Quantitative Easing and the yield curve : has the ECB successfully lowered the yield curve?

University of Amsterdam

Faculty of Economics and Business

Quantitative Easing and the yield curve.

Has the ECB successfully lowered the yield curve?

Sebastiaan Hermanus Sollie 11157143

1st supervisor: prof. dr. S.J.G. van Wijnbergen 2nd supervisor: dr. W.E. Romp

Abstract:

In the aftermath of the 2008 financial crisis and with Europe’s economy in a slump the ECB had to take unconventional policy measures in order to push inflation back to the targeted level. In line with other central banks all over the world, the ECB started to conduct a Quantitative Easing policy. This paper explores the effect that the Quantitative Easing policy has on the government yield curve. In order to estimate this effect this paper employs a Nelson-Siegel model to estimate the level, slope and curvature of the yield curve. After this a regression is used to distillate the effect that Quantitative Easing had on the yield curve. A large group of control variables is added in order to control for movements in the macro economy that influence the yield curve over time. This paper finds that QE successfully lowered en flattened the yield curve, aking it for now a successful unconventional policy tool.

Introduction ... 3

Section 1: Initial data exploration ... 4

Section 2:Literature overview, channels via which QE influences the yield curve ... 7

A “little” critique ... 9

Section 3: The model ... 10

The Nelson-Siegel model ...10

Estimating the model ...12

The regression ... 13

Section 4: Data ... 14

Variables determining the yield curve ...15

Section 5: Results ... 17

Robustness ...21

Different data ...21

Inflation expectations ...23

Section 6: Conclusion ... 25

Acknowledgments and future research References ... 25

Introduction

The 2008 financial crisis pushed the world economy in a recession from which it yet has to recover. In the direct aftermath of the collapse of Lehman Brothers central banks acted firmly by lowering their policy rates to virtually zero. By doing so they expected to push the economy out of the negative spiral were it was in due to the collapse of the financial system. However, these low interest rates were not sufficient enough to push the economy out of the recession and back to full employment. Large central banks from all over the world faced a feared scenario that is known as the liquidity trap.

With policy rates at zero and inflation expectations decreasing the European Central Bank (ECB) as well as the FED, BoE and the BoJ, was forced to look for other policy measures that could increase economic activity. They took a wide variety of measures from expending the lending facilities to buying bonds first solely government bonds and later also corporate bonds. These purchases of bonds came to be known as quantitative easing (QE), which is till today a highly controversial topic in economics because the effectiveness and side effects of this policy is still unexplored territory.

Krugman (2000) describes the importance of policy alternatives that can be used when central banks face a liquidity trap. Quantitative easing and unconventional market operations are the main policy alternatives according to Krugman. Those two alternatives are used today by the large central banks all over the world. Both the FED and the ECB use an asset purchase programme in order to expand the monetary base, or to put it in the words of Krugman those central banks use unconventional market operations to ensure quantitative easing. Krugman points to different channels via which QE has an effect on the economy one of those is a decrease in interest rates because of bond purchases by the central bank.

While the paper by Krugman was published fifteen years ago the topic of quantitative easing is more relevant than ever. Since the outburst of the financial crisis and the increased amount of unconventional monetary policy measures a large amount of literature has been written on the topic. Kehmraj and Yu (2016) find that the FEDs QE program was effective in reducing the corporate bond spread and increased short-term investments. This contradicts the findings from Bowmand et.al. (2015) that QE is not very effective in stimulating the economy. They used Japanese bank data from the period 2000-2006 during which Japan had a QE program in place. The researchers found that QE was not able to significantly increase bank lending and therefor was unfit to stimulate the economy. Putnam (2013) presents a more nuanced view on the topic he argues that QE was most effective when it was used to stabilize the banking sector and less effective during the period after this. Putnam also points to a new challenge that comes with QE namely an exit strategy, which can have negative effects on the just recovered economy.

This paper adds to the literature by examining the effect of QE on the yield curve. Central banks argue that by buying bonds they increase the price of these assets that lowers the yield curve thereby stimulating investments. The hypothesis that is tested in this paper is formulated as follows

The rest of this paper is structured as follows: section one explores some initial data on the effectiveness of QE. Section two gives an overview of the current literature on quantitative easing and the yield curve. In section three a model is presented that is used to estimate the yield curve after that the model that is used to distillate the effect of QE is presented. In section four the used data is presented. Section five shows the results, acknowledgments and shortcomings. Finally section six concludes.

Section 1: Initial data exploration

On January 22, 2015 the ECB(2015) announced an extended asset purchasing program to push the struggling Euro-area economy out of the financial crisis. By buying large amounts of euro denoted government bonds the ECB aimed on lowering interest rates and flattening the yield curve, since then the ECB bought a large amount of euro denoted government debt. Figure 1 shows the amount of bonds bought under the PSPP as a percentage of both total euro denoted government debt as well as all euro denoted debt. At the end of March 2016 the ECB had bought almost 10% of the total government debt in the Eurozone.

Note: Figure 1 shows the magnitude of the amount of bonds bought by the ECB under the QE-policy so far. While total debt stayed more or less constant over the period, the percentage of purchased debt increased to 9% of all

euro denoted government bonds in march 2016. Source ECB, 2016a and 2016b.

Figure two displays the movement of the Euro-area yield curve since the implementation of the PSPP. In order to avoid the suspicion of data selection I took the yield curve data on the 20th of every month or the closest available date. The figure does not show all the yield curves from December 2014 up to May 2016 because some were virtually equal. The following yield curves were basically equal,

 February 2015: March 2015  May 2015: June 2015

 July 2015: August, September, October, November, December  January 2016: February, March, April

0 2 4 6 8 10 12 11950 12450 12950 13450 13950 14450

Apr/15 May/15 Jul/15 Aug/15 Oct/15 Dec/15 Jan/16 Mar/16

Per ce n tage o f d e b t Eu ro d e n o te d d e b t (x1b ill io n )

Figure 1: The QE policy takes a large part out of the total amount of debt

Total euro denoted debt (LHS)

PSPP holdings % of Total Gov debt (RHS)

Note: Every line represents a yield curve as published by the ECB for the 20th _{of the month or the nearest published day. For}

some days the lines where virtually equal, those lines are combined and represented by the earliest curve. Source: ECB, 2016C

The December yield curve is the base reference line because this was exactly one month before the Asset Purchase Programme (APP) was announced. The 15 year spot rate was the highest in December ’15 and decreased over the rest of the sample period. This decrease of the 15 year spot rate is a first tentative sign that the QE program was effective in lowering the yield curve at longer maturities.

Two days before the announcement, on January the 20th there was already a lot of speculation in the market in anticipation of the ECBs board meeting. These expectations lowered the yield curve in advance of the policies implementation. The curve kept lowering over February and March hitting the lowest point in April. At that point the assets purchasing program appeared to be very effective in lowering the yield curve. However, in May 2015 the yield curve moved up again to levels comparable to the situation in January. After that, in July the yield curve moved up even further and remained at that level up to January 2016. During that time markets expected the ECB to increase her asset purchase program, which she eventually did on 10th March 2016. (ECB, 2016e) On that day the ECB announced an expansion to 80 billion of her monthly bond purchases. In the months after that the yield curve lowered to the level of May 2016.

Just observing the yield curve over time does not provide compelling evidence on the effectiveness of the asset-purchasing program but also not for the ineffectiveness. It is important to keep in mind, that the yield curve is influenced by a wide variety of variables that change over time and that it would be shortsighted to judge the effect of the program solely on the level of the yield curve.

-0.7 -0.2 0.3 0.8 1.3 1.8 2.3 0 20 40 60 80 100 120 140 160 180 Sp o t rate

Months till maturity Figure 2: The movement of the yield curve during the QE-policy

19-Dec-14 20-Jan-15 20-Feb-15

20-Apr-15 20-may-15 20-Jul-15

Note: Figure 3 shows the movement of the expected future rate of inflation over the time period that the QE-policy was active. The figure shows that especially short-term inflation expectations picked up over the period the policy is in place.

In order to assess the effectiveness of the program from a different perspective figure 3 shows the inflation expectations in the Euro area since the first of January 2015. The program was implemented in a time when inflation expectations were at an all time low, especially the one and five year inflation expectations have increase after the introduction of QE. Both the decreased yield curve as the increased inflation expectations show initial evidence on the effectiveness of the unconventional policy measures.

In order to determine the effect of quantitative easing, section 3 describes a regression model that employs a variety of variables that can be used to adjust for changing macroeconomic variable. By adding a QE dummy this paper tries to filter the effect that quantitative easing had out of yield curve data.

-1 -0.5 0 0.5 1 1.5 2

Jan/15 Apr/15 Jul/15 Oct/15 Feb/16 May/16

Ex pe ct ed infl at io n le ve l

Figure 3: Movement of inflation expectation over the QE period

1 year expectations 5 year expectations 10 year expectations

Section 2:Literature overview, channels via which QE influences the yield curve The European Central Bank (ECB) did not start her QE program in the direct aftermath of the 2007 financial crisis as the FED and BoE did. Despite this difference in timing the channels via which the QE program influences the yield curve remain the same. Altavilla, Carboni & Motto (2015) describe four different channels: the signaling channel, asset scarcity channel, duration channel and credit premium channel.

The signaling channel has an effect even before the actual policy is implemented. Due to the forward looking nature of the yield curve agents incorporate expectations of lower interest rates in the future into the current yield curve. Gern et.al. (2015) state that via this channel a central bank can increase the power of her forward guidance channel. If a central bank buys bonds of a longer maturity she commits to keeping short-term interest rates low for a long period of time. If the central bank decides to increase the short term interest rate while holding long term bond the central bank suffers a lose due to a decrease in value of the long term bonds. Eggertsson & Woodford (2003) elaborate on this effect. According to them both inflation as well as real output are unaffected by QE. (Which is contradicted by the latest literature.) However, they still emphasize the importance and effectiveness of an asset purchase programme in case of deflation risk.

In order to lower the yield curve the central bank has to convince markets that she is willing to keep interest rates low even if this by the Taylor rule is no longer required. Only then markets lower the term structure thereby stimulating the economy, which eventually boosts inflation. While in theory this sounds straight forward in practice this more chanllanging, after all financial markets doubt and test the commitment of the central bank to keep interest rates low. By purchasing long term assets and therewith putting skin in the game the central bank can convince financial markets that she is determined to keep interest low according to Eggertsson & Woodford. The effect of this signaling channel is stronger for medium term bonds than for long term bonds because interest rates are expected to increase in the future at some point. Eggertsson & Woodford argue that the size and composition of QE does not affect the yield curve, as the signaling channel is the only channel via which QE can influence the yield curve, making it impossible to effect long maturity spot rates. This confirmed by Bomfim (2003) who found that the medium length of the yield curve changes almost one-for-one with the expectations on the future stance of monetary policy. This is in line with the expectations expressed by Eggertsson & Woodford.

The asset scarcity channel, assumes investors who have preferred habits. Such a model with preferred habits was introduced by Modigliani & Sutch (1966) to explain the behavior of the yield curve. They combine multiple theories into one more realistic theory. The authors argue that the yield is influenced by three assumptions. The first assumption is based on the existence of arbitrageurs whose actions create a yield curve in which expected returns are equal over all maturities. The second assumption is about the forward looking nature of the yield curve due to investors who take into account future capital gains or losses due to changes in future interest rates. The third assumption takes into account the personal preferences of investors. Individual investors have their own objectives and might be only interested in an investment that pays of after n-periods. When this is the case their payoffs are certain if they invest in a n-period bond. Those investors would not have an incentive to

larger transaction costs as well as the increased risk, those investors would be tempted to engage in such a portfolio. The existence of preferred habit investors creates segregations between different maturities within the yield curve.

Altavilla, et.al.(2015) argue that if there are sufficient preferred habit investors in the market and arbitrageurs are largely risk averse, the central bank has the possibility to lower the longer term of the yield curve by buying bonds of longer maturity. This is in line with the work of Greenwood & Vayanos (2014) who tested the effect of the US government’s debt composition on the yield curve. Their results were in line with a model in which different investors have different preferred habits and therefor create a segregated yield curve. The authors argue that a central bank can exploit this segregation by buying bonds of longer maturity and thereby lowering the longer term of the yield curve.

The duration channel is in contradiction to the previous channel driven by the arbitrageurs and preferred habit investors together. For the asset scarcity channel arbitrageurs were assumed to be too risk averse which made them unable to integrate the yield curve. However, in a model set up by Vayanos & Vila (2009) the yield curve is determined by the interaction between arbitrageurs and preferred habit investors. In this model, when a demand shock hits for a given maturity, arbitrageurs have the ability to spread out the effect over the entire yield curve. So when the central bank shocks demand by buying longer-term bonds this lowers the entire yield curve. Altavilla, et.al.(2015) make use of this insight in formulating the duration channel. When a central bank successfully lowers the maturity structure by buying longer-term bonds, thereby lowering the total duration of the yield curve this reduces duration risk. The reduction of the risk spreads out over the yield curve due to arbitrageurs.

The final channel described by Altavila et.al. (2015) is the Credit premium channel. While the authors call this channel the credit premium channel, it is similar (equal) to the portfolio rebalancing model described in the literature. The portfolio rebalance model finds it origin in the work of Tobin (1969) in which he point out that a change in the supply of an asset group does not only alter the yield of that asset group but also the relative price towards other asset groups. This change in the relative price forces investors to rebalance their portfolios. Not only Keynsian economists like Tobin had a firm believe in the portfolio rebalance channel, also monetarists like Friedman & Schwartz (1965) believed in this mechanism.

The main message of the portfolio rebalance model model is that a demand or supply shocks to a group of bonds with a similar maturity influences the entire yield curve. Christensen (2016) and Bernanke & Reinhart (2004) describe a reserve induced effects of QE. They argues that when central banks buy government bonds directly from non-banks this eventually effects the balance structure of banks, this forces banks to rebalance their portfolios creating an upward pressure on asset prices with different maturities. Some authors argue that the new reserves would stay on banks balances sheets in the form of deposits (Herbst, Wu & Ho, 2014). This might be true in the short run but is not in the medium run according to Bernanke & Reinhart (2004)

The authors above focus on the financial intermediaries and how they are affected by a central bank that buys government bonds. Joyce et.al. (2012) and Bowdler & Radia (2012) address the portfolio rebalancing model from the perspective of non-financial corporations and investors. When a central bank buys assets directly from the public investors portfolio changes from one with long-term bonds to one

for duration reasons as well ass for risk to return of their portfolio and thus start buying term assets in order to rebalance. This increased demand for longer-term assets increases their prices, which in the case of bonds equals a lower yield.

Another channel often mentioned in the literature that is not specified by Altavila et.al. (2015) is the liquidity channel. Kiyotaki & Moore (2012) describe a model in which the central bank stimulates the economy via quantitative easing by buying illiquid assets from the public. By replacing those assets with liquidity (cash) the relief the liquidity constraint faced by the economy. According to the authors this lowers the effect of a liquidity shock on output and consumption. This is why the FED initially bought large amounts of MBSs in order to provide liquidity to a market that was almost completely dried out. This liquidity was necessary in order to keep the link between the FEDs policy rate and the market interest rate in place, in other words to keep the short term interest rate low. In the Eurozone the ECB tried to accomplish this via a different policy. The ECB (2015b) states that in order to overcome a liquidity shortage, which banks all over the Eurozone faced, the ECB decided to enhance her refinance facilities. The objective of the ECB after June 2014 was both to enhance the transmission mechanism and increase the accommodative monetary policy stance.

A “little” critique

The channels mentioned above are the main channels according to a large part of literature. However, Krishnamhurthy & Vissing-Jogrensen (2011) mention another channel, namely the default risk channel. They argue that bonds with higher default risk have higher prices. The default level of these riskier bonds will decrease after the QE program is in place while this program stimulates the economy therewith lowering the default risk. The described mechanism lowers the yield curve even further. Despite the fact that Krishnamhurthy & Vissing-Jogrensen argue that this is channel is important for the effectiveness of the QE program I object to the importance of this channel. In their assessment of the default risk channel the authors claim that the increased economic environment lowers default risk and thereby lowers the default premium. I would like to emphasize that during a time of economic recovery the economic environment enhances almost by definition and therefore the default risk would decrease even without QE. In my opinion this channel shows the willingness in the literature to find ways that explain why QE has to be a success. While the papers are written by highly respected economists and published in important journals it is important to point out that most of the researchers have to the more or lesser extend a relationship with a central bank. This could make them a little biased towards papers that applaud the effects of QE.

Martin & Milias (2012) address this problem. The authors point out that the current literature is mainly based on highly frequency data, this makes it impossible to include important macroeconomic variables such as inflation and output in the equation. Besides the lack of longer periods of data also the period over which the data is gathered is criticized by the authors. QE is a policy measurement that is only undertaken in times of severe economic downturn, therefor there is only data available gathered during crisis times. The lack of available data combined with the lack of any counterfactual testing, that is not uncommon in economics, wories Martin & Milias. Despite their critique on the economic research they could not deny that QE probably has some positive impact on the economy.

Martin & Milias have some genuine concerns regarding the academic research on the effects of QE. However, still most of the literature finds a positive effect of QE that enhances economic recovery, via a lower yield curve, and helps to reach a sustainable growth path. In the remaining part of this paper it is important to keep in mind that this paper preforms research on the frontier of monetary policy in a non ideal situation. Precisely because of this challenge researching the effect of QE is important and necessary.

Section 3: The model

In order to assess the effectiveness of QE, this paper uses a model that estimates the shape of the yield curve. After that a regression model is used in an attempt to preform something as close as possible to a counterfactual. This section describes the models used in order to estimate the effectiveness of the ECBs quantitative easing policy. A model is presented that is used to estimate the shape, level, slope and curvature of the yield curve. After that a regression model is presented that can be used to estimate the effect of quantitative easing on the yield curve.

The Nelson-Siegel model

A good estimation of the yield curve is beneficial both for academics in order to enhance their research as well for central banks who can employ the information within the yield curve in order to assess the effectiveness of policy measures. McCulloch(1979), Vasicek and Fong (1982) and Steeley (1991) are some authors who tried to set up yield curve estimation model using a variety spline estimation techniques. These models are often criticized for having parameters with no straightforward economic value or intuition behind them and for being black boxes. (Ioannides, 2003).

Svensson (1994) presents a much simpler filter that can extract a yield curve out of data containing different spot rates. Svensonn builds on a similar model from Nelson and Siegel (1987) this is why the model is often referred to as the Nelson-Siegel-Svensson (NSS) model. While being slightly less accurate than the model by McCulloch (1979) the NSS-model is often used because of its simplicity. A report from the BIS (2005) confirms the popularity of the models, 8 out of 13 central banks reporting to the BIS used either a NS-model or a NSS-model. The simple and straightforward interpretation of the betas makes these models most useful for the purpose of this paper.

In the Nelson and Siegel (1987) paper two models are presented, first a model shown in equation (1) that connects forward rates over all maturities ().

𝑓(𝜏)= 𝛽₁+ 𝛽₂(𝑒− 𝜏 𝜆_1)+𝛽 3( 𝜏 𝜆1 𝑒− 𝜏 𝜆₁₎ (1) Equation (2) shows the spot rate that is an average of all forward rates.

𝑦(𝜏)=1

𝜏∫𝑓(𝜏)𝑑𝑥 𝜏

0 (2)

Inserting the equation (1) into equation (2) and solving the intergral gives the equation as shown in equation (3). Appendix 2 shows the intermediate steps.

𝑦(𝜏) = 𝛽₁+ 𝛽₂(1 − 𝑒 −𝜏 𝜆₁ 𝜏 𝜆1 )+𝛽₃(1 − 𝑒 −𝜏 𝜆₁ 𝜏 𝜆1 − 𝑒− 𝜏 𝜆1) (3)

In the Svensson (1994) paper a fourth variable is added to the model that enables to model to fit the data even better. Equation (4) bellows shows the final Nelson-Siegel-Svensson model (NSS-model)

𝑦(𝜏) = 𝛽₁+ 𝛽₂(1 − 𝑒 −𝜏 𝜆1 𝜏 𝜆1 )+𝛽₃(1 − 𝑒 −𝜏 𝜆1 𝜏 𝜆1 − 𝑒− 𝜏 𝜆_1)+𝛽 4( 1 − 𝑒− 𝜏 𝜆2 𝜏 𝜆2 − 𝑒− 𝜏 𝜆₂₎ (4)

The NSS-model as shown in equation (4) represents the spot rates at a given time τ. In the equation y(τ) is the spot rate for a given maturity over τ periods in years. The two lambdas have no direct informative value but are necessary to make the model fit the data. The Betas nonetheless do have informative value regarding the shape and level of the yield curve. The first beta is also called the level and is the asymptote to which the yield curve converges. The second beta is the Slope of the yield curve and determines the overall steepness of the curve. The third and fourth beta measure the curvature of the yield curve and give it a hump-shape over the medium term.

Note: both figures show the effect of the different shape estimators according to the NSS-model for different days, as published by the ECB (ECB, 2016c). A comparison between both graphs shows the large movement of both curvature parameters. This can

large movement make them less valuable in order to estimate the effect of QE. Therefore, in this paper the NS-model is used.

Figure four and five show two different yield curves as calculated by the ECB for the Eurozone. It comes clear that the shape of the both curvature variable can vary over time which enables the yield curve to take both an U-shape as well as an S-Shape over the medium run. The slope is a more constant factor that mainly makes sure that the longer term of the yield curve is higher than the shorter term. Morales (2010) points out that first beta can be seen as an long term factor because it does not change over time. The second beta on the other hand is a short-term factor that starts at one but decays monotonically and quickly to zero. Both curvature betas are on the medium side, they start at zero and in the limit will move back to zero.

Although the NSS-model has a better fit to the data than the Nelson-Siegel model (NS-model) it also has one disadvantage regarding the curvature betas. The first model has two curvature betas that can vary largely over time, making it more flexible. These two betas make it at the same time more difficult to interpret the meaning of the betas. The first curvature beta can go up while the other goes down, thereby canceling each other out. The second model on the other hand fits the data a little less but is more useful regarding the objective of this paper. The effect of QE on the medium run is easier assessed if there is only one beta that describes the behavior of the yield curve on the medium run. Figure 6 shows the shape the curvature beta can take on depending on the level of lambda. In the graph the betas are shown for lambda 1, 2, 5 and 10 from the top to the bottom.

Note: Figure 6 shows the shape of the of the curvature beta for different values of lambda. The figure shows that the shape of the curvature beta is sensitive for the value of lambda. This makes the yield curve sensitive to the value of lambda. It is important

that the true value of lambda is estimated because a false value can bias the results by a lot.

Estimating the model

Before the betas can be estimated it is necessary to estimate the lambdas or shape parameters. Nelson and Siegel (1987) estimate these shape parameters by estimating the model for a range of lambdas and picking the lambda that has the best fit to the data. In this paper the same technique is used. In order to determine the best fit, the model is estimated for each date in the dataset and the R-Squared is calculated after this the average R-Squared is calculated for each lambda after which the lambda with the highest squared is picked. Figure 7 and Appendix 1 shows the average R-squares for all values of lambda between 0,1 and 6,0 for both datasets of bond yield of all governments and yield of triple-A rated bonds. The highest R-squared is reached with a lambda of 2,2 for the triple-A bonds and at 3,48 for all government bonds. These shape parameters are used for the estimation of the betas from the Nelson-Siegel model. The parameters are used to estimate the model for each date in the dataset. This gives a set of betas for each date in the dataset, these betas are then used as depended variables in order to estimate the effect of quantitative easing on the yield curve. 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0 5 10 15 20 25 30 Eff ect o f the cur va tur e be ta Years to maturity

Figure 6: Shape of the Curvature beta

Note: Figure 7 shows the results of the grid search for the best fitting lambda parameter. For every set of yield a regression was run for all lambda’s between 0 and 6 (interval 0.1) for all days the R-squared was calculated. After this the avcrage R-squared was calculated for every lambda. The highest average lambda for the dataset containing the triple-A bonds is at a lambda of 2.2.

The largest average R-squared for the data containing all bonds is lies at 3.48 however this is not a clear a maximum as for the triple-A bonds. Making the estimated results more sensitive for measurement errors.

The same technique is used by Fabozzi. Martellini and Priaulet (2005) and Diebold, Rudebusch and Aruoba (2006) those authors find a lambda of 3 and 1,37 respectively. This implies that the maximum of the curvature parameters lie at 2,5 for the first and 5,4 years for the later, for this paper the maximum level of the curvature beta lie 3,95 year for the triple-A bonds and at 4,22 for all bonds.

The regression

The estimated level, slope and curvature parameter are used as depended variables in the final regression. This regression is used to determine the effect of the ECBs QE program, the regression adjust for changes in macroeconomic variables that influence the yield curve. Equation (8) shows the regression.

𝐴_𝑡𝑖 = 𝛽_1𝑡𝑖 + 𝛽_2𝑡𝑖 𝑄𝐸𝐷𝑈𝑀𝑀𝑌𝑡+ 𝛽3𝑡𝑖 𝑴𝑡+ 𝝐𝒕𝒊 (8)

For

𝐴𝑡𝑖 = (𝐿𝑡, 𝑆𝑡, 𝐶𝑡)

The variable of interest is the QE-dummy that measures the effect of the QE policy on the yield curve. The matrix M (variables stated below) contains all the control variables that influence the effect QE has on the yield curve.

 Consumer Price index  Producer Price Index

 Harmonized index of consumer prices  Unemployment

 Employment growth  Output gap

 Manufacturing utilized capacity  Production growth rate

 Balance of Payments

 ECB policy rate  FED policy rate

 Risk aversion as computed by ECB  M1, M2 and M3 growth rate  EURONIA

 EURIBOR 3months  Oil price

 Exchange rates € to $ and the £  Inflation Swap Rates 1, 2, 5, 10, 15,

20, 25 and 30 years 0.98 0.985 0.99 0.995 1 0.985 0.986 0.987 0.988 0.989 0.990 0.991 0.992 0.993 0 1 2 3 4 5 6 A ve rag e R -sq uar ed for all b on ds A ve rag e R -sq uar ed for t ri pl e-A b on ds Lambda

Figure 7: the average R-squared for a range of lambdas

Triple-A (LHS) All bonds (RHS)

The data section below explores the foundation, as found in the literature, for using this data. Further descriptions of the variables can be found in appendix 5 and the online appendix.

The different estimations of the level, slope and curvature are used in order to check for robustness of the findings. For all regressions robust standard deviations are used in order to prevent hetroscedasticity from influencing the results. In total four datasets are used namely, triple-A bonds and all government bonds combined with both estimation techniques described above make four different sets of levels, slopes and curvature betas.

Section 4: Data

This section describes the data used to estimate the yield curve as well as the data used for the final regression. The data that is used to estimate the nelson-siegel model are the daily yield curve estimations as published by the ECB (2016c) for the triple-A bonds a maturity vector containing 1, 5, 10, 20 and 30 year spot rate is used, while for the NS-model containing all bonds a maturity vector that contains the 1, 2, 5, 10, and 14,75 year spot rates is used.

The ECB uses bond data that is provided by Euro MTS ltd and the ratings as published by Fitch Ratings. The ECB selects all bonds issued by central European governments. After that all bonds with special features are excluded, so that only fixed coupon bonds with finite maturity and zero coupon bonds remain in dataset. Only government bonds that are actively traded and have a maximum spread of three basis point between the bid and ask price are used in the calculation of the yield curve. In order to make sure that the bonds are traded in markets with sufficient market debt only bonds with a remaining maturity between 3 months and 30 years are included. The remaining bonds then are used to calculate the spot rates for that day. After the bond selection an outlier removal mechanism is applied to remove all bonds with a yield that deviates more than two standard deviations from the mean. (ECB, 2016d)

When all bonds are included the euro-area yield curve is a weighted average of the individual euro-area countries yield curves, weighted by the amount of outstanding government bonds. Because of the restriction imposed on the used bonds, it is likely that for the later time period no Greek bonds are included because those are not traded actively and deviate from the mean. For the triple-A rated bonds this is mainly the German yield curve in combination with a little influence of the Netherlands and an even smaller influence of Luxembourg. Those are at this moment the only three euro-area countries with a triple-A rating according to Fitch.

Table 1, Descriptive statistics of the spot rate data for triple-A rated bonds

1 Year 5 Years 10 Years 20 Years 30 Years

Mean 1,344395 2,089701 2,841297 3,366456 3,395890

Median 0,741502 2,422745 3,305643 3,723839 3,786511

Min -0,559712 -0,386434 0,097205 0,419161 0,527714

Max 4,539553 4,730363 4,776331 4,984930 5,175029

Count 3010 3010 3010 3010 3010

Note: the spot rate data that is used is an average over all triple-A rated euro-zone countries at the day of measurement. Source ECB, 2016c

Table 2, Descriptive statistics of the spot rate data for all bonds

1 Year 2 Years 5 Years 10 Years 14,75 Years

Mean 1,671534 1,918389 2,570918 3,371678 3,749281

Median 1,368967 1,889895 2,803460 3,763688 4,081454

Min -0,365623 -0,319417 0,004168 0,700843 1,012885

Max 4,559675 4,769203 4,810006 5,068630 5,437936

Count 3010 3010 3010 3010 3010

Note: the spot rate data that is used is an average over all euro-zone countries. Source ECB, 2016c

Table one and two show the descriptive statistics of the triple-A and all bond yield data respectively. The dataset consist out of a large variety of data with the one year spot rate varying form 4.5 percent in September 2008 to -0.56 and -0.37 in June 2016. It also appears that there is almost no difference in maximum values while there are larger differences in minimum values between the two datasets. This difference can be explained by the changing composition of both groups. The maximum values were reached at times when almost all Eurozone countries rated triple-A while the minimum values were reached in a time in which this was not the case. Currently only Germany, the Netherlands and Luxembourg have a triple-A status at the three largest rating agencies. (Trading economics, 2016)

Variables determining the yield curve

According to Ang and Piazzesi (2003) 85% of the variation in the yield curve is explained by macroeconomic variables. This is good news regarding the objective of this paper, that is to identify the effect of quantitative easing. When the right control variables are added it is possible to distillate the effect of QE out of the regression. Ang and Piazzesi use two categories of macro variables namely real activity variables and inflation variables. Unemployment, growth rate of employment and the growth rate of industrial production are part of the first and the CPI and PPI are part of the later. The real activity variables are added while an increase in economic activity increases demand for capital, this increases interest rates over all maturities. The inflation variables are added because higher inflation means a higher depreciation pressure for which capital owners want to be compensated.

Diebold, Rudebusch and Aruoba (2006) use the level of inflation, the central bank’s policy rate and manufacturing utilization capacity in order to determine the state of the economy. Paccagnini (2016) adds consumption to these variables in order to determine the state of the economy. Both papers argue that the state of the economy determines the level shape and curvature of the yield curve. Bernhardsen (2000) emphasizes the importance of the variables mentioned above but for another reason. Bernhardsen point out that according to the no arbitrage condition the yield curve takes in account the future stance of monetary policy and the variables used by Diebold et.al. (2006) are indicators on this future stance. Bernhardsen further adds real income growth and inflation expectations ass indicators for the future stance of monetary policy. Additionally Bernhardsen uses the debt-to-GDP ratio because the higher the debt-to-GDP ratio the larger the risk premium on government bonds will be, this pushes up the yield curve.

Chadha and Waters (2014) use 31 different variables divided over five categories, of which inflation, real activity and policy variables are also used by the authors above. The new categories introduced by Chada and Waters (2014) are foreign variables and financial market variables. The foreign category includes exchange rates and policy rates from the FED and the BoE while the financial category consist out of the LIBOR rate, VIX volatility, gold- and oil-prices. The authors don’t provide real economic intuition for their choice of these variables but test post-hoc if the variables are of any significance.

The used dataset consists out of daily data on all days at which the ECB published their yield curve data. For variables for which no daily data is available an log interpolation technique is used in order to estimate the missing data. While this data is used to complete the data set, the estimated variable are not used in the conclusion and no explanatory value is attached to them. Appendix two and three shows the descriptive statistics and the correlation matrix respectively.

Table 3 shows the correlation matrix for the interest rates that are used in the paper of Chada and Waters (2014). It appears that there is a large correlation between the different interest rates that apply on the Euro. This can be expected by the connection all interest rates have to the ECB policy rate. All the interest rates below are short term riskless rates, the large similarity and the conection to the policy rate explain the large correlation shown in table 3. The table also shows the high correlation between the policy rate of the FED, ECB and the BoE this can be explained by an increasing globalized financial system. Because of the correlation and in regard of the multicollinearity issue only the ECB policy rate is included in the regression.

Table 3: correlation matrix of the different interest rates

ECB FED BoE EURONIA EURIBOR

ECB 1

FED 0,7701 1

BoE 0,8900 0,8955 1

EURONIA 0,9738 0,8444 0,9454 1

EURIBOR 0,9899 0,7742 0,8977 0,9823 1

Note: as can be expected there is a large correlation between the different interest rates. Therefor in the rest of this paper only the most important interest rate, with regard on the effect of QE, the ECB policy rate is used.

The same multicollinearity problem arises for the inflation expectations data as shown in table 4. Therefor each regression only uses one variable containing inflation expectations. The expectation with the best fit according the regression is used.

Table 4: correlation matrix of the different inflation expectation maturities

1Year 2Year 5Year 10Year 15Year 20Year 25Year

1Year 1 2Year 0,9692 1 5Year 0,8548 0,9400 1 10Year 0,7572 0,8632 0,9691 1 15Year 0,7470 0,8420 0,9443 0,9782 1 20Year 0,7611 0,8477 0,9374 0,9629 0,9868 1 25Year 0,7733 0,8506 0,9249 0,9438 0,9738 0,9884 1 30Year 0,7947 0,8580 0,9104 0,9193 0,9555 0,9761 0,9892

Note: table 4 shows the correlation between the different time spans of inflation expectations. The largest correlation between the different expectation variables is between those for who the maturity is close togheter. Because the variables for a large part

In an attempt to prevent a suspicion of data mining and because of the large correlation between the variables CPI, PPI and HICP as well as between the variables employment growth and unemployment and the variables utilized manufacturing capacity, output gap and production growth rate, these variables are pooled in a standardized variable. Respectively standardize inflation, standardized employment and standardized output. The method used in order to pool these variables to one single variable can be found in appendix 5.2.

Section 5: Results

This section describes the results from the regression in equation (8). Before the regression output can be interpreted it is important to realize what the estimated betas explain regarding the level, slope and curvature of the yield curve. The level is the long term level towards which the yield curve converses and therefore the easiest to interpret, when the QE variable lowers the level beta this means that the long term interest rates decreases. The shape of the slope and curvature are determined by the shape parameter lambda, figure 8 shows the shape of the shape and slope when the lambda is fixed at the level calculated above.

Note: Figure 8 shows the shape of the slope and curvature beta for he estimated values of lambda. The both curves do not differ much. The largest difference is between the both slope betas in the medium run.

The lines shown in figure 8 are flipped over the x-axes by making them negative. This is because the slope and curvature are expected to have negative values and therefor are flipped over the x-axes. While the shape of both curves are fixed over time the regression only tests for a movement in the line.

Table 5 shows the regression output for the level parameter estimated with all bond data using the NS-model. The first row shows the effect of QE on the level beta and shows a significant decrease due to the QE policy, of about 2.5%. Unfortunately for the ECB who would be happy with such a results, other variables needs to be included that also influence the yield curve parameter and the interaction between the parameter and the QE policy. First in regression (2) the policy rate is added to the regression. The policy rate needs to be included fot two reasons, first the policy rate determines at the short term the level of the interest rate and second in line with the forward looking nature of the yield curve the curve takes in account all future levels of the policy rate. In order to adjust for the stance of monetary policy more precisely the growth rate of the monetary aggregate M1 is added. According to the quantity

-0.9 -0.8 -0.7 -0.6 -0.5 -0.4 -0.3 -0.2 -0.1 0 0 5 10 15 20 25 30 Ef fec t on th e yi el d cu rv e

Years till maturity

Figure 8: The shape of the slope and curvature parameters

Slope All bonds Curvature All bonds Slope tripple-A bonds Curvature tripple-A bonds

inflation by the same percentage. This gives to reasons for including the M1 growth rate. For starters the M1 growth rate is expected to change long term inflation and thereby it influences the longer term of the yield curve. Second as mentioned above it gives a better view on the stance of monetary policy.

Regression (4) adds the standardized variables for inflation, employment and production in order to correct for the stance of the economy, adding all these variables lowers the effect of QE to a decrease of -0.845 of the level parameter. Finally regression (5) corrects for the changed risk aversion. Regression output column (4) and (5) combined show that the estimated effect of the QE policy on the level beta lies between 0.845 and 0.786 for this dataset. Further on in this section the effect of the QE-policy on the slope and curvature is assessed after that figure 9 and 10 give more insight in the movement of the entire yield curve.

Table 5 (β1) – Regression results for all bonds

(β1) (1) (β1) (2) (β1) (3) (β1) (4) (β1) (5)

QE dummy -1.943*** -2.080*** -1.829*** -0.641*** -0.769***

(0.0293) (0.0340) (0.0381) (0.0389) (0.0457)

ECB policy rate -0.0796*** -0.0928*** 0.0392*** -0.00739 (0.00910) (0.00839) (0.00790) (0.00941) M1 growth rate -0.0554*** -0.0475*** -0.0437*** (0.00211) (0.00162) (0.00160) Std. inflation 0.302*** (0.0108) (0.0118) Std. employment -0.535*** (0.00786) (0.00997) Std. production (0.00654) Risk aversion 0.0313*** (0.00480) Constant 5.174*** 5.315*** 5.740*** 5.333*** 5.395*** (0.0115) (0.0210) (0.0236) (0.0179) (0.0205) N 3010 3010 3010 3010 3008 adj. R-sq 0.532 0.544 0.609 0.818 0.821

Note: Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001, The table shows the effect that the independent variables have on the level beta (β1) all bonds NS-model. All variables but the QE dummy are control variables and no explanatory value can be attached to the estimated betas. The description of the

variables can be found in appendix 5.

The signs of the standardized employment and the standardized production are different as expected. For the unemployment variable this means that an increase in unemployment or a decrease in the employment growth increases the long-term interest rate were a decrease is expected. Lower employment growth or increased unemployment equals lower economic activity and therefor a decrease in the interest rate is expected. The same applies for the standardized production variable for which the output gap, the utilized production capacity and the industrial production growth are the base variables. Appendix 7 contains a regression table that shows the different base variables for the standardized variables. The table shows a consequent picture in which higher activity and a lower interest rate go hand in hand. Therewith it is not said that higher activity leads to a lower interest rate but for this dataset this is the

The variables proposed by Chada and Waters (2014) are not used in the regression because of multicollinearity problems. As described above are most of the interest rates highly correlated. Appendix 8 contains all the VIF values of the reported regressions. The variables Euro/Dollar exchange rate and the oil price are used as robustness test variable further on the in this section. Euro/Dollar exchange rate is added because the US is the largest trade partner of Eurozone. (ECB, 2016f)

Table 6 (β2) - Regression results for all bonds

(β2) (1) (β2) (2) (β2) (3) (β2) (4) (β2) (5)

QE dummy 0.179*** 2.200*** 0.161*** 0.0350 0.608***

(0.0446) (0.0417) (0.0466) (0.0475) (0.0546)

ECB policy rate 1.170*** 0.731*** 0.709*** 0.885***

(0.0145) (0.0130) (0.0122) (0.0114) Std. inflation -0.269*** -0.140*** (0.0133) (0.0133) (0.0151) Std. employment 0.968*** 0.967*** (0.0128) (0.0114) (0.0140) Std. production 0.162*** 0.112*** (0.0131) (0.0129) (0.0100) M1 growth rate 0.0565*** (0.00220) Risk aversion -0.130*** (0.00528) Constant -3.576*** -5.647*** -4.720*** -5.115*** -5.029*** (0.0336) (0.0301) (0.0226) (0.0223) (0.0244) N 3010 3010 3010 3010 3008 adj. R-sq 0.001 0.712 0.899 0.915 0.916

Note: Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001, The table shows the effect that the independent variables have on the level beta (β2) all bonds NS-model . All variables but the QE dummy are

control variables and no explanatory value can be attached to the estimated betas. The description of the variables can be found in appendix 5.

Table 6 shows the regression results with depended variable B2. The same intuition is used as for B1 however, for B2 the standardized production variable is no longer significant. This might be because production is stickier than inflation and employment, therefor it has no influence on the effect that QE has on the short-term interest rate. This final added regression does not change the QE variable by much.. Also when regressions that are not reported are estimated containing oil prices or the Euro/Pound exchange rate, the QE beta remain close to 0.3. While this is good news from an econometrics point of view this is not from a policy point of view. If 0.3 is the true value of the beta, quantitative easing increases the short-term interest rate and thereby has the opposite effect than was intended by the ECB. However, the slightly positive value can be misleading due to the monopoly that the ECB has over the short-term interest rate. The ECB decreased the interest rate from 4.25% in October 2008 to 1% in May 2009. This decreased the short-term side of the interest rate already even before the QE program was in place. Since the slope parameter is most active on the short term, this affects the estimated effect of the QE policy. It is also important to take in account the effect on the level, slope and curvature beta combined.

Table 7 (β3) - Regression results for all bonds

(β3) (1) (β3 (2) (β3 (3) (β3 (4) (β3 (5)

QE dummy -2.452*** -1.912*** 0.394*** 0.202* 0.369***

(0.0452) (0.0688) (0.101) (0.0924) (0.108)

ECB policy rate 0.313*** 0.817*** 0.785*** 0.841***

(0.0259) (0.0348) (0.0339) (0.0364) Standardized inflation 0.349*** 0.547*** 0.559*** (0.0410) (0.0377) (0.0378) Standardized employment -1.053*** -1.055*** (0.0279) (0.0251) (0.0288) Standardized production -0.286*** -0.363*** (0.0265) (0.0271) (0.0273) M1 growth rate 0.0864*** 0.0796*** (0.00556) (0.00608) Risk aversion -0.0453** (0.0140) Constant -1.335*** -1.888*** -2.949*** -3.553*** -3.613*** (0.0311) (0.0614) (0.0651) (0.0804) (0.0805) N 3010 3010 3010 3010 3008 adj. R-sq 0.211 0.258 0.472 0.506 0.507

Note: Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001, The table shows the effect that the independent variables have on the level beta (β3) all bonds NS-model. All variables but the QE dummy are control variables and no explanatory value can be attached to the estimated betas. The description of the

variables can be found in appendix 5.

Table 7 states the regression output for the curvature beta In order to do so an “average”1

yield curve is calculated using the average value of the variables used in the regressions in table 5, 6 and 7. Figure 9 and 10 show these “average” yield curves, figure 9 shows the estimated yield curves without the risk aversion variable while figure 10 shows the curves with the risk aversion variable included.

Note: Figure 9 and 10 show the “average” yield curve with and without the QE policy. The dotted lines show the difference between both yield curves. The increasing shape of the difference curve proves the flattening effect that QE has on the yield

curve.

Both figures also show a line that describes the difference between both curves. This is to test if the ECB succeeded in her objective to flatten the yield curve. When the risk aversion variable is included it seems that the ECB was successful in flattening the yield curve, however, when the risk aversion variable is not included this is not the case. In a model without risk aversion, the difference between both yield curves decreases over time so the yield curve becomes steeper and not flatter only for 30 year maturity bonds the difference is larger than for short term bonds.

Robustness

In order to prove the robustness of the findings this section uses the variables euro-dollar exchange rate and the oil price to show the sensitivity of the QE variable.

Table 8 – Robustness test of the estimated effect of QE, all bonds

β1 (1) β1 (2) β2 (3) β2 (4) β3 (5) β3 (6)

QE dummy -0.678*** -0.704*** 0.460*** 0.522*** 0.103 0.0652 (0.0360) (0.0375) (0.0532) (0.0526) (0.0851) (0.0853) ECB policy rate 0.491*** 0.276*** 0.454*** 0.729*** 0.920*** 0.607*** (0.0109) (0.0126) (0.0132) (0.0171) (0.0366) (0.0460) Std. inflation 0.052*** 0.0222*** 0.035*** -0.007 (0.00187) (0.00201) (0.00711) (0.00733) Std. employment 0.198*** 0.381*** -0.308*** -0.536*** 0.069 0.336*** (0.0110) (0.0120) (0.0153) (0.0146) (0.0356) (0.0421) Std. production -0.404*** -0.387*** 1.007*** 0.933*** -0.887*** -0.861*** (0.0102) (0.0104) (0.0121) (0.0136) (0.0324) (0.0309) M1 growth rate -0.0265** -0.0265** 0.136*** 0.146*** (0.00883) (0.00865) (0.0118) (0.0109) Risk aversion 0.00139 -0.017*** -0.097*** -0.099*** -0.0172 -0.044*** (0.00428) (0.00398) (0.00528) (0.00478) (0.0131) (0.0122) €/$ exchange 1.955*** 3.952*** 0.499*** -2.284*** -2.846*** 0.0664 (0.112) (0.112) (0.115) (0.131) (0.416) (0.488) Oil price -0.015*** 0.0142*** -0.021*** (0.00054) (0.00053) (0.0016) Constant 0.181 -0.733*** -4.060*** -1.956*** -0.281 -1.614** (0.151) (0.129) (0.153) (0.142) (0.550) (0.554) N 3008 3008 3008 3008 3008 3008 adj. R-sq 0.878 0.913 0.892 0.913 0.437 0.476

Note: Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001, The table checks for robustness of the QE-dummy by adding the Exchange rate and the oil dollar price (β1,2 and 3) all bonds NS-model . All variables but the QE dummy are control variables and no explanatory value can be attached to the estimated betas. The description of the variables can

be found in appendix 5.

Table 8 shows the estimated parameters when those variables are included. The added variables give an insight in the robustness of the model, the estimated betas for QE are similar to the estimated column (5) betas for QE in table 5, 6 and 7 column (5). This proves that the model is robust and can be assumed to estimate the value of the QE beta with a fair degree of certainty.

Different data

The section above explored the effect of QE on the yield curve parameters using the NS-model and Euro-area spot rate data containing bonds of all ratings. In this section the data on triple-A bonds is used as well as the estimated level, slope and curvature parameters that are estimated using the model that adjusts for multicollinearity. In this section only the regressions from column (4) and (5) are reported because the section above provides sufficient evidence on the validity of the model. Appendix 9 reports the robustness test using the same variables as above.

Table 9 - Regression results for triple-A rated bonds

Table 9 β1 (1) β1 (2) β2 (3) β2 (4) β3 (5) β3 (6)

QE dummy -0.843*** -0.800*** 0.0436 0.400*** 0.277*** 0.243** (0.0321) (0.0349) (0.0418) (0.0483) (0.0750) (0.0833) ECB policy rate 0.576*** 0.588*** 0.353*** 0.472*** 0.769*** 0.760***

(0.00947) (0.00758) (0.0138) (0.0132) (0.0278) (0.0270) Std. inflation 0.152*** 0.150*** -0.337*** -0.307*** 0.124*** 0.125*** (0.0115) (0.0117) (0.0131) (0.0154) (0.0329) (0.0329) Std. employment -0.492*** -0.505*** 1.081*** 0.986*** -0.744*** -0.736*** (0.00791) (0.00855) (0.0101) (0.0124) (0.0203) (0.0217) Std. production -0.00401 0.207*** 0.137*** (0.00816) (0.0143) (0.0121) M1 growth rate 0.028*** 0.0253*** 0.0118*** 0.0712*** 0.073*** (0.00151) (0.00147) (0.00197) (0.00463) (0.00505) Risk aversion -0.0125** -0.0961*** 0.00905 (0.00441) (0.00521) (0.0130) Constant 2.813*** 2.808*** -3.294*** -3.436*** -4.073*** -4.068*** (0.0197) (0.0179) (0.0233) (0.0244) (0.0599) (0.0588) N 3010 3008 3010 3008 3010 3008 adj. R-sq 0.867 0.866 0.879 0.891 0.426 0.425

Note: Standard errors in parentheses * p<0.05, ** p<0.01, *** p<0.001, The table shows the effect that the independent variables have on the level beta (β1, 2 and 3)triple-A bonds NS-model . All variables but the QE dummy are control variables and no explanatory value can be attached to the estimated betas. The description of the variables can be found in

appendix 5.

Table 9 shows the regression output for the estimated NS-parameters using the dataset containing triple-A bond data. The results are similar to the output for the data containing all bonds. The largest difference is the significance of third beta parameter, this parameter is insignificant for the previous data but is significant when the triple-A bonds are used.

Note: Figure 11 and 12 the “average” yield curve with and without the QE policy. The dotted lines show the difference between both yield curves. The increasing shape of the difference curve proves the flattening effect that QE has on the yield curve.

Figure 11 and 12 show the effect that QE has on the yield curve. The changes in both yield curves are larger for the triple-A bonds than for all bonds in figure 9 and 10. For these types of bonds the QE policy has a flattening effect on the yield curve. These larger effects can be explained by taking a look at the targeted bonds of the QE

results in the figures 9 and 10 are smaller because they also consist out of non-targeted bonds that are issued by Euro-area peripheral countries. While the differences are small, the largest difference comes from the regression without risk aversion for a ten-year maturity and is only about 0.3%. The asymmetric effect over the euro area can create new imbalances in the currency area, that make it harder for the ECB to reach her targets in a currency area without a fiscal transfer system. While a difference of 0.3% appears small it is important to keep in mind that this is only a part of the real difference. Considering that the dataset containing all bonds also includes triple-A rated bonds making the 0.3% an underestimation of the real difference in effect between targeted and peripheral countries.

The area inside the little dashed box is shown in appendix 6 in order to assess the 95% confidence interval of both yield curves. Because of the small values of the standard errors those confidence intervals are not visible in figure 11.

Inflation expectations

The signaling channel described in the literature overview assumes forward looking investors that take the future stance of monetary policy into account as well as the future level of inflation. The future level of inflation can also help to predict the future stance of monetary policy. When the central bank conducts monetary policy via a Taylor rule (Taylor, 1993) an investor can estimate the future stance of monetary policy using his or her expectations on future inflation. Chun (2011) shows evidence for the sensitivity of the yield curve to changes in expectations on both future inflation as well as the future output level. By a lack of reliable data on expectations on the future output level this section only assesses the effect of inflation expectations. Because of the lack of reliable inflation expectations data from before 2008, this section uses employs a dataset that contains data from July 22 2008.

The section above already proves that the different datasets give similar results regarding the effect of QE. In order to prevent an information overload this section only shows the effect from inflation expectation on the effectiveness of QE for the estimated NS-parameters for all bonds data. The regression output table can be found in appendix 10. The variable M1 growth is omitted because of multicollinearity issues. Figure 13 and 14 show the effects of QE without and with inflation expectations.

Note: Figure 18 and 19 show the “average” yield curve with and without the QE policy. The dotted lines show the difference between both yield curves. The increasing shape of the difference curve proves the flattening effect that QE has on the yield

The biggest difference between graphs 13 and 14 in comparison with other similar graphs above is the lack in impact on the short run from the QE policy. This can be explained by the employed dataset, where the sections above employs a datasets that runs from 2004-2016 the dataset used to assess the effect of inflation expectations only runs from 2008-2016. During this period ECB interest rates where equal or below 1%. Considering the monopoly power that the ECB has over the short-term interest rate it seems logical that the lower end of the yield curve did not change much over this sample period.

The effect that QE has on the yield curve does not change much when the regression incorporates an inflation expectation variable, the difference is about 0.1% over each maturity. The lack of impact from the inflation expectation can be explained by the way the ECB conducts monetary policy. The ECB actively targets inflation expectations, therefore, the inflation expectation variable is not independent from the ECB policy rate as well as the QE dummy. This relation had an effect on the influence of the inflation expectations variable.

Section 6: Conclusion

References

Altavilla, C., Carboni, G. & Motto, R. (2015) Asset purchase programmes and financial markets: lessons from the euro area. European central bank working paper series. No 1864 Ang, A. & Piazzesi, M. (2003) A no-arbitrage vector autoregression model of term structure dynamics with macroeconomics and latent variables. Journal of monetary economics. Vol.50 745-787

Anneart, J,. Cleas, A.G.P., De Ceuster, M.J.K. & Zhang, H. (2013) Estimating the spot rate curve using the Nelson-Siegel model: A ridge regression approach. International review of

economics and finance. Vol.27 482-496

Bernanke, B.S. & Reinhart, V.R. (2004) Conducting monetatery policy at very low short-term interest rates. American economic review. Vol.94(2) 85-90

Bernhardsen, T. (2000) The relationship between interest rate differentials and macroeconomic variables: a panel data study for European countries. Journal for

international money and finance. Vol.19(2) 289-308

BIS (2005) Zero-Coupon Yield Curves: Technical Documentation. Conducted on June 16, 2016. < http://www.bis.org/publ/bppdf/bispap25.htm >

Bomfim, A.N. (2003) Monetary policy and the yield curve. Federal Reserve Board.

Bowdler, C. & Radia, A.(2012) Unconventional monetary policy: the assessment. Oxford

review of economic policy. Vol.28(4) 603-621

Chadha, J.S. & Waters, A. (2014) Applying a macro-finance yield curve to UK quantitative easing. Journal of banking and finance. Vol.39 68-86

Christensen, J.H.E. & Krogstrup, S. (2016) Transmission of quantitative easing: the role of central bank reserves. Federal Reserve Bank of San Fransisco. Working paper 2014-18 Diebold, F.X., Rudebusch, G.D. & Aruoba (2006) The macroeconomy and the yield curve: a dynamic latent factor approach. Journal of econometrics. Vol. 131 309-338

ECB (2015a) ECB announces expanded asset purchase programme, 22 January 2015. Conducted on June 6, 2016.

< https://www.ecb.europa.eu/press/pr/date/2015/html/pr150122_1.en.html >

ECB (2015b) The transmission of the ECB’s recent non-standard monetary policy measures.

ECB economic bulletin. Vol.2015(7)

ECB (2016a) Debt securities, outstanding amounts and transactions of euro-denominated debt securities issued by euro area residents, by sector of issuer and original maturity. Conducted

on May 25, 2016

<https://www.ecb.europa.eu/stats/money/securities/debt/html/index.en.html>

ECB (2016b) Asset purchase programme, Public Sector Purchase Programme. Conducted on May 24, 2016 <https://www.ecb.europa.eu/mopo/implement/omt/html/index.en.html>

ECB (2016c) Euro area yield curve. Conducted on June 8, 2016

ECB (2016d) The ECB’s Directorate General Statistics releases euro area yield curves every TARGET working day at 12 noon Central European Summer Time (or Central European

Time). Conducted on July 13, 2016 <

https://www.ecb.europa.eu/stats/money/yc/html/technical_notes.pdf?update=20160513 > ECB (2016e) Monetary policy decisions, 10 March 2016 Conducted on June 6, 2016. <https://www.ecb.europa.eu/press/pr/date/2016/html/pr160310.en.html>

ECB (2016f) Structure of the euro area economy, External trade. Conducted on June 30, 2016.

< https://www.ecb.europa.eu/mopo/eaec/trade/html/index.en.html >

Eggertsson, G. & Woodford, M. (2003) The zero bound on interest rates and optimal monetary policy. Brookings paper on economic activity. Vol.34(1) 139-211

Fabozzi, F.J., Martellini, L. & Priaulet, P. (2005) Predictability in the shape of the term structure of interest rates. Journal of fixed income. Vol.15(1) 40-53

Friedman, M. & Schwartz, A.J. (1965) Money and Business cycle.The state of Monetary

economics. 32-78

Gern, K.J., Jannsen, N., Kooths, S, & Wolters, M. (2015) Quantitative easing in the Euro area: Transmission channels and risks. Intereconomics. Vol.50(4) 206-212

Greenwood, R.M. & Vayanos (2014) Bond supply and excess bond returns. The Review of Financial Studies. 2014, Vol. 27(3) 663-713

Hair, J. F., Jr., Anderson, R. E., Tatham, R. L. and Black, W. C. (1995) Multivariate Data

Analysis, 3rd ed, Macmillan Publishing Company, New York.

Herbst, A.F., Wu, J.S.K. & Ho, C.P. (2014) Quantitative easing in an open economy- Not a liquidity but a reserve trap. Global finance journal. Vol.25 1-16

Ioannidess, M. (2003) A comparison of yield curve estimating techniques using UK data.

Journal of banking and finance. Vol. 27(1) 1-26

Joyce, M., Miles, D., Scott, A. & Voyanos, D. (2012) Quantitative easing and unconventional monetary policy, an introduction. The economic journal. Vol.122 F271-F288

Khemraj, T. & Yu, S. (2016) The effectiveness of quantitative easing: new evidence on private investments. Applied econometrics. Vol. 48(28) 2625-2635

Kiyotaki, N. & Moore, J. (2012) Liquidity, business cycles and monetary policy. National

Bureau of Economic Research. Working paper 17934

Krishnamhurthy, A. & Vissing-Jogrensen, A. (2011) The effect of quantitative easing on interest rates: channels and implications for policy. Brookings paper of economic activity. 43(2) 215-287

Krugman, P.R. (2000) Thinking about the liquidity trap. Journal of Japanese and

international economies. Vol. 14(4) 221-237