what makes the s&p 500 jump? - econ.au.dk · what makes the s&p 500 jump? november 10, 2014...
TRANSCRIPT
What Makes the S&P 500 Jump?
November 10, 2014
Abstract
Using intraday transaction prices and a non-parametric test, we study
jumps in the S&P 500 index. We show that jumps are low probability, high
impact events. Using a comprehensive set of newswires, we explore the events
that trigger jumps. We are able to relate around 80% of jumps to scheduled
and, more importantly, unscheduled news. Interestingly, unscheduled events
frequently trigger the largest jumps, indicating that unexpected events truly
“shock” the market. Analyzing scheduled events, we find little evidence to
suggest that announcement surprises and the level of uncertainty surrounding
scheduled news can explain why some announcements trigger jumps while
others do not. Collectively, our results challenge common beliefs about
jumps, suggesting that the relationship between news and jumps may be more
complex than originally thought.
JEL classification: G12, G13
Keywords: Jumps, News, Intraday, S&P 500
I Introduction
The stock market is increasingly prone to sudden and sharp fluctuations and
identifying these large variations in asset prices has become an active area of
research.1 Notwithstanding the recent developments in the econometric literature
on jump detection, it is surprising that we know only very little about the economic
drivers of jumps especially during the recent financial crisis (2008–2012). How many
jumps occur during the crisis period? What events trigger jumps? Why do some
events trigger jumps while other similar events do not? These are some of the
questions that we seek to answer.
In this paper, we conduct a comprehensive analysis of jumps in the S&P 500
market. In doing so, we make three important contributions to the literature. First,
we present model-free evidence of jumps in the S&P 500 during the recent crisis
period. To accomplish this goal, we implement the non-parametric jump detection
methodology of Lee and Mykland (2008) on 15-min transaction returns. We find
that jumps in the S&P 500 are rare events that occur with a low probability of
0.22%. We detect 238 jumps during our sample period, most of which are negative.
Notwithstanding their rare nature, jumps are high impact events. Our analysis
reveals that the mean absolute jump return is 9 times larger than the mean absolute
return observed in our sample.
Second, and more important from an economic point of view, we investigate the
events that cause jumps. Rather than preselecting, and thus restricting ourselves
to, a handful of scheduled events as is done in extant studies, we present and
implement an approach that enables us to pursue our investigation in a more
general manner.2 Importantly, our approach allows us to distinguish between
scheduled and unscheduled events that “shock” the market. To do this, we draw
on a comprehensive set of newswires obtained from four leading news agencies: The
1See the important econometric contributions of Pan (2002), Eraker et al. (2003), Barndorff-Nielsen and Shephard (2004), Eraker (2004), Huang and Tauchen (2005), Johannes (2005),Barndorff-Nielsen and Shephard (2006), Andersen et al. (2007), Jiang and Oomen (2008) andLee and Mykland (2008).
2For example, Andersen et al. (2007), Dungey et al. (2009) and Evans (2011) focus on a fewscheduled macroeconomic announcements.
1
Associated Press, The Dow Jones Newswires, The New York Times and The Wall
Street Journal. These newswires are widely followed by market participants and
have the great advantage of displaying accurate time-stamps (Bradley et al., 2014),
which crucially allow us to accurately link the occurrence of jumps to information
events. Our approach consists in searching newswires to identify important news
around individual jumps. In conducting our investigation, we pay a great deal of
attention to the issue of stale news, which we expunge from our set of important
news. When we consider only scheduled macroeconomic announcements, as is done
in extant studies, we are able to relate less than 40% of jumps to news.3 However,
by analyzing not only scheduled announcements but also unscheduled events, we
uncover a much stronger link between jumps and news. More specifically, we are able
to relate close to 80% of jumps to both scheduled macroeconomic announcements
and unscheduled events. Our analysis reveals that unscheduled events such as
unconventional (monetary/fiscal) policy and political announcements play a key role
in our understanding of jumps. We find that they frequently trigger the largest jumps
in the S&P 500. In particular, the mean absolute jump return related to unscheduled
news (1.1%) is 50% higher than that of scheduled events. The intuition behind this
result is simple. Since the timing of scheduled macroeconomic announcements is
perfectly known in advance, they are widely anticipated, making them less likely to
trigger the largest fluctuations in the stock index. In contrast, unscheduled news
takes the market by surprise and truly “shocks” market participants, leading to
sharp movements in the S&P 500.
Third, we investigate why some scheduled (and recurring) events trigger jumps
while other similar events do not. To achieve this goal, we test two commonly held
views. The first posits that financial markets react strongly to large announcement
surprises, defined as the absolute value of the difference between the announced and
expected values. Consequently, large announcement surprises should trigger the
majority of jumps. The second argues that announcements made at times of high
uncertainty, proxied by the cross-sectional standard deviation of forecasts, account
3Dungey et al. (2009), Evans (2011) and Bradley et al. (2014) study the link between scheduledevents and jumps and report similar results.
2
for most jumps. Using analyst forecasts from Bloomberg, we construct empirical
proxies for announcement surprises and uncertainty and test both hypotheses. For
each macroeconomic series, we sort announcements into three surprise portfolios and
study the proportion of jumps associated with each portfolio. If the first hypothesis
is valid, we expect to see that a disproportionately large fraction of jumps is classified
into the high surprise group. We find little evidence of this, thus indicating that
there is little truth in the “theory” of jumps according to which announcement
surprises explain the occurrence of jumps. We pursue our analysis by sorting all
announcements of a specific economic series into three portfolios of uncertainty:
low, median and high. The proportion of jumps linked to each portfolio appears
difficult to reconcile with the belief that announcements made at times of high
uncertainty are strongly associated with jumps. Digging deeper, we double-sort the
announcements first on surprise and then on uncertainty, yielding a total of nine
portfolios. Holding the level of surprise fixed, we find little evidence to suggest
that announcements occurring during periods of high uncertainty trigger the largest
proportion of jumps. Overall, our results challenge the conventional wisdom that
announcement surprises and uncertainty can explain why some announcements of
the same economic variable trigger jumps, while other similar announcements do
not. Our results survive a number of robustness checks, including different sources
of macroeconomic forecasts (Reuters instead of Bloomberg) and alternative proxies
for announcement surprises and uncertainty.
Our work contributes to several strands of the literature, including studies on
modeling time-series of asset prices. Merton (1976), Bates (1996), Pan (2002),
Eraker et al. (2003), Eraker (2004), Johannes (2005) and Broadie et al. (2007)
propose and estimate various jump-diffusion models to capture the dynamics
of asset prices. However, their studies are cast in a parametric and therefore
model-dependent framework. That is, they assume that the asset price follows
a specific data-generating process. In contrast, we do not make any assumption
regarding the underlying dynamics. Rather, we implement a state-of-the-art and
non-parametric jump detection test that allows us to detect jumps at the high
3
frequency level. Besides identifying jumps using a model-free test, we conduct a
comprehensive analysis of the economic events that trigger jumps. Learning about
these triggers could lead to important extensions of existing models of asset prices.
In particular, one could propose and estimate richer jump-diffusion models where
jumps are explicitly linked to the occurrence of expected events along the lines of
Piazzesi (2005).
Our study also relates to the empirical literature on scheduled announcements
and large fluctuations in asset prices. Andersen et al. (2007) analyze the connection
between macroeconomic announcements and jumps in financial markets. In a similar
vein, Bradley et al. (2014) study the impact of company specific announcements such
as earnings reports on the likelihood of jumps. Relative to these interesting papers,
we make two important contributions. First, we go beyond documenting the link
between scheduled events and jumps and investigate the important question of why
some announcements trigger jumps, while other similar announcements do not. In
doing so, we document an interesting puzzle on the disconnect between the perceived
importance of some announcements and asset price jumps. We show that, contrary
to popular beliefs, some apparently important scheduled announcements are not
associated with large price movements, echoing the findings of Roll (1988). Second
and perhaps more important, we develop a framework that allows us to identify
the events, scheduled and unscheduled, that ex-post likely trigger jumps. This is
different from extant studies that typically restrict their attention to a small number
of scheduled news, implicitly assuming that unscheduled events play a marginal role
in the dynamics of jumps. We document that this implicit assumption is not borne
out in the data. Quite the contrary, unscheduled events trigger a sizable proportion
of jumps in the stock market. The large magnitude of jumps related to unscheduled
news confirms that unexpected news plays an important role in the dynamics of
jumps. Thus, it is important to study both scheduled and unscheduled events to
improve our understanding of jumps.
Finally, we build on and extend the works of Cutler et al. (1989), Cornell (2013)
and Bajgrowicz et al. (2013). Cutler et al. (1989) and Cornell (2013) explore the
4
daily press to identify the events that trigger the largest changes in (daily) stock
prices. However, they do not use a formal jump detection test but simply specify
an ad hoc threshold to identify large returns, which could materially affect their
findings.4 Bajgrowicz et al. (2013) improve on these studies by using a formal jump
detection test (Barndorff-Nielsen and Shephard, 2006) that identifies jump days.
Although these studies connect news to jumps occurring on the same trading day,
none of them pays attention to the exact timing of events. This is a potentially
severe shortcomings since focusing only on jump days and ignoring the exact time
of events may significantly distort the results. This is because an information event
occurring in the morning might be (mistakenly) linked to a jump occurring in the
afternoon, even though the two events, i.e. information and jump, clearly take
place at very different points in time. This limitation casts doubts on previous
findings and calls for a more precise analysis at the intraday level. We advance in
this direction. We use a cutting-edge jump detection method that precisely detects
jumps at the intraday level, enabling us to identify not only the days but also the
exact times when jump events occur. We then exploit a large database of accurately
time-stamped newswires to investigate the connection between intraday news and
jumps, resulting in a precise analysis of the events that shock the market.
This paper proceeds as follows. Section II introduces our data and presents
the jump detection test. Section III describes the dynamics of jumps. Section IV
explores the events that cause jumps. Section V tests several hypotheses aimed at
understanding the characteristics of scheduled announcements that trigger jumps.
Finally, Section VI concludes.
II Data and Methodology
In this section, we provide a detailed account of our research design. We proceed
in two steps. To begin with, we describe our dataset of intraday prices. Next, we
4An important limitation of the approach that identifies jumps based on an unconditionalthreshold is the acute clustering of “pseudo” jumps during volatile episodes of the market. SeeCornell (2013).
5
present the jump detection methodology.
A. Data
We obtain intraday transaction prices for the S&P 500 futures market from
TickData. Our sample covers the period from January 1, 2008 to July 12, 2012.
The futures contracts expire in March, June, September, December and the following
three Decembers. S&P 500 futures contracts trade on two venues at the CME, pit
and electronic. Trading hours on both platforms have no overlap and collectively
span 22:45 hours. Pit trading takes place between 9:30 AM (ET) and 4:15 PM (ET).5
Electronic trading starts at 4:30 PM (ET), pauses at 5:15 PM (ET) for 45 minutes,
resumes at 6:00 PM (ET) and stops the following day at 9:15 AM (ET).6 The dataset
contains the following information: trading day, transaction time (stamped to the
second), transaction price, transaction volume and an indicator to determine the
trading venue.7
We use both pit and electronic transaction records and process the dataset
as follows. First, we discard all transactions with prices lower than or equal to
zero. Second, we expunge all trades with time-stamps that are inconsistent with the
exchange’s trading hours. Finally, we retain the futures contract with the highest
number of transactions only.8
Following existing studies, e.g. Lee and Mykland (2008) and Bradley et al.
(2014), we sample our data at the 15-min frequency.9 To construct regularly spaced
transaction prices, we use the most recent transaction price up to the end of each
15 minutes interval. Figure 1 displays the time-series of prices. The index price
(expressed in index points) starts at around 1,400 in January 2008, plunges to 700
5See the CME website for further details: www.cmegroup.com.6Note that TickData uses the Chicago Time (CT) as the reference time zone. To facilitate
comparisons with earlier studies, we adopt Eastern Time (ET) as our reference time zone andadjust all time-stamps accordingly.
7TickData provides volume data for all trades executed on the electronic platform only. Hence,we do not have volume data during the pit hours (9:30 AM–4:15 PM).
8Usually, this is either the first or second nearby futures contract.9We also analyze the variance signature plot. To this end, we estimate the variance of the stock
market using different sampling frequencies. Plotting the variance estimates against the samplingfrequencies supports the choice of sampling frequency.
6
in 2009, steadily rises to 1,300 in 2011, dips and recovers to finish at around 1,400
in July 2012. Table 1 reports summary statistics of the log returns. Column 1
reports the total number of return observations. Columns 2 through 7 report the
mean, minimum, maximum, standard deviation, skewness and kurtosis, respectively.
Overall, our sample includes 106,205 (15-min) returns. The average return is small
and negligible (−9.97 × 10−6). Turning to the higher order moments, we observe
a slightly negative skewness coefficient (−0.06) and a high level of kurtosis (47.98),
providing an early indication that the stock market may exhibit jumps.
B. Jump Detection Methodology
We implement the jump detection test of Lee and Mykland (2008). This
methodology presents several advantages. Unlike the model-specific approach of
Bates (1996), Eraker et al. (2003), Eraker (2004) and Johannes (2005), it is
non-parametric. This is important, as it makes our analysis robust to model
misspecification risk. Moreover, Lee and Mykland’s method allows us to precisely
identify jump events at the intraday level.10 This constitutes an important advantage
over the procedures of Huang and Tauchen (2005), Barndorff-Nielsen and Shephard
(2006) and Jiang and Oomen (2008), that only identify jump days and offer little
insights into the sign, size and timing of intraday jumps.11 Furthermore, Dumitru
and Urga (2012) compare existing jump detection tests in detail and find strong
evidence in support of the Lee and Mykland (2008) jump detection method.
The intuition behind the jump test of Lee and Mykland (2008) is simple. We can
think of jumps as large returns compared to a local estimate of variance. Hence, to
detect jumps, one can simply compare the magnitude of individual returns relative to
local estimates of volatility. Lee and Mykland emphasize that the volatility estimate
must be robust to jumps and suggest using the square root of the bipower variation
10The Lee and Mykland (2008) method can also be applied at lower frequencies. See Schneideret al. (2010) and Pukthuanthong and Roll (2011) for some interesting applications.
11Huang and Tauchen (2005) and Barndorff-Nielsen and Shephard (2006) build on the theoryof quadratic variation (Barndorff-Nielsen and Shephard, 2004), which decomposes the quadraticvariation of returns into a continuous and discontinuous components. Jiang and Oomen (2008) usethe insights of variance swaps to devise their jump test.
7
(BV) of Barndorff-Nielsen and Shephard (2006). Equation (1) presents the jump
test:
LMi − Cn
Sn
≥ ε (1)
LMi, Cn, Sn and ε are computed as follows:
LMi =|ri|
BV1
2
i
(2)
ri = logPi
Pi−1
(3)
BVi =1
K − 2
i−1∑
j=i−K+2
|rj−1||rj| (4)
Cn =(2 logn)
1
2
c−
log π + log (log n)
2c(2 logn)1
2
(5)
Sn =1
c(2 logn)1
2
(6)
ε = − log (− log (−x)) (7)
where LMi indicates the Lee and Mykland test statistic. ri denotes the log return at
time i. Pi and Pi−1 are the asset prices at times i and i− 1, respectively. Similarly,
BVi refers to the bipower variation computed just before observing return i. We
calculate the bipower variation over a rolling window of length K. Lee and Mykland
(2008) recommend a window of 156 observations for a sampling frequency of 15-min.
We follow their advice. Cn and Sn are constant for a given sample of n total
observations. c denotes the normalization factor√
2
π. ε refers to the critical value
calculated at the x confidence level. As is standard in the literature, we use a strict
confidence level and set x to 99.9%.
III Jump Dynamics
This section discusses the dynamics of jumps. We begin by illustrating the workings
of the test on two important days. We then extend the methodology to all trading
8
days. Furthermore, we study asymmetries in the dynamics of jumps.
A. Illustration
Before applying the jump detection test to all 15-min returns, it is important to
ascertain that the methodology yields sensible results. A simple way to achieve this
is to focus on two important days, namely September 29, 2008 and May 6, 2010.
The former marks the Fed’s decision to increase the size of its Term Auction Facility
and the defeat of the $700 billion financial rescue bill by the US Congress, adding to
concerns about the state of the US financial sector. The latter relates to the “flash
crash”, when the market dives several percentage points at around 2:45 PM and
recovers a few minutes after.
Figure 2 plots the price-series of the S&P 500 for each of the two dates. The
top plot shows the price-series on September 29, 2008. The bottom panel displays
prices on May 6, 2010. Starting with the top plot, we can see that the S&P 500
displays small variations until around 10:00 AM, when it falls by 1%. We observe
an even larger crash of 3.16% at 1:45 PM. Similarly, the bottom plot highlights
small variations in the market portfolio until around 2:45 PM when it tumbles by
4% and immediately rebounds by 2.64%. Clearly, both plots point to large intraday
variations in the stock index.
We implement the jump test on both days. If it works well, we expect to
flag jumps at 10:00 AM and 1:45 PM on September 29, 2008. Similarly, the jump
detection test should identify jumps at 2:45 PM and 3:00 PM on May 6, 2008.
Moreover, it should not identify jumps when there were none. Recall that in order
to detect a jump, the ratio LMi−Cn
Snmust exceed the threshold ε, which is equal
to 6.91.12 Starting with September 29, 2008, our jump test-statistic takes value
8.52 and 45.66 at 10:00 AM and 1:45 PM, respectively. On May 6, 2010, the jump
statistic is equal to 114.50 and 36.76 at 2:45 PM and 3:00 PM, respectively. These
values suggest that Lee and Mykland’s methodology correctly identifies all jumps
12To arrive at this figure, recall that the confidence level x is 99.9%. Plugging this value inEquation (7), we obtain − log(− log(99.9%)) = 6.91.
9
on both days. Moreover, the jump detection test does not flag other jumps on both
days, thus providing us with the necessary confidence to pursue our investigation.
B. Unconditional Analysis
Having illustrated our approach with the examples discussed above, we now extend
the analysis to all trading days. Table 2 summarizes our main findings. It comprises
of 4 parts. From top to bottom, we report information pertaining to the sample of
observations, the jump intensity, the jump returns and the absolute jump returns.
How often does the stock market jump? The second part of Table 2 sheds light
on this question. “# Jumps” reveals that the S&P 500 jumps on 238 occasions
between 2008 and 2012. To better appreciate this result, we compute the likelihood
of jumps as the total number of jumps (238) divided by the total number of
observations (106,205). This ratio indicates that only 0.22% of returns are jumps,
lending more credence to the notion that jumps are rare events.
How large are jump returns? To answer this question, we turn to the entries
reported in the third part of Table 2. Standing at −0.15%, the average jump
return is negative and economically large.13 It is statistically significant, too.14
Interestingly, jump returns vary within a wide range from −4.41% to 3.83%. These
bounds are particularly revealing. For instance, the lower boundary (−4.41%)
implies that the “flash crash” of May 6, 2010 is not the sharpest drop experienced
by the S&P 500.15 The higher order moments of jump returns are also noteworthy.
Jump returns display little skewness (−0.20) and a modest level of kurtosis (5.50).
Collectively, the higher order moments provide some support to the assumption of
normally distributed jump returns implicit in standard time-series models of equity
prices (Merton, 1976; Bates, 1996; Eraker et al., 2003; Broadie et al., 2007).
To better appreciate the magnitude of jumps, we proceed in two steps. First,
13Recall the evidence of Table 1, which indicates that the average return in our entire sampleis negligible.
14We also repeat the above analysis using the median (rather than the mean) and obtainvirtually identical results. These are not reported for brevity.
15As Section IV shows, the largest intraday drop in our sample period occurs on October 29,2008, when the FED unveils swap lines worth up to $120 billion with other central banks.
10
we analyze the absolute value of jump returns. Analyzing the absolute (rather than
the signed) value of jump returns allows us to control for the offsetting effects of
positive and negative jumps. Second, we compute the relative size ratio, which tells
us how the mean absolute jump return compares to the average absolute return
observed in our sample:
Rel Size =
∑m
k=1|Jk|
m∑n
i=1|ri|
n
(8)
where Jk denotes the jump return k. ri denotes the return i. Overall, we observe m
jumps and n returns. Third, following Huang and Tauchen (2005) and Barndorff-
Nielsen and Shephard (2006), we estimate the contribution of jumps to the total
variation of the S&P 500. Specifically, we sum all squared jump returns, which we
divide by the sum of all squared returns included in our sample:
Jump Ratio =
∑m
k=1J2k
∑n
i=1r2i
(9)
where all variables are as previously defined.
We find that the average absolute jump return is economically large: 0.90%.
The relative size ratio reported in the penultimate entry of Table 2 shows that the
mean absolute jump return is more than 9 times larger than the average absolute
return, indicating that jumps are high impact events. Furthermore, the jump ratio
indicates that jumps account for 9.35% of the total variance of the market index.16
In short, jumps are low probability, high impact events. Although they occur
only 0.22% of the time, they have large effects on the stock market. Indeed, our
results indicate that these rare events collectively account for more than 9% of the
total variance of the S&P 500.
16This is somewhat higher than the figure reported by Huang and Tauchen (2005), who study adifferent time period than ours. The difference in results is most likely due to our more turbulentsample period.
11
C. Jump Asymmetries
We sharpen our analysis by distinguishing between crashes and surges. We define
“crashes” as negative jump returns. Conversely, “surges” correspond to positive
jump returns. Organizing jumps along this dimension helps us to pin down the
source(s) of the negative average jump return, which may originate from asymmetries
in jump intensity, jump returns, or both.
We begin by plotting the histogram of jump returns (see Figure 3). This plot
allows us to quickly ascertain which of the positive and negative jumps occur more
frequently. We notice that a greater mass of the histogram lies in the negative
domain, implying that most jumps are negative. We then turn our attention to
Table 3, which analyzes crashes and surges separately. The first part of this table
deals with the intensity of jumps. It shows that 96 and 142 (of the 238) jumps are
positive and negative, respectively. In other words, 60% of jumps correspond to
crashes, corroborating the visual evidence of Figure 3. The second part of Table
3 relates to the absolute value of jump returns.17 The average absolute return of
surges (0.92%) is roughly similar to that of crashes (0.88%), implying that there is
very little to distinguish between the magnitudes of crashes and surges.
Overall, this analysis reveals that the negative unconditional jump return is
mainly due to the fact that crashes occur more frequently than surges.
IV Events that Make Traders Jump
Beyond the identification and statistical analysis of jumps, it is fundamentally
important to go a step further and understand their economic drivers. This section
explores the events that ex-post likely cause jumps. We begin by presenting our
simple and intuitive approach that allows us to accurately connect news and jumps.
We then discuss our main findings.
17Note that, since we examine positive and negative jumps separately, there are no offsettingeffects of positive and negative jumps. The upshot of this is that we do not need to distinguishbetween jump returns and absolute jump returns as in Table 2. As a result, we report only thesummary statistics of absolute jump returns (see Table 3).
12
A. Methodology
A growing literature analyzes the effects of scheduled announcements on jumps in
asset prices. Recent contributions along these lines include Andersen et al. (2007),
Dungey et al. (2009) and Evans (2011), who typically relate less than 30% of jumps
to scheduled news.
While focusing on scheduled news greatly simplifies the empirical work, there
are severe challenges inherent to this approach.18 First, several economic data are
regularly released to market participants and it is not immediately obvious which
economic series should be included in the analysis. Hence, selecting the set of
economic variables to analyze is very often a subjective choice. Second, it is very
likely that the “relevant” economic variables change across asset classes and market
conditions. For instance, consumer confidence is likely to be more important during
recessions than expansions. As a result, one needs to tailor the list of economic
series to different market conditions and financial assets. Alas, it is not entirely clear
how to do this in an objective manner. Third, as economies become increasingly
interconnected, it is unclear whether one should focus on US or both US and foreign
scheduled announcements. Fourth, and most important, focusing exclusively on
scheduled announcements implies that unscheduled events are of minor importance.
This is somewhat counterintuitive since unscheduled events are exactly the type of
events that are most likely to truly “shock” market participants and trigger jumps.
The issues discussed above motivate us to propose an intuitive and general
procedure that enables us to identify the scheduled and, more importantly,
unscheduled events that trigger jumps. Our approach is simple. For each jump, we
manually sift through the news database FACTIVA, in order to identify important
news around the jump event.19 In using FACTIVA, we access all reports published in
English by four prominent news agencies, i.e. The Associated Press, The Dow Jones
Newswire, The New York Times and The Wall Street Journal. Accessing newswires
18The simplification mainly arises from the fact that scheduled macroeconomic announcementsare well-defined. Their timing is “precisely” known in advance, facilitating the scale of datacollection.
19Formerly known as the Dow Jones Interactive Service, FACTIVA is considered as the premierdatabase for news. It has been used in several recent studies, e.g. Bradley et al. (2014).
13
from all four agencies allows us to observe the public signals available to market
participants in real-time. This is crucial in order to precisely identify the information
shocks that ex-post likely cause jumps. We focus specifically on two hourly news
briefings published by FACTIVA: “Top Equity Stories” and “Top Economic Stories”.
As their titles suggest, these briefings contain the most important news stories up
to the time of publication.20 For each jump, we retain only the top news stories first
reported in the four-period time window centered around the jump time, i.e. in the
window [−30 min.; +30 min.].21,22
Two issues deserve further discussion. First, we need to be mindful of stale news.
It may be that the identified news report merely quotes another report published
earlier. Although recent studies such as Gilbert et al. (2012) document that investors
significantly react to stale news, we discard these old news events in order to make
our analysis robust to data-mining concerns. For each “relevant” news event, we
manually search the database to identify the earliest time when the information was
first released. If this time-stamp does not fall within our four-period time window,
20It is important to emphasize that focusing on these news briefings allows us to address theconcern that we may spuriously associate minor news to jumps. Since “Top Equity Stories” and“Top Economic Stories” summarize the most important events at the time of publication, it isunlikely that our news events are irrelevant for market participants.
21Some news events are first broadcasted on TV or via other audiovisual channels. It is onlyafter that newswires are written. Congressional testimonies by key decision-makers are a primeexample of such events. Since most market participants have access to the live video/audio feeds,it appears important to account for this in our analysis. Thus, we analyze news released up to twoperiods after the jump event.
22Similar to Boudoukh et al. (2013) and Dzielinski and Hasseltoft (2013), one could alsoimplement a full-fledged data-mining approach such as textual analysis. A successful applicationof this approach will most likely result in a stronger association between news and jumps than thatreported in this study. The reason for this is that textual analysis can be used to conduct a broadersearch of newswires beyond those provided by The Associated Press, The Dow Jones Newswire,The New York Times and The Wall Street Journal. However, a pure textual analysis approachis fraught with difficulties. For example, it is not entirely clear what key words should be usedfor the systematic search. Although, it is tempting to use words such as “GDP” and “Inflation”,doing so implies that one expects these scheduled events (and nothing else) to trigger jumps. Asa result, the textual analysis approach will miss all news stories that are related to importantevents other than the search words. Since theory is relatively silent on the events that are likelytrigger jumps, it is crucial to remain agnostic in the analysis. An interesting way to achieve thisconsists in searching for “Top Equity Stories” or “Top Economic Stories”, as we do. However,by doing so one runs the risk that the software will highlight instances where “Top”, “Economic”and “Stories” appear either as separate words or as a bag-of-words, yielding too many “false” hits.For example, the software might mistakenly flag a news report that includes the expression “topeconomic advisor” as a valid news story because the first two words, i.e. top and economic, areincluded in our key words.
14
we discard the report. As a result, we only retain “new” (as opposed to stale) news.
The second issue arises from market closures such as weekends. For all jumps that
occur during the first 15 minutes following the opening of the market, we consider
only the newswires released during the break.23
B. What Events Trigger Jumps?
The online Appendix lists all jumps and their causative events. The first column
shows the jump returns. The second column displays the jump times. The third
column reports the days when jumps occur. The last column briefly summarizes the
events that trigger jumps.
We find that the stock market jumps in response to both scheduled
macroeconomic announcements and unscheduled news. The causative scheduled
macroeconomic announcements mainly originate from the US. Three macroeconomic
series are closely associated to jumps: non-farm payroll, consumer confidence and
construction spending. Our analysis reveals that 32.73%, 18.18% and 18.18% of
non-farm payrolls, consumer confidence and construction spending announcements
cause jumps, respectively.
More importantly, we find that the S&P 500 reacts strongly to unscheduled news
such as unconventional policy, political and regulatory news. The unconventional
policy events encompass the extraordinary measures taken by governments to fight
the longest and deepest recession of the past two decades. These measures include
unscheduled fiscal and monetary policy decisions, coordinated central banks actions
and financial rescue packages. The effect of such decisions is best illustrated by the
4.41% crash on October 29, 2008 (see Figure 2), when the FED unveils swap lines
worth up to $120 billion with Brazil, Mexico, South Korea and Singapore. Another
example is the 1.21% fall in the S&P 500 on February 10, 2009 (11:15 AM), in
response to the Obama administration’s unveiling of its $1 trillion plan to combat
23These jumps represent a small fraction, more precisely 13%, of our sample of jump events.
15
the economic downturn.24
Political news, e.g. speeches by political leaders, moves the S&P 500 too.25
At 1:45 PM on September 29, 2008, the S&P 500 jumps by −3.16% as the US
Congress rejects the financial rescue bill.26 Intriguingly, the stock market surges
3.28%, on November 21, 2011, following reports of nominations for roles within the
US government. Overall, these examples tie in well with studies on the interplay
between politics and the stock market. Recent theoretical works by Pastor and
Veronesi (2012) and Pastor and Veronesi (2013) show that government policies can
trigger jumps in equity prices.27 We provide direct empirical evidence in support of
some of their theoretical predictions.
We can see that regulatory news also triggers jumps in the S&P 500. For
instance, the stock market consecutively jumps at 10:45 AM (−0.32%) on April 16,
2010, when the Securities Exchange Commission (SEC) charges Goldman Sachs and
its staff with defrauding investors about mortgage backed securities.
Overall, we are able to associate 186 (out of 238) jumps with news events. In
other words, we are able to relate almost 80% of jumps to information shocks. This
result is in stark contrast with the 30% typically reported in the literature, e.g.
Evans (2011). We attribute our superior results to our more general framework,
which allows us to uncover the effect of not only scheduled but also unscheduled
events.
24Our list of events could be used to study the impact of unconventional monetary policies onasset prices at the intraday level. It complements the work of Aıt-Sahalia et al. (2012), who identifyimportant unconventional policies based on the amount of daily press coverage they receive. Weidentify these unconventional policy actions at the intraday level, thus allowing for a much sharperanalysis of the response of asset prices at the intraday level. We leave this for future work.
25Although political speeches and congressional hearings are typically scheduled in advance,these events last for long time periods. This means that the timing of the information shock is not“precisely” known in advance. Hence, we include these events under the unscheduled set of news.
26The S&P 500 also jumps on October 3, 2008, when the US House of Representatives approvesthe financial rescue bill. See Veronesi and Zingales (2010) for an interesting discussion of the“Paulson’s Gift” with reference to the financial rescue package orchestrated by the then USTreasury Secretary Henry Paulson.
27The authors develop a general equilibrium model featuring policy and political uncertainty.Policy uncertainty refers to the impact of a specific policy on the profitability of firms. Agents learnabout this impact over time. Political uncertainty relates to the fact that governments may changetheir headline policy. The policy choice is guided by both political (likelihood to win elections) andeconomic (social welfare) considerations. Once a policy change occurs, investors gradually learnabout the impact of the new policy on the profitability of firms.
16
C. The Relative Importance of Scheduled v.s. Unscheduled
News
Having identified the events that trigger jumps in the equity market, we now conduct
an in-depth comparison of scheduled and unscheduled jumps.28 Undertaking this
analysis allows us to better appreciate the importance of distinguishing between
scheduled macroeconomic news and unscheduled events. We distinguish between
four categories: No News, Scheduled News, Unscheduled News and Mixed News.
The “No News” group comprises all jumps which we cannot connect to any
important information event. “Scheduled” jumps include all jumps related to
scheduled macroeconomic announcements, whose timing is “precisely” known ex
ante.29 “Unscheduled” jumps refer to jumps that are linked to news events other
than scheduled macroeconomic announcements. Finally, we group under “Mixed”
all instances where both scheduled and unscheduled news are related to the same
jump.
Table 4 displays the summary statistics of each category. Specifically, columns
3 through 6 relate to the No News, Scheduled, Unscheduled and Mixed categories,
respectively. This table contains four components. From top to bottom, the entries
relate to the jump returns, absolute jump returns, positive jump returns and negative
jump returns, respectively.
It is interesting to study the proportion of jumps linked to each category. “Prop”
divides the number of jumps associated with category [name in column] by the total
number of jumps in our sample (238). We can see that scheduled jumps account
for around 37% of jumps detected during our sample period. The percentage of
news-matched jumps soars to almost 80%, when we include unscheduled news in
our analysis.30 This result underscores the importance of analyzing both scheduled
and unscheduled news and helps explain the weak relationship between (scheduled)
28With a slight abuse of terminology, we denote by scheduled and unscheduled jumps, all jumpsthat are related to scheduled and unscheduled news events, respectively.
29Note that even though the timing of a scheduled macroeconomic announcement is “precisely”known in advance, its content is not known to market participants in advance.
30To see this, we add together the “Prop” of scheduled (36.97%), unscheduled (32.35%) andmixed (8.82%).
17
news and jumps documented in earlier studies.
The average jump return is negative for both scheduled and unscheduled jumps.
We can see that unscheduled jumps display an average jump return (−0.19%) that
is 4 times larger than that of scheduled jumps. Similarly, the absolute unscheduled
jump return (1.10%) is on average 50% larger than that of scheduled jumps. These
findings are intuitive. Scheduled events are highly anticipated, making them less
likely to move the market by a large amount. In contrast, the unanticipated nature
of unscheduled news makes such an event more likely to truly “shock” the market,
thus triggering large movements in the S&P 500.
Following this reasoning, we also expect unscheduled jumps to account for most
of the jump risk of the stock market.31 To verify this, we compute the contribution of
each category of jumps to the total jump variation of the S&P 500. Specifically, we
divide the sum of squared jump returns related to category [name in column] by the
sum of all squared jump returns observed in our sample. If our intuition is correct,
we expect to see a high jump ratio for the unscheduled events. The figures reported
in the row “Cont” confirm our intuition. Scheduled jumps contribute around 21%
to the total jump variation of the S&P 500. Strikingly, unscheduled jumps alone
account for over 51% of the total jump variation of the market index, demonstrating
that they trigger the largest movements in the stock market.
In summary, this section presents several novel and intuitive results. First, we
uncover a much stronger relationship between news and jumps than that reported
in existing studies. Close to 80% of jumps occur around scheduled and, more
importantly, unscheduled news events. Second, unscheduled jumps play a pivotal
role. They account for an important proportion of jumps in the market portfolio
and trigger the biggest changes in the stock market, which is in accordance with the
intuition that unanticipated events truly shock market participants.
31This conjecture flows directly from the bigger size of unscheduled jumps, which feeds backinto their contribution to the jump variation of the S&P 500.
18
V Why Do Some Announcements Trigger
Jumps?
The previous section documents that jumps occur around some scheduled news
announcements. Given that these announcements convey similar information to
the market at regular time intervals, it is natural to wonder: why do some
announcements trigger jumps in the S&P 500, while other similar announcements
do not? Consider the monthly non-farm payroll data releases as an example.
The announcements of January 2008 and March 2008 trigger jumps. However,
the S&P 500 does not jump around the announcement of February 2008. Why
does the market jump in January and March but not in February? What are the
informational differences between the announcements that trigger jumps and those
that do not? These are the questions we investigate in this section.32
A. Hypotheses
It is often argued that “big” macroeconomic news leads to jumps. Therefore, we
test the following two commonly held beliefs.
H1: A very large proportion of jumps occurs around announcements with
large surprise levels This intuitive hypothesis draws on the large literature on
the effect of macroeconomic announcements on the stock market (Andersen et al.,
2003), which documents that high announcement surprises lead to high (absolute)
stock returns. Since jumps are large returns, it is natural to conjecture that a
substantially large proportion of jumps occurs around announcements with the
largest surprises.
32Ideally, one would aim to answer similar questions for unscheduled events too. However, doingso is fraught with difficulties. For instance, such analysis would entail the difficult task of collectingall unscheduled news released in Factiva. This is computationally challenging. Furthermore,unscheduled events are so dissimilar from one another that such analysis may not be particularlyinformative.
19
H2: A very large proportion of jumps occurs around announcements made
at times of high uncertainty This hypothesis emerges from the literature
on heterogeneous beliefs (Shalen, 1993; Banerjee and Kremer, 2010), which shows
that higher uncertainty among agents leads to larger price movements once the
uncertainty is resolved. The intuition behind these findings is that agents are more
cautious during times of high uncertainty and delay their trading activity until the
public announcement is made. The announcement resolves the uncertainty and
agents resume their trading activity, resulting in potentially large price moves, i.e.
jumps.
B. Announcement Data
We focus on three prominent macroeconomic series, i.e. non-farm payroll, consumer
confidence and construction spending, that are closely associated with jump
events.33,34 We obtain economic forecasts data from Bloomberg.35,36 On average,
Bloomberg polls 84, 70 and 49 experts before each announcement of non-farm
payroll, consumer confidence and construction spending, respectively. We also
download vintage announcement data from Bloomberg. That is, we obtain the
initial (rather than revised) announcement data for each economic series. This is
important, since it aligns the econometrician’s information set with that of market
participants at the announcement time.37
33Non-farm payroll data are released at around 8:30 AM (ET). Consumer confidence andconstruction spending data releases occur at 10:00 AM (ET).
34As previously discussed, 33%, 18% and 18% of non-farm payroll, consumer confidence andconstruction spending announcements trigger jumps, respectively.
35The Bloomberg mnemonics for non-farm payroll, consumer confidence and constructionspending are “NFP TCH Index”, “CONCCONF Index” and “CONST TMOM Index”, respectively.
36We also investigate the robustness of our results to the source of economic forecasts. To thisend, we obtain economic forecast data from Reuters and repeat our main analysis. Section V.D.shows that the source of economic forecasts does not change our conclusions.
37See Christoffersen et al. (2002) and Aruoba (2008) for interesting discussions of the effects ofdata revisions on empirical studies.
20
C. Empirical Results
Announcement Surprises Equipped with our announcement dataset, we set
out to test the first hypothesis (H1). To achieve this goal, we compute a proxy for
surprise as:
St = |At − Et| (10)
where St, At and Et denote the surprise, announced data and median forecast at time
t, respectively.38,39 Since we are interested in the occurrence of jumps, irrespective
of their signs, we work with the absolute value of the difference between announced
and expected data.40
To investigate the economic relationship between announcement surprises and
the occurrence of jumps, we perform a simple and intuitive sorting exercise. For
each economic series, e.g. non-farm payroll, we sort all announcements into tercile
portfolios of announcement surprises: small, median and large. We then compute
the fraction of jumps that correspond with each tercile portfolio.41 We obtain this
number by counting the number of jumps related to a given portfolio, which we
divide by the total number of jumps across all three portfolios.42
Table 5 presents our results. It comprises three panels, A, B and C, reporting
the results for the consumer confidence, construction spending and non-farm payroll
series, respectively. The last column of each panel sheds light on the validity of H1.
From top to bottom, we present the fraction of jumps related to the small, median
and large surprise portfolios, respectively. If H1 is valid, we expect to see that an
overwhelming proportion of jumps corresponds with the large surprise portfolio.
38We also consider alternative surprise proxies. Section V.D. shows that our findings are robustto this.
39As a robustness check, we repeat this analysis with the average (rather than the median)forecast. The results of this exercise are virtually the same as those obtained when the median isused. These analyses are not reported for brevity.
40We repeat our analysis without the absolute value function and reach similar results.41As a further robustness check, we investigate whether our results depend on the overall stock
market conditions. To this end, we condition our analysis on the VIX level. Again, we obtainsimilar results. Section V.D. contains further details on this.
42Our sorting exercise is similar, in spirit, to the approach used by studies interested in predictingdefault events. Typically, these studies sort their data into portfolios based on an explanatoryvariable. They then analyze the proportion of defaults associated with each portfolio.
21
Starting with Panel A, the small, median and large portfolios contain 10%,
40% and 50% of jumps related to consumer confidence, respectively. Clearly, there
is very little to distinguish between the proportion of jumps associated with the
median and large surprise portfolios. This is difficult to reconcile with H1, which
predicts a substantially large proportion of jumps in the largest surprise portfolio.
Turning to construction spending (Panel B), we find that the proportion of jumps is
roughly the same across all surprise portfolios. Specifically, we observe 30%, 30%
and 40% for the small, median and large surprise portfolios, respectively. These
figures confirm that a “theory” of jumps that relies on announcement surprises as a
trigger mechanism cannot successfully explain the observed pattern of jumps.
Overall, there is little to distinguish between the proportion of jumps in the
median and large surprise portfolios. Clearly, this finding is difficult to reconcile
with the conventional wisdom that announcement surprises can explain why some
announcements create jumps while other similar announcements do not.43
The Role of Uncertainty We proxy uncertainty by the monthly cross-sectional
standard deviation of individual forecasts:
Ut = σ(Ft) (11)
where Ut and Ft denote the uncertainty, and a vector of all forecasts at time t,
respectively.
We sort all announcements of a specific series into three portfolios of
uncertainty: low, median and high. The last row of each panel of Table 5 presents
our results. Moving from left to right, we report the proportion of jumps for the low,
median and high uncertainty portfolios, respectively. To fix ideas, if H2 holds, we
expect to see that a substantial proportion of jumps falls into the highest uncertainty
portfolio.
43One may wonder if the results also hold when we combine all macroeconomic series andanalyze them jointly rather than separately. In Section V.D., we introduce a new surprise proxythat controls for the unit of measurements of different series, thus allowing us to pool together alleconomic series and analyze them jointly. We reach very similar conclusions.
22
We empirically observe that the high uncertainty portfolio does not account for
the largest fraction of jumps. For example, 28%, 39% and 33% of jumps linked
to non-farm payroll news fall in the low, median and high uncertainty portfolios,
respectively. The results for consumer confidence and construction spending also
paint a similar picture, suggesting that the prediction of H2 is not borne out in the
data.
The analysis above may be criticized on the grounds that it does not control
for the level of surprise. In particular, one may argue that, holding the level of
surprise constant, announcements made at times of high uncertainty may be more
closely associated with jumps. To investigate this, we double-sort announcements
first on surprise (three portfolios) and then on uncertainty (three portfolios). In
doing so, we obtain 9 portfolios. Analyzing the large portfolio of surprises in
Panel A, the penultimate row reveals that 30%, 20% and 0% of jumps related
to consumer confidence correspond with the low, median and high uncertainty
portfolios, respectively. The negative relationship between the uncertainty level and
the proportion of jumps run contrary to the intuition underpinning H2. Turning
to the large surprise portfolio of both construction spending and non-farm payroll,
we observe a non-monotonic relationship between uncertainty and the proportion
of jumps. Clearly, these findings are very difficult to reconcile with the second
hypothesis (H2): there is little evidence to suggest that scheduled announcements
released during periods of high uncertainty trigger the majority of jumps.
Overall, our results challenge the commonly held views that announcement
surprises and the level of uncertainty can convincingly explain why some
announcements trigger jumps while other similar news does not.
D. Robustness Checks
In the following, we conduct a variety of robustness tests to validate our findings.
First, one may argue that the effect of a large surprise differs during periods of
good and bad economic conditions. We use the volatility index (VIX) to proxy
for these conditions and repeat our sorting exercise. Second, we consider several
23
alternative proxies for surprises and repeat our analysis. Finally, one may argue
that the economic forecasts provided by Bloomberg do not represent the views of
market participants. To address this concern, we turn to an alternative data provider
(Reuters).
The Role of Stock Market Conditions For each series, we sort announcements
on surprises and also on the VIX level, which we download from Bloomberg. We
introduce a new measure for surprise S2t :
S2t =
∣
∣
∣
∣
At −Et
Ut
∣
∣
∣
∣
(12)
where S2t , At, Et and Ut denote the surprise, announced data, (median) expectation
and uncertainty at time t, respectively. This proxy serves two purposes. First, it
jointly captures the effects of the classical announcement surprise (Equation (10))
and the level of uncertainty (Equation (11)) in a simple and parsimonious way.
Second, this proxy is unit-free, allowing us to pool together all three economic
series, i.e. non-farm payrolls, consumer confidence and construction spending, and
analyze them jointly.
Table 6 shows our findings. It contains four panels. Panels A through C deal
with consumer confidence, construction spending and non-farm payroll, respectively.
Panel D presents results for the pooled analysis. The layout of the individual panels
resembles that of Table 5. Looking at Panel D, the last column shows that the
small, median and large surprise portfolios account for 18%, 42% and 39% of jumps,
respectively. These findings echo the conclusions of our baseline analyses: there is
very little evidence to suggest that announcements characterized by a high degree of
surprises cause the largest proportion of jumps. The last row of Panel D also reveals
a relatively flat relationship between stock market conditions and the proportion of
jumps. This result indicates that stock market conditions do not change our main
conclusions. Panels A–C paint a similar picture for individual macroeconomic series.
24
Surprise Proxies Although inspired by the large literature on macroeconomic
announcements and stock returns, our analysis can be criticized on the grounds
that the baseline proxy for surprises (see Equation (10)) focuses on the absolute
(rather than the relative) level of announcement surprises. To better understand this
point, consider the following two scenarios. On the one hand, the market expects an
announcement of 10% and the released figure is 11%. On the other hand, the market
expects 1% and the released figure is 2%. In both cases, the absolute deviation
between the expected and announced figures is 1 percentage point. However the
absolute relative surprise, which we compute by dividing the absolute deviation by
the absolute value of the market consensus, paints a different picture. The deviation
represents 10 and 100 percentage points of the market consensus in the first and
second cases, respectively. The relative surprise proxy suggests that investors are
more surprised in the latter case than the former. It is possible that this subtle
argument drives our results.
To tackle this concern, we consider another proxy for surprises:
S3t =
∣
∣
∣
∣
At −Et
Et
∣
∣
∣
∣
(13)
where S3t , At and Et denote the surprise proxy, announced data and (median)
expectation at time t, respectively. Obviously, we cannot compute this ratio when
the denominator, i.e. the median forecast, takes the value zero. We discard these
instances.44
The results for consumer confidence (Panel A) and construction spending (Panel
B) of Table 7 are remarkably similar to those based on the baseline surprise proxy
(see Table 5), suggesting that our main findings are robust to this new proxy.45
Table 8 repeats the above analysis using a fourth proxy for surprises, S4t , which
scales the baseline proxy for surprise by the previously announced figure (rather
44This occurs on two instances, one for the consumer confidence and another for constructionspending series.
45We also pool observations across economic series and repeat the analysis. Our conclusions areunchanged.
25
than the contemporaneous market consensus):
S4t =
∣
∣
∣
∣
At −Et
At−1
∣
∣
∣
∣
(14)
where S4t , At and Et denote the surprise, announced value and (median) expectation
at time t, respectively. Using this proxy yields results that are similar to our
baseline findings, suggesting that the findings are very robust to the computation
of announcement surprises.
Different News Source One may argue that the true expectation of the market
differs somewhat from the consensus data obtained from Bloomberg. It may be that
agents do not rely on Bloomberg but rather follow a different news provider such as
Reuters.
We repeat our analysis based on Reuters analysts estimates. To do this, we
download the median, high and low forecasts from Thomson Reuters.46 We filter
out erroneous entries. In particular, we discard announcements where the lowest
and highest forecasts are equal and yet the median is different from either value.
Additionally, we expunge instances where the lowest forecast is higher than the
highest forecast.
Tables 9 through 12 repeat our previous analysis based on the economic
forecasts downloaded from Reuters.47 The results of this exercise are equally difficult
to reconcile with the two hypotheses, suggesting that the source of economic forecasts
has little bearing on our findings.
46The Reuters tickers for the non-farm payroll, consumer confidence and construction spendingare “USMEMPALO”, “USMCNFCOQ” and “USMNEWCOB”, respectively. We obtain the “high”and “low” forecasts by replacing the letter “M”, short for median, immediately after the prefix“US”, by “H” and “L”, respectively.
47Reuters does not provide us with individual forecast data. Hence, we cannot compute ourusual proxy for uncertainty, i.e. the cross-sectional standard deviation of forecasts. We use thespread between the highest and lowest forecasts as a proxy for dispersion.
26
VI Conclusion
This paper studies the dynamics of jumps in the S&P 500 between 2008 and 2012.
Using high frequency transaction data, we implement the non-parametric jump
detection test of Lee and Mykland (2008). We find that jumps are low probability,
high impact events. Although they occur with a low probability (0.22%), the average
absolute jump return is 9 times larger than the mean absolute return observed in
our sample.
We then investigate the events that trigger jumps. To this end, we present a
general approach that consists in searching an extensive database of newswires to
identify the scheduled and, more importantly, unscheduled events that “shock” the
market. Contrary to earlier studies, which focus only on scheduled news, we are able
to relate around 80% of jumps to important news events. The causative news events
include not only scheduled macroeconomic announcements but also unscheduled
events such as unconventional policy announcements and political events. Our
analysis highlights the influence of unscheduled events. Indeed, unscheduled news is
closely associated with the biggest jumps in the S&P 500. Compared to scheduled
news, they account for a much larger share of the jump variation of the stock index,
indicating that they truly surprise market participants.
Finally, we test the commonly held views that announcement surprises and
the level of uncertainty can explain why some announcements trigger jumps in the
equity market, while other similar announcements do not. We use a comprehensive
survey dataset to construct various empirical proxies for announcement surprises
and uncertainty levels. Double-sorting announcements first on surprise and then on
uncertainty, we find that there is very little evidence in support of these “theories” of
jumps. Overall, our results challenge common beliefs regarding jumps and indicate
that the relationship between news and jumps may be more complex than previously
thought.
27
References
Aıt-Sahalia, Y., Andritzky, J., Jobst, A., Nowak, S., and Tamirisa, N. (2012).
Market response to policy initiatives during the global financial crisis. Journal of
International Economics, 87(1):162–177.
Andersen, T., Bollerslev, T., and Diebold, F. (2007). Roughing it up: Including jump
components in the measurement, modeling, and forecasting of return volatility.
Review of Economics and Statistics, 89(4):701–720.
Andersen, T. G., Bollerslev, T., Diebold, F. X., and Vega, C. (2003). Micro effects of
macro announcements: Real-time price discovery in foreign exchange. American
Economic Review, 93(1):38–62.
Aruoba, S. B. (2008). Data revisions are not well behaved. Journal of Money, Credit
and Banking, 40(2-3):319–340.
Bajgrowicz, P., Scaillet, O., and Treccani, A. (2013). Jumps in high-frequency data:
Spurious detections, dynamics, and news. Swiss Finance Institute Research Paper.
Banerjee, S. and Kremer, I. (2010). Disagreement and learning: Dynamic patterns
of trade. Journal of Finance, 65(4):1269–1302.
Barndorff-Nielsen, O. and Shephard, N. (2004). Power and bipower variation with
stochastic volatility and jumps. Journal of Financial Econometrics, 2(1):1–37.
Barndorff-Nielsen, O. and Shephard, N. (2006). Econometrics of testing for jumps in
financial economics using bipower variation. Journal of Financial Econometrics,
4(1):1–30.
Bates, D. (1996). Jumps and stochastic volatility: Exchange rate processes implicit
in Deutsche Mark options. Review of Financial Studies, 9(1):69–107.
Boudoukh, J., Feldman, R., Kogan, S., and Richardson, M. (2013). Which news
moves stock prices? A textual analysis. WFA Meetings Paper.
28
Bradley, D., Clarke, J., Lee, S., and Ornthanalai, C. (2014). Are analysts’
recommendations informative? Intraday evidence on the impact of time stamp
delays. Journal of Finance, 69(2):645–673.
Broadie, M., Chernov, M., and Johannes, M. (2007). Model specification and risk
premia: Evidence from futures options. Journal of Finance, 62(3):1453–1490.
Christoffersen, P., Ghysels, E., and Swanson, N. R. (2002). Let’s get “real” about
using economic data. Journal of Empirical Finance, 9(3):343–360.
Cornell, B. (2013). What moves stock prices: Another look. Journal of Portfolio
Management, 39(3):32–38.
Cutler, D. M., Poterba, J. M., and Summers, L. H. (1989). What moves stock
prices? Journal of Portfolio Management, 15(3):34–12.
Dumitru, A. and Urga, G. (2012). Identifying jumps in financial assets: A
comparison between nonparametric jump tests. Journal of Business & Economic
Statistics, 30(2):242–255.
Dungey, M., McKenzie, M., and Smith, L. (2009). Empirical evidence on jumps in
the term structure of the US Treasury market. Journal of Empirical Finance,
16(3):430–445.
Dzielinski, M. and Hasseltoft, H. (2013). Why do investors disagree? The role of a
dispersed news flow. AEA Meetings Paper.
Eraker, B. (2004). Do stock prices and volatility jump? Reconciling evidence from
spot and option prices. Journal of Finance, 59(3):1367–1404.
Eraker, B., Johannes, M., and Polson, N. (2003). The impact of jumps in volatility
and returns. Journal of Finance, 58(3):1269–1300.
Evans, K. (2011). Intraday jumps and US macroeconomic news announcements.
Journal of Banking & Finance, 35(10):2511–2527.
29
Gilbert, T., Kogan, S., Lochstoer, L., and Ozyildirim, A. (2012). Investor inattention
and the market impact of summary statistics. Management Science, 58(2):336–
350.
Huang, X. and Tauchen, G. (2005). The relative contribution of jumps to total price
variance. Journal of Financial Econometrics, 3(4):456–499.
Jiang, G. and Oomen, R. (2008). Testing for jumps when asset prices are observed
with noise–A “swap variance” approach. Journal of Econometrics, 144(2):352–
370.
Johannes, M. (2005). The statistical and economic role of jumps in continuous-time
interest rate models. Journal of Finance, 59(1):227–260.
Lee, S. and Mykland, P. (2008). Jumps in financial markets: A new nonparametric
test and jump dynamics. Review of Financial Studies, 21(6):2535–2563.
Merton, R. C. (1976). Option pricing when underlying stock returns are
discontinuous. Journal of Financial Economics, 3(1):125–144.
Pan, J. (2002). The jump-risk premia implicit in options: evidence from an
integrated time-series study. Journal of Financial Economics, 63(1):3–50.
Pastor, L. and Veronesi, P. (2012). Uncertainty about government policy and stock
prices. Journal of Finance, 67(4):1219–1264.
Pastor, L. and Veronesi, P. (2013). Political uncertainty and risk premia. Journal
of Financial Economics, 110(3):520–545.
Piazzesi, M. (2005). Bond yields and the federal reserve. Journal of Political
Economy, 113(2):311–344.
Pukthuanthong, K. and Roll, R. (2011). Internationally correlated jumps. AFA
Working Paper.
Roll, R. (1988). R2. Journal of Finance, 43(2):541–566.
30
Schneider, P., Sogner, L., and Veza, T. (2010). The economic role of jumps and
recovery rates in the market for corporate default risk. Journal of Financial and
Quantitative Analysis, 45(6):1517–1547.
Shalen, C. T. (1993). Volume, volatility, and the dispersion of beliefs. Review of
Financial Studies, 6(2):405–434.
Veronesi, P. and Zingales, L. (2010). Paulson’s gift. Journal of Financial Economics,
97(3):339–368.
31
2008 2009 2010 2011 2012600
700
800
900
1000
1100
1200
1300
1400
1500
Year
Pric
e(I
ndex
Poi
nts)
Time−Series of S&P 500
Figure 1: Time-Series of the S&P 500 Index
This figure displays the dynamics of the stock market between January 2008 and July 2012.
The graph shows the price-series during our sample period. All data are sampled at the 15-min
frequency.
32
0 2 4 6 8 10 12 14 16 181110
1120
1130
1140
1150
1160
1170
1180
1190
1200
1210
Time
Pric
e(I
ndex
Poi
nts)
S&P 500
0 2 4 6 8 10 12 14 16 181080
1090
1100
1110
1120
1130
1140
1150
1160
1170
Time
Pric
e(I
ndex
Poi
nts)
S&P 500
Figure 2: Examples of Jumps
This figure shows the intraday price of the S&P 500 on two distinct trading days. The top graph
shows the dynamics of the S&P 500 on September 29, 2008. Around 10:00 AM, the FED announces
an increase in the size of its Term Auction Facility. At 1:45 PM, the US Congress rejects the $700
billion financial rescue bill. The bottom graph represents the dynamics of the stock market on May
6, 2010, when the “flash crash” occurs between 2:30 PM and 3:00 PM.
33
−5 −4 −3 −2 −1 0 1 2 3 40
10
20
30
40
50
60
70
Returns
Num
ber
of J
umps
S&P 500
Figure 3: Histogram of Jump Returns
This figure depicts the histogram of S&P 500 jump returns for the period 2008–2012. The horizontal
axis shows jump returns, expressed as a percentage (%). The vertical axis indicates the number of
jumps. We implement the non-parametric test of Lee and Mykland (2008) to identify jumps. We
use the 15-min sampling frequency and a significance level of 0.1%.
34
Table 1: Summary Statistics of Returns
This table presents summary statistics of 15-min log returns (expressed in %) on the S&P 500
between 2008 and 2012. “Nobs” shows the total number of 15-min returns in our sample. “Mean”,
“Min”, “Max” and “Std Dev” report the average, minimum, maximum and standard deviation of
returns, respectively. “Skew” and “Kurt” denote the skewness and kurtosis, respectively.
Nobs Mean Min Max Std Dev Skew Kurt
S&P 500 106,205 −9.97× 10−6 -4.41 3.83 0.18 -0.06 47.98
Table 2: Summary Statistics of Jumps
This table summarizes the dynamics of jumps in the S&P 500. We compute 15-min percentage
returns between 2008 and 2012 and identify jumps following the method of Lee and Mykland (2008).
We use a significance level of 0.1%. “# Obs” is the total number of returns in our sample. “#
Jumps” reports the number of jumps. “Prop Jump Returns” reports the number of jumps as a
fraction of the total number of observations in our sample. “Mean”, “Std”, “Min”, “Max”, “Skew”
and “Kurt” are the average, standard deviation, minimum, maximum, skewness and kurtosis,
respectively. “Rel Size” is the relative size ratio computed as the average absolute jump return
divided by the average of all absolute returns in the sample (see Equation (8)). Finally, “Jump
Ratio” captures the contribution of jumps to the total risk of the asset (see Equation (9)).
S&P 500
Sample # Obs 106,205
Jump Intensity# Jumps 238Prop Jump Returns (%) 0.22
Jump Returns
Mean -0.15Std 1.13Min -4.41Max 3.83Skew -0.20Kurt 5.46
Abs Jump Returns
Mean 0.90Std 0.70Min 0.19Max 4.41Skew 2.74Kurt 11.62Rel Size 9.13Jump Ratio(%) 9.35
35
Table 3: Jump Asymmetries
This table presents the summary statistics of positive and negative jumps in the S&P 500. Our
analysis builds on 15-min percentage returns observed between 2008 and 2012. We identify jumps
following the method of Lee and Mykland (2008). We use a significance level of 0.1%. “# Jumps”
reports the number of jumps. “Prop Jump Returns” reports the number of jumps as a fraction of the
total number of observations in our sample. “Mean”, “Std”, “Min”, “Max”, “Skew” and “Kurt”
denote the average, standard deviation, minimum, maximum, skewness and kurtosis, respectively.
“Rel Size” is the relative size ratio computed as the average absolute jump return divided by the
average of all absolute returns in the sample (see Equation (8)). Finally, “Jump Ratio” captures
the contribution of jumps to the total risk of the asset (see Equation (9)).
Positive Negative
Jump Intensity# Jumps 96 142Prop Jump Returns (%) 0.09 0.13
Abs Jump Returns
Mean 0.92 0.88Std 0.66 0.73Min 0.27 0.19Max 3.83 4.41Skew 2.37 2.91Kurt 9.34 12.53Rel Size 9.38 8.96Jump Ratio (%) 3.71 5.64
36
Table 4: Scheduled v.s. Unscheduled NewsThis table reports the summary statistics of jumps for each of the following four categories: no news, scheduled macro, unscheduled and mixed.
“No News” includes all jumps that we cannot tie to news. “Scheduled Macro” collects all jumps related to US and international scheduled
macroeconomic announcements. “Unscheduled” contains jumps that are linked to events other than scheduled macroeconomic news. Finally,
“Mixed” encompasses all jumps that are linked to both scheduled and unscheduled news. “Mean”, “Std”, “Min”, “Max”, “Skew” and “Kurt”
denote the average, standard deviation, minimum, maximum, skewness and kurtosis, respectively. “Prop” refers to the ratio of the number of
jumps in category [name in column] over the total number of jumps across all categories. “Cont” refers to the sum of squared jump returns from
category [name in column] over the sum of all squared jump returns (across categories). “Prop Pos” (“Prop Neg”) shows the number of surges
(crashes) in category [name in column] as a proportion of the total number of jumps that belong to category [name in column]. To detect jumps,
we implement the non-parametric test of Lee and Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
No News Scheduled Macro Unscheduled Mixed
Jump Returns
Mean -0.40 -0.05 -0.19 0.12Std 1.10 0.86 1.42 0.95Min -4.16 -2.16 -4.41 -0.96Max 2.64 3.83 3.45 1.72Skew -0.32 0.95 -0.36 0.34Kurt 5.09 6.15 4.34 1.58
Abs Jump Returns
Mean 0.93 0.72 1.10 0.86Std 0.70 0.47 0.92 0.38Min 0.27 0.19 0.32 0.40Max 4.16 3.83 4.41 1.72Skew 2.58 3.82 1.97 0.93Kurt 10.93 24.07 6.47 2.87Prop (%) 21.85 36.97 32.35 8.82Cont (%) 22.53 20.87 50.68 5.92
Positive Jump Returns
Mean 0.86 0.78 1.09 1.02Std 0.58 0.58 0.79 0.46Min 0.27 0.29 0.38 0.48Max 2.64 3.83 3.45 1.72Skew 1.86 3.98 1.64 0.26Kurt 6.45 21.35 5.01 1.50Prop Pos (%) 30.77 43.18 41.56 47.62
Negative Jump Returns
Mean -0.96 -0.67 -1.10 -0.70Std 0.75 0.36 1.01 0.21Min -4.16 -2.16 -4.41 -0.96Max -0.29 -0.19 -0.32 -0.40Skew -2.63 -1.81 -2.03 0.11Kurt 10.80 7.36 6.47 1.73Prop Neg (%) 69.23 56.82 58.44 52.38
37
Table 5: Surprises, Uncertainty and Jumps (Bloomberg)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Bloomberg. For each announcement, we measure the announcement surprise St as:
St = |At − Et|
where St, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the cross-sectional standard
deviation of all forecasts. We sort all announcements of a specific economic series into portfolios
and report the fraction of jumps that corresponds to individual portfolios. We obtain this number
by counting the number of jumps related to a given portfolio, which we divide by the total number
of jumps across all portfolios. The last column of each panel shows the proportion of jumps
associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 0% 10% 10%Median S 10% 10% 20% 40%Large S 30% 20% 0% 50%Uncertainty 50% 20% 30%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 0% 10% 20% 30%Median S 0% 10% 20% 30%Large S 10% 20% 10% 40%Uncertainty 0% 60% 40%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 6% 0% 0% 6%Median S 17% 6% 22% 44%Large S 17% 22% 11% 50%Uncertainty 28% 39% 33%
38
Table 6: Market Conditions, surprises and Jumps (Bloomberg)
This table studies the relationship between stock market conditions, information surprises and
jumps. We obtain non-farm payroll, consumer confidence and construction spending data between
2008 and 2012 from Bloomberg. We also retrieve the VIX data series from Bloomberg. For each
announcement, we measure the announcement surprise S2
tas:
S2
t=
|At − Et|
Ut
where S2
t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We compute the uncertainty proxy, Ut, as the cross-sectional standard
deviation of all forecasts. We sort all announcements of a specific economic series into portfolios
and report the fraction of jumps that corresponds to individual portfolios. We obtain this number
by counting the number of jumps related to a given portfolio, which we divide by the total number
of jumps across all portfolios. The last column of each panel shows the proportion of jumps
associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows the
proportion of jumps linked to each of the 3 portfolios based on the VIX. The remaining figures
indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We first sort
announcements on surprise and then on the level of the VIX. Panels A, B and C show the results
pertaining to consumer confidence, construction spending and non-farm payroll announcements,
respectively. Panel D shows the results when we pool together announcements of all three economic
series. To detect jumps, we implement the non-parametric test of Lee and Mykland (2008) using
the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low VIX Median VIX High VIX Total
Small S 0% 0% 10% 10%Median S 10% 20% 10% 40%Large S 10% 30% 10% 50%VIX 10% 50% 40%
Panel B: Construction Spending
Low VIX Median VIX High VIX Total
Small S 10% 0% 20% 30%Median S 20% 0% 10% 30%Large S 10% 10% 20% 40%VIX 40% 10% 50%
Panel C: Non-Farm Payroll
Low VIX Median VIX High VIX Total
Small S 0% 6% 0% 6%Median S 17% 28% 0% 44%Large S 22% 11% 17% 50%VIX 39% 44% 17%
Panel D: Pooled
Low VIX Median VIX High VIX Total
Small S 8% 5% 5% 18%Median S 16% 16% 11% 42%Large S 13% 13% 13% 39%VIX 32% 39% 29%
39
Table 7: Alternative Surprises, Uncertainty and Jumps (Bloomberg)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Bloomberg. For each announcement, we measure the announcement surprise S3
tas:
S3
t=
∣
∣
∣
∣
At − Et
Et
∣
∣
∣
∣
where S3t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the cross-sectional standard
deviation of all forecasts. We sort all announcements of a specific economic series into portfolios
and report the fraction of jumps that corresponds to individual portfolios. We obtain this number
by counting the number of jumps related to a given portfolio, which we divide by the total number
of jumps across all portfolios. The last column of each panel shows the proportion of jumps
associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 0% 0% 0%Median S 20% 0% 30% 50%Large S 30% 20% 0% 50%Uncertainty 50% 20% 30%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 0% 11% 11% 22%Median S 0% 11% 22% 33%Large S 0% 33% 11% 44%Uncertainty 0% 56% 44%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 6% 0% 6% 11%Median S 0% 17% 17% 33%Large S 28% 11% 17% 56%Uncertainty 28% 39% 33%
40
Table 8: Alternative Surprises, Uncertainty and Jumps II (Bloomberg)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Bloomberg. For each announcement, we measure the announcement surprise S4
tas:
S4
t=
∣
∣
∣
∣
At − Et
At−1
∣
∣
∣
∣
where S4t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the cross-sectional standard
deviation of all forecasts. We sort all announcements of a specific economic series into portfolios
and report the fraction of jumps that corresponds to individual portfolios. We obtain this number
by counting the number of jumps related to a given portfolio, which we divide by the total number
of jumps across all portfolios. The last column of each panel shows the proportion of jumps
associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 0% 0% 0%Median S 20% 0% 30% 50%Large S 30% 20% 0% 50%Uncertainty 50% 20% 30%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 0% 20% 10% 30%Median S 0% 40% 0% 40%Large S 0% 10% 20% 30%Uncertainty 0% 60% 40%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 0% 0% 6% 6%Median S 12% 18% 29% 59%Large S 18% 12% 6% 35%Uncertainty 24% 41% 35%
41
Table 9: Surprises, Uncertainty and Jumps (Reuters)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Thomson Reuters. For each announcement, we measure the announcement surprise St
as:
St = |At − Et|
where St, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the spread between
the highest and lowest forecasts. We sort all announcements of a specific economic series into
portfolios and report the fraction of jumps that corresponds to individual portfolios. We obtain
this number by counting the number of jumps related to a given portfolio, which we divide by the
total number of jumps across all portfolios. The last column of each panel shows the proportion of
jumps associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 11% 11% 22%Median S 11% 22% 11% 44%Large S 22% 0% 11% 33%Uncertainty 22% 44% 33%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 11% 11% 11% 33%Median S 11% 11% 11% 33%Large S 22% 11% 0% 33%Uncertainty 44% 33% 22%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 12% 18% 18% 47%Median S 12% 12% 6% 29%Large S 6% 6% 12% 24%Uncertainty 35% 29% 35%
42
Table 10: Market Conditions, Surprises and Jumps (Reuters)
This table studies the relationship between stock market conditions, information surprises,
uncertainty and jumps. We obtain non-farm payroll, consumer confidence and construction
spending data between 2008 and 2012 from Reuters. We also retrieve the VIX data series from
Bloomberg. For each announcement, we measure the announcement surprise S2
tas:
S2
t=
|At − Et|
Ut
where S2
t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. Ut denotes the level of uncertainty, computed as the difference between
the highest and lowest forecasts. We sort all announcements of a specific economic series into
portfolios and report the fraction of jumps that corresponds to individual portfolios. We obtain
this number by counting the number of jumps related to a given portfolio, which we divide by the
total number of jumps across all portfolios. The last column of each panel shows the proportion of
jumps associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on the VIX. The remaining figures
indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We first sort
announcements on surprise and then on the level of the VIX. Panels A, B and C show the results
pertaining to consumer confidence, construction spending and non-farm payroll announcements,
respectively. Panel D shows the results when we pool together announcements of all three economic
series. To detect jumps, we implement the non-parametric test of Lee and Mykland (2008) using
the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low VIX Median VIX High VIX Total
Small S 0% 11% 11% 22%Median S 22% 22% 0% 44%Large S 0% 0% 33% 33%VIX 33% 22% 44%
Panel B: Construction Spending
Low VIX Median VIX High VIX Total
Small S 22% 0% 22% 44%Median S 11% 0% 11% 22%Large S 22% 0% 11% 33%VIX 44% 11% 44%
Panel C: Non-Farm Payroll
Low VIX Median VIX High VIX Total
Small S 12% 24% 6% 41%Median S 18% 12% 0% 29%Large S 6% 18% 6% 29%VIX 41% 41% 18%
Panel D: Pooled
Low VIX Median VIX High VIX Total
Small S 17% 14% 11% 43%Median S 14% 9% 6% 29%Large S 9% 9% 11% 29%VIX 40% 29% 31%
43
Table 11: Alternative Surprises, Uncertainty and Jumps (Reuters)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Thomson Reuters. For each announcement, we measure the announcement surprise S3
t
as:
S3
t=
|At − Et|
Et
where S3
t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the spread between
the highest and lowest forecasts. We sort all announcements of a specific economic series into
portfolios and report the fraction of jumps that corresponds to individual portfolios. We obtain
this number by counting the number of jumps related to a given portfolio, which we divide by the
total number of jumps across all portfolios. The last column of each panel shows the proportion of
jumps associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 11% 11% 22%Median S 0% 33% 11% 44%Large S 22% 0% 11% 33%Uncertainty 22% 44% 33%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 0% 13% 13% 25%Median S 0% 0% 13% 13%Large S 38% 25% 0% 63%Uncertainty 38% 38% 25%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 6% 25% 6% 38%Median S 19% 6% 13% 38%Large S 13% 13% 0% 25%Uncertainty 38% 25% 38%
44
Table 12: Alternative Surprises, Uncertainty and Jumps II (Reuters)
This table studies the relationship between information surprises, uncertainty and jumps. We
obtain non-farm payroll, consumer confidence and construction spending data between 2008 and
2012 from Thomson Reuters. For each announcement, we measure the announcement surprise S4t
as:
S4
t=
|At − Et|
At−1
where S4t, At and Et denote the surprise, announced data and the median expected figure computed
at time t, respectively. We also compute a proxy for uncertainty, Ut, as the spread between
the highest and lowest forecasts. We sort all announcements of a specific economic series into
portfolios and report the fraction of jumps that corresponds to individual portfolios. We obtain
this number by counting the number of jumps related to a given portfolio, which we divide by the
total number of jumps across all portfolios. The last column of each panel shows the proportion of
jumps associated with each of the 3 surprise portfolios. Similarly, the last row of each panel shows
the proportion of jumps linked to each of the 3 portfolios based on uncertainty. The remaining
figures indicate the proportion of jumps associated with each of the 9 double-sorted portfolios. We
first sort announcements on surprise and then on the level of uncertainty. Panels A, B and C
show the results pertaining to consumer confidence, construction spending and non-farm payroll
announcements, respectively. To detect jumps, we implement the non-parametric test of Lee and
Mykland (2008) using the 15-min sampling frequency and a significance level of 0.1%.
Panel A: Consumer Confidence
Low U Median U High U Surprise
Small S 0% 11% 11% 22%Median S 11% 22% 11% 44%Large S 11% 11% 11% 33%Uncertainty 22% 44% 33%
Panel B: Construction Spending
Low U Median U High U Surprise
Small S 0% 11% 22% 33%Median S 0% 11% 0% 11%Large S 33% 22% 0% 56%Uncertainty 44% 33% 22%
Panel C: Non-Farm Payroll
Low U Median U High U Surprise
Small S 13% 13% 6% 31%Median S 6% 13% 6% 25%Large S 6% 19% 19% 44%Uncertainty 31% 31% 38%
45