Quantifying Quantitative Easing
2.2 Our event study
The event study presented here draws on work by the author in both Caglar et al (2011) and Meaning and Zhu (2011). We …rst identify a series of events relevant to each quantitative easing programme. These events are selected on the basis of when new information became publicly available to markets so comprise of signi…cant policy announcements and speeches by the Federal Reserve and Bank of England relating to their purchases of government securities.5 We then
measure the change in variables over a two day window around each event, from the close of business the day before the event to close of business the day after it. This is consistent with Joyce et al (2010) and other studies and is preferred by this author as it allows for some slight lag in impact as markets internalise the information, or in case there exist frictions in the process of settling transactions. Evidence of such frictions is provided in the later ‡ow analysis.6
We then analyse movements in a range of variables over these identi…ed win- dows and compare the changes in these periods to those over the average two day window in the full sample period concerning quantitative easing.
Figure 1 shows the cumulative impact of each programme on the yields of government securities of di¤erent maturities and ten-year corporate bonds of di¤erent grades. This is de…ned as the sum of changes on days associated with QE events. It is clear that each asset purchase programme had a signi…cant negative impact on yields at all maturities. For the UK this got stronger the further along the yield curve one looks, lowering yields at the long end by over 100bps. Both US programmes on the other hand exhibit a peak around 5 year rates of over 30bps. This di¤erent pattern can be explained both by the shorter average maturity of the US government debt portfolio and also the di¤erence in the maturity of the purchases carried out by the Federal Reserve and the Bank of England. The second US programme had a stronger impact on ten-year rates
5A full list of events and their rationale for inclusion can be found in Table 1.
than the …rst. This is to be expected as the Federal Reserve began buying more longer-dated securities under US QE2 so as not to withdraw too much of any one area of the market and to take advantage of duration e¤ects.
Table 2 presents a breakdown of the cumulative impact by event. In the UK the …rst two announcements are clearly dominant, contributing much of the …nal value. This is not surprising as they can be assumed to have had much more informational content than later announcements as they outlined not only the quantity of assets to be bought, but also their type and the relevant mecha- nisms by which this unprecedented policy for the UK economy was to be applied. They also re‡ect the larger size of purchases announced in March 2009, £75bn compared to later extensions of £50bn and eventually £25bn. Looking at the announcement on 6th August, one can also clearly see the market reaction to the Bank of England’s shift towards purchasing longer-dated securities re‡ected in a larger fall in 20 year rates.
Corporate yields fell over the event windows by a comparable amount to gov- ernment securities. This implies spillover e¤ects as news of central bank purchases of government bonds encouraged demand for their corporate counterparts. This may suggest a degree of substitutability between the two asset classes, supporting the portfolio rebalancing theory of transmission, or just that the asset purchase programmes increased con…dence more widely and that moved investors into asset classes which had been seen as more risky, such as corporate bonds.
Aside from the limited size of the event window, designed to keep out non- policy related factors, we also look to con…rm that the fall in yields is not just due to a downward trend in the data. A limit of the event study methodology is that such a trend would cause a fall across event windows and non-event days similarly. To combat this we construct a primitive counterfactual. Tables 2-4 also show the average two-day change for yields over the whole sample period.7 In all
cases they are less than one basis point in absolute terms, and for government securities in excess of one year, they were actually increasing making the falls observed over the event windows even starker.
the full sample, suggesting that, to the small extent QE did in‡uence UK equity markets, it only increased volatility. US asset purchases had a much stronger e¤ect on equities, at least for the …rst programme where associated two day event windows experienced an average change of over 1% compared to just 0.2% for the full sample. As with corporate bonds, it seems that US QE served to bolster investment sentiment in other assets. It also reduced implied volatility by an average of 2.6%.
Initial asset purchases resulted in signi…cant currency depreciations for both the US and UK. QE in the UK reduced the sterling nominal e¤ective exchange rate by an average of 0.68% per event. In the US this was slightly higher, at 0.76%. However, taking into account the counterfactual of the full sample where the UK experienced a very slight appreciation and the US a small depreciation, the exchange rate implications of the two programmes are almost identical. This result may be thought to correspond to a monetary view of the exchange rate where increases in the money supply devalue one currency against another, or it may be that markets envisaged asset purchases being successful at lowering interest rates, both long and real, and adjusted the exchange rate in a manner consistent with interest rate parity.
Whilst this paper is not directly concerned with the transmission mechanisms behind asset purchases, it is interesting to note that one year OIS rates fell sig- ni…cantly, especially in the US. As it is based on a geometric average of the overnight rates over the period of the contract, the OIS rate is often considered a good proxy for market expectations of the policy rate. This fall then lends support to the transmission mechanism put forward by Eggertson and Woodford (2003) and others that unconventional policies work by manipulating agents’ ex- pectations of the future path of policy rates. Combined with our other results this suggests that both portfolio rebalancing and the expectations channels have a role to play.
These results must be viewed with some caveats in mind which derive from the limitations of event study methodology itself. Firstly, events may be ’muddy’ to varying degrees as it is impossible in some instances to completely isolate when information reaches markets concerning purchases from the release of other information. A clear example is the 5th March 2009 announcement by the Bank of England where they simultaneously announced the asset purchase programme and a cut in Bank Rate. The March event window will inevitably capture the market absorbing both these pieces of information, although the cut in Bank Rate was widely anticipated so to a large extent would have already been built in to
agents’ behaviour.
This highlights another limitation of event studies. If the announcement is anticipated then it will be priced in by markets prior to an event window which straddles the announcement. This was not as much of a problem at the incep- tion of QE as its novelty meant that people had very limited knowledge of how it would be implemented and, crucially, of the central bank’s reaction function concerning asset purchases. However, as time has progressed the reaction func- tion has arguably become clearer, to the extent that it should be almost explicit in the case of forward guidance, and thus the power of event studies is weaker. This would lead to an underestimation of the impact of the policy and is a po- tential explanation for at least part of the diminished impact observed in later announcements in Tables 2-4.
More generally, an e¤ective event study relies on correctly identifying all the moments where data was internalised by the market. This is hard to do as agents may change their behaviour in anticipation of a policy, when it is announced, or perhaps not until it is completed. It could also be anywhere in between. Failure to capture a relevant event may bias the estimated impact of the policy.
3
The D’Amico and King Framework
A more complete methodology is that of D’Amico and King (2010). They propose two estimations. The …rst is a panel regression on daily data at the individual security level designed to capture the ‡ow e¤ects associated with the physical act of making purchases. The second is a cross-sectional regression, again at the individual security level, but this time looking at the change in variables from the day prior to the inception of the policy, to the day of the last physical purchase has been made. This cross-sectional approach aims to elucidate on the more permanent "stock" e¤ect that results from reducing the supply of a security in the publicly available market, relative to the existing demand for it.
maturity of 90 days or less to avoid contamination by the di¤ering behaviour of investors as a security approaches its redemption date. The sample consists of 239 securities for US QE2 and 40 for the Bank of England’s …rst QE programme.
For each security, tranches of substitute securities within the sample are con- structed. Any security with a remaining maturity within two years of that of security i is deemed its near substitute. A remaining maturity of between two and six years from that of security i is considered a mid-substitute whilst a dif- ference of more than 6 years makes it a far-substitute.
The raw purchase data, which is sourced directly from the Federal Reserve and the Bank of England, is normalised by dividing each purchase amount by the total outstanding supply of the security and its near substitutes. This has a number of bene…ts. It means the data now shows the fraction of the supply of a given security type that the central bank buys and removes from the market in each operation. The depth of a particular part of the market is thus controlled for. After all, £10mn of purchases in a part of the market where total supply is only £20mn could reasonably be expected to have a di¤erent impact than the same £10mn of purchases in an area of the market where the total supply is £20bn. The normalisation also controls for changes in the total supply caused by new issuance by the …scal authority.8 Lastly, it makes the coe¢cients easily
comparable.
A …nal comment to make is that, for the sake of simplicity, all of the analy- sis is conducted in price space9. The implied impact is then converted into an
equivalent change in yields using the modi…ed duration of the security.
Ultimately, we are left with a set of time-series on the price, supply and supply of substitute securities for a large number of securities which we use to build both our panel, and also our cross-sectional database. For the panel we bring together the individual time series and maintain the daily frequency of the underlying data. For the cross-section we simply take the change in each time series from the …rst day of the policy to the last, reducing each time series to a single data point.
8This is likely to a¤ect the price and highlights the monetary-…scal interaction inherent in
these policies.
9Though D’Amico and King (2010) …nd that their results are little changed if the same