Measuring Euro Area Monetary Policy

We study the information flow from the ECB on policy dates since its inception, using tick data. We show that three factors capture about all of the variation in the yield curve but that these are different factors with different variance shares in the window that contains the policy decision announcement and the window that contains the press conference. We also show that the QE-related policy factor has been dominant in the recent period and that Forward Guidance and QE effects have been very persistent on the longer-end of the yield curve. We further show that broad and banking stock indices' responses to monetary policy surprises depended on the perceived nature of the surprises. We find no evidence of asymmetric responses of financial markets to positive and negative surprises, in contrast to the literature on asymmetric real effects of monetary policy. Lastly, we show how to implement our methodology for any policy-related news release, such as policymaker speeches. To carry out the analysis, we construct the Euro Area Monetary Policy Event-Study Database (EA-MPD). This database, which contains intraday asset price changes around the policy decision announcement as well as around the press conference, is a contribution on its own right and we expect it to be the standard in monetary policy research for the euro area.


Introduction
Monetary policy, for better or worse, has been on the forefront of cyclical policymaking in the past two decades, especially during the Great Recession and the sovereign debt crisis. While there is an extensive literature on the US monetary policy, a comprehensive understanding of the ECB policy and its nancial market eects are yet to materialize. This is partly due to the lack of a systematic database of high frequency, intraday data for a broad class of asset prices in the euro area of the kind that has been employed in the US for more than a decade. This paper helps remedy both deciencies by constructing a long (by euro area standards) time series of policy-date event studies for the euro area, which will be kept up to date, and by using this data to measure and assess the ECB monetary policy. While our substantive analysis focuses on the nancial market eects of monetary policy in the euro area, the dataset we construct to carry out the analysis is of independent interest. Our euro area monetary policy event-study database (EA-MPD) features price changes for a broad class of assets and various maturities, including Overnight Index Swaps (OIS), sovereign yields, stock prices, and exchange rates. We have exercised considerable care in studying each event window to clean outliers and misquotes, of which there were plenty early in the sample, to make sure that what we report accurately reects the market reaction. We provide a description of the data, and timing of events on ECB policy dates in section 2 and oer details and event-by-event analysis in an accompanying online appendix.
Monetary policy surprises in the euro area are not only multi-dimensionalwe show the existence of perceived policy Target, Timing, Forward Guidance, and Quantitative Easing (QE) surprisesthey are also revealed in a multi-step structure. A press release provides information on the policy decision with no rationale and discussion, later followed by a press conference.
The information markets extract from these two events are distinct, and intraday data allows us to measure these dierent types of information separately. Our event-study database utilizes the intraday data to form event windows bracketing about 10 minutes before the release of monetary policy information, to about 10 minutes after. Employing tick data allows us to separately compute asset price changes around the release of the monetary policy decision at 13.45, and the reading of the statement and the following Q&A in the press conference beginning at 14.30. The separate releases of the policy decision and the narrative information dierentiates euro area monetary policy information revelation from that of the US, where the two happen simultaneously.
After briey presenting the event-study database, we use it to study monetary policy surprises in the euro area and their eects on various nancial market segments. To do so, we employ the methods developed by Gürkaynak et al. (2005) and Swanson (2017) for the US and use them in the two intraday windows of the euro area policy communication. The dierence in the nature of information released in these two windows provides a new understanding of market participants' interpretation of monetary policy communication.
Our results show that, naturally, the preponderance of surprises in the press release window are about the current setting of the policy rates, Target surprises, with no other statistically signicant policy surprise factors. In the press conference window, as expected, there are no Target surprises and the Path and QE surprises dominate. Interestingly, while our analysis suggests that a Target factor does not even appear as a statistically signicant factor in the press conference window, a dierent factor, Timing emerges. This type of surprise captures the revision of policy expectations by shifting the expected policy action between the current meeting and the next or the one following, in a way that leaves longer-term policy expectations about unchanged. In essence, market participants extract two distinct types of guidance from the press conference. One that is informative about the medium run, peaking at about two yearsForward Guidanceand one for the near future, peaking at about six months maturity what we call Timing.
The methodology we employ to extract QE surprises yields continuous measures of the market surprise. Hence, in studying the asset price responses to QE, we are able to condition on the size of the surprises rather than only on a binary variable that shows when a QE announcement took place. This allows for a substantially more precise understanding of QE eects and helps distinguish QE from Forward Guidance surprises, which were also frequent during the Zero Lower Bound (ZLB) period.
1 We nd that both QE (after 2014) and Forward Guidance surprises are active in the press conference window and that while Forward Guidance aected the middle of the yield curve most heavily, with a peak eect at about two years, QE eects get larger as maturity increases, peaking at the 10-year maturity. Surprises about the current setting of monetary policy, which were present in the pre-ZLB period, never had noticeable eects on the long-end of the yield curve. Having the quantied QE surprises also allows us to study the persistence of their eects better. Unlike in the US, where QE eects were short lived, with a half-life of three months as estimated by Wright (2012), we estimate a half-life of about one year in the euro area. Thus, ECB QE not only had substantial immediate eects on yields, it also had long lasting eects.
We also study the eects of ECB policy surprises on dierent sovereign yields, exchange rates, and stock prices. Some of these were studied previously in the literature using the combined press release and press conference windows (such as Andrade and Ferroni, 2016) and in the separate windows but not including the QE surprises in the analysis (such as Brand et al., 2010 andLeombroni et al., 2017). This paper is the rst to look at intraday data to separately study the press release and press conference windows and extract both conventional and unconventional monetary policy communication surprises from both. It is also the rst not to assume only Target surprises take place in the press release and broadly dened communication surprises in the press conference windows but to ask statistically how many factors are present in each and to estimate factors that can be attributed to specic types of communication. Importantly, this paper is also the rst in presenting a market-based identication of QE surprises in the euro area that shows QE has narrowed spreads, rather than identifying QE surprises by assuming that these have narrowed spreads.
In the limited cases where both the event window coverage and the monetary policy surprise denitions overlap with the existing literature, our ndings are in line with what is already knownsuch as eects of Target surprises that are signicant for the short end of the yield curveand instills condence for the new results we report on the dierence between Timing and Forward Guidance, on the eects of QE, on persistence of these eects, on information and stock market reactions, on nonlinearity, and nally on using our methodology for the analysis of policy news that do not come out on Governing Council policy dates.
On stock prices, we nd that the reaction of broad and banking indices can only be understood when the genuine policy surprises (perceived deviations of interest rates from the policy rule) are separated from information eects (perceived information signaled by the central bank on the current and future state). When studied using this lens, monetary policy has had signicant eects on stock prices but there was signicant time variance in the perceived variances of genuine policy and information surprises. We further nd that the policy surprise eects on stocks were persistent as were the eects of these surprises on longer-term interest rates, especially for Forward Guidance and QE surprises.
Our ndings on nonlinearity are noteworthy. We study whether the market responses to positive and negative surprises are dierent. A nascent literature is suggesting that in the US, monetary policy has asymmetric real eects (Tenreyro and Thwaites, 2016; Barnichon and Matthes, 2017). We nd that in the euro area nancial market participants do not perceive monetary policy eects to be asymmetric with respect to positive surprises and negative in providing asset price responses.
Lastly, we use our estimated factors of monetary policy surprises and show how to use them to decompose any policy-related news eects into these four factors. In two illustrative examples we show that policymaker speeches and newswire reports can have substantial effects on yields and that in the recent past there have been cases where the nancial market participants have extracted information on QE-related policy from such news.
The paper is organized as follows. In section 2 below, we discuss the ECB policy communication and our event-study database. Then, in section 3 we construct our surprises in terms of Target, Timing, Forward Guidance, and QE factors of monetary policy in the two policy communication windows, and discuss their features and plausibility. Section 4 presents asset price responses to the monetary policy surprises using a linear model. In section 5 we turn to stock price reactions and information eects. Section 6 studies the dynamics of the identied surprises in a daily VAR framework. In Section 7 we relax the linearity assumption and allow for asymmetries. Lastly, in section 8, we show how to adapt our methodology to speeches and other monetary policy events and in section 9 we oer concluding thoughts.

Euro Area Monetary Policy Event-Study Database
One of the contributions of this paper is to develop a Euro Area Monetary Policy Event-Study Database (EA-MPD). This section is an introductory manual for the dataset, which will be periodically updated and made available. The underlying tick data come from the Thomson Reuters Tick History database, and we report our transformations of those data rather than the underlying data. Here, we rst describe the monetary policy communication process in the euro area and then concentrate on the features of the dataset that we develop to measure the monetary policy surprises, delegating more technical details on the dataset construction to online appendices.

2.1
A Primer on euro area monetary policy communication At its inception in 1999, the ECB Governing Council took policy decisions twice a month, whereas a press conference took place only once a monthon the rst meeting of the month.
After November 2001 only one meeting per month was a policy meeting, taking place on the rst Thursday of the month, regularly accompanied by the press conferences, with some exceptions (Ehrmann and Fratzscher 2009 discuss this policy communication structure). As of January 2015, the frequency of monetary policy meetings has moved to a six-week cycle. will be announced during the press conference. As of March 2016, the content of the decisions on non-standard policy measures has been included in the press release.
In constructing the database, we rst cleanse the data of misquotes, 3 then discretize the data within each window by taking the last quote of each minute within the window, and then use the median price in the 13:25-13:35 interval as the pre-press-release quote, and the median price in the 14:00-14:15 as the post-press-release quote. Similarly, we take the median price in the 14:15-14:25 interval as the pre-conference quote and the median price in the 15:40-15:50 interval as the post-conference quote.
4 We make use of an interval rather than selecting a particular minute to measure the pre-and post-event quotes in order to minimize the risk of selecting a quote that is not representative. The changes reported in the database are changes from the pre-event quote to the post-event quote for each communication window. We dene 2 The Appendix is available online and is structured as follows: Appendix A provides details on the Governing Council meeting frequencies, the information release structure on policy dates and other relevant information since the inception of the ECB. It also contains a 3 Such misquotes, which did not reect actual market pricing, were prevalent especially early in the sample.
4 Note that while the post-press release and pre-conference quote intervals partially coincide (the minute the monetary policy event as the union of the press release and conference, and measure changes in asset prices due to this event as the change from the pre-press release quote to the post-conference quote. Figure 1 shows this timeline in stylized form. The EA-MPD reports the asset price/yield changes we construct for the three event windows in separate worksheets. In each worksheet a policy date is in the rst column on each row, and the following columns show changes in selected asset prices/yields. The assets covered are: OIS rates with 1, 3, 6 month, 1 to 10, 15, and 20 year maturities; German bund yields with 3 and 6 month, 1 to 10, 15, 20, and 30 year maturities, French, Italian, and Spanish sovereign yields with 2, 5, and 10 year maturities, the STOXX50E and the stock price index comprising banks (SX7E), and the exchange rate of the euro. The EA-MPD is made available as a supplement to this paper and will be regularly updated.
In our substantive analysis we start the sample in 2002 because from 1999 (the beginning of the single currency in the Euro Area) to the end of 2001 our intraday OIS data are very noisy, with large spikes and sparse quotes. To provide an illustration of the (cleansed) intraday data, Figure 2 shows the changes in the 2-year Overnight Index Swap (OIS) rate around the publication of the press release and around the press conference on four dierent policy meeting dates: 4 July 2013, 4 September 2014, 3 December 2015, and 7 September 2017. We select the 2-year rate as it is of suciently long maturity to display movements in response to announcements of non-standard as well as standard measures.
These four panels are illustrative of the dierent cases in which the monetary policy surprises may arise within the policy meeting day. Panel (a) displays no reaction of the 2-year OIS rate in the press release window and a reaction in the conference window. This episode corresponds to the ECB announcement in the press conference, rst time ever, of formal Forward Guidance on the future path of its policy rates, by stating that policy rates are expected to remain at present or lower levels for an extended period of time. Panel (b) shows a reaction in the press release window, with no further news aecting the OIS rate in the conference window.
This episode corresponds to the announcement of a cut in the ECB deposit rate announced in the press release. Panel (c) depicts a policy date in which there are sizable movements in both windows. This episode captures the nancial markets' disappointment following the ECB decision to increase the size of its QE program: markets evidently were expecting a larger increase of QE, as well as a cut in the policy rate, as suggested also by survey expectations among nancial analysts gathered ahead of the policy meeting. Lastly, panel (d) shows a day in which there is no surprise, either in the press release or the conference windows. Policy dates like these are surprisingly rare; there is usually some news for the nancial markets, especially in the press conference window. While this discussion of the policy surprises on the basis of intraday changes in the 2-year OIS rate is suggestive of a Target/Path/QE decomposition, formally carrying out this analysis requires the simultaneous study of interest rates at dierent maturities, which are covered by our event-study database and analyzed in section 3 below.
Before turning to the measurement of surprises it is useful to briey note some of the internal consistency and robustness checks we have carried out on the data. The Appendix contains details and further checks. The raw changes and surprises we measure, and the results of the analysis we do using these are mostly insensitive to changes in the measurement windows. Our estimates would have been more or less the same if we had taken the last quote instead of the median within the window or used wider or narrower windows. We have also veried that the changes and surprises are independent across the two windows, showing that we are indeed measuring reactions to unanticipated news in the two windows and not momentum eects that carry over from one window to the other.
Another important issue has to do with the US data release calendar, which makes the initial unemployment claims release often overlap with the event windows in the Euro Area.
In the window that contains both an ECB policy communication and the US initial claims release, 5 although the two will be uncorrelated, if the short-term euro OIS rates respond to the initial claims surprise the monetary policy surprise measures will be subject to measurement error. We studied the impact of US initial claims surprises on short-term euro OIS rates and found that even in the very rare cases where there are statistically signicant coecients, the R 2 coecients are in single digits, implying that the simultaneous release of US initial claims does not introduce noticeable measurement error into the Euro Area monetary policy surprises that we measure. For completeness we include these as control variables in the statistical work we present in section 4. It is reassuring that whether we include this control or not has no bearing on the results we report.

Measuring Policy Surprises in the Euro Area
Understanding the eects of monetary policy requires identifying orthogonal changes in the policy stance. These changes may be orthogonal to the state of the economy, as in VAR analysisin which case they are usually called policy shocksor they may be orthogonal to the 5 In the window that does not overlap with the initial claims release there is no correlation between the US surprise and the changes and surprises we measure for the Euro Area, again verifying that we are measuring proper surprises arising from monetary policy communication.
information set of nancial market participantsin which case they are called (market-based) policy surprises.
Using monetary policy surprises measured in daily or higher frequency one can study asset price responses to monetary policy in a meaningful way. The very reasonable identifying assumption here is that monetary policy does not respond to asset price changes within the day, hence causality goes from monetary policy to asset prices and nancial markets' reaction to monetary policy can be studied. Work on related questions for the euro area have been done All of these papers had to construct their own event-study database to carry out similar analyses. The EA-MPD we provide, by virtue of being regularly updated, will eliminate this sizable xed cost of doing research on Euro Area monetary policy. In terms of the exercises we carry out, our value added will be in estimating rather than assuming the number of dierent types of policy surprises market participants perceive and in naming these. The results of this exercise are interesting and dierent from what has been assumed so far. Our further value added will be in covering the crisis period and measuring the eects of ECB non-standard monetary policy measures over a range of nancial assets and comparing their transmission with that of standard policy measures, as well as measuring the persistence of responses and possible asymmetry of these responses to positive and negative surprises. In particular, we will be estimating the persistence of QE eects in the euro area using a precise, continuous measure of QE surprises for the rst time.

Identifying the surprises
We rst measure monetary policy as a potentially two dimensional process with possible Target/Timing and Path (Forward Guidance) components, and then allow for a third dimension after the onset of the nancial crisis so as to capture the information about non-standard measures and especially QE. As noted above, the ECB policy communication extends over two separate windows where the policy action is rst announced in a press release with no motivation, and then a statement is read by the President, followed by a question-and-answer session. Market participants may update their beliefs about the current stance and future path of monetary policy in response to the press release as well as the press conference; hence 8 Electronic copy available at: https://ssrn.com/abstract=3402078 on each policy date, using intraday data, we measure two sets of surprises.
We construct the Target and Forward Guidance surprises following the methodology employed in Gürkaynak et al. (2005) and Gürkaynak (2005), who in turn build on the work of Kuttner (2001), using Federal Funds futures quotes to measure policy surprises in the US.
The methodology of constructing the QE factor follows that of Swanson (2017). In particular, we extract factors from changes in yields of risk-free rates at dierent maturities, spanning one-month to ten-years, in each of the two windows (press release and conference). Ideally, the risk-free rate curve in the euro area would be proxied by the term structure of the OIS.
Unfortunately, however, at maturities longer than 2 years high-frequency data on the OIS rates is only available after August 2011. Therefore, prior to that date we use yields on the German sovereign yields as proxy for the risk-free rates. Using the German yields throughout the period makes no signicant dierence.
To extract monetary policy surprises that admit economic interpretation from these asset price changes we estimate latent factors from changes in yields and rotate these factors. The matrix X j , j = {press release, press conference }, has changes in 1, 3, and 6-month and 1, 2, 5, and 10-year yields in its seven columns, with each row corresponding to a policy date. This matrix is taken directly from the EA-MPD. The factor structure is where F are the common latent factors, Λ are the factor loadings, and are the idiosyncratic variation of yields at dierent maturities. We analyze the press release and press conference windows separately and estimate the factorscommon drivers of yield changesby principal components.
We test the number of statistically signicant factors in the two windows over the full sample and the pre-QE samples. As shown in Table 2, we consistently nd a single signicant factor in the press release window in both periods, but nd two factors in the press conference window before QE and three in the full sample, suggesting the presence of a new factor in this window in the QE period.

6
The latent factors F j do not have clean interpretations as monetary policy surprises, so 6 The full sample press conference window having three factors might also have been due to the two subsamples having two factors each, but one of these being dierent across the pre-QE and QE periods. But the rst two factors of the pre-QE and full samples are exactly the same, ruling out this interpretation. Further, although utilizing the test for only the QE sample is undesirable due to the very low degrees of freedom one would have, doing that nonetheless nds three factors in the press conference window in the QE period.
we rotate the factors to make them interpretable. This method has been used frequently since Gürkaynak et al. (2005) and was rst applied to Euro Area data by Brand et al. (2010) using intraday data. For our purposes, including the analysis of QE, the orthogonal factors are identied by imposing the following restrictions on the rotation matrix: (1) the second and third (when the third factor is present) factors do not load on the one-month OIS; (2) the rotation is such that the third factor has the smallest variance in the pre-crisis period (2 January 2002 -7 Aug 2008). Essentially, we are forcing two factors not to be correlated with the one-month OIS (the standard measure of the immediate policy setting surprise) and allowing for one of them to aect the yield curve such that this factor was not important in the pre-crisis period.
7 This factor will turn out to contribute only to the movements in the longend of the yield curve, and only be active post-2014, naturally leading to the QE factor label.
Extraction of this last factor is an application of the Swanson (2017)  The sequential nature of these orthogonalizations is important to keep in mind to understand the rotated factors. One factor is dened by orthogonality to the 1-month OIS change.
Another is dened by orthogonality to this, such that the two explain most of the variance of the yields. The third factor is dened such that it is orthogonal to the rst two and the 1-month OIS, and explains the minimal share of the pre-crisis variance.

8
With these rotations, the rst and only statistically signicant factor in the press release window turns out to be the Target factor, which loads only on the short rates. As expected, we nd a Forward Guidance (Path) factor that is always present in the press conference window. The existence of guidance by central banks predates the global nancial crisis as well as the explicit designation of Forward Guidance and even the adoption of statements accompanying policy decisions. The Forward Guidance factor captures the revision in market expectations about the future path of policy rates that are orthogonal to the current policy surprise.
7 Note that this would be a somewhat imperfect measure prior to November 2001, as over that period the Governing Council took policy decisions at a fortnightly frequency. (Note also that, as discussed above, although we make the event study entries for all dates available in the dataset, the analysis in this section and beyond uses data beginning in 2002.) In that case, the one-month OIS covers the next meeting as well, hence may capture changes in expectations about the next meeting in addition to the immediate policy surprise.
There are very few, if any, quotes for the one-week OIS on many event dates. Otherwise utilizing that measure would have solved the problem of frequency of meetings being higher than maturity of contract used for the period prior to November 2001. 8 A more elaborate and formal presentation of the factor rotations is presented in the Appendix.
The other factor that is always present in the press conference window does not have factor loadings that lead to a Target interpretation. That factor does not load on the 1month OIS much. This is mechanically possible because the rotation forces one factor to be orthogonal to 1-month OIS but does not force the other factor to closely follow it. The other factor in the press conference window turns out to be a Timing factor that captures the shifts in market expectations over the next few meetings that leaves longer-term interest rates essentially unchanged. It is therefore important to note that estimating a Target factor in the press release window and a Timing factor in the press conference window are ndings, not assumptions. Similarly, the QE factor that loads only on longer-term yields is a nding interpretable this way, not ex-ante assumed.
The interpretable surprises we measure using the factor rotation are suitable for our research We have also experimented with alternative identifying assumptions. Since we found a single factor in the press release window, identifying the Target factor by orthogonalizing with respect to a second factor may have been misleading so we used the change in the one-month OIS directly as the Target factor, which unsurprisingly made no dierence. Similarly, not using the one-month OIS in the conference window (where it has negligible variance) and using orthogonality to three-month OIS to do the factor rotation gave qualitatively the same results. Many other ways of identifying surprises are possible and will surely be used by other researchers.
3.2 Policy surprises in the euro area Figure 3 shows the loadings of the rotated factors over the seven maturities in the analysis.
As the factors are identied up to scale, we scale them such that Target has unit eect on one-month OIS, Timing has unit eect on six-month OIS, Forward Guidance has unit eect on the two-year and QE has unit eect on the ten-year yields. This normalization has no eect on the variance shares and statistical signicance of the results we report.
Note again that these factors are dened by the factor loadings, the shapes are not assumed but estimated. The assumptions we made are the orthogonality conditions discussed above.
Our statistical tests tell us to look for a single factor in the press release window and three factors in the press conference window (the third factor is only statistically present in the QE period but the factor loadings of all rotated factors are the same throughout the period. The QE factor has the same denition in the pre-QE period but does not show activity, consistent with its label.) Those factors are rotated according to the orthogonality conditions assumed so that they may be interpretable. In the event, they turn out to have factor loadings shown in the gure and admit labels that have economic meaning. Notice in particular that the eect of QE was consistent with its design: the eect is greater the longer the maturityrecall that only the eect on the current-month OIS are set (to zero) by construction, whereas the eects on all other maturities are estimated. 9 Table 3 shows the relative contribution (variance shares) of our identied factors in explaining changes across the risk-free rate curve in the press release and conference windows. This is a striking table showing that in the press release window yield curve variance is greatest in the shortest maturity to begin with, and the Target factor captures about all of the short-end volatility. Since there is no other statistically signicant factor in this window and the Target factor does not load on the long-end at all, the volatility of the long-end is entirely in the idiosyncratic residual, but this volatility is small. Conversely, in the press conference window, the overall volatility is much higher and is concentrated on the longer-end of the yield curve, with a peak in volatility at two to ve years, the maturities that are aected by both Forward Guidance and QE. Thus, in the cases where much of the variance of a particular maturity is attributed to the residual, such as 10-years in the press release window or one-month in the press conference window, these have the lowest variances across the yield curve and the maturities with largest variances, one-month in the press release and two and ve-years in the press conference windows are fully explained by the measured policy surprises.
As noted above, we nd that (a) changes of interest rates at short maturities in the press release window are mainly explained by the Target surprise, while in the conference window the Target surprise is absent and a Timing surprise emerges; (b) the Path (Forward Guidance) factor is present in the conference window and aects interest rates across all maturities above one month, with a typical hump-shaped pattern, and (c) the QE factor, present in 9 ECB used a variety of maturities for its QE program, with an average maturity of almost eight years.
Hence, nding largest eects of QE on the long-end of the yield curve is as reasonable in the euro area as in the US. the conference window, is dierentiated from Forward Guidance by having an impact that is stronger the longer the maturity.
This nding is interesting and speaks to both the methodology of extracting the factors and the structure of monetary policy communication in the Euro Area. The methodology forces two of the factors, which we call Forward Guidance and QE factors, to be uncorrelated with the current interest rate surprise, in this case the change in the one-month OIS rate. The other factor is not forced to display the features of a Target factor (that is, not forced to be highly correlated with the change in the one month OIS rate) but in the press release window it turns out endogenously to be so, suggesting that Target is a dominant feature in the data.
Recall that in the US policy window Target is always a dominant policy surprise as the policy decision and statement are simultaneously released; therefore this identication strategy in the US leads to nding the rst factor to be Target.
In the euro area, in the conference window we nd instead that the rst factor does not look like a Target factor; it is not correlated with the one-month OIS at all. This has to be expected because once the current setting of policy is announced in the press release, updates to beliefs about Target should no longer be a dominant source of variation in the conference window. Indeed, principal components and the factor rotations that dene Forward Guidance (Path) and QE lead to nding a remaining factor that does not display the features of a Target factor, but the features of a factor that resembles the Timing factor discussed by Gürkaynak et al. (2007). This is a factor that loads on the very near term (within six months) OIS rate changes and captures shifts in the perceived Timing of policy actions, such as beliefs that although a policy change did not happen today it is now more likely that it will happen in the next meeting or two. That is, the press conference makes the relative probabilities shift across very near-term meetings, as well as farther out. Forward Guidance captures changes in policy expectations aecting interest rates about two years out and Timing captures changes in expectations at a shorter horizon, aecting only to a small extent interest rates of longer maturities. Although we name these factors Timing and Forward Guidance, it is important to recognize that in essence these are two separate Forward Guidance factors, one with a peak eect at shorter maturities and the other at relatively longer maturities.
One can do the analysis with these factors, as we will, and keep in mind that the interpretations of factors across windows will be dierent: the rst factor is Target in the press release window, and Timing in the conference window. Of course, it is possible to force the rst factor to consistently be Target in both windows by imposing a large positive correlation of this factor with the change in the one-month OIS rate. In that case, the factor interpretations will be symmetric across windows, but this will come at the cost that less of the variance of the conference window will be explained: Timing, which we have shown is important, would have to turn into a residual. We continue with allowing for the Timing factor in the conference window. Figure 4 shows the Target/Timing, Forward Guidance and QE factors over the press release and the conference windows as time series, normalized as in Section 3.2. We compare these identied surprises with ECB's relevant policy actions to make sure that we have an accurate mapping. While we do not comment on each one of these for brevity, it instills condence in our method to observe that each of the largest surprises we extract using market-based methods coincide with market commentary that accords with the measured surprise.

Surprises and ECB policy announcements
For instance, the largest realization for the Target factor occurred on 3 November 2011 when the ECB cut its policy rates by 25 basis points, which came unexpectedly in part because it was the rst meeting of Mario Draghi as ECB President. In addition, we nd a sizeable negative realization on 4 September 2014 when the ECB cut the deposit facility rate to more negative values, which was unexpected as also suggested by surveys among analysts ahead of the policy meeting.
10 Turning to the Timing factor, large realizations occur on 3 March 2011 and 5 June 2008. In these episodes the policy rates were left unchanged, as expected; and indeed we nd that the Target factor hardly moved in the press release window. But the Introductory Statement read out at the beginning of the press conference contained expressions (strong vigilance in one episode, and state of heightened alertness as well as readiness to act in a rm and timely manner in the other episode) meant to signal high likelihood of a rate hike in the subsequent policy meetings.
The largest realization of the Forward Guidance factor in the conference window occurred on 3 July 2008 and was a negative realization. That is interesting because on that date policy rates were increased by 25 basis points, and this was fully expected, our Target factor did not register a surprise. (A fact also supported by survey evidence gathered ahead of the policy meeting). However, the press conference was taken as signaling that no more hikes were intended, leading markets to price out further increases. Finally, the largest realization of the QE factor is on 22 January 2015; this is a negative reading (an easing surprise) and 10 Bloomberg News on 4 September 2014 reported that Of 57 economists surveyed by Bloomberg News, 51 said [today] the ECB will keep its key interest rate unchanged.
corresponds to the announcement of the ECB's asset purchase program, which was made in the press conference.
A more detailed description of the time series of these surprises is relegated to the appendix.
While we present the surprises and the statements that led to these, a formal mapping between quantied words, as in Hansen and McMahon (2016) and the market perceptions we measure here has to be a separate paper.

Asset Price Response to Policy
An important contribution of this paper is dening the interpretable policy surprise factors for the euro area. Once the policy surprises are measured as described in Section 3, causality is established, and the response of other asset prices can be studied via OLS. We estimate the impact of monetary policy on sovereign yields, the exchange rate, and the stock market.
In the conference window we continue to use the surprise in the US initial jobless claims as an additional control (labeled IJC) whenever we run regressions. Including or excluding the initial claims makes essentially no dierence to our coecients of interest. Table 4 shows the intraday reaction of euro area risk-free rates to the Target surprise we have identied in the press release window (panel A) and the three surprises in the conference window (panel B). This table presents the information in Figure 3 with numbers, the normalized factor loadings using an OLS regression framework and controlling for the US Initial Jobless Claims surprises in the press conference window. The loadings, which are essentially unchanged compared to Figure 3  shows that reactions to these have not changed markedly over time.

Sovereign yields
Having veried the consistency of our surprise measures across subsamples for the euro area risk-free rates, we now turn to their their eects on the sovereign yields of Italy and Spain.
11 Tables 8 and 9 show the sovereign yield responses and document that these factors aected the sovereign yields the same way they aected the safe rates. Importantly, QE, when active, has aected Spanish and especially Italian long-term yields more than the safe rate, narrowing spreads on impact, consistent with the evidence in Altavilla et al. (2015).
Also once again note the high R 2 s across tables 4 to 9. The monetary policy surprises we extracted capture large fractions of yield changes for risk-free and sovereign rates, especially in the conference window for the last subsample for long rates, as the QE factor helps explain the 10-year yield changes. In the earlier samples the R 2 for long rates are lower but the variance of these rates are also lower without QE surprises that move them. This is in contrast to Leombroni et al. (2017), who nd that spreads were widened due to Forward Guidance surprises in a study that did not separately assess QE eects. We therefore nd that the rst stage of monetary policy transmission is not dierent across the largest euro area countries. The real eects may be dierent due to non-nancial dierences across them, as argued by Corsetti 11 Eects on German rates are indistinguishable from eects on OIS rates when both are available and eects on French rates very closely follow German ones. We do not separately report these for brevity.

16
Electronic copy available at: https://ssrn.com/abstract=3402078 et al. (2018) but that is a separate topic of study.  However, the two types of surprises have opposite eects on stocks. Lower rates are good news (lower discount rates and higher demand) but learning that the cyclical state is worse than previously thought is bad news (lower dividends). One can see similar eects on ination expectations; a positive policy surprise should decrease ination compensation implied by indexed securities unless the surprise is perceived to be signalling information about high inationary pressures, in which case ination compensation will increase. The macroeconomic impacts of these policy surprises are also found to dier depending on whether the policy surprise triggers a positive or negative response of the stock market (Jaroci«ski and Karadi, 2018) or, similarly, a positive or negative response of ination-linked swaps (Andrade and Ferroni, 2016), as measured at high-frequency around policy events.

Exchange rate
12 Therefore, the presence of these two types of policy can make the response of the stock market, on average, insignicant and can produce the results reported in Table 11.
To assess whether there is evidence of these two types of surprises in our dataset, for each policy event we compare the sign of the response of the stock market and ination-linked swaps. If nominal rates, stock prices and ination-linked swaps all move in the same direction, this would suggest that information shocks may be prevailing. To overcome the problem that data on ination compensation are not reliably available at intraday frequency, we adapt our analysis to daily frequency and measure the interest rate reaction as the tted value of the one-day change (around the policy events) of the 2-year OIS regressed on the intraday factors.
We do the same for the one-day change of ination-linked swaps and the one-day log-dierence of stock prices.
12 Jaroci«ski and Karadi (2018) use intraday data employing stocks whereas Andrade and Ferroni (2016) use daily data due to lack of reliable intraday data for ination-linked swaps. Figure 5 shows the dates for which stock prices and ination swaps move in the same direction so that the market perception of policy can be clearly interpreted as information (Delphic) surprises or pure policy (Odyssean) surprises.
13 The gure shows that there is a marked dierence across sub-samples in terms of the information market participants extract from policy surprises. Whereas in the post-2014 sub-sample (bottom panel) information shocks (dened as nominal interest rates, stock prices, and ination-linked swaps moving in the same direction) are rare, these are frequent during the crisis sub-sample (top panel). That is, in the crisis period market participants attributed more of the surprises they perceived to information they thought the ECB has, consistent with a similar nding for the US by Lunsford (2018).
To delve deeper into this issue we build a daily VAR. The identication strategy is based on the idea of using high-frequency monetary policy surprises to isolate the variation in the reduced-form residuals in the VAR due to monetary policy shocks. The use of external instruments for identication in macroeconometric models goes back to Stock and Watson (2012) and Mertens and Ravn (2013). We start the analysis by estimating the reduced form VAR as described in equation (2).
where Y t is a vector consisting of the 2-year OIS, the log EUR-USD exchange rate, the log of the stock market index (Euro Stoxx 50), and the 2-year ination linked swap (ILS2Y). As we are only interested in the eects of monetary policy shocks, our objective is to identify the column of the matrix A 0 corresponding to the contemporaneous eect of the monetary policy shock. The instrument must satisfy the relevance and exogeneity assumptions: where Z t is the instrument, and u m t and u o t denote the monetary and the non-monetary shocks.
Our instruments will be the policy surprises we have identied. The external instruments are the Target factor (press release), the Timing, the Forward Guidance, and the QE factor (press conference). As the VAR residual is a linear combination of structural shocks, we instrument the residuals of the VAR with one instrument at a time, as we aim to extrapolate the component correlated with the Target, Timing, Forward Guidance and Quantitative Easing surprises 13 In about 80% of the policy dates stock prices and ination-linked swaps move in the same direction. In the other 20%, in almost all of the cases the stock price reaction is about zero and the ination-linked swap reaction is very small. Hence, these two measures almost never meaningfully disagree. Therefore, incidentally, our results verify Andrade and Ferroni (2016) and Jaroci«ski and Karadi (2018) methods against each other and nd that they are in agreement.
respectively. As a dierent exercise, one can also include all the instruments simultaneously, however, this will only help to identify a single monetary policy shock which reects a linear combination of the four. To be consistent with the previous analysis, we use the intraday factors as external instruments for the VAR estimated on all sub-samples. When analyzing Quantitative Easing, the instrument is used on the period 2014-2018. Overall, these results suggest that there is evidence of information shocks and that stock prices and ination-linked swaps may both help in telling apart these shocks from the more traditional policy shocks. We nd that these results are fairly persistent, they do not only manifest themselves on policy dates then disappear. The results also suggest that it is possible to extend and generalise the macroeconomic analysis of Andrade and Ferroni (2016) and Jaroci«ski and Karadi (2018). They use the overall policy surprise as measured by the high- year for the QE eects is much longer than that of Wright (2012) for the USat about three monthswhich was obtained using QE announcement dates in a heteroskedasticity-based estimation setting. It is also much longer than Swanson (2017), who estimates a persistence in the US similar to Wright, using local projections. , using a shorter sample, also nd persistemt eects of QE in the euro area.
To further investigate these results, we check whether these ndings are robust to dierent VAR specications. Figure 8 shows the IRFs from a Combinatoric VAR (CVAR Guidance surprises aected the 10-year yields for a long period of time.
It is interesting to nd the persistence of unconventional policy in the euro area on longterm interest rates to be much higher than what is found for the US. In interpreting this result it is important to keep in mind the dierence in methodologies. We employ a VAR with a continuous measure for Forward Guidance and QE surprises as instruments for the euro area, in contrast to heteroskedasticity-identied VARs and local projections for the US. We note the high and robust persistence of unconventional policy eects we nd with the VAR but caution that these results are not directly comparable to those in the literature for the US.
15 As 10-year yields we include the German, Italian, Spanish and French yields, the OIS yields, the EA GDP-weighted yield, and the EA yields derived by the ECB using the Svensson (1994)  16 The more recent literature, using Romer and Romer (2004) identication and Jordà (2005) local projections, nds that monetary policy tightening has stronger eects than loosening (Tenreyro and Thwaites, 2016;Barnichon and Matthes, 2017).
In this section we ask a related but distinct question and study whether market participants Tables 12 to 14 report the results. These results are striking in their lack of asymmetry.
Across the three tables for the eects of policy surprises on euro area risk free rates and Italian and Spanish sovereign yields, only very few interaction terms are statistically signicant.

17
Importantly, even in the few cases where the interaction term is signicant, we nd positive interaction eects for negative surprises, indicating stronger eects of easing surprises on the yield curve. Hence, we nd no evidence that the yield curve responses in the euro area may lead to, or are consistent with, weaker eects of monetary policy when policy is expansionary.
We leave studying the possible asymmetry in real eects of monetary policy in the euro area for future work but note the apparent dierence between the results in the literature for 16 The small literature on non-linear eects of monetary policy focuses on the possible asymmetric responses of macroeconomic, rather than nancial, variables. Thoma (1994) shows that monetary policy is more eective in expansions than recessions. Weise (1999) nds using a Smooth-Transition VAR (ST-VAR) model that monetary policy does not have any power in recessions. Peersman and Smets (2002), Garcia and Schaller (2002) and Lo and Piger (2005), using a two-state Markow-Switching Model (MSM) nd that monetary policy is more powerful in recessions than in expansions. the US real eect asymmetry and the symmetry we nd in nancial market eects in the euro area.

Extension: Decomposing Market Reactions to Other
Policy News Policy events that are not covered by our data setpolicy communications that are not Governing Council policy decisionscan also be analyzed using our methodology. Policymaker speeches, releases of minutes and the like are also policy communication and have nancial market eects. We can treat these events as if they are Governing Council policy dates and, given the factor loadings we estimated for our factors, use the changes in the OIS yields in those event windows to decompose the market perception into Target, Timing, Forward Guidance, and QE surprises. This is an exercise in nding the combination of monetary policy surprise factors that best t the change in yields around the relevant window.
The methodology is straightforward; for a particular event ie.g., a speech we take a window long enough to bracket the beginning and the end of the event, and compute the change in the 1M, 3M, 6M, 1Y, 2Y, 5Y, and 10Y OIS. We collect these yield changes in a vector OIS i , and, given the (rotated) factor loadingsΛ we estimated from the EA-MPD, we nd the factorsF i which minimize the sum of squared residuals of OIS i − F iΛ . That iŝ The solution of this minimization problem can be recognized as the OLS estimator of F i , in a regression of OIS i onto the space spanned byΛ.
We apply our methodology to two illustrative events that elicited noticeable market reac-  Figure 9 shows the yield changes around these events. From the chart, it is clear that both events moved the long end of the curve, as it was perceived that new information about QE was released. Figure 10 then shows our estimated factors for these two events, in terms of the Target, Timing, FG, and QE decomposition. To make the factors comparable to what were shown earlier in this paper, we rescale them as in section 3, and we report them as a fraction of the average absolute value of the in-sample surprise of each type. Note, as before, that the dierent scalings imply that magnitudes are comparable across dierent dates for the same type of surprise, but not across dierent types of surprises.
As expected, our exercise nds a large QE factor in both events. For the rst event, we found a relatively large Forward Guidance component as well. Large here is with respect to the in-sample average size of these surprises. The normalization we used imply that a reading of any factor above unity means that factor was larger than its average absolute reading insample.
A few notes on this exercise are in order. First of all, we show that it is possible to map our understanding of identied policy surprises based on the analysis of Governing Council policy announcements into any other kind of policy news. This is of independent interest.
Secondly, the Target factor is also estimated here. One can of course treat these news as analogous to press conferences and limit the possible factors to Timing, Forward Guidance, and QE but to the extent that market participants update their beliefs about outcomes of policy meetings within a month, one may measure a reaction interpretable as Target.
Lastly, this is a good place to discuss methodological choices. Our methodology is based only on the changes in safe rates of dierent maturities: the factors, including QE, do not load on individual country yields or spreads. Thus, the QE surprise we identify is one that lowers the long-term euro area safe rate and we can show, as a nding, that this QE surprise also lowers spreads. One should think of this as a macroeconomic easing QE. An alternative is QE that is perceived to particularly aect the Italian and Spanish yields, as in Rogers et al. (2014). Measuring this requires having spreads in the matrix from which one extracts the factors; we chose not to follow this path. For example, as spreads narrowed sharply around the whatever it takes speech, this second type of QE factor would have signaled a very large QE easing surprise by denition.

19
19 The Whatever it takes speech by Mario Draghi is perhaps the most famous of euro area policy communications. We do not use it as an example here because we do not have data on the intraday ve-year OIS yield for that event. However, the minor increase in the 10-year OIS yield around that speech suggests that our methodology would have interpreted this event, if anything, as a small QE tightening surprise. Identifying QE surprises by utilizing spreads would not lead to this nding for this speech but in that case QE would narrow spreads by assumption, and we would not be able to discuss eects of QE on spreads.
Using our methodology, we are able to show that the macroeconomic easing QE factor that we identify also narrows spreads but we are silent about QE that may be perceived to dierentially aect sovereign yields rather than providing overall stimulus. Clearly it is possible to measure a second type of QE and indeed to measure the two simultaneously, by allowing for two dierent types of QE surprises. Our research questions are best answered by the rotated factors we identied, but work on speeches of the whatever it takes type, or on the particular experiences of periphery countries may require alternative surprise identications and associated factor rotations.

Conclusions
Our aims in writing this paper were twofold. First, we wanted to explain and make available what we expect to be the standard euro area monetary policy event-study database, the EA-MPD. The paper rst explains the construction of this data set. This is data we carefully analyzed and cleaned so that misquotes that were prevalent in the early period of ECB policymaking do not lead to spurious ndings. The data will be continuously updated and made available online. Our second aim was to show the use of this data in constructing euro area monetary policy surprises, to decompose these into surprises in policy actions, policy path, and QE news.
The two-stage nature of ECB policy news dissemination turns out to be very helpful in both increasing the number of data points and making statistical analysis more precise, and also in providing a validity check for our surprise measures. Comfortingly, we nd that the Target surprises were dominant in the announcement window but that in the conference window this factor does not even exist, with news about future path of policy the main driver of yield changes before QE; and nally, that QE news were present in the latter part of our sample, only in the conference window. These all dovetail with the general understanding of how the ECB policy communication is designed to operate. We show that in practice policy communication has indeed worked as designed. We studied the response of various asset prices to these surprises and learned that surprises in the immediate setting of monetary policy, Target surprises, have eects only on the short end of the yield curve while Timing, Forward Guidance, and QE surprises both aect longer-term yields but in dierent ways.
It is important to emphasize again that rather than assuming the presence of predened surprises in dierent windows, we estimated and identied these, nding a multifaceted information structure in the press conference window. We further learned that sovereign yields and exchange rates respond to ECB policy communication, in interpretable ways, based on the source of the surprise. Importantly, QE turns out to have lowered all yields and narrowed spreads. Also importantly, we show that this eect was long-lived, with a half life of about a year. This is much longer than what was found earlier in the literature, when QE surprises were not quantied, and eects were based only on the dates of QE announcements.
Among other extensions, we showed that measuring the information surprise in ECB policy employing stock prices or ination compensation yield similar results and this decomposition is needed to understand stock price reactions to monetary policy. Importantly, we also showed how to use the identied policy surprise factors to analyze any policy communication, not only Governing Council policy releases and statements on meeting dates.
We hope that the event-study dataset that we have compiled and will regularly update will foster more research on monetary policy and its eects in the euro area. Questions about the eects of monetary policy on markets in dierent countries, how these dier by the fundamentals of those countries, using the surprises to help identify VAR-based real eects in the Gertler and Karadi (2015) fashion, the transmission of ECB policies to non-euro area countries are some of the questions that continue to be important for academics and policymakers alike.
While we shed more light than before on measuring and assessing monetary policy in the euro area, there certainly is much to be done.     Robust standard errors in parentheses; ***, **, and * denote statistical signicance at the 1%, 5% and 10% levels, respectively. ***, **, and * denote statistical signicance at the 1%, 5% and 10% levels, respectively.   ***, **, and * denote statistical signicance at the 1%, 5% and 10% levels, respectively.    1 to 4), and the reaction of the euro area bank stock market sub-index over the same samples (columns 5 to 8) to surprises in monetary policy using intraday data.

Tables
Coecients are expressed in percentage points per standard deviation change in the factors.
Robust standard errors in parentheses; ***, **, and * denote statistical signicance at the 1%, 5% and 10% levels, respectively.     QE Note: The gure shows the factor loadings for the press release (rst row) and the conference windows (second row), in basis points. In each window, for each maturity the loadings are obtained by regressing the surprises onto the factors also controlling for the standardized surprise associated with the release of the US initial jobless claims. The Target and the Timing factors are normalized to have unit eect on the 1-month and 6-month OIS, respectively. The Forward Guidance and QE factors are normalized to have unit eect on the 2-year and on the 10-year yields, respectively. The shaded areas indicate the 90%, 95% and 99% condence intervals.

QE Factor
Note: The gure shows the estimated factors over time reported in basis points. As the factors are identied up to scale, we scale them such that target has unit eect on one-month OIS, timing has unit eect on six-month OIS, forward guidance has unit eect on the two-year and QE has unit eect on the ten-year yields.