13

12/12

06/12

12/11

06/11

12/10

06/10

12/09

06/09

12/08

06/08

12/07

0

US 10y Yield - 50 BP

Fig. 7. Test of known beta, M 10Y recomputed for modified US interest rate data. Notes: This is a modified form of Fig. 6. The lowest line is identical to the one t in Fig. 6, labelled R1. Other lines correspond to the use of modified US interest rate paths.

Inspection of Fig. 7 shows that a permanent reduction in US rates of only 25 basis points would have resulted in our test statistics signaling a structural break at the 10% level by early 2009 and at the 5% level by the end of that year. One should keep in mind that 25 basis points used to be the smallest step when central banks moved their policy interest rates. The next line (labelled 50 bps) shows that a permanent reduction of the 10-year yield would have forced one to reject the null hypothesis of no (differential) impact of QE already by early 2009, using the conventional 5% level. Had the impact of QE been even larger, as claimed by some, our test would have picked this up even earlier, as one can see from the upper most line in this figure. We thus conclude that our approach would have been powerful enough to detect even minor changes to the long-term relationship between long-term interest rates across the Atlantic. 4.2.4. Robustness II: VAR analysis Cointegration analysis is based on the assumption that our time series are I(1). As mentioned in Section 4.1, there are several (theoretical) arguments why interest rates can not contain a stochastic trend. If one assumes that interest rates are zero, it is necessary to follow Thornton (2014a). Therefore, we proceed by estimating bilateral VARs in order to generate further evidence which is not based on the assumption that interest rates are (I(1)). This paper is not the first to test the effectiveness of QE by seeing whether it affected interest rate differentials. Thornton (2014a) used a similar approach by looking at the effect of QE on the difference between the US 10-year Treasury yields and the 10-year sovereign yields for Germany, France and the UK. We have concentrated so far on German interest rates as the proxy for the euro-area riskless rate because most of the literature focuses on the impact of QE on this type of rate. However, it might also be interesting to check for the impact on other sovereign bond yields as this will shed light on the impact of QE on the risk premia embedded in lower-rated countries, such as Italy. We proceed by using VAR analysis including inflation, real GDP growth, bond yields and the exchange rate for country pair analysis. We regard each country relative to the US with data stated in country differences.27 The additional variables are included in order to account for other factors that can be expected to affect interest rates. A sudden increase in real GDP growth or the inflation rate in one country might, ceteris paribus, increase the national interest rate and thereby change the interest rate differential (Bridges and Thomas (2012), Kapetanios et al., 2012). An additional advantage of VAR analysis is the possibility of using standard approaches (e.g. Bai-Perron) to test for structural breaks. We estimate models based on monthly28 and quarterly models. For our monthly estimations, our sample contains data between 1999:01 and 2015:12.29 The quarterly estimations are based on data between 1984:01 and 2015:04. We estimate VARs for the following nine countries: Australia, Canada, France, Germany, Italy, Japan, Sweden, Switzerland and the United Kingdom. The following variables are used: 27

We are grateful to an anonymous referee for pointing out the necessity of doing this. We use the temporal disaggregation method presented by Litterman (1983) in order to generate monthly GDP measures which is robust against nonstationary variables. The time series used in order to disaggregate quarterly GDP to monthly GDP measures are industrial production and retail sales. 29 For the Italian model, the sample is reduced to 1999:01 to 2010:12 due to large evidence of autocorrelation and ARCH-effects when the whole sample is used. The Australian model including core inflation starts in 2003 due to data limitations. 28

346

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

(1) the difference between the year-on-year growth of real GDP in the US and the country under observation, (2) the difference between the year-on-year change of inflation in the US and the country under observation. We use headline as well as core inflation in separate models,30 (3) the logarithm of the exchange rate between the USD and the national currency, and (4) the difference between 10-year interest yields of the US and the country under observation. This variable represents the interest-rate differential. The results of the unit root tests confirm that none of the variables is integrated of order two. However, the exchange rate variables and the inflation variables are predominantly integrated of order one. We chose to estimate the VARs in levels instead of using (stationary) first difference). Sims et al. (1990) show that OLS estimates of VAR coefficients are consistent under a broad range of circumstances even if the variables are non-stationary. According to Hamilton (1994), taking the first differences appears to perform only better when samples are small (and smaller compared to our sample). However, if variables are cointegrated in levels, first difference estimates are biased, because the error correction term is omitted. Because we cannot rule out the possibility of cointegrating relationships and do not aim at explicitly modelling the long-term relations, we estimate the VAR model in levels. As we are interested in the effects on the interest rate differential, we focus our analysis on the corresponding equation in each model. First of all, the inclusion of the additional variables appears to be important as F-statistics for the GDP growth and inflation differentials are significant at the 5% level in every model. The exchange rate variables are also significant in a majority of models. Therefore, these variables explain movements of the interest rate differentials, which should help with the identification of structural breaks. Strictly following Thornton (2014a) and Troug and Murray (2015), we first apply a Bai-Perron test to detect (multiple) structural breaks of the interest rate differential. This methodology is valid under fairly general assumptions on regressors and disturbances (Bai and Perron, 2003).31 The determination is based on (global) information criteria (BIC,32 LWZ33) as well as the sequential and global F-tests proposed by Bai and Perron (1998). The results of the Bai-Perron tests can be summarized succinctly: in none of the estimated equations (country pairs) included in our analysis do we find evidence that QE1 has generated a significant structural break in our monthly VAR models. Our quarterly VAR estimations reveal only evidence for an impact of QE1 on the US-Canadian interest rate differential, as we find evidence of a structural break in 2009:01. However, we find evidence of an impact of the European debt crisis for France and Italy. Based on the global information criteria, we find large evidence for a structural break in the French interest rate equation in mid-2010. Similar evidence is generated for Italy, but it fails to reach significance. Based on these results, we can confirm the results of Thornton (2014a), as we find no evidence (except for Canada) of a structural break during the announcement/execution of QE1, but we do find evidence of a structural change for France during the beginning of the European debt crisis. As a second testing procedure, we present the results of the Chow forecast tests of our monthly VAR models, which test for a structural break at a pre-specified period of time in Figs. 8 and 9. The graphs are obtained by repeating the test procedure for every time period between 2007:06 and 2009:07 and show the corresponding p-values. The horizontal line indicates the 10% significance level. Therefore, a graph below the black line indicates evidence of a structural change/break in our interest rate equation at the 10% level. There seems to be little difference between the results using core inflation (Fig. 9) or headline inflation (Fig. 8). It is apparent that some lines are below the, generous, 10%, threshold only during 2007/early 2008, i.e. long before QE1 was announced. This means that we find evidence of structural instability between mid-2007 and the beginning of 2008 for Australia, Canada, France and Italy. After 2008:04 (still seven months before QE1), we find significant evidence of structural change only for the Australian VAR model, and only for the one using core inflation. Apart from the case of Australia, we find very similar results across both sets of inflation measures, indicating no evidence of a structural break around the announcement of QE or its implementation. We subsequently performed the Andrews-Ploberger test, which is used to test for a single structural break at an unknown point of time within the sample (Andrews and Ploberger, 1994). The results for our monthly VARs using headline inflation indicate only evidence for Canada for which the test propagates a structural break in 2008:12. For Australia, we find evidence of a structural change in 2008:11 when using core inflation. For the remaining and therefore large majority of countries, we find no evidence of a structural change between 2008:10 and end-2009. For completeness we mention that, additionally, we estimated VARs based on quarterly data between 1988:01 and 2015:04. The results (available upon request) were the same: there is no evidence of structural breaks over this period of time.

30 Our measure of core inflation equals headline inflation without energy and food prices for the monthly data. We use the GDP deflator data for the quarterly VARs. 31 Antoshin et al. (2008) argue that the Bai-Perron test performs quite well in samples that exceed 50 data points. This is the case for both our monthly as well as quarterly models. 32 Schwarz-Bayes information criteria. 33 Modification of the Schwarz-Bayes information criteria presented by Liu et al. (1997).

347

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

1.0

0.8

0.6

0.4

0.2

10% Significance Level France Japan

Australia Germany Sweden

06/09

03/09

12/08

09/08

06/08

03/08

12/07

09/07

06/07

0.0

Canada Italy Switzerland

Fig. 8. Chow forecast test (monthly VAR using headline inflation). Source: Own calculations.

1.0

0.8

0.6

0.4

0.2

10% Significance Level France Japan

Australia Germany Sweden

06/09

03/09

12/08

09/08

06/08

03/08

12/07

09/07

06/07

0.0

Canada Italy Switzerland

Fig. 9. Chow forecast test (monthly VAR using core inflation). Source: Own calculations.

5. Conclusions Did the various large-scale asset purchase programs of the Federal Reserve, which started in 2008 (and continued until 2013) generate a structural break in the relationship between European and US long-term interest rates? We consider this the relevant question to ask if one wants to isolate the impact of US QE on US interest rates. The answer we found is that there is little evidence that the Fed’s asset purchases did have a statistically significant impact on the trans-Atlantic interest rate relationship. This result was not due to a low power of the tests. Repeating our procedure with counterfactual paths for US interest rates, we found that our procedure would have signaled a structural break if, starting with the announcement of QE1, US (long-term) interest rates had been lower by 25–50 basis points. Our procedure would thus have picked up even a modest impact of QE1 on the trans-Atlantic interest rate relationship. We also find that the euro area debt crisis resulted in a structural break. It thus had a more destabilizing impact on transatlantic interest rate differentials than QE. VAR tests which take into account the macroeconomic feedback between interest rates, inflation and growth confirm this result: we do not find evidence of structural breaks in cross-country relationships around the time of QE.

348

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

These results are compatible with two views: (a) QE has been ineffective, and, (b) the impact of QE, possibly non-zero, has been exactly the same on both sides of the Atlantic. The latter might be interpreted as a ‘‘perfect interest rate spill-over”.34 However, perfect spill-over is not compatible with the portfolio balance approach, which is the dominant view of how QE is supposed to work. There might be some euro area investors who hold both US treasuries and some US investors who hold euro area government bonds. A change in US rates might thus have some spill-over impact into the euro interest rates. But the biggest institutional investors are usually required to invest in their home currency, and government and most households rarely diversify internationally. It is thus difficult to rationalize from a portfolio-balance point of view that bond-buying by the US Federal Reserve should have exactly the same impact on safe euro area long-term interest rates as in the US. As Taylor (2016) puts it: ‘‘this finding raises doubts about some of the rationales for the effect of QE, including portfolio balance arguments, which would not be expected to have such large global effects.” It is also certainly possible that US QE has strong international effects. But there is a difference between a significant, but partial spill-over effects and a co-movement, which does not seem to be affected at all by QE in one country alone. Moreover, interest rates differentials are the main drivers of short-term currency movements. The observation that QE did not affect transatlantic interest rate differentials makes it difficult to maintain the general presumption that US QE led to a depreciation of the US dollar (at the time of US QE, some spoke of a ‘currency war’). The observation that (US) QE has not affected transatlantic interest rate co-movements is, however, compatible with the view that QE could have been the result of negative shocks to global inflation and/or demand. In this view, QE should be viewed as a (predictable) reaction of central banks to negative inflation and/or demand shocks. These shocks could be global in nature as suggested by the recent empirical work showing that inflation has a strong global component (e.g. Belke et al., 2010; Ciccarelli and Mojon, 2010). But even if the shocks were national in nature (i.e. there could be stronger deflationary forces in the euro area than in the US), their impact on long-term interest rates could be uniform if they are long-term themselves, given that long-term shocks to demand in any large economy would tend to be distributed across the global economy. This view of QE as endogenous would still be compatible with the results from event studies, which generally find some reduction of interest rates around dates when major asset purchase plans were announced. The fact that the Federal Reserve felt it necessary to adopt this unconventional policy tool could be interpreted by investors as new information about how the Federal Reserve views the state of the economy. Given that the Federal Reserve can well be assumed to have a very good view of the US economy, this could motivate investors to modify their own views as well. Acknowledgements We gratefully acknowledge valuable comments from two anonymous referees and our discussant John B. Taylor at the Annual International Conference on Macroeconomic Analysis and International Finance (ICMAIF), Rethymno/Crete, May 28–30, 2016. Moreover, we thank Kenneth West, Makram El-Shagi, Arie E. Gozluklu, David-Jan Jansen and participants in the 2nd HenU/INFER Workshop on Applied Macroeconomics, March 18–19, 2016, Henan University/China, the 20th Annual International Conference on Macroeconomic Analysis and International Finance (ICMAIF), Rethymno/Crete, May 28–30, 2016, and the World Finance Conference (WFC), New York, July 29–31, 2016, for their comments on different versions of this paper. References Antoshin, S., Berg, A., Souto, M., 2008. Testing for structural breaks in small samples. In: IMF Working Paper WP/08/75. International Monetary Fund, Washington, D.C., March. Andrews, D., Ploberger, W., 1994. Optimal tests when a nuisance parameter is present only under the alternative. Econometrica 62, 1383–1414. Bai, J., Perron, P., 1998. Estimating and testing linear models with multiple structural changes. Econometrica 66, 47–68. Bai, J., Perron, P., 2003. Critical values for multiple structural change tests. Econometrics J. 6, 72–78. Bauer, M.D., Neely, C.J., 2014. International channels of the Fed’s unconventional monetary policy. J. Int. Money Finance 44, 24–46. Belke, A., Orth, W., Setzer, R., 2010. Liquidity and the dynamic pattern of asset price adjustment: a global view. J. Bank. Finance 34, 1933–1945. Borio, C, Zabai, A., 2016. Unconventional monetary policies: a re-appraisal. BIS Working Paper No. 570. Bank for International Settlements, Basel. Bridges, J., Thomas, R., 2012. The impact of QE on the UK economy – some supportive monetarist arithmetic. Bank of England Working Paper 431. Bank of England, London. Campbell, J.Y., Shiller, R.J., 1987. Cointegration and tests of present value models. J. Political Econ. 95 (5), 1062–1088. Chen, Q., Filardo, A., He, D., Zhu, Feng, 2015. Financial crisis, US unconventional monetary policy and international spillovers. BIS Working Papers 494. Bank for International Settlements, Basel. Christensen, J., Rudebusch, G., 2012. The response of interest rates to US and UK Quantitative Easing. Econ. J. 122 (564), F385–F414. Chung, H., Laforte, J.-P., Reifschneider, D., Williams, J.C., 2012. Have we underestimated the likelihood and severity of zero lower bound events? J. Money Credit Bank. 44 (1), 47–82. Ciccarelli, M., Mojon, B., 2010. Global inflation. Rev. Econ. Stat. 92 (3), 524–535. Cloyne, J., Ryland, T., Tuckett, A., Wills, S., 2015. A sectoral framework for analysing money, credit and unconventional monetary policy. Bank of England Staff Working Paper 556. Bank of England, London. D’Amico, S., King, T.B., 2013. Flow and stock effects of large-scale Treasury purchases: evidence on the importance of local supply. J. Financ. Econ. 108 (2), 425–448. 34 Existing studies focus on developments within the country undertaking QE, mostly event studies. We find that bond purchases in the US alone do not disturb co-movements of US and euro area interest rates. This implies either that QE is ineffective or that QE is very effective as it depresses interest rates everywhere. But then the impact on the exchange rate should be zero.

A. Belke et al. / Journal of International Money and Finance 73 (2017) 335–349

349

D’Amico, S., English, W., López-Salido, J.D., Nelson, E., 2012. The Federal Reserve’s large-scale asset purchase programs: rationale and effects. Econ. J. 122, F415–F446. Dees, S., di Mauro, F., Pesaran, M.H., Smith, L.V., 2005. Exploring the international linkages of the Euro area: a global VAR analysis. CESifo Working Paper No. 1425. Center for Economic Studies, Munich. De Santis, R.A., 2016. Impact of the asset purchase programme on euro area government bond yields using market news. ECB Working Paper 1939. European Central Bank, Frankfurt/Main. Deutsche Bundesbank, 2016. The macroeconomic impact of Quantitative Easing in the Euro Area. Monthly Reports, June, Frankfurt/Main. Dickey, D.A., Jansen, D.W., Thornton, D.L., 1991. A Primer on Cointegration with an Application to Money and Income. Federal Reserve Bank of St. Louis, St. Louis, MO, pp. 58–78. March/April. Doh, T., 2010. The efficacy of large-scale asset purchases at the zero lower bound. Econ. Rev. - Fed. Reserve Bank Kansas City (January), 5–34. Financial Times, 2014. The problem with QE is that it works in practice, but it doesn’t work in theory, October 13. URL: (last accessed on 9.12.16). Fratzscher, M., Lo Duca, M., Straub, R., 2013. On the international spillovers of US Quantitative Easing. DIW Discussion Paper, No. 1304. German Institute for Economic Research, Berlin. Gagnon, J., Raskin, M., Remache, J., Sack, B., 2011. The financial market effects of the Federal Reserve’s large-scale asset purchases. Int. J. Cent. Bank. 7 (1), 3– 43. Gilchrist, S., Zakrajsek, E., 2012. Credit spreads and business cycle fluctuations. Am. Econ. Rev. 102 (4), 1692–1720. Gros, D., Aldici, C., De Groen, W.P., 2015. Lessons from Quantitative Easing: much ado about so little?. CEPS Policy Brief No. 330. Centre for European Policy Studies, Brussels. Hamilton, J.D., 1994. Time Series Analysis. Princeton University Press, Princeton, NJ. Hamilton, J.D., Wu, J.C., 2012. The effectiveness of alternative monetary policy tools in a zero lower bound environment. J. Money Credit Bank. 44 (1), 3–46. Johansen, S., Juselius, K., 2006. The Cointegrated VAR model: Methodology and Applications. Oxford University Press, Oxford. Kapetanios, G., Mumtaz, H., Stevens, I., Konstantinos, T., 2012. Assessing the economy-wide effects of Quantitative Easing. Econ. J. 122 (564), F316–F347. Krishnamurthy, A., Vissing-Jorgensen, A., 2011. The effects of Quantitative Easing on interest rates: channels and implications for policy. National Bureau of Economic Research Working Paper No. 17555, Cambridge, MA. Li, C., Wei, M., 2013. Term structure modelling with supply factors and the Federal Reserve’s large-scale asset purchase programmes. Int. J. Cent. Bank. 9, 3– 39. Litterman, R.B., 1983. A random walk, Markov model for the distribution of time series. J. Bus. Econ. Stat. 1 (2), 169–173. Liu, J., Wu, S., Zidek, J.V., 1997. On segmented multivariate regression. Stat. Sin. 7, 497–525. Liu, P., Mumtaz, H., Theodoridis, K., Zanetti, F., 2014. Changing Macroeconomic Dynamics at the Zero Lower Bound. Bank of England, London, mimeo. Meyer, L.H., Bomfim, A.N., 2010. Quantifying the effects of Fed asset purchases on Treasury yields. Macroeconomic Advisers Monetary Policy Insights: Fixed Income Focus, 17 June. Meaning, J., Zhu, F., 2011. The impact of recent central bank asset purchase programmes. BIS Quarterly Review, December, pp. 73–83. Neely, C.J. (2010), ‘‘The large-scale asset purchases had large international effects”, Federal Reserve Bank of St. Louis Working Paper 2010–018G, St. Louis, July. Neely, C.J., 2015. How persistent are unconventional monetary policy effects?. Federal Reserve Bank of St. Louis Working Paper 2014-004B, July, update. Neely, C.J., 2015a. Unconventional monetary policy had large international effects. J. Bank. Finance 52, 101–111. Perron, P., 1989. The great crash, the oil price shock, and the unit root hypothesis. Econometrica 57 (6), 1361–1401. Pesaran, H.M., Smith, R., 2012. Counterfactual analysis in macroeconometrics: an empirical investigation into the effects of Quantitative Easing. IZA Discussion Papers 6618. Institute for the Study of Labor, Bonn. Rey, H., 2013. Dilemma not trilemma: the global financial cycle and monetary policy independence. In: Presented at the Federal Reserve Bank of Kansas City Economic Policy Symposium on Global Dimensions of Unconventional Monetary Policy, Jackson Hole, 22–24 August. Saeidinezhad, E., 2015. The Impacts of QE on US Mortgage Market. Institute for New Economic Thinking, April 7. Sims, C.A., Stock, J.H., Watson, M.W., 1990. Inference in linear time series models with some unit roots. Econometrica 58 (1), 113–144. Stock, J., Watson, M.W., 1988. Testing for common trends. J. Am. Stat. Assoc. 83 (404), 1097–1107. Stroebel, J., Taylor, J.B., 2012. Estimated impact of the Fed’s mortgage-backed securities Purchase Program. Int. J. Cent. Bank. 8 (2), 1–42. Taylor, J.B., 2016. New test finds no impact of QE on long-term interest rate. URL: (last accessed on 9.12.16). Thornton, D.L., 2013. An evaluation of event-study evidence on the effectiveness of the FOMC’s LSAP program: are the announcement effects identified?. Federal Reserve Bank of St. Louis Working Paper No. 2013-033A, October. Thornton, D.L., 2014a. Has QE been effective?. Federal Reserve Bank of St. Louis Economic Synopses, No. 3. Thornton, D.L., 2014b. QE: Is there a portfolio balance effect? Fed. Reserve Bank St. Louis Rev. 96 (1), 55–72. Troug, H.A., Murray, M., 2015. Quantitative Easing in Japan and the UK - An Econometric Evaluation of the Impacts of Unconventional Monetary Policy on the Returns of Aggregate Output and Price Levels. MPRA Paper No. 68707, Munich. Wright, J.H., 2012. What does monetary policy do to long-term interest rates at the zero lower bound? Econ. J. 122 (564), F447–F466. Zivot, E., Andrews, D.W., 1992. Further evidence on the Great Crash, the oil-price shock, and the unit-root. J. Bus. Econ. Stat. 10 (3), 251–270.