1
The change in GDP growth during the quantitative easing policy
in the UK from 2009 up until 2014
Bachelor thesis
By: Nicole Zomerhuis
Studentnumber: 10180931 Supervisor: Rutger Teulings
Date: 3-7-2014
Faculty of Economics and Business Track: Economics and Finance Field: Economics
2
Content
Abstract page 3
Introduction page 4-5
Section 2: Theoretical framework page 5-13
2.1 The central bank balance sheet page 5-6
2.2 Macroeconomic model page 6-7
2.3 Monetary policy tools page 8-10
2.4 Quantitative easing and transmission mechanisms according to BoE page 10-11
2.5 UK QE policy page12-13
Section 3: Literature review page 13-17
3.1 Interest rate reduction page 13-15
3.2 Effect of QE on economic growth page 15-17
Section 4: Empirical model and Methodology page 17-20
Section 5: Data page 20-22
Section 6: Results page 22-27
Section 7: Robustness page 27-28
Section 8: Conclusion page 28-30
Bibliography page 31-33
3
Abstract
This thesis tests if quantitative easing caused more growth in the UK from 2009 up until 2014. Quantitative easing(QE) is an unconventional monetary policy which is used when
conventional monetary policy is at the zero lower bound and cannot increase demand anymore. QE policy is a policy whereby government bonds are bought in large quantities. This increases the price of the bonds which in turn decreases the yields on these bonds. Investors will seek more risky investments and this will increase trading. More trading means more investment which should in theory increase GDP growth. Current research is divided between QE causing more growth and not causing any increased growth. There is no single consensus on this matter. This thesis will try and investigate the relation between QE asset purchases and increased GDP growth. The model used to test if quantitative easing increased growth, measures the effectiveness of central bank monetary policy tools and intermediate targets. The model is estimated using OLS and the empirical methodology used is the general-to-specific econometric modelling methodology. From the results obtained, this thesis
concludes that the QE policy did not cause more GDP growth and that it even decreased growth. Thereafter it concludes that QE has a lagged effect on growth of six to nine months. This lag effect is large.
4
1 Introduction
In 2008 the world was hit by a severe crisis. This crisis decreased the aggregate demand of countries steeply. To counter the crisis and try and increase aggregate demand the central banks used different monetary policy tools. One of these tools was quantitative easing (QE).
Quantitative easing is an unconventional monetary policy tool. Unconventional tools are used when the conventional tools no longer have an effect on demand. This is the case when the conventional tools are at a zero lower bound which happened during the 2008 crisis. During the time that the conventional monetary policy tools were at their zero lower bound the aggregate demand in countries like the UK was decreasing (Joyce, 2011, p 200). When demand falls, the wealth of a country falls too, which is an undesirable situation. The central bank of the UK, the Bank of England (BoE), chose QE as an unconventional policy measure to increase GDP growth. This tool had first been used in Japan in 2001 (Breedon et al., 2012, p.703). The Bank of England started quantitative easing in March 2009 (Joyce et al., 2011, p.200).
Quantitative easing is a policy whereby government bonds called gilts are bought by the central bank of a county. In theory, this would decrease the yields on these bonds and investors would replace the gilts they sold to the Bank of England with other securities. This would then in turn decrease those yields etc. Investors would seek more risky investments and trading would increase. This trading increase theoretically would increase investment and therefore aggregate demand.
From the perspective of economic research it is important to know if quantitative easing indeed produces growth. Furthermore, it is vital to find a solution because when unconventional measures are needed to improve the state of the economy, the crisis is severe (Joyce, 2011, p 200). When a crisis is severe the economy is highly unlikely to recover on its own. Nevertheless, the economy needs to recover because wealth is decreasing. It is important to know if the big capital injection used with quantitative easing causes an acceptable increase in growth. Society benefits from the growth because wealth will increase.
This thesis will focus on the quantitative easing policy of the United Kingdom. It will determine if the policy worked or not. The research question will be: did quantitative easing cause increased growth in the UK from 2009 up until now? This question will be answered by doing a linear OLS regression and analysing if the factor quantitative easing was a significant determinant of economic growth. The analysis reflects that quantitative easing did not cause
5
more growth and even deteriorates it. Also, it has a lagged effect on GDP growth of six to nine months. This long-run effect is large.
This thesis will consist of 6 sections. First in section 2 there will be an explanation of the theoretical framework needed to understand quantitative easing. After that a review of the literature in section 3 will be given. This review will reflect the current debate and research on quantitative easing. In addition, section 4 will present the methodology and models used for the analysis. The data used for the regression will be explained in section 5 and in section 6 the results of the analysis will be described and analysed. The robustness of the models used will be tested in section 7. To conclude, the answer of the research question will be given in section 8. To add to this section the flaws of the research and suggestions for a follow-up study will be defined.
2 Theoretical framework
This section will explain a selection of theoretical models that are important to understand the concept of quantitative easing. In paragraph 2.1 the central bank balance sheet in general will be clarified. Paragraph 2.2 will describe the theoretical macroeconomic models which are important to understand the dynamics of central bank policy. Furthermore in paragraph 2.3 the conventional monetary policy tools will be looked at. Also, in paragraph 2.4 quantitative easing will be explained in detail. Finally, paragraph 2.5 will give an overview of the
quantitative easing policy of the Bank of England.
2.1 The central bank balance sheet
The central bank of a country has the prime goal to insure economic stability. Theoretically, this is created by using monetary policy. When the central bank conducts monetary policy this changes its balance sheet and as a result the money supply.
The balance sheet of a central bank consists of two parts: assets and liabilities, which can be seen in figure 1.
6
The assets side comprises of government securities and loans to banks. The government securities part is an amount of government bonds bought by the central bank. This transaction increases the money supply in a country, because the assets are bought with money. Loans to banks can also increase the money supply, because when a loan is given out to a bank the bank has more money to lend to consumers. Banks can borrow at central bank interest rates which in the case of the Bank of England is called the Bank rate (Joyce et al., 2011, p.200). Currency in circulation (money) and reserves are aspects of the liabilities side of the central bank balance sheet. The sum of all liabilities is called the monetary base. The currency in circulation is all money which is in the hands of the public (Bank of England, 1981). The reserves part of the balance sheet of the central bank is money that has been deposited at the central bank by regular banks. These reserves can both be required and excess reserves. Required reserves are a percentage of all deposits that banks have to keep in funds and are required by the central bank. Excess reserves are reserves held by banks to insure that in an economic downturn no unnecessary costs such as: borrowing costs from loans from other banks, selling securities, selling off loans and borrowing from the central bank occur (Frost, 1971). The amount of excess reserves increases when the likelihood of financial difficulty is larger (Frost, 1971).
When the central bank changes the balance sheet composition this effects the economy as a whole. How the economy is effected is elaborated on below.
2.2 Macroeconomic models
To help understand the central bank system, supply and demand curves can be used. This model helps to understand how the monetary policy tools affect the wider economy.
Figure 2 shows the supply and demand curves. The y-axis is the
overnight interbank lending rate. This is the rate banks charge each other when lending from each other. The x-axis is the
quantity of reserves. Above was stated that the reserves consist of required and excess reserves. These excess reserves can be deposited at the central bank for a depository interest
Figure 2: Supply and demand curves for understanding the central bank system. Source: Giuliodori et al., 2012, p.328
7
rate. Banks can also lend from the central bank for a lending rate. The lending rate (iL) is reflected in the figure by the horizontal part of the supply curve. The deposit rate (iD) is reflected by the horizontal part of the demand curve. These two rates are both set by the staff of a central bank. The deposit and lending rates can let the interbank lending rate vary between the bounds. Accordingly, the Bank of England staff can manipulate the interbank lending rate by changing the deposit and lending rates.
In figure 2 the demand curve is labelled by Rd and is a downward sloping curve. This is caused by the fact that when the interbank interest rate is above the depository interest rate of the central bank, the opportunity costs of holding excess reserves at the central bank are positive. The more the interbank lending rate increases, the higher the opportunity costs become and therefore fewer excess reserves will be held. The contrary is also true. When the interbank interest rate is below the deposit rate, the return on the deposits is higher at the central bank than in the interbank market so the amount of excess reserves deposited at the central bank will go to infinity (Frost, 1971).
The supply curve Rs has a prominent vertical line. This line projects the non-borrowed reserves which is the total monetary base minus the total banks borrowings (Bank of England, 1981). The horizontal part is created by the lending rate iL. When the interbank lending rate is below the central bank lending rate, banks will not borrow from the central bank. On the contrary, when the rate is above the lending rate, there will be infinite central bank borrowing by banks because it is cheaper to lend from the central bank than from other banks (Frost, 1971).
The lending and depository rates keep the interbank lending rate between two bounds. When the demand for reserves suddenly has a positive shock, this means the demand curve would shift to the right. If the demand curve shifts up enough, it will intersect the supply curve in the horizontal part. Therefore, the interbank lending rate will never be higher than the central bank lending rate. Also, when the central bank increases its non-borrowed reserves by buying government securities the intersection eventually ends up in the horizontal part of the demand curve. And therefore the interbank rate will never be below the deposit rate.
In this way, the interbank lending rate can be manipulated by changing the lending and deposit rates. Changing the lending rate shifts the horizontal part of the supply curve.
Changing the deposit rate, shifts the horizontal part of the demand curve. By manipulating these rates the central bank can target the interbank rate and as a result GDP growth.
The last important factor in the model is point of intersection: i*. This is the equilibrium interbank lending rate which can influence GDP growth.
8
2.3 Monetary policy tools
Monetary policy tools all target interest rates. When the economy is in a downturn the tools are used to influence the interest rate between banks. In a downturn, the tools are used to decrease the interest rate. In a steep upturn, the growth is softened by increasing the interest rate. A low interest rate makes borrowing more appealing and a high one less appealing. This causes spending and accordingly growth to increase and decrease (Gärtner, 2012). By looking at the different monetary policy tools it will become clear how exactly the interest rates and therefore GDP growth is affected. There are three types of conventional monetary policy tools: open market operations, the interest rate and reserve requirements.
Open market operations are operations where the central bank buys or sells government bonds to increase and decrease the money supply. This is the most important monetary policy tool (Bank of England, 1981). The advantages of open market operations are that it is precise, easily reversed and implemented quickly.
An expansionary open market operation is an operation where the central bank buys government bonds. This causes the money supply to increase which is graphically shown by shifting the vertical NBR line to the right. This causes interbank lending rate to fall provided that the initial intersection was in the downward sloping part of the demand curve. The intersection in the downward sloping part of the demand curve means that the interbank lending rate is between the lending and deposit rate bounds. Therefore, the interest rate can fluctuate with changes in reserves. When the amount of reserves in the economy is increased, relative demand for money decreases. This causes demand for loans to decrease. As a result, the interbank interest rate will decrease. This decrease will never be more than the central bank deposit rate. A contractionary operation does the opposite and the interbank interest rates increase. For this reason, in order to increase GDP growth the interest rate has to fall and as a result an expansionary open market operation is needed. The contrary is also true.
The second tool is the interest rate. The central bank can decide on their depository (iD) and lending rates (iL). The Bank of England calls these rates the Bank rates (Joyce et al., 2011). When the central bank changes the lending rate, this only has an effect if the demand curve intersected the supply curve in the horizontal part. A lending rate decrease will decrease the interbank rate because banks will not be willing to borrow from each other at a higher rate than what the central bank offers (Bank of England, 1981). A lending rate increase will increase the interbank rate, because banks who lend the money to other banks can achieve a better return by raising the interest rate. When the initial intersection was not in the horizontal
9
but vertical part of the supply curve a change in the lending rate would not have effect. This can be explained by the fact that the interbank rate is lower than the lending rate and
accordingly no borrowing will take place at the central bank unless the lending rate is decreased sufficiently. This situation could arise when there is high competition between banks.
When the central bank changes the deposit rate, this only has an effect if the supply curve intersects the horizontal part of the demand curve. An increase in the deposit rate will increase the interbank rate. This is because otherwise the deposits will all be at the central bank and the banks would not have any deposits left. A decrease in the deposit rate will decrease the interbank rate. Banks that offer deposit accounts will want to decrease deposit rate costs and thus decrease the rate. On the other hand, when the initial intersection is in the downward sloping part of the demand curve a change in deposit rate does not have effect. The deposit rate is lower than the interbank rate and no commercial bank will deposit funds at the central bank. The deposit rate needs to increase up until point of current intersection to influence the interbank rate.
The advantages of the Bank rates are that it brackets the interbank lending rate into a certain range and central bank staff can manipulate the interbank lending rate. To increase GDP growth the lending rate needs to be reduced to provide cheaper borrowing. Moreover, the deposit rate needs to decrease otherwise consumers will save and not spend which decreases growth. The contrary is also true.
The third and final tool is the reserve requirement. When the reserve requirements are adjusted this affects the downward sloping part of the demand curve. When the requirements are increased this immediately raises the demand for reserves. This causes the demand curve to shift outward (to the right) and the interbank interest rate to increase given the supply curve intersects the downward sloping part of the curve. This increase in the interbank interest rate is caused by the fact that the demand for loans will increase to meet the increased reserve requirements. This increase in demand makes loans more expensive and as a result the interest rate increases. A higher interest rate causes GDP growth to decrease. But when the initial intersection of the supply curve is in the horizontal part of the demand curve this will not affect the interbank rate. In this situation the supply of money is higher than the demand for money. This means that the demand needs to increase past the intersection to have an impact on interest rates. In the United Kingdom no reserve requirement exists (Bank of England, 1981).
10
To conclude, all monetary policy tools try to influence the interest rate to change GDP growth. The initial intersection of the two curves indicate whether a policy will have effect on the interest rate or not.
2.4 Quantitative easing and transmission mechanisms according to BoE
Quantitative easing is an unconventional monetary policy. It is a policy where the central bank buys government and private sectors bonds in large quantities. This policy was implemented after the conventional tools reached their zero lower bound and did not work anymore (Kapetanios et al., 2012 p. 316).
The theory behind quantitative easing has to do with transmission channels.
Transmission channels are channels whereby the pass through of effects is analysed (Clementi et al., 2005, p. 3). Here, the pass through from the point of buying bonds to the point where theoretically demand should increase is analysed. Joyce et al. (2011, pp. 201-202) considers 5 channels. These are: portfolio balance effects, liquidity premium effects, bank lending effects, confidence effects and policy signalling effects. The first three all have an effect on interest rates which either increases investment or borrowing. The final two have an effect on confidence.
The portfolio balance effect of quantitative easing is the effect of the increasing price of assets after the Bank of England has purchased gilts. This transmission channel is
considered the most important when deciding on using quantitative easing or not (Joyce McLaren &Young, 2013, p.672). The demand for gilts has gone up and the amount of gilts outstanding has remained constant which, in turn, causes the price of the gilts to rise. The sellers of the gilts to the Bank of England now hold money. This money is directly reinvested in other assets. Demand for other assets increases and the prices of these assets increase too. The price increase causes the yield on the assets to fall. Borrowing for firms and households will become cheaper because the yields have fallen. Cheaper borrowing could increase spending and thus demand (Joyce et al., 2011 , p. 201). What is more, the lower yields will encourage investors to seek more risky investments to try and increase their returns (Joyce et al., 2011 , p. 201). As a consequence, investment increases which causes growth of GDP.
This channel will only work when the increased money holdings of the sellers of the gilts to the central bank is not a perfect substitute of the sold gilts (Joyce, McLaren &Young, 2013, p 674). When it is a prefect substitute, this means the sellers will hold the money and not rebalance their portfolio’s (Joyce, McLaren &Young, 2013, p 674). When it is not a
11
perfect substitute, investors seek a better investment than money balances and buy assets like corporate bonds (Joyce, McLaren &Young, 2013, p 674).
The magnitude at which this portfolio rebalance effect works is partly dependent on the economic situation. When there is a bad economic outlook and investors become more risk averse, this means the money balances received from gilt purchases come increasingly more substitutable for assets. This means investors do not rebalance their portfolio’s and the transmission channel fails (Joyce, McLaren &Young, 2013, p 678).
The liquidity premium transmission channel is encouraged by quantitative easing because the policy increases trading of assets (Joyce et al., 2011 , p. 202). Trading is
increased when the amount of assets bought and sold every day rises. This means that assets become more liquid, because it is easier to sell the assets and therefore easier to convert them into money. When an asset is illiquid it is an asset which is not easily traded and not easily converted into money. This is an undesirable trait and thus traders expect to be compensated for this inconvenience by a higher yield. The prices on these illiquid assets are lower than the liquid assets. Therefore, when liquidity is increased less compensation is needed for the undesirable illiquidity trait and the price of the asset increases. The price increase forces yields down which has the same effect as was stated above. (Joyce et al., 2011 , p. 202)
The bank lending effect originates from having more money in the system due to conduction of QE. Some money will be deposited at banks. This will increase bank liquidity and they will be more willing to lend money to firms and households. This increases spending (Joyce et al., 2011 , p.202).
The confidence effect of quantitative easing is another transmission channel and is caused by the belief that the measure betters the state of the economy which increases consumer confidence. This increases spending immediately (Joyce et al., 2011, p. 202).
The policy signalling effects of quantitative easing lie in the fact that the asset
purchase announcements could indicate the central bank is serious about hitting inflation rate targets. The definition of inflation is an increase in the price level of a country. By committing to an inflation target, expectations on inflation will be near to central bank targets. This will raise confidence in the central bank. This confidence will increase spending.
To sum up, quantitative easing can work through several different channels. The portfolio rebalancing effect is believed to be the main transmission channel through which QE works (Joyce et al., 2011 , p. 202).
12
2.5 UK QE policy
Quantitative easing is an unconventional monetary policy tool which was adopted by the Bank of England from 2009 till 2012 (Joyce et al., 2011 , p. 200). To be precise it adopted the measure on 1 March 2009 until the end of November 2012 (Joyce et al., 2012, p. 671).
In 2009 the United Kingdom was experiencing a decrease in demand because of the financial crisis. This called for conventional monetary policy and the interest rate was lowered significantly. On 5 February 2009, the interest rate was reduced to 1% (Joyce et al., 2011 , p. 204). This did not increase aggregate demand enough so a month
later it was further decreased to 0.5% (Joyce et al., 2011 , p. 204). In figure 3, the deposit rate has been lowered to the zero lower bound, mimicking the situation in the UK in 2009. When expanding the balance sheet of the Bank of England, it is clear the overnight lending rate will not be affected anymore (Bank of Engand, 1981). Generally, such a significant reduction in interest rates causes cheaper borrowing. When it is cheaper to borrow, the consumer usually spends more and demand increases. During this crisis, this mechanism did not cause the growth of the U.K. to be positive (Joyce et al., 2011, p. 204). The reaction of the Bank of England was to conduct quantitative easing a month after the last lowering of the interest rates (Joyce et al., 2011 , p. 204).
The quantitative easing policy implemented by the Bank of England was done by primarily buying governments bonds. The government bonds are called gilts. The purchase of these assets is a continuous process over several months.
Similarly, the Bank of England also bought a small portion of private sector bonds (Joyce et al., 2012 p 672). These private bonds had to be good quality (meaning low risk of default) commercial paper (Joyce et al., 2011 , p. 202).
Quantitative easing was performed in three periods (Joyce et al., 2012 p 671). The first period was called QE1 and was carried out from March 2009 till January 2010. During this period 200 billion (bln) pounds worth of assets were bought, 198bln of which were
government bonds and 2 bln corporate bonds (Joyce et al., 2012 p 671). The first asset purchase announcement was made on 5 March and announced a purchase of 75 bln pounds worth of assets. On 7 May, the Bank of England announced an extension of the purchases worth 50 bln pounds. On the 6 of August another 50 bln was announced and on 5 November a
Figure 3: Supply and demand curves for with the deposit rate at the zero lower bound Source: Giuliodori et al., 2012, p.336
13
further 25 bln. The 200 bln worth of assets bought was 14% of UK GDP (Kapetanipos et al., 2012, p 316).
The second period was called QE2 and was carried out from October 2011 till May 2012. During this period another 125 bln worth of government bonds were bought. The
contradicting part is that the Bank of England announced beforehand that purchases would not increase past the 200 bln pound mark (Joyce et al., 2011 , p. 204).
The final period was called QE3 and was from July 2012 till November 2012. The amount of central bank spending on assets was another 50 bln pounds. This gives a total of 375 bln pounds of assets bought. At the time, this was 25% of GDP and 35% of all total issued gilts (Joyce et al., 2012 pp. 671-672).
3 Literature review
The current literature seems divided between two fields of research. First, the research on yield reduction on gilts caused by QE and second the increase in GDP growth because of this yield reduction. Accordingly this literature review will first discuss the views of different researchers on the yield reduction effect of QE and second the effect of QE on GDP growth.
3.1 Interest rate reduction
There is uncertainty about what transmission channels caused the yield reduction and how much of the yield reduction is a direct result of the asset purchases of the Bank of England.
Joyce et al. (2011, pp. 201-205) suggests that the biggest impact of the asset purchases is on the gilt yields. This is due to the fact that the Bank of England bought gilts the most out of all assets. The immediate reaction on the announcement of the QE policy was a reduction of 0,75% in gilt yields. After the first actual purchases the yield reduced further to 1%. They also find that the portfolio rebalancing transmission channel is probably the most important channel through which QE causes a positive reaction to economic growth. These results were later emphasized again by Joyce, McLaren & Young (2013, p 680). Joyce, McLaren & Young (2013, p 680) also concluded that for every 1 billion pounds unexpected extra investment, the yields would fall by 0.69%. Therefore, when the actual purchases were carried out and these were more than expected, the yields fell by 1.25%.
14
Another paper that supports both the conclusions made in the Joyce et al. (2011) paper is Gagnon et al., (2011, pp. 28-30). Gagnon et al., (2011, p.28-30) provide evidence for the US that is similar to that of the UK. The yields on 10 year treasury bonds in the US declined by 0.91%. Furthermore, the portfolio rebalancing transmission channel was also proven to be the most important transmission channel. Christensen & Rudebush (2012, p 385) argue that the US and UK both bought about the same amount of government bonds relative to the size of the economy and outstanding government debt. As a result the Gagnon et al. (2011) and Joyce et al. (2011) conclusions can be compared. Hence, these papers can be viewed as evidence that the conduction of QE in two differing countries has the same effect and can therefore be generalized.
Conversely, Christensen & Rudebush (2012) found that when the Bank of England and the Federal Reserve (US central bank) announced their quantitative easing programmes the overnight index swap rate reactions differed. The overnight index swap rate is an
interbank lending rate. In the US the government bond yields fell jointly with the swap rates. The UK, however, saw swap rates only fall by a small margin. This difference could indicate that there is a different pass through of effects of QE in the UK and the US. This could mean that the portfolio rebalance transmission channel is not the main transmission channel through which QE causes a reduction in yields (Christensen & Rudebusch, 2012, p 386). This
challenges the conclusions made by Gagnon et al. (2012) and Joyce et al. (2011).
Christensen & Rudebusch (2012, p 386) suggest the signalling transmission channel is a dominant factor in the US. The signalling channel causes expectations of short term interest rates to fall which in turn causes all long interest rates to fall (Christensen & Rudebusch, 2012, p 386). This pass through of effects reduces the swap rates which is responsible for 60% of the total change in 10 year US treasury interest rates. This 60% was responsible for a 0.53% decline in treasury yields (Christensen & Rudebusch 2012, p. 399). The portfolio rebalancing channel only accounts for 0.29% decline in US government bond yields
(Christensen & Rudebusch, 2012, p. 399). Thus, the signalling effect has a bigger impact than the rebalance effect in the US.
Nevertheless, in the UK, gilts and swaps were imperfect substitutes and did not move in sync (Christensen & Rudebusch, 2012, p 386). This suggests that the signalling effect was not a significant transmission channel in the UK which was also found by Breedon et al. (2011, pp. 718-719). The portfolio rebalance transmission channel is thought to be the most important channel by reducing yields by 0.43% (Christensen & Rudebusch, 2012, pp. 404-406).
15
Chadha & Waters (2014, p. 84) however, concluded that at least 0.50% yield reduction in the UK was caused by the signalling channel when quantitative easing was first announced. Yet, when the asset purchases actually began the portfolio rebalance transmission channel caused as much as 1.25% of reduction in UK gilts (Chadha & Waters, 2014, p.85). This was also predicted by Joyce, McLaren & Young (2013, p 680) and on average by Breedon et al (2012, p. 722). Chadha & Waters (2014, p.85) are confident that their analysis shows that QE has indeed had a significant impact on the lowering of UK gilt yields and that the portfolio rebalance transmission channel was the most prominent channel.
To conclude the above, the UK yield reduction was primarily caused by the portfolio transmission channel. The amount of yield reduction is still a point of debate and varies between 0,43% and 1,25% reduction. On the other hand, the transmission channel that caused more than half of the reduction in US government bond interest was the signalling channel. For this reason, when analysing the effects of QE, the transmission channel through which QE causes the yields to decrease is not certain and is probably country specific.
3.2 Effect of QE on economic growth
The effects of QE on the wider economy are hard to analyse because of three difficulties. First, it is difficult to specify the counterfactual (Breedon et al., 2012, p. 718). The
counterfactual is defined as the economic situation that would have arisen when QE was not implemented. Second, there are a lot of factors that influence the economy which are hard to isolate from QE effects (Joyce et al., 2011, p. 209). Third, the models do not capture all transmission channels through which QE works (Joyce et al., 2011,p. 210).
To correct for these problems Kapetanious et al. (2012, p328) created a model to assess what would have happened had quantitative easing not been implemented and compare this outcome with a model which includes quantitative easing. This model is called the
multiple time-series model (Joyce et al., 2011, p.209).
In order to test the effect of QE on economic growth, three models where estimated. First, a BVAR-model where structural change in the economy is allowed (Kapetanious et al., 2012, p317). This model estimates interrelationships well when a large data set is used . The interrelationship that needs to be assessed with QE is that of the gilts yield spread and GDP growth. Second, a SVAR model where a change in parameters is allowed to mimic policy changes (Kapetanious et al, 2012, p317). Third, a TVP-SVAR model that allows for time variation in parameters (Kapetanious et al., 2012, p317).
16
Kapetanious et al. (2012, p317) uses the Joyce et al. (2011) conclusion that medium to long-term gilt yields were reduced by 1% because of quantitative easing. This reduction in yields is used to estimate the model that includes quantitative easing. The counterfactual model, a model without the conduction of QE policy, consists of gilt yields that are increased by 1%.
The results of this analysis were provided by looking at the slope differences between the no QE policy model and the QE policy model (Kapetanious et al., 2012, p328).
Kapetanious et al. (2012, pp. 329-330) concluded that the maximum effect of QE on real GDP occurs after six to nine months. Looking at the BVAR- model where the Bank rate was set to 0.5% and the yield on gilts had a 1% decline due to QE, this caused a 0.72% increase in real GDP level compared to a policy without QE (Kapetanious et al., 2012, p330-332). The MS-SVAR model predicts that the increase in real GDP level will be 2.75% (Kapetanious et al., 2012, p340). Finally, the TVP-SVAR model predicts a 0.86% increase in GDP
(Kapetanious et al., 2012, p340). This gives a model average of 1.42%. One of the most important caveats with this analysis of the impact of QE is that QE might have influenced other variables within the model (Kapetanious et al., 2012, p. 341).
The model average of Kapetanious et al (2012, p. 340) comes close to what was concluded by Bridges & Thomas (2012, p.39). They suggest an average of 1.75% increase in GDP over the first part of QE. Bridges & Thomas (2012) constructed a model where the asset prices and spending was calculated that was needed in order to make money demand
consistent with the increase in the money supply due to the asset purchases. There is still a lot of uncertainty about the model outcome, because of the problem of identifying the
counterfactual, isolation of the QE effect on GDP growth and uncertainty of capturing all transmission channels (Bridges & Thomas, 2012).
Joyce et al. (2011, p. 210) also constructs a simple model but uses Q models. These models are constructed in a way that the amount invested is determined by the market value of assets relative to the cost to replace them. The outcomes of these models suggest that the real level of GDP growth will be between 1.5% and 2.5%.
However, Chen et al (2012, p. 291) suggests the growth rate is far lower. From a Bayesian analysis the effect on GDP growth was concluded to be 0.13%. Nevertheless, this effect is long-lasting and will last for 6 years .Within these six years the growth will be 0.07% above the growth that would have existed without the quantitative easing policy.
On the other hand there is also a lot of literature on how other factors than the asset purchases were the cause of economic growth during the time that QE was conducted. Curdia
17
& Woodford (2011), Eggertsson & Woodford (2003) and Wallace (1981) were all sceptical of an increase in growth by asset purchases. Eggertsson & Woodford (2003, p.165-166) thought that the asset purchase programme was not the cause of the GDP growth. The paper
concluded that changing the expectations on interest rates was the cause of GDP growth. What was not considered in the paper is that the conventional monetary policy tools are at their zero lower bound and cannot be used to produce more growth and change expectations. Consequently, a triggering event is needed to change expectations (Chada & Waters, 2014). This triggering event was the asset purchase programme. Therefore asset purchases indirectly cause an increase in growth and the conclusion in Eggertsson & Woodford (2003) is
challenged.
In conclusion, it is difficult to test if QE indeed produces growth because of the identification of the counterfactual, the influence of other factors than QE and capturing all transmission channels. Most recent research believes the asset purchases of 2009 did indeed produce more growth than without the purchases but research has still not reached a single consensus on this matter.
4 Empirical model and Methodology
This section will elaborate on the empirical model and methodology for the analysis to test if quantitative easing did indeed cause significantly more growth in the UK than without the policy.
The model that will be used is a multiple linear regression model with various control variables. A multiple linear regression model has one dependant and several independent variables. The independent variables have high explanatory power regarding the dependent variable. This gives the following form:
Yi = β0 + β1xi +…..+ βnxi + ε i (3.1)
The model and analysis in this thesis will be similar to that of Lyonnet & Werner (2012). In their paper a linear multiple regression model was estimated with OLS to test if quantitative easing was a significant factor in causing growth. As a result the dependent variable in the model is nominal GDP growth. The independent variables are all monetary policy instruments that influence nominal GDP growth. To be precise, the variables chosen in the model by Lyonnet & Werner (2012, p 95) are potential and actual monetary policy tools and intermediate targets that were or could have been used by the Bank of England to
18
influence growth. In other words, this means that all tools that the central bank can use to influence growth have been included in the model. This approach should solve the three problems of testing QE identified in section 3.
The monetary policy tools included as variables are both on the asset and liabilities side of the Bank of England balance sheet (Lyonnet & Werner, 2012, p99). In a financial and economic crisis the central bank changes its balance sheet to counter the crisis and cause more nominal GDP growth. Both sides of the balance sheet are affected. The asset side is affected by for example, asset purchases of government securities. The liabilities side is affected because, for instance, banks hold more excess reserves as a buffer for economic bad times and deposit them at the Bank of England. Thus the model has to include both asset and liability side monetary policy tools (Lyonnet & Werner, 2012, p99).
The paper of Lyonnet and Werner (2012) identifies six variables that represent the asset and liability side monetary policy tools. The first variable is the UK Bank Rate and this is the interest rate influenced by the lending and deposit rates set by the Bank of England. These rates affect the amount that is borrowed and deposited at the central bank. Therefore it affects the asset and liability sides. The second variable is the bank reserves. Bank reserves are all reserves of commercial banks held by the central bank. This is a liability side tool. The third variable is total assets. The total assets are all assets held by the central bank and
therefore are obviously part of the asset side of the balance sheet. The fourth variable is quantitative easing and is the variable of interest. The variable quantitative easing is included because the policy expands the asset side of the balance sheet. It differs from the variable total assets because quantitative easing is a policy where long term assets were purchased (see paragraph 2.5). As a consequence, Lyonnet and Werner (2012, p.99) created the variable qualitative easing and included this in the model. This variable is a ratio of long-term assets to total assets. The qualitative easing variable captures a shift from less risky and more liquid assets to more risky and less liquid assets. This variable thus tests quantitative easing (Lyonnet & Werner, 2012, p. 100). The fifth variable is the M4 money aggregate . M4 is a broad money aggregate and consists of notes and coins, deposits, commercial paper, bonds and estimated holdings of sterling bank bills (Hussain & Maitland-Smith, 2010). This aggregate captures the monetary supply in the economy which is also part of the liabilities side of the central bank balance sheet. M4 is chosen over all other monetary aggregates because M4 is the broadest aggregate (Hussain & Maitland-Smith, 2010). The sixth variable is the bank credit to the real economy. This is central bank lending to households, non-financial corporations and non-profit making institutions. Bank lending (credit) has shown to
19
be important (Bernanke & Gertler, 1995). Lyonnet and Werner (2012, p 100) created a bank credit variable that filters out the sectors that do non-GDP financial transactions. In earlier papers of Werner (1992,1997,2005) this variable has proven to be theoretically and empirically superior and for this reason is included. All these variables are included in the model in the following form:
GDPGROWTHi= β0+β1BANKRATEi+ β2BANKRESERVESi+ β3TOTALASSETSi + β4QUALITATIVEEASINGi + β5M4i+ β6M4BANKCREDITi + εi (3.2) To add to the six variables explained above lag variables will be included in the second part of the analysis. The lag variables are included because the effect of QE has a lag of six to nine months on GDP growth according to Kapetanios et al., (2012). To test this conclusion 3 QE lag variables will be added with lags of 1,2 and 3 periods respectively. The model will change into the following form:
GDPGROWTHi= β0+β1BANKRATEi+ β2BANKRESERVESi+ β3TOTALASSETSi + β4QUALITATIVEEASINGi + β5M4i+ β6M4BANKCREDITi + β7QElag1i + β8QElag2i +
β9QElag 3i + εi (3.3)
It will be assumed that the standard errors are homoscedastic, there are no omitted variables and the model has serial correlation. It is assumed there is serial correlation, because there is persistence in the data in the sense that part of the current GDP growth is explained by GDP growth from the past period. This is what causes serial correlation and using a dynamic model1 could correct for this. A GDP lag variable with 1 period of lag will be added to the model to capture the persistence effect and therefore serial correlation. This will give the following empirical model:
GDPGROWTHi= β0+ β1GDPlag1i+ β2BANKRATEi+ β3BANKRESERVESi+ β4TOTALASSETSi + β5QUALITATIVEEASINGi + β6M4i+ β7M4BANKCREDITi +
β8GDPLAGn i + β9 QElag1 i + β10QElag2 i + β11QElag3 i + εi (3.3)
When the analysis of the models with and without lags is run the econometric method used to estimate the coefficients of this model will be ordinary least squares (OLS). The OLS method calculates the best fit to the actual data points. The fit is estimated by minimizing the sum squared residuals (Stock & Watson, 2012, p. 156). The advantage of using OLS to estimate the coefficients in the multiple linear regression model is that it is simple to perform (Stock & Watson, 2012, pp. 159-161). Also, the estimated coefficients are easy to interpret (Stock & Watson, 2012, p. 157). Thereafter, theoretically it has a desired property in the sense that when the three OLS assumptions are met, the OLS estimator is unbiased and consistent (Stock & Watson, 2012, p. 161). The first OLS assumption is that the conditional distribution 1
20
of the error term given the independent variables has a mean of zero (Stock & Watson, 2012, p. 164). This means that the omitted factors are not correlated with the independent variables in the model but only correlated with the error term. The second assumption is that the variables are independently and identically distributed (Stock & Watson, 2012, p. 166). Which means the model has homoscedastic standard errors and no serial correlation. The third assumption says that large outliers are unlikely (Stock & Watson, 2012, p. 167). These data points are far out of the usual range of data.
The empirical methodology to test the estimated OLS variables is the general-to-specific model selection methodology (Lyonnet & Werner, 2012, p. 99). This methodology has shown to be robust when using time series models which is the case here (Bauwens & Sucarrat, 2010: Voutsinas & Werner, 2010: Werner, 2005). This methodology tests the
significance of all variables and excludes the variable that has the highest level of significance (Campos et al., 2005). After excluding the variable another regression is run. This sequential downward reduction to a form, which is called the parsimonious form, will allow an analysis of the policy models (Campos et al., 2005). The cut-off point is at a 5% significance level (Lyonnet & Werner, 2012, p. 100). When the variable which represents a policy has to be dropped from the model, it means the policy was not an important factor in causing growth (Campos et al., 2005). Conversely, when the variable is significant for 5%, it has most likely influenced nominal GDP growth (Lyonnet & Werner, 2012, p. 99).
5. Data
For the analysis of the models in section 4 a time period of the first quarter of 2001 till the first quarter of 2014 was chosen. This period is chosen to capture all effects of quantitative easing which was last conducted in November 2012. The literature described a six to nine months lag which means to capture this lag at least the data up until the third quarter of 2013 needs to be included. Two extra quarters after 2013 are added to the data to test if the lag was not more than the predicted nine months. The starting point of 2001 is chosen because this gives enough data before the start of the crisis to make the analysis more robust. Also, this will allow to test for a significant increase in GDP growth for the entire QE period (2009-2012). However, Lyonnet & Werner used data from 1985 till 2012. The paper was finished before later data was published and as a result only includes data from the QE1 period. The starting point of 1985 is too far back in time and therefore may not test well for the QE period thus in this thesis more recent data is used.
21
All data of the independent variables can be found in the interactive database of the Bank of England. The dependent variable: nominal GDP growth can be found in the database of the OECD. GDP growth is only available in quarterly data hence all other data has to be adjusted accordingly. All data that was provided in weekly or monthly periods have been averaged to quarterly data. To ease interpretation of the estimated coefficients and because the GDP variable is a growth rate all variables except the Bank Rate have to be converted into growth rates. The bank rate is excluded because this is the only variable which is not a process and is determined by Bank of England staff. The coefficients will now represent the percentage change in GDP growth when the growth of the variable changes by 1%.
The six variables were constructed in different ways. On 18 May 2006 there was a money market reform which altered the calculation method of the Bank of England for certain data and caused the variables to have discontinued data sets (Lyonnet & Werner, 2012,
p.103). Therefore the calculation of these variables differed prior to June 2006 from post June 2006. From the variables elaborated on in section 4, nominal GDP growth and Bank Rate data from 2001 till 2014 is readily available. On the other hand Bank reserves, total assets,
qualitative easing, M4 and bank credit to the real economy variables were not.
The bank reserves data can be found by simply looking at the bank reserves account on the Bank of England balance sheet. Post 2006 this account existed but prior to June 2006 the Bank Reserves account did not. Therefore the data had to be constructed by taking the M0 money aggregate (which is the most narrow money aggregate (Bank of England (1981)) and subtracting all notes and coins outside the Bank of England. Subtracting this amount will leave only the deposits at the bank and thus the bank reserves (Bank of England, 2006).
The data of the total assets variable can be found by looking at the assets side of the Bank of England balance sheet which comprises of the issue and banking department. By adding up the total assets of these departments the total assets of the Bank of England post June 2006 is constructed. Prior to June 2006 the total assets of both departments are not available. The total assets have to be constructed by adding up the securities 1, advances and other premises equipment and other, notes and coins, securities 2 and other securities accounts on the balance sheet (Lyonnet & Werner, 2012, p.103).
The qualitative easing variable is a ratio that consists of the long term assets divided by the total assets of the BoE. The long term assets can be found on the balance sheet and are divided up into long term issue and long term banking department assets. To calculate total Bank of England long term assets the long term assets of these two departments have to be added up. The total assets of the BoE were calculated as shown above.
22
The data for the M4 money aggregate can also be found on the bank of England balance sheet and consists of the M4 holdings of other financial corporations , private non-financial corporations and the household sector balance sheet accounts. All these accounts are available in the database of the Bank of England and this data is free from discontinuation.
Lastly, the M4 lending aggregate to the real economy had to be constructed. This variable is also free from discontinuation of data. It consists of the following balance sheet accounts: lending to private non-financial corporations, secured lending to the household sector, unsecured lending to the household sector and lending to unincorporated businesses and non-profit making institutions This excludes non-GDP influence lending (Lyonnet & Werner, 2012, p.103).
All the data for the construction of the variables that need to be constructed can be easily found in the Bank of England database by looking for the data code. In the appendix all data codes for all balance sheet accounts needed will be given in figure 1.
6. Results
In this section the results of the regressions of the models explained in section 4 will be given. The results will be presented in 2 tables. First, a model without a GDP lag variable will be tested. The results will be given in table 1 Second, dynamic models with a GDP lag variable will be tested. The results of all models with lags will be presented in table 2 and analysed.
First a static model without GDP lag was estimated. This model was presented in section 4 and is similar to Lyonnet & Werner (2012). The results are shown in table 1.
Table 1 shows all the coefficients for all models estimated. The standard errors are given below the coefficients. Model 1 in table 1 did not allow for time variables and thus no lags were added. The other two models allowed for QE lag variables of up to 3 periods. Model 2 did not delete the lag variables when insignificant and model 3 did. All models are the parsimonious models estimated by using the general-to-specific method. The models were tested for serial correlation which was detected in this model. The serial correlation was corrected by using a Orcutt correction with a Prais-Winsten option. Chocrane-Orcutt corrects for serial correlation, because it estimates a parameter when regressing the difference of the untransformed error terms and newly created lagged error terms. This parameter is used when testing the coefficients and corrects serial correlation.
23 Parsemonius model 1 static model(s.e. corrected for serial correlation) Pasemonius model 2 with QE lags (s.e. corrected for serial correlation)
Pasemonius model 3with QE lags, insignificant lags deleted (s.e. corrected for serial correlation) Constant 0.4374 (0.1740) 0.5579 (0.3179) 0.4374 (0.1740) UK Bank Rate - - - Total Assets -2.0862 (0.7122) - -2.0862 (0.7122) QE -4.1298 (1.3574) 0.1950 (0.6256) -4.1298 (1.3574) Bank Reserves -0.0424 0.0189 - -0.0424 (0.0189) M4 - - - M4LA - - - QE lag 1 - 0.7728 (0.8552) - QE lag 2 - 0.1748 (0.8598) - QE lag 3 - -0.6141 (0.6327) -
Table 1:Regression output of the models: coefficients and standard errors(s.e.), obs no. 53
What is striking in model 1 is that the QE factor, total assets and bank reserves are all significant at a 5% significance level. This would mean that in this model these policies are indeed a significant factor in influencing growth. The significance of the variable total assets is in line with expectations because quantitative easing increased total assets held by the central bank. Thus, if QE is significant, it is logical for total assets to be significant too. Moreover, in paragraph 2.4, a bank lending transmission channel was identified. QE would increase the money supply and people would deposit more at banks. Banks will become more liquid and will be more willing to give out loans which increases spending and as a result GDP growth. This pass through of effects could have increased central bank reserves, because of the increased amount of funds available to banks and made the bank reserves variable significant at a 5% level
When considering model 2 and 3 the following results can be seen. In model 2 no variable is at a 5% significance level. As a result, this model does not say anything. Model 3 has exactly the same coefficients as model 1. This is the case because all lag variables in
24
model 3 are insignificant and dropped from the model. Therefore, it can be concluded that in these models QE does not have a lagged impact on GDP growth.
After looking at the significance levels of the models the coefficients are analysed which are all negative. The coefficient for QE in model 1 is -4.1298 which means that for every percentage of long term asset purchases the GDP growth will decrease by 4.13%. For total assets the coefficient is -2.0862 which means for every percentage increase in total assets a decrease of 2.09 % in GDP growth will occur. Last, the Bank Reserves coefficient is
-0.0424 which says that for every extra percentage of bank reserves held at the central bank the economy would shrink by 0.042%.
These outcomes are strange, because for the total asset and QE variable it would mean that an increase in quantitative easing asset purchases would negatively influence GDP growth. According to theory presented in paragraph 2.4, the asset purchases of QE should increase GDP growth. However, these negative coefficients for QE and total asset were also found by Lyonnet & Werner (2012, p 100). The possible explanation for the QE and total assets variables to be negative is that the variables might be distorted. This distortion could be because prior to June 2006 the data for the assets were made up from advances and other accounts. The accounting for these accounts changed with the introduction of the euro Lyonnet & Werner (2012, p100). This is similar to a measurement and error could be the cause of bias in the model.
What is also strange is that contrary to the result above bank reserves needed to be dropped from the Lyonnet & Werner model. The inclusion of the variable in the model in this thesis could result from adding more data points to the regression after quantitative easing was conducted in comparison to the Lyonnet & Werner (2012) analysis. Furthermore, the coefficient is negative. This could indicate that the bank lending transmission channel did not work as was predicted. The negative sign could show that at the time that QE was conducted the money supply did increase but, because of the financial difficulty at the time, banks wanted to hold more excess reserves and deposit them at the central bank instead of giving that money out as loans. This could potentially harm economic growth, hence the negative sign in the model.
What is encouraging about the analysis of the model is that the R² of the model is reasonably high (0.2111). This illustrates the model reasonably explains the variance in the dependent variable GDP growth. Furthermore, the F-statistic in the model is significant enough to reject the H0 joint hypothesis which says that the coefficients in the model are equal to zero. Thus the coefficients are significantly different from zero. This makes it likely
25
that the model is a model which explains the situation well and quantitative easing did indeed cause some influence on GDP growth because the variable is significant in the final model.
In the second analysis a lag variable of GDP growth was added. Three different analyses were run. Model 1 without QE lag variables, model 2 with QE lag variables which are not deleted when insignificant and model 3 with QE lag variables which are deleted when proven insignificant. In table 2 only the final parsimonious models of the analyses will be given. The table gives coefficients and standard errors.
Model 1 with GDP lag 1 period Model 2 with GDP lag 1 period and QE lags
Model 2 with GDP lag 1 period and QE lags only significant variables Constant 0.1607 (0.0867) 0.1459 (0.1002) 0.1137 (0.0861) GDPlag 1 0.6552 (0.0979) 0.6640 (0.0965) 0.6707 (0.0946) UK Bankrate - - - Total assets -1.205 (0.4714) -1.3183 (0.4838) -1.2078 (0.4530) QE -1.5484 (0.7482) -1.5632 (0.7481) -1.4863 (0.7198) Bank Reserves - - - M4 - - - M4LA - - - QE lag 1 - 0.1478 (0.5595) - QE lag 2 - 0.4507 (0.5854) - QE lag 3 - -1.1139 (0.5381) -1.2749 (0.4873)
Table 2: Coefficients and standard errors for the dynamic model. No. obs: 49
In the models in table 2 the GDPlag 1 variable is positive and significant for a level of 5%. This means that part of all three models can be explained by lagged growth. This helps correct the persistence in the model. This way serial correlation is corrected. To be sure of this a Breush- Godfrey test was run and it became clear that there was no serial correlation in all three models. Also a White test for heteroskedasticity was which was only detected in model 1. This was corrected by using robust standard errors.
When looking at the coefficients of model 1 QE and total assets are negative like the models above. A possible reason for these negative coefficients was therefore given above. To
26
compare the coefficients of the different models a formula needs to be extracted. This is extracted from (excluding all variables that are not related to QE) GDPgrowth = βGDPlag1 + αQE+ αQElag1 + αQElag2+ αQElag3. This gives: α+ α(n-1)β+ α(n-2)β²+ α(n-2)β3 α(n-4)β4 which can be rewritten as: α ∑ . This knowledge helps to compare a static model with a dynamic model coefficient.
When comparing model 1 with GDP lag with the model without GDP lag the QE coefficients coincide to: (α/(1-β))= (-1.5484/(1-0.6552)= -4.4907, because only 1 QE variable and thus 1 alpha is in the model. The result is encouraging because the coefficients of the dynamic and static model 1 are similar. This could mean these models explain the current situation well. The coefficient predicts that a 1% increase in asset purchases will decrease GDP growth by 4.49%.
In model 2 the three QE lag variables are added. The only significant lag variable is QE lag variable 3. Hence QE has a lagged effect on GDP growth of 3 quarters which is between six to nine months. Therefore the prediction of Kapetanious et al. (2012) is correct. Thereafter, this lagged effect is quite large. The QE coefficient now coincides with: (α+α
(n-1))+α(n-2)+α(n-3))/(1-β), because 4 QE variables (4 alpha’s) are included in the model. This
gives (-1.5632 +0.1478 +0.4507 - 1.1139)/(1-0.6640) = -6.1863. This is 2.0565 lower than the coefficient in the model 1 without lags.
Last, in model 3 only the significant lag variables are regressed in the model. This model yet again concludes that QE was a significant factor in causing growth. Also it can be concluded again that QE has a lag of 3 periods. The long run effect of QE can now be better estimated by not including the insignificant variables. The coefficient is now (α+α(n-3)) (1-β0) =(1.275-1.486)/(1-0.6707)= -8.3858. This coefficient is now more than double the size of the coefficient in the model without lags. Accordingly, the long run effect of QE is large.
The three models with the GDP lag variable also have good explanatory power, because the R² statistic is high. The R² of model 1 with GDP lag is 0.5832 that of model 2 is 0.6445 and model 3 is 0.6394. Moreover, the F-statistic of model 1 is 57.48 that of model 2 12.69 and of model 19.51. All these F-statistics reject the joint H0 hypothesis and hence it can be said that the estimated coefficients are significantly different from zero. This makes a case for the models being good estimators of the current situation and thus having meaningful results.
All in all, from the results above, it becomes clear that the variable of interest the qualitative easing variable, which measures the quantitative easing effect, is significant for a 5% significance level in all models. The coefficient for the factor, however, indicates that the
27
policy caused a decrease in economic growth. This is the opposite of what the policy is predicted to do. Thereafter, the effect of QE on GDP growth is lagged by six to nine months. This lag causes QE to have a high long run effect and this should not be underestimated.
7 Robustness
This section will analyse the models used in section 6. The models used were tested for heteroskedasticity and serial autocorrelation.
As was mentioned in section 6, the static model without lags would be analysed. The only problem that arose when testing this model was one of serial correlation. This was tested by using a Breusch-Godfrey test and can be found in figure 3.
Test Breusch-Godfrey White test Ramsey Reset H0 H0: No serial correlation H0: Homoskedastic errors H0: No omitted variables Test statistic χ²= 5.215 χ²= 25.05 F(3,44)=0.22 P-value 0.0224 0.1588 0.8820
Figure 3: Results of several models tests.
This test rejected the H0 hypothesis. Thus there is serial autocorrelation and there had to be a correction for this. This correction was conducted using a Prais-Winsten correction, an extension of Cochran-Orcutt .Cochrane-Orcutt is chosen because this estimation method adjusts the serial correlated error terms in linear regression models (Cochrane & Orcutt, 1949). The Prais-Winsten option is chosen because this method does not delete the first observation and therefore the estimator is more efficient than other serial correlation correction estimators (Prais & Winsten, 1954). All models in table 1 of section 6 were regressed with this correction, because all had serial correlation.
Moreover, the static model showed no signs of heteroskedasticty. In figure 3 the White test for homoscedasticity is shown. The H0 hypothesis is not rejected, because the p-value of the White-test is 0.1588. The H0 hypotheses states that the error terms are homoscedastic. Thus, the lack of rejection of H0 means the errors are not heteroskedastic and no correction had to be added to the model.
Last, a Ramsey Reset test is performed to examine if there are any omitted variables. In figure 3 the results suggest that H0 need not be rejected which means there are no omitted variables in the model which cause bias.
28
The second part of the analysis contained 3 regressions of dynamic models. The dynamic nature of the models should correct for serial correlation this was tested and the results can be found in figure 4.
Test Breusch-Godfrey model 1 Breusch-Godfrey model 1 Breusch-Godfrey model 1 H0 H0: No serial correlation H0: No serial correlation H0: No serial correlation Test statistic χ²= 2.090 χ²= 1.100 χ²= 0.930 P-value 0.1483 0.2943 0.3347
Figure 4: Results of several models tests. (dynamic models)
H0 is not rejected in all three models and as a result there is no serial correlation. The three models were also tested for heteroskedastic standard errors. In figure 5 the results of the White test are given. The only model where the H0 is rejected is model 1. This model was therefore corrected for these errors by using robust standard errors.
Test White test White test White test
H0 H0: Homoskedastic errors H0: Homoskedastic errors H0: Homoskedastic errors Test statistic χ²= 26.16 χ²= 27.59 χ²= 0.930 P-value 0.0019 0.0917 0.3347
Figure 5: Results of several models tests. (dynamic models)
To conclude, all issues that arose when analysing the models were dealt with. The dynamic models corrected for serial correlation as was expected.
8 Conclusion
The purpose of this thesis was to examine if the unconventional monetary policy tool,
quantitative easing, was a significant factor in causing increased growth in the UK from 2009 up until now.
Researchers are divided by how to test quantitative easing. Part of the researchers chose to examine the yield reduction of gilts caused by the quantitative easing asset
purchases. The assumption made with this analysis, is that this yield reduction theoretically would increase GDP growth. This way researchers concluded that the interest rate reduction had a negative correlation with GDP growth. Many researchers believed that this reduction of yields was indeed reached and that quantitative easing worked.
29
The other part of the researchers tried to test if GDP growth was increased more with the use of quantitative easing than without. Older papers concluded that QE did not cause more GDP growth but newer papers concluded the contrary.
This thesis constructed a model that was similar to Lyonnet & Werner (2012) and tried to test the relationship between QE and GDP growth. The model focuses on all possible monetary policy tools and intermediate targets. The variable of interest is the qualitative easing variable which represents the quantitative easing policy. From the regression analysis of the model it becomes clear that the coefficient for the QE variable is -4.1298 and is
significant. This result means that QE policy did indeed influence GDP growth and that a 1% increase in asset purchases under the QE policy will decrease GDP growth with 4.1298%. This result is the opposite of the increase in GDP growth which was expected and the
opposite of why QE was implemented in the first place. From the analysis with lag variables it can be concluded that QE has a lagged effect on GDP growth of six to nine months. In
addition, this lagged long run effect of QE more than doubles the QE coefficient to -8.3858, which means the long run effect is large. This means that when testing for QE the long run effect should not be underestimated. There are, however, some comments to be made on these conclusions.
The OLS estimation method used may be too simplistic and does not estimate the coefficients well. This could be caused by the independent variables being influenced by other effects than the QE asset purchases. As was seen in the literature, more advanced econometric methods can better separate QE effects from other effects. The more advanced models can capture structural changes in the economy and test interrelationships of QE and the yields on gilts well. These methods could therefore capture the true effect of QE better than a simple OLS model.
The potentially wrong estimation of the coefficients of the OLS model can also be explained by not including all variables needed to explain GDP growth. Even though, the Ramsey Reset test estimated that there were no omitted variables, the model could be improved by allowing for more explanatory variables.
Something that can also influence the coefficient estimation is the time period chosen. The longer the time period estimated the better the estimation. Lastly, the data of the Bank of England was changed due to calculation and accounting changes. This could also have an impact on the odd outcomes of the estimated coefficients.
Further research on the significance of the factor QE in the relationship between QE and GDP growth needs to thus consider adding variables to the model used in this thesis.
30
Moreover, the data period needs to be extended. What also can be considered is contacting the Bank of England and discussing how to solve the change in accounting problems and create an even better proxy for the variables affected than was done in this thesis. On the other hand when it is desired to try and estimate the exact amount of GDP growth caused by QE, it would be wise to choose a more advance econometric method.
31
Bibliography
Bank of England (1981). The monetary base — a statistical note. Bank of England Quarterly Bulletin, March , pp. 59–65.
Bank of England (2006). The implications of money market reform for data published in Monetary and Financial Statistics Bank of England Quarterly Bulletin, June, pp. 59– 65.
Bauwens, L. & Sucarrat, G. (2010). General-to-specific modelling of exchange rate colatility: A forecast evaluation. International Journal of Forecasting, 26, pp. 885-907.
Bernanke, B.S., Reinhart, V.R. & Sack, B.O. (2004). Monetary alternatives at the zero bound: an empirical assessment, Brookings papers on economic activity, vol 35, pp.1-100. Borio, C., & Disyatat, P. (2010). Unconventional monetary policies: an appraisal, The
Manchester school, pp. 53-89.
Breedon, F., Chadha, J. S., & Waters, A. (2012), The financial market impact of UK quantitative easing, Oxford Review of Economic Policy, 28(4), pp. 702–728. Bridges, J. & Thomas, R. (2012). The impact of QE on the UK economy-some supportive
monetarist arithmetic. Bank of England Working Paper. Vol 442, pp. 1-51. Budd, A., Buiter, W., Clementi, D., George, E., Goodhart, C., Julius, D., King, M. &
Plenderleit, I. (2005). The transmission mechanism of monetary policy, The bank of England, pp. 1-12.
Campos, J., Ericsson, N. R. & Hendry D.F. (2005). General-to-specific Modeling:
An Overview and Selected Bibliography. International Finance Discussion Papers, vol 838, pp. 2-92.
Campos, J., Ericsson, N. R., & Hendy, D. F. (2005). General-to-specific modelling. Cheltenham: Edward Elgar.
Chadha, J.S., Waters, A. (2014). Applying a macro-finance yield curve to UK quantitative easing. Journal of Banking & Finance, vol. 39, pp.68-86.
Christenen, J. H. E., & Rudebusch, G. D. (2012), The response of interest rates to US and UK quantitative easing, The economic journal, 122, pp. 385-414.
Cochrane, D. & Orcutt, G. H. (1949). Application of least squares regression to relationships containing autocorrelated error terms. American Statistical Association Journal.
Curdia, V. & Woodford, M. (2011). The central bank balance sheet as an instrument of monetary policy. Journal of Monetary Economics, vol. 58, pp. 54-79.
