when it cannot get better or worse: the asymmetric impact of …mbrandt/papers/published/... ·...
TRANSCRIPT
-
When It Cannot Get Better or Worse:
The Asymmetric Impact of Good and Bad News
on Bond Returns in Expansions and Recessions∗
Alessandro Beber Michael W. Brandt
Amsterdam Business School Fuqua School of BusinessUniversity of Amsterdam Duke University
and NBER
Abstract
We examine empirically the response of bond returns and their volatility togood and bad macroeconomic news in economic expansions and recessions. Wefind that the information content of macroeconomic announcements is mostimportant when it contains bad news for bond returns in expansions and, toa lesser extent, when it contains good news for bond returns in contractions. Inparticular, we observe the strongest bond market response to bad news in therelease of non-farm payrolls in expansions. During recessions, inflation news arerelatively more important when they contain good news. We also document thatmacroeconomic news impacts substantially the volatility of bond returns at allmaturities by increasing jump intensities and by altering the distribution of thejump size. While a large proportion of employment news results in a jump inexpansions at all maturities, inflation news becomes more important in recessionsand particularly at medium maturities.
∗We thank two anonymous referees, Andrew Ang, Jean Boivin, Jean-Pierre Danthine, Joost Driessen,Marc Giannoni, Pascal St-Amour, Josef Zechner (the editor) and seminar participants at Cass BusinessSchool, Columbia University, and University of Lausanne for helpful comments. We also thank Clara Vegaand Informa Global Markets for providing some of the announcement data.
-
1. Introduction
Investors receive every day a wealth of information, ranging from news about corporate
cash flows to news about the economic fundamentals. Investors process this information
and update their expectations of the economy’s future growth, inflation, and discount rates,
thus affecting asset prices. However, some macroeconomic information is more relevant than
other, and the same information can have a different importance at different stages in the
business cycle. Moreover, the same information can have a different impact on different
assets and, even for the same asset, a different impact on returns than on volatility.
This paper studies these different aspects of information jointly. In particular, we examine
empirically the response of U.S. Treasury bond returns and their volatility to good and bad
macroeconomic news in economic expansions and recessions. We will uncover interesting
asymmetric effects when investors and monetary authorities are confronted with different
type of news at different times.
The design of our analysis is straightforward. We first study the effect of macroeconomic
announcements on bond returns using a regression specification that allows for an asymmetric
impact of good and bad macroeconomic news in different phases of the business cycle. We
find that the information content of the announcements is most important for bond returns
when it contains bad news for the bond market in expansions and, to a lesser extent, good
news for the bond market in contractions. We observe the strongest bond market response
to bad news in the release of non-farm payrolls during expansionary periods. This effect is
apparent at all maturities. During recessionary periods, inflation news are relatively more
important when they contain good news and particularly at short maturities.
Motivated by the strong economic significance of these results, we then formulate and
estimate a statistical model of bond returns incorporating jumps that depend explicitly
on the timing and information content of scheduled macroeconomic news. The estimates
of this model show that macroeconomic news has a substantial impact on bond volatility
by increasing jump intensities and by affecting the distributions of the jump size. The
relation between macroeconomic news and jumps, however, depends on the combination
of the announcement type, the maturity of the bond, and the phase of the business cycle.
Specifically, the proportion of releases resulting in a jump decreases dramatically for short
maturities and is generally lower in recessions. Jumps on announcement days are associated
predominantly with non-farm payroll releases, but inflation news becomes relatively more
associated with jumps in recessions and at medium maturities.
The estimation of the model of bond returns with jumps confirms the findings from the
regression analysis that the response of the term structure to the macroeconomic release
1
-
depends on the multifaceted aspects of information. More importantly, given the rich
structure of the model that we hypothesize, the estimates provide a detailed account of
jump dynamics as the conduit through which macroeconomic information enters the bond
market.
A number of papers investigate the effect of macroeconomic announcements on
financial markets along different lines. Fleming and Remolona (1999) study the effect of
macroeconomic news on bond returns, while Jones, Lamont and Lumsdaine (1998) document
the effect on bond volatility.1 However, they do not condition their analysis on different
phases of the business cycle and they do not disentangle the effect of good versus bad news.
Recent work by Boyd, Hu and Jagannathan (2005) investigates the response of stock and
bond market returns to civilian unemployment rate news in expansions and recessions, but
they do not consider the release of non-farm payrolls and, most importantly, they do not
separate the effect of good versus bad news.2 Andersen, Bollerslev, Diebold and Vega (2003)
show that the responses of foreign exchange rates to macroeconomic news vary with the
sign of the news. However, their sample period is all in an expansion phase of the business
cycle. The contribution of our paper is thus to show that the effect of macroeconomic
announcements on bond market returns and volatility depends jointly on the type of news,
on the sign of the news, and on the phase of the business cycle.
Recent theoretical work describes the asymmetric response of stock prices to news in
good or bad times (e.g., Veronesi, 1999). The driving mechanism is the agents’ uncertainty
about the state of the economy and the associated learning process. Specifically, investors
believe that the economy follows a two-state regime-switching process, with low and high
states corresponding to recessions and expansions. Agents solve a signal extraction problem
to determine the probability of being in the high state. Suppose that investors believe
that the high state almost surely prevails. Then if bad news arrives, two things happen:
expected future asset values decrease, and second, state risk increases. Risk-averse investors
require additional returns for bearing this additional risk; hence they require an additional
discount on the asset price, which drops by more than it would in a present-value model.
In the bond market, the response to macroeconomic news is peculiar for the key role of
the monetary authority that has direct effects on short-term yields and some influence on
1Other papers document the effect of macroeconomic news on bond returns or volatility for differentsample periods, announcement sets, underlying assets. Recent work by Christiansen and Ranaldo (2007)and Brenner, Pasquariello, and Subrahmanyam (2006) focus instead on the effect of macroeconomic newson comovements of stock and bond markets.
2Andersen, Bollerslev, Diebold and Vega (2007) study a more comprehensive set of macroeconomicannouncements with respect to Boyd, Hu, and Jagannathan (2005), but their sample period includes onlyone short recession. Furthermore, they also do not separate the effect of good versus bad news.
2
-
long-term yields via inflation expectations. We can think, for example, of an asymmetric
Fed reaction function by which, at the beginning of an expansion with low rates, positive
economic news raise the likelihood of future increases in the target rate by more than bad
news raise the likelihood of future decreases in the target rate. In this setting, the response
of bond prices to macroeconomic news depends critically on the content of the news and on
the phase of the business cycle.
The paper proceeds as follows. In Section 2, we describe the announcements and bond
market data. In Section 3, we present the results of our regression specification. In Section
4, we describe and estimate the jump model. Section 5 concludes.
2. Data
Our data has a number of key features that allow us to address our research questions.
First, it covers a long sample period including several expansions and recessions. Second,
it contains detailed information on the timing and content of macroeconomic news. Lastly,
our bond market data is observed at a daily frequency, so that we can precisely relate asset
prices to macroeconomic releases. This section describes the announcement data, the data
characterizing the stages of the business cycle, and the U.S. Treasury bond market data.
2.1. Survey and Announcement Data
We obtain data on the dates, release times, actual released figures, and median forecasts for
five of the most important U.S. macroeconomic information releases from Money Market
Services (MMS) and Informa Global Markets, covering the period from February 1980
through December 2003. MMS conducts a survey of about 40 money market managers
on the Friday of the week before the release of each economic indicator. MMS reports the
median forecast from the survey, which is made available to the market and the business
press immediately after the survey is taken. The availability of a direct measure of the
expected release – the median forecast – allows us to avoid potential biases of using time-
series econometric models to infer market expectations.
We consider a set of four key announcements describing inflation dynamics by the
consumer price index (CPI) and the producer price index (PPI) and labor market
dynamics by the civilian unemployment rate (CUR) and non-farm payrolls (NFP).3 These
3In a previous version of this paper, we have also considered the announcements of the Federal OpenMarket Committee federal funds target rate (FOMC) to describe the conditions of the money market.However, we have not found any significant result and thus we have decided to leave it out. The FOMCcould be irrelevant in our setting because of the potential inaccuracy of the MMS forecast of FOMC withrespect to market based forecasts (e.g., Bernanke and Kuttner (2005) use the prices of the Federal FundsFutures). Another reason could be that changes of the target rate occur sometimes outside the scheduledeight meetings per year.
3
-
announcements are released widely and virtually instantaneously at a precise scheduled
time. The statistical agencies impose lock-up conditions to ensure that the information
is not released to the public before the scheduled time (see Fleming and Remolona, 1999).
These announcements are all timed at 8:30am ET and are monthly. A majority of the
announcements occur on a Friday and the employment report (CUR and NFP) is normally
the first government information release concerning economic activity in a given month.
Panel A of Table 1 describes in more detail the macroeconomic announcements in our sample.
Several studies have examined the accuracy of the MMS forecasts. We find strong
evidence that the MMS median forecast has predictive ability for the actual release. We
also find that the median forecast is usually an unbiased predictor. We observe only one
case in which the median forecast is not an efficient predictor. This happens whenever the
prediction of the CPI, which is always announced after the PPI, is collected before the PPI
actual release. In these few cases, the market learns about inflation from the PPI release and
the median forecast of the CPI should take this into account.4 We thus obtain the expected
CPI from a regression on the median MMS CPI forecast and the actual PPI release.
2.2. Expansions and Recessions
We measure expansions and recessions using the experimental coincident recession index
(XRI-C) constructed by Stock and Watson (1989) and available for all our sample period.
The XRI-C is a monthly estimate of the probability that the economy was in a recession in
that month, constructed using four series of leading indicators such as industrial production,
real personal income less transfer payments, real manufacturing and trade sales, and total
employee-hours in non-agricultural establishments. Several papers in the literature used
XRI-C to measure the state of the economy (e.g., Boyd, Hu, and Jagannathan (2005); Hong,
Tourous, Valkanov (2007)).
An important feature of the XRI-C is that it establishes a real-time coincident public
forecasting record. It is thus different from other indexes constructed by Stock and Watson,
such as the experimental leading index (XLI) and the experimental recession index (XRI),
which are based on leading indicators and forecast the growth of the economy in the following
six months. Furthermore, the XRI-C uses only information that is publicly available
at a certain point in time. In contrast, the NBER business cycle dates for expansions
and recessions make use of information that becomes available later. Stock and Watson
discontinued the publication of the index after the end of our sample, but the Federal
Reserve Bank of Chicago now publishes a monthly expansion index, the Chicago Fed National
4We find that in our sample period the actual PPI has strong predictive ability for the actual CPI andthus PPI information is relevant in determining CPI predictions.
4
-
Activity Index (CFNAI), which is based on an extension of the methodology used to construct
the original XRI-C. Positive (negative) values of the CFNAI index indicate that the economy
is expanding above (below) its historical growth trend. For our sample period, the XRI-C
and the CFNAI index have a large correlation equal to −0.84.Since we aim to describe the different response of the bond markets to macroeconomic
news according to the perceived state of the economy, we use the XRI-C in our empirical
analysis. The correlation between a dummy variable reflecting the NBER business cycle
dating and the XRI-C is equal to 0.78 during the period 1980–2003. It is also interesting to
note that the economy was in a recession phase, the XRI-C indicator was above 50%, for
about 15% of the time during our sample period 1980–2003.
2.3. Futures Data
We focus on the futures market as opposed to the cash market because futures contracts
are typically used to speculate on events such as macroeconomic news and because futures
market prices are not affected by the liquidity biases of the cash market (e.g., the on-the-run
versus off-the-run spread).5 We obtain daily prices of a set of U.S. Treasury bond futures
that require delivery of bonds with different maturities. The contracts mature in March,
June, September, and December. More specifically, we obtain data for the 30-year Treasury
bond futures (from February 1980 through December 2003), for the 10-year Treasury note
futures (from June 1982 through December 2003), for the 5-year Treasury note futures (from
May 1988 through December 2003), and for the 2-year Treasury note futures (from June 1990
through December 2003) traded at the CBOT. We also obtain data for the 90-day Treasury
bill futures (from February 1980 to August 2003), traded at the CME.
In Table 1, Panel B and C, we show summary statistics for daily returns computed on
the settlement prices of all these futures contracts.
3. Regression Model
In this section, we study the effect on bond returns of good and bad macroeconomic news
in expansions and recessions using a regression model. This is the simplest approach to
investigate whether there is any news effect and to what extent this effect is asymmetric
along two dimensions (different type and different times of the news). The regression model
is unsophisticated and does not suggest a specific mechanism for macroeconomic news to be
impounded into bond prices.
3.1. Methodology
5Other papers studying the effect of macroeconomic news on the bond market have used futures marketprices for the same reasons. See, among others, Ederington and Lee (1993), Bollerslev, Cai and Song (2000),Andersen, Bollerslev, Diebold, and Vega (2007).
5
-
To gauge the extent to which an announcement contains new information, we construct the
following standardized measure of surprise:
Skt =Akt −Xkt
σk, (1)
where Akt is the value of the main statistic released in announcement k at time t, Xkt
denotes the corresponding median survey forecast, and σk is the (unconditional) empirical
standard deviation of the innovations Akt − Xkt. Standardizing the surprise by σk allowsus to compare the regression coefficients across different announcement types. We allow
different announcement surprises to have different effects on the bond returns and we also
allow these effects to be different in expansion versus recession phases. We thus estimate for
each announcement type the following regression:
rt,t−1 = αk + βexp,k (1−XRIt) Sk + βrec,kXRItSk + ek, (2)
where rt,t−1 represents the day-to-day return for each of the Treasury bond futures and XRItis the probability of being in a recession phase at time t as indicated by the value of the
recession index XRI-C.6
We further examine whether the announcement effects vary with the sign of the surprise.
For this, we generalize equation (2) by allowing for different slope coefficients depending on
the good and bad news as follows:
rt,t−1 = αk + βGexp,k (1−XRIt) SktGkt + βBexp,k (1−XRIt) SkBkt+ βGrec,kXRItSktGkt + βBrec,kXRItSkBkt + ekt,
(3)
where Gkt = 1 and Bkt = 0 if the information released in announcement k at time t is
good news (for the Treasury market) and Gkt = 0 and Bkt = 1 otherwise. Whether an
announcement is good or bad news for the bond market is ultimately an empirical question.
However, to facilitate interpretation and consistency of our results, we define as good (bad)
news all negative (positive) surprises, as defined in equation (1). It is generally the case in
our sample that negative (positive) surprises generate positive (negative) bond returns.
3.2. Results
Table 2 presents the results of estimating equation (2) on bond returns at five different
maturities. The importance of allowing for different effects of macroeconomic news in
6For the announcements considered in this paper, CUR and NFP are always released jointly in theEmployment Report. In this case, we include both the CUR and the NFP surprise in the regression ofequation (2) to isolate the marginal effect of each announcement type.
6
-
expansions and recessions is apparent. Macroeconomic news matters more in expansions, as
the statistical significance of the coefficients readily shows. The explanatory power of this
model with respect to a non state-dependent model is in all cases substantially higher, with
an average increase in the R2 ranging between 19% for the short-end of the yield curve to a
31% increase for the long-end (results not tabulated). In most cases, a statistical test rejects
the restricted model with no state-dependence at least at the 10% significance level.
This asymmetric effect is striking also economically. For the release of NFP, which is
generally significant both in good and bad times, the effect of a surprise in expansions is on
average 30% larger than in recessions.
Another general result in Table 2 is the negative signs of all the statistically significant
coefficients, with the exception of the coefficient on the civilian unemployment rate (CUR).
The release of a higher than expected CPI, NFP, PPI or a lower than expected CUR are all
bad news for the bond market, since they lead to negative bond returns. These effects are
most relevant for the release of NFP.
In particular, our results that employment report surprises matter for bond returns in
recessions are in contrast with the non-significant findings of Boyd, Hu, and Jagannathan
(2005). This discrepancy could be due to several reasons. First, our two sample periods
do not coincide. We study the period 1980 through 2003, whereas Boyd et al. (2005) look
at 1972 through 2000. About one third of their sample, including the 1973-1975 recession,
is affected by the U.S. monetary policy of the pre-Volker period, which has been shown
to respond less strongly to inflation expectations (e.g., Boivin and Giannoni, 2002) and
to depart from Taylor-type rules (see, Primiceri, 2005). Second, Boyd et al. (2005) focus
on the release of CUR and ignore the contemporaneous announcement of NFP. The NFP
release is considered the most relevant of all scheduled macroeconomic announcement for
bond returns by a number of papers (e.g., Balduzzi, Elton, Green, 2001). Furthermore, since
the release of the employment report occurs at the same time, failing to control for all the
news component can affect the results.7 Finally, we measure the expected outcome of the
macroeconomic release as the median analyst forecast from MMS, as it is now standard in
the announcement literature. Boyd et al. (2005) use instead an econometric model, which
is typically estimated using data that were not available in real-time. In contrast, survey
median forecasts are widely available before the release. Furthermore, survey forecasts seem
to be more efficient and less biased than econometric models (e.g., Ang et al., 2007), at least
for inflation forecasting.
7We rerun our analysis ignoring NFP to have an indication of the importance of this effect for our sample.We find, like Boyd et al. (2005), that the release of the CUR has no significant effect in recessions for bondreturns of the US, TY, FV and TU. We do find a significant effect, however, for TB returns.
7
-
The results of Table 2 do not allow to disentangle the effect of good and bad news
in expansions and recessions. We may have reasons to believe, however, in an asymmetric
response of bond prices in good and bad times. For example, we could think of an asymmetric
Fed reaction function by which, at the beginning of an expansion with low rates, positive
economic news raise the likelihood of future increases in the target rate by more than bad
news raise the likelihood of future decreases in the target rate.
For this purpose, we estimate the richer specification of equation (3) and present the
results in Table 3. We observe that the explanatory power of this specification basically
doubles for the release of the CPI and slightly increases for the employment report and the
PPI. Accordingly, the restriction that the coefficients on good and bad news are the same in
expansions and in recessions is always rejected at least at the 5% level for the CPI, is never
rejected for the PPI, and is rejected in most cases for the release of the employment report.
However, when the NFP is considered in isolation (separately from the CUR), we always
reject the restriction of equal coefficients for good and bad news.
The overwhelming evidence in Table 3 is that bad news is what matters in good times
and, to a slightly lesser extent, good news is what matters in bad times. When signals of
overheating of the economy materialize, because of higher than expected releases of CPI or
NFP with the economy in an expansion phase, bond returns show a negative response at all
maturities. In contrast, signals of a deteriorating economy in a recession, e.g., the release
of lower than expected NFP, determine positive bond returns. Bond returns also seem to
have a positive response to news of a decelerating economy in expansions arising from lower
than expected NFP or lower than expected PPI. Notice that we defined a negative surprise
as good news. As a result, all the estimated coefficients on good news are multiplied by
negative quantities and thus have an opposite effect on bond returns. Finally, news of an
accelerating economy in recessions do not seem to impact bond returns at any horizon.
The interpretation of these results and the comparison of the economic significance across
maturities is more meaningful if we translate the effect of a surprise on returns to the effect
of a surprise on yields. We map returns into yields using a simple duration adjustment for
each segment of the yield curve, using the standard formula for the modified duration:
dP
P= −DmdY (4)
dY = −dPP
/Dm (5)
where dP/P is the return on the underlying, Dm is the modified MacCauley duration, and dY
is the change in yields. We obtain the return on the underlying dP/P as the return generated
8
-
by a one-standard deviation announcement surprise. We then calculate the duration of each
Treasury futures contract as the modified duration of a deliverable bond with the average
maturity in the maturity range.8 Using the measure for bond returns suggested above and
the appropriate modified duration, we obtain a measure of the effect on yields.
In Figure 1, we use the formula in (4) and the estimates of Table 3 to plot the effect of
a one standard deviation surprise in CPI, NFP, CUR, and PPI on bond yields at different
maturities. The upper plot shows that in good times the effect on yields of bad news is
greatest at the shorter horizon and declines for longer maturities. The NFP, and the CPI
to a lesser extent, are the only announcements that, besides showing the strongest effect on
short-term yields, have economically significant effect on yields at longer maturities. The
central plot shows that the effect of good news in bad times is dominated by CPI and CUR
at the short-horizon, whereas only the release of NFP has consistent significant impacts on
longer yields. The bottom plot shows that the effect of good news in good times is weaker
in economic terms with respect to the previous cases and is concentrated in the short-end of
the yield curve.
We repeat all the analysis carried out in this section using two alternative definitions of
expansions and recessions. First, we use a dummy variable reflecting the NBER business
cycle dates, set equal to zero in expansions and equal to one otherwise. Although the NBER
dates are not an ideal indicator because the dates of the business cycle were not known
in real-time, it is still useful for comparison with some of the literature. Second, we use
the CFNAI expansion index, which is considered the natural successor of the XRI-C and is
constructed from a broader set of indicators. We find that using a NBER dummy does not
change the results. More specifically, we obtain the same significant coefficients across the
board, although the explanatory power is always lower with the NBER dummy. Moreover,
when a coefficient is not significant with XRI-C is also not significant with the NBER dummy.
When we use the CFNAI expansion index, more precisely (-CFNAI), we obtain the same
results of using XRI-C and a similar explanatory power.
4. State-Dependent Jump Model
In this section, we extend the analysis of the effects on bond returns of good and bad
macroeconomic news in expansions and recessions using a state-dependent jump model. This
model puts a detailed structure on the conduit through which macroeconomic information
enters the bond market. We start by specifying the model and then present the results of
the estimation.8We obtain the following modified durations for each Treasury futures: 11.82 for the 30-year Treasury
bond futures, 6.09 for the 10-year Treasury note futures, 3.93 for the 5-year Treasury note futures, 1.71 forthe 2-year Treasury note futures, 0.24 for the Treasury-bill futures.
9
-
4.1. Methodology
Our econometric specification for the volatility of bond returns builds on the model of Maheu
and McCurdy (2004) for individual stock returns. We model innovations to the daily log
return as the sum of a heteroskedastic shock associated with diffusive information flow and
a jump term capturing the effect of a sudden release of important news.9 However, the
Maheu and McCurdy’s autoregressive conditional jump intensity parametrization is not well
suited for the bond market, because anecdotal evidence shows that jumps tend to occur on
macroeconomic announcement days and these information releases are not autocorrelated.
Therefore, we explicitly allow the jump arrival rate to increase on announcement days. We
also let the jump size mean on announcement days to depend on announcement surprises.
We start by specifying the log futures return as the sum of a time-varying conditional
mean, a normal innovation ²1,t+1 representing diffusive information flow, and a jump
innovation ²2,t+1 representing the effect of sudden information arrival:
xt+1 = ln Ft+1 − ln Ft = µt+1 + ²1,t+1 + ²2,t+1. (6)
We assume that both innovations have a conditional mean of zero and are contemporaneously
independent. Furthermore, the first innovation is distributed:
²1,t+1 = σt+1zt+1, where zt+1 ∼ N[0, 1]. (7)
The second innovation is governed by a Poisson jump distribution with a time-varying
conditional jump intensity λt. That is, the conditional probability of observing nt+1 = j
jumps between dates t and t + 1 is:
Prob[nt+1 = j
∣∣Φt]
=exp (−λt+1) λjt+1
j!, for j = 1, 2, . . . . (8)
The jump innovation can then be expressed as:
²2,t+1 =
nt+1∑
k=1
Yt+1,k − θt+1λt+1, (9)
where Yt+1,k represents the size of the kth jump, if it occurs, drawn independently from a
normal distribution with time-varying conditional mean θt+1 and constant volatility δ.
9Some other papers in the literature have used econometric specifications with time-varying jumpintensities. For example, Bekaert and Gray (1998) and Neely (1999) both model probability of realignmentsfor target zones exchange rates as a function of interest rate differentials and other economic variables. Das(2002) allows the jump intensity of Fed fund rates to vary with the day of the week or with FOMC meetings.
10
-
4.1.1 Conditional Variance
The conditional variance of returns is also divided into two components. First, we assume
that the volatility of the normal innovation ²1,t+1 follows a standard GARCH process:
σ2t+1 = ω + α²2t + βσ
2t , (10)
where ²t denotes the sum of the two return innovations ²1,t + ²2,t. This specification allows
both return innovations to affect future volatility through the coefficient α and for volatility
shocks to be persistent through the coefficient β. Second, the conditional volatility of the
jump innovation ²2,t+1 is governed by the jump process and is given by:
var[²2,t+1
∣∣Φt+1]
= λt+1(θ2t+1 + δ
2), (11)
where the information set Φt+1 contains all the macroeconomic information up to time t + 1
but not the financial information, i.e. the bond return innovation ²2. The contribution of
jumps to the total conditional variance is time-varying with the conditional intensity λt+1
and the square of the jump-size mean θ2t+1.
The conditional variance function of equation (10) constrains the normal and the jump
innovations to affect future volatility in the same way through the parameter α. However,
the two components of ²t could have in general a different effect. Since we cannot perfectly
disentangle the two innovations from observed returns, we use a proxy for the importance
of the jump innovation. We thus estimate the ex-post expected number of jumps occurred
during period t and allow this estimate to affect the feedback of past innovations on expected
future volatility:
σ2t+1 = ω +(α + αjE
[nt
∣∣Φt])
²2t + βσ2t , (12)
where E[nt
∣∣Φt]
is the ex-post estimate of the expected number of jumps occurred between
time t− 1 and t using period t information. We provide further details on the methodologyused to estimate the ex-post number of jumps in section 4.1.4 that describes the properties
of the distribution of returns implied by the model.
Given the focus of this paper on asymmetric effects of good versus bad news across the
business cycle, we also estimate richer specifications of the conditional variance function of
11
-
equation (12). More specifically, we model the conditional variance as follows:
σ2t+1 = ω + g(Φt
)²2t + βσ
2t (13)
g(Φt
)=
(1−XRIt
)(αexp + αexpj E
[nt
∣∣Φt]+ αexpa I (²t) + α
expaj I (²t) E
[nt
∣∣Φt])
+ (14)
XRIt(αrec + αrecj E
[nt
∣∣Φt]+ αreca I (²t) + α
recaj I (²t) E
[nt
∣∣Φt])
where E[nt
∣∣Φt]
is again the ex-post estimate of the expected number of jumps occurred
between time t− 1 and t using period t information, I (²t) is an indicator variable equal toone when ²t < 0, and XRIt is the recession index XRI-C at time t.
10 The function g(Φt
)
needs to be positive to have a well-specified GARCH process. However, this constraint is
never binding with the parameter estimates obtained in the empirical section of this paper.
Expected volatility responds to past innovations through the function g(.). This function
allows the feedback to be unconditionally different in expansions versus recessions (αexp and
αrec). The feedback can also be different depending on whether the observed innovation
has likely been a jump, through the additional effect of αj. Furthermore, observed negative
innovations also generate an additional response of volatility, through the parameters αa, in
case of normal innovations, and αaj, in case of expected jump. The jump and asymmetric
feedback effects are allowed to be different in expansion versus recession phases.
4.1.2 Effect of Macroeconomic News
The ultimate goal of our model is to capture changes in the futures returns in response to
macroeconomic announcements for different phases of the business cycle. To accomplish this,
we allow the macroeconomic announcements to affect the return dynamics in three distinct
ways. First, we let the mean log return µt be different in expansions and recessions:
µt+1 = µexpt+1(1−XRIt+1) + µrect+1XRIt+1, (15)
where, following our notation above, XRIt is the probability of being in a recession phase
at time t as indicated by the value of the recession index XRI-C.
The second and more prominent role of macroeconomic announcements is through the
conditional jump intensity. Consistent with the empirical fact that the majority of extreme
Treasury market moves have occurred on announcement days (e.g., Fleming and Remolona,
1999), we allow the jump intensity to increase on announcement days from λ0 to λ0 + λk,
where k denotes either the CPI release, the Employment Report (CUR and NFP) release,
10Maheu and McCurdy (2004) use a similar specification for the conditional variance of stocks and stockindices. However, they do not distinguish between expansion and recession phases.
12
-
or the PPI release. More concretely, we model the conditional jump intensity as:
λt+1 = λ0 +∑
k=cpi,er,ppi
λkDk,t+1, (16)
where Dk,t+1 is a dummy variable signaling whether the announcement k is scheduled between
dates t and t + 1. Notice that unless λ0 = 0, which the model accommodates, jumps can
occur on non-announcement days, as they do in the data. We do not expect the intensity
of the jump process to be different in expansion versus recession phases. Our intuition is
supported by a likelihood ratio test that shows no significant improvement for more general
versions of equation (16) (results not reported).
Finally, we model the conditional mean of the jump size distribution on announcement
days as a function of the new information released. To gauge the extent to which an
announcement contains new information, we use the standardized measure of surprise
constructed in equation (1). We allow different announcement surprises to have different
effects on the jump size mean and we also allow these effects to be different in expansion
versus recession phases:
θt+1 = (1−XRIt+1)(
θexp0 +∑
k=cpi,cur,nfp,ppi
θexpk Sk,t+1
)+
XRIt+1
(θrec0 +
∑
k=cpi,cur,nfp,ppi
θreck Sk,t+1
).
(17)
The intuition underlying this specification is that the sign and magnitude of the immediate
market reaction to the announcement depends on the news content. However, this link is far
from deterministic in the data, which leads us to model the mean jump size, as opposed to
the jump itself, as a function of the surprise. On non-announcement days, the mean jump
size is equal to θexp0 in expansions and θrec0 in recessions.
4.1.3 Asymmetric Effects of Macroeconomic News
Given the focus of the paper on asymmetric effects of macroeconomic news, we also estimate
our model with a more general version of equation (17) that allows good and bad news to
have different effects in expansions and recessions. More concretely, we express the mean of
13
-
the jump size as:
θt+1 = (1−XRIt+1)(
θexp0 +∑
k=cpi,cur,nfp,ppi
θexpGk Gk,t+1Sk,t+1 + θexpBk Bk,t+1Sk,t+1
)+
XRIt+1
(θrec0 +
∑
k=cpi,cur,nfp,ppi
θrecGkGk,t+1Sk,t+1 + θrecBkBk,t+1Sk,t+1
) (18)
where Gkt = 1 and Bkt = 0 if the information released in announcement k at time t is a
negative surprise, and Gkt = 0 and Bkt = 1 otherwise, similarly to the equation (3) for bond
returns.
We allow the mean of the jump size to be different than zero during non-announcement
days and when the actual release is equal to the median forecast.11 Furthermore, the intercept
will absorb the effects of macroeconomic news on the jump size mean that are not related
to the surprise.
4.1.4 Distribution of Returns
Conditional on j jumps occurring, our model implies the following distribution of log returns:
f (xt+1|nt+1 = j, Φt) =1√
2π(σ2t+1 + jδ
2) exp
(−(xt+1 − µt+1 + θt+1λt+1 − θt+1j)
2
2(σ2t+1 + jδ
2)
).
(19)
The unconditional distribution is then obtained by forming an expectation over the number
of jumps for each date:
f (xt+1|Φt) =∞∑
j=0
f (xt+1|nt+1 = j, Φt) Prob (nt+1 = j|Φt) . (20)
Finally, the likelihood function is the product of the T unconditional log return distributions.
Note that the likelihood function involves infinite summations over the number of possible
jumps. In practice, we truncate the summations at 20. It turns out that, for our parameter
estimates, the conditional Poisson distribution has zero probability for more than ten jumps.
We can use the unconditional distribution of returns implied by our model to construct
an ex-post probability distribution for the number of jumps occurred at t+1, conditional on
11For all the model specifications, we use a likelihood-ratio test to reject the restriction of a zero interceptin the equation for the jump size mean. In all cases, we reject the restriction and, consistently, we find thatthe intercept is statistically different from zero, either in expansions, or in recessions, or in both.
14
-
the information at time t + 1:
Prob (nt+1 = j|Φt+1) = f (xt+1|nt+1 = j, Φt) Prob (nt+1 = j|Φt)f (xt+1|Φt) . (21)
The probability that at least one jump has occurred at time t+1 can then easily be obtained
as the complement to one of the probability of having had no jumps at all:
Prob (nt+1 ≥ 1|Φt+1) = 1− Prob (nt+1 = 0|Φt+1) . (22)
4.2. Results
We estimate the state-dependent jump model by standard maximum likelihood using daily
data on the returns of Treasury bond futures for contracts with different maturities.
Table 4 shows the estimates of the model where we allow for the effect of macroeconomic
news on the jump dynamics in expansions and contractions, retaining however the simpler
specifications for the conditional variance in equation (12) and for the mean of the jump size
in equation (17).12 The results on the conditional variance parameters are intriguing. The
feedback of the current innovation and volatility on future conditional volatility is similar
at all maturities, except for the shortest, where current innovations are relatively more
important. When we allow the feedback of the innovation on future volatility to depend
on the ex-post probability of a jump, we infer that jump innovations are not persistent.
More specifically, we observe that the feedback coefficient α is almost completely offset by
αj when the observed innovation was very likely to be a jump. This pattern holds for all
the maturities. Our evidence is consistent with the results of Jones, Lamont and Lumsdaine
(1998), but it is more general, since we identify jump innovations endogenously rather than
imposing that only announcement day innovations can have a different feedback on future
volatility.
The conditional variance of the estimated state-dependent jump model leads to
interesting volatility dynamics for different segments of the yield curve. In Figure 2, we plot
the conditional variance implied by the model for the 30-year Treasury bond futures and for
the 5-year Treasury-note futures over the recent period 2001 to 2003. The total conditional
variance is the sum of a GARCH component, as in equation (12), and a jump component, as
in equation (11). The plot illustrates well that the GARCH component describes the smooth
changes in daily volatility, whereas jumps take care of abnormal episodes of volatility.
12The main results of the model are unaffected by the use of these simpler specifications. We return tothe most general version of the model later in the paper.
15
-
Table 4 also shows that in general the mean return µ is statistically positive both in
expansions and in contractions, but it is always more economically significant in recessions
for all maturities. This is evidence of higher risk premiums in contraction phases and is
consistent with previous results on counter-cyclical risk-aversion (e.g., Engle and Rosenberg,
2002; Gordon and St-Amour, 2004).
The significant results on the jump parameters confirm the importance of augmenting a
basic GARCH specification. The unconditional jump intensity λ0 is significantly positive for
all the maturities. This confirms our intuition that jumps can be triggered by macroeconomic
announcements that are not included in our sample or by other categories of events. However,
the macroeconomic announcements we consider play a leading role in determining the
jump dynamics. We obtain a significantly positive estimate of λk for the CPI release, the
Employment Report, and the PPI release, indicating that the jump intensity rises sharply
on these announcement days for all the maturities. The Employment Report is the most
important announcement in this respect. In contrast, the CPI release is usually the least
important in increasing the intensity of a jump. When we compare jump intensities across
maturities, we observe that the CPI is most relevant for the jump intensity of the five-year
futures, but it is not significant for the Treasury-bill. The employment report has equally
important effects for all maturities, less so for the two-year futures. Finally, the PPI release
affects less the intensity of a jump for shorter maturities, but it turns out to have always
a stronger impact than the CPI announcement. This last result is likely to depend on the
timing of the announcements, where the PPI is always released at least one day before the
CPI in our sample.
When we look at the parameters describing the jump magnitudes, we observe that,
both in expansions and recessions, the size of the jumps is drawn unconditionally from a
distribution with a negative mean, except for the shortest maturity in recessions, and a
volatility decreasing with shorter maturities. Again, the macroeconomic announcements
play an important role. In expansion phases of the business cycle, the estimates that are
significant are generally negative, that is positive (negative) surprises in the announcements
have a negative (positive) impact on the mean of the jump size. This effect is highly
statistically significant for the release of NFP: when payrolls are higher than expected in
periods of economic growth, the size of jumps in bond returns is drawn from a distribution
with a lower mean, i.e. there is a greater probability of observing higher yields. We register
the opposite effect when payrolls are lower than expected in periods of economic growth.
Economically, a one standard deviation positive NFP surprise in expansion decreases the
jump-size mean by 0.3%. Considering that the average number of jumps induced by the
employment report in one year is 1.4 (0.34+1.06), this corresponds to a more negative mean
16
-
of the jump size distribution of 0.42%. The release of the CPI in expansions has similar
effects on the long maturity, it is not significant for intermediate maturities, and it switches
sign in the short term.
In contraction phases of the business cycle, however, the effects of economic releases are
less clear. When the NFP is higher than expected, the mean of the jump size increases and
viceversa for lower than expected NFP. The effect of CPI on bond returns at the longer
maturities is also reversed. The release of the PPI has always a negative effect, significant
however only at the two and five year maturities. This evidence suggests that it is indeed
important to allow for different responses to the same announcements in expansions versus
contractions. We can nevertheless obtain further insights and a deeper interpretation of these
kind of results by disentangling the effect of good and bad news in good and bad times. We
will devote Section 4.2.2 to this exercise.
Table 5 presents the characteristics of the jump and volatility dynamics implied by the
estimated state-dependent jump model. Panel A shows summary statistics for the jump
intensity λt, for the jump size mean θt, and for the part of the total variance induced by the
jumps (var[²2,t+1
∣∣Φt]/
var[rt+1
∣∣Φt]), using the estimated model on all the days in our sample,
considering the actual timing of scheduled macroeconomic news, and separating expansions
versus contractions. We conduct non-parametric tests on the samples of λt, θt, and the
variance induced by the jumps for the different segment of the yield curve and we find in all
cases that the samples come from distributions with different medians.13
We observe that the jump intensity is highest for the longest maturity, it is about at the
same level for the ten and five year maturity, and it decreases for shorter maturities. These
intensities range from 0.34 jumps per year (about one jump every three years) at the long
end to 0.13 jumps per year (about one jump every seven and a half years) at the short end
of the yield curve. The median jump intensities clearly resemble the unconditional jump
intensity estimates presented in Table 4. The mean of the jump size distribution is always
negative, less so for shorter maturities. There are no appreciable differences of these statistics
between expansion and recession phases, except for the jump-induced variance that is lower
in recessions at the very long- and short-end of the yield curve.
The proportion of the total variance induced by the jumps is highest for the two-year
futures, where jumps generate on average more than half of the variance, and lower for longer
maturities. This evidence on the prominence of the middle maturity range for the variance
induced by the jumps is consistent both with a macroeconomic news explanation and with an
order flow explanation. In fact, Fleming and Remolona (1999b) find that the high-frequency
13The results of the non-parametric tests are not reported, but are available on request from the authors.
17
-
responses of 5-minutes bond yields to macroeconomic announcements over the period 1991 to
1995 have a hump-shaped pattern that peaks around two to three years. We do not observe
this pattern in our analysis of returns in the previous section for our longer sample at the
daily frequency. There would thus be a closer link between high-frequency returns dynamics
and jump-induced variance then there is with daily returns. However, Brandt and Kavajecz
(2004) observe that, on non-announcement days, the strongest response of yields to order
flow imbalance is at the two-five-year maturity range. We argue that the jump innovations
to the price process, triggered either by macroeconomic announcements or price discovery,
are most important for the maturity range containing securities that are more universally
held and used for hedging.
Panel B of Table 5 investigates more closely the pattern of jumps implied by the estimated
model. We use the ex-post probability of having at least one jump derived in equations (21)
and (22) to determine the average frequency of jumps in our sample. We observe the highest
number of average jumps per year for the five-year futures. The comparison of expansions
and recessions shows that for all the different futures contracts the average number of jumps
is lower in recessions. We obtain further insights when we relate the jump dynamics to the
pattern of macroeconomic announcements. For example, we observe that the proportion of
jumps occurring on the same day of a macroeconomic release varies a lot with the maturity.
Less than 35% of the total number of jumps occurs on announcement days for the two-year
futures, i.e. the majority of observed jumps are unrelated to our set of macroeconomic
releases, but more than 60% of the jumps occurs on announcement days for the longest
maturity futures.14 The analysis across business cycle phases shows that in recessions jumps
are slightly less related to macroeconomic news than in expansions for all maturities, except
at the short-end of the yield curve, where jumps on announcement days are more frequent
in recessions than in expansions.
We can investigate the same issues from a different perspective, that is the proportion
of macroeconomic announcements that results in a jump. We observe that less than 30%
of the macroeconomic releases in our sample causes a jump for the shortest maturities -
the two-year futures and the Treasury bill -, whereas more than 60% of the announcements
generates a jump for the five-year and longer maturity securities. For all maturities except
at the short-end of the yield curve, the proportion of macroeconomic releases resulting in
14Note that this descriptive analysis has three limitations. First, we cannot talk about a causal linkbetween macroeconomic news and jumps in a strict sense. We are only arguing that, since we observenews and jump on the same day, they are very likely to be related. Second, we are only considering fourmacroeconomic releases, but the set of influential announcements is certainly wider. If this is the case, thereported proportions are biased downward. Lastly, we consider that a jump has occurred when the ex-postprobability is greater than 0.5 (see Table 3, Panel B), but this threshold is arbitrary.
18
-
a jump is lower in recessions, indicating that unconditionally learning about the state of
the economy from news is less than a surprise when the economy is in contraction. We can
further develop this reasoning by looking at each macroeconomic release separately. The
release of the CPI is more frequently a cause of jumps for the intermediate maturities and
particularly for the two-years in recessions. The employment report (ER) is always the most
relevant release and increases in importance with shorter maturities, whereas the PPI exhibit
the opposite pattern of decreasing importance with maturity.
In Figure 3, we summarize these results. We observe more clearly that announcements
are relatively more important in causing jumps in expansions rather than in recessions and
for longer than for shorter maturity securities. The release of the employment report is
invariably the most influential, both in expansions and recessions. The only exception seems
to be the two-year horizon, where the role of inducing jumps of the three releases that we
consider is more balanced.
4.2.1 Specification Tests
To ensure that the model is reasonably well specified with respect to the inclusion of the
timing and information of macroeconomic news, we conduct a sequence of specification tests.
Specifically, starting with a version of the model with constant jump intensity (i.e., λt = λ0),
we use likelihood ratio tests to sequentially examine each restriction with respect to the
general model estimated in Table 4.
Table 6 shows the results of three specification tests. First, we find that, for all the
Treasury futures, a constant jump intensity is strongly rejected in favor of the most general
version of the model. It is thus statistically important to allow the intensity of the jump to
be time-varying. The second tested restriction is to allow the intensity of the jump to be
different on announcement days, but with no distinction of the announcement type. Again,
we strongly reject this restriction for all the different segments of the yield curve, except
for the two years. It seems therefore essential to allow different announcements to affect
differently the intensity of the jump process. Lastly, we constrain the mean of the jump size
to be independent on the information content of the macroeconomic release. This restriction
is again strongly rejected for all the maturities. From these specification tests, we conclude
that the features of the proposed state-dependent jump model are all statistically important
to accurately describe the returns of the bond futures.
4.2.2 Asymmetric Effect of Macroeconomic News
We argue that the information content of macroeconomic news can have different effects on
bond markets depending on observing good news versus bad news, in good or bad times.
19
-
To disentangle these asymmetric effects, we use the general specification for the jump size
mean formalized in equation (18).
Table 7 presents the results for the effects on the mean of the jump size. The results
on the other characteristics of the model are not different from the estimates of the simpler
specification presented in Table 4 and are thus not reported. A likelihood ratio test shows in
all cases a significant statistical improvement in generalizing the specification for the mean
of the jump size. More specifically, it is statistically important to allow good and bad news
to have different effects in good or bad times and this characteristic of the model seems
especially crucial for the shorter maturities.
We observe that the parameter estimates are mostly significant for bad news in expansion
phases of the business cycle. In contrast, good news in good times do not seem to
systematically affect the distribution of the size of the jump in bond returns. We register
the most relevant effects of the information in macroeconomic news on jump dynamics when
the release of the NFP is higher than expected in an expansion phase of the business cycle.
In this case, the distribution of the size of jumps in Treasury bond futures returns has a
negative conditional mean for all maturities. We observe a very similar pattern for PPI
announcements, with higher than expected releases leading to more negative jump size
means. The announcement of CPI has the opposite impact, except for the longest maturity:
a positive CPI surprise leads to a distribution of the jump size with a higher conditional
mean.
In recession phases, the pattern of the relevant information is less clear, but good news
(negative surprises) tend to be more important in affecting jump dynamics. More specifically,
when the released NFP is lower than expected, the size of the jump in returns for shorter
maturities is drawn from a distribution with a negative mean.15 This pattern is similar,
although less significant, for the CPI. In contrast, jump after good news in the PPI release
are drawn from a distribution with a higher mean, although this effect is significant only for
intermediate maturities.
The interpretation of these results is again more meaningful, if we translate the effect of
a surprise on returns to the effect of a surprise on yields. The mapping between returns and
yields goes again simply through the duration of each segment of the yield curve as shown
before in equation (4). In this case, we obtain the return on the underlying dP/P as the
mean size of a jump triggered by a one-standard deviation announcement surprise occurring
with intensity λt. We then calculate the duration of each Treasury futures contract as the
15We observe a negative impact on the mean of the jump size, because the parameter estimate θrecG,nfp ispositive, but the surprise is always negative.
20
-
modified duration of a deliverable bond with the average maturity in the maturity range.
Using the measure for bond returns suggested above and the appropriate modified duration,
we obtain a measure of the effect on yields.16
Figure 4 plots the effect of a one standard deviation surprise in NFP, PPI and CPI on the
mean of the jump size distribution of yields at different maturities. We also plot two standard
deviation bands, obtained from the standard errors of the parameter estimates.17 Panel A
shows the most significant case, the effect of positive surprises in an expansion phase of the
business cycle. The exception is the release of the CUR, for which we actually show the effects
of negative surprises, since the sign of the surprise has the opposite implications with respect
to the other macroeconomic announcements. The release of a lower than expected CUR has
similar implications for the labor market than the release of an higher than expected NFP.
Yields for all maturities rise in response to higher than expected numbers for NFP and PPI
releases. In contrast, the results for the release of CUR are never significantly affecting the
mean of the jump size in expansion. We interpret the results for NFP and PPI as evidence
that an unexpected pickup in economic growth when the economy is rising is perceived as
indication of overheating, which leads short and long term interest rates to increase. The
effect of a positive CPI surprise is less intuitive. We observe a general reduction in yields,
with a higher yield only at the longest maturity, implying a steeper yield curve. Either an
expected pickup in economic growth or higher expected future inflation would tend to raise
long-term rates relative to short rates, steepening the yield curve. It is more difficult to
explain why a higher than expected CPI release would reduce the mean of the jump size in
yields at the intermediate maturities. In fact, the market seems to interpret positive shocks
to CPI differently than PPI. However, any interpretation of the CPI results should be taken
cautiously, because the CPI is often released the day after the PPI and, although we control
for the PPI release in forming the CPI expectation, the results could still be affected by
overreaction or underreaction to PPI news.
Panel B shows the effect of negative surprises in the NFP, PPI, CPI release and positive
surprise in the CUR release, in a recession phase of the business cycle. When the release of
PPI is lower than expected, we observe a general reduction in yields. The monetary policy
16The measure of bond returns is the effect of a one standard deviation surprise in the macroeconomicannouncement obtained as an expected jump size times an intensity measure. As such, the change in yieldscan be interpreted in the same logic as the jump effect of a macroeconomic surprise.
17We obtain the error bands of the effect on yields using the bond return dP/P computed as the jump sizemean parameters plus/minus two standard errors and the procedure outlined above to convert bond returnsinto bond yields. This methodology delivers a good approximation of the true confidence bands, since thecovariance between the intercept and the slope term of the jump size mean is negligible and negative (e.g.,the covariance between θexp0 and θ
expB,nfp for the 30-year Treasury bond futures is -1.01e-08 when θ
expB,nfp is
1.35e-02) and the standard error of the jump intensity term exerts a second order effect.
21
-
has the expected expansionary bias, since the economy is in recession and inflation shows
indications of being below analyst estimates. Furthermore, during a recession, the negative
inflation shock is also likely to lower investors’ inflation expectation. This latter effect could
explain the drop in long term yields. The release of weaker than expected NFP numbers
in a recession phase does not have very significant effects. We observe a flattening of the
yield curve that is consistent with future rate hikes being less likely when the economy is in
a recession and payrolls are showing no sign of recovery. Alternatively, lower than expected
NFP numbers could lower payroll pressure on production costs, resulting in lower inflation
expectations, and therefore lower long term yields. The release of higher than expected CUR
has significant effects mostly on the ten and five-year yields, where we observe reduction in
yields consistent with signs of slow growth in recessions. It is interesting to see that CUR has
some significant effects in recession, but was never relevant in expansions, indicating that
the bond market looks at different indicators to gauge the state of the economy in different
phases of the business cycle. The release of a lower than expected CPI does not significantly
affect the mean of the jump size, except that it implies higher short term yields. It is hard
to interpret this result in isolation, it only makes sense when we observe that the yield curve
becomes flatter. However, the release of the PPI very shortly before implies again that all
inferences should be taken with caution.
4.2.3 Jumps and Asymmetric Feedbacks on Conditional Volatility
We further examine the effect of macroeconomic news on volatility dynamics using a
richer specification for conditional volatility. More specifically, Table 8 presents the results
of estimating the state-dependent jump model with the GARCH component specified in
equation (13), where we allow for different feedback on future volatility of normal versus
jump innovations in expansions versus recessions, and asymmetries of positive and negative
innovations. We do not report the estimates for the parameters of the model that are not
related to volatility, because they are very similar to the general estimation results described
in Table 4.
A likelihood ratio test indicates that the richer specification for the conditional volatility
function is especially important for the longest and shortest maturity. In general, we observe
that the feedback of innovations on future conditional volatility is stronger in recessions. We
attribute the low or non-significant αrec for the five and two year maturity to the shorter
sample that contains only two recessionary periods. The feedback of the jump innovations
αexp,recj is always negative and statistically significant, both in expansions and in recessions,
indicating that the effect of jumps on future volatility tends to not persist. We find evidence
of asymmetric volatility for the longest and shortest maturity, but only in expansions. In
22
-
both cases, if the negative innovation is likely to be a jump, the higher persistence αexpa is
almost completely offset by the asymmetric jump feedback term αexpaj .
Figure 5 exploits the parameters estimated in Table 8 to plot the effects of positive and
negative innovations in expansions and recessions, separating normal innovations from jump
innovations, i.e. the innovations that were jumps with a probability of 50%. We observe
the lower persistence of positive normal innovations in expansions versus recessions, but the
higher persistence of negative normal innovations in expansions versus recessions because of
asymmetric volatility. We also notice how jump innovations dampen persistence and this is
true in expansions, recessions, positive and negative innovations.
5. Conclusion
We analyze the effects of good and bad macroeconomic news on bond market returns and
volatility, during economic expansions and economic contractions. Specifically, we identify
a high-frequency relation between yields and the economy, whereas the macroeconomic
literature has typically modeled interactions at a lower frequency.
We find that the information content of the announcements is most important for bond
returns when it contains bad news for the bond market in expansions and, to a lesser extent,
good news for the bond market in contractions. Furthermore, the response of the bond
market depends on the announcement type, on the bond maturity, and on the business
cycle. We also find that macroeconomic news impact substantially the volatility of bond
returns at all maturities through increased jump intensities and different distributions of the
jump size. The proportion of the total variance induced by the jumps is greatest for the
medium term maturities of two and five years. Finally, we observe that the impact of jumps
on volatility does not persist, consistent with Jones, Lamont, and Lumsdaine (1998).
In bond markets, monetary authorities play a key role, with direct effects on short-
term yields and some influence on long-term yields via inflation expectations. Our results
are consistent, for example, with an asymmetric Fed reaction function by which positive
economic news induce a strong increase in the likelihood of higher future target rates in
expansions, but only a marginal increase in contractions. In this setting, the response of
bond prices to macroeconomic news depends critically on the content of the news and on
the phase of the business cycle.
Furthermore our paper shows the importance of modeling volatility in the bond market
using both a diffusive and a jump component and incorporating the timing and information
of macroeconomic news, allowing for different dynamics in expansions and recessions. A
challenging area for future study is to incorporate these results in developing no-arbitrage
models of the term structure of Treasury bond yields.
23
-
References
Andersen, T.G., Bollerslev, T., Diebold, F.X., and Vega, C. (2007) Real-time pricediscovery in stock, bond and foreign exchange markets, Journal of International Economics,forthcoming.
Andersen, T.G., Bollerslev, T., Diebold, F.X., and Vega, C. (2003) Micro effects of macroannouncements: Real-time price discovery in foreign exchange, American Economic Review93, 38–62.
Ang, A., Bekaert, G., and Wei, M. (2007) Do macro variables, asset markets or surveysforecast inflation better?, Journal of Monetary Economics 54, 1163–1212.
Bekaert, G. and Gray, S.F. (1998) Target zones and exchange rates: An empiricalinvestigation, Journal of International Economics 45, 1–35.
Bernanke, B.S. and Kuttner, K.N. (2005) What explains the stock market’s reaction toFederal Reserve policy?, Journal of Finance 60, 1221–1257.
Bollerslev, T., Cai, J., and Song, F.M. (2000) Intraday periodicity, long memory volatility,and macroeconomic announcement effects in the U.S. Treasury bond market, Journal ofEmpirical Finance 7, 37–55.
Boyd, J.H., Hu, J., and Jagannathan, R. (2005) The stock market’s reaction tounemployment news: Why bad news is usually good for stocks, Journal of Finance 60,649–672.
Brandt, M.W. and Kavajecz, K.A. (2004) Price discovery in the U.S. Treasury market: Theimpact of orderflow and liquidity on the yield curve, Journal of Finance 59, 2623–2654.
Brenner, M., Pasquariello, P., and Subrahmanyam, M. (2006) Financial markets and themacro economy, Working Paper, University of Michigan.
Christiansen, C. and Ranaldo, A. (2007) Realized bond-stock correlation: Macroeconomicannouncement effects, Journal of Futures Markets, forthcoming.
Das, S.R. (2002) The surprise element: Jumps in interest rates, Journal of Econometrics106, 27–65.
Ederington, L.H. and Lee, J.H. (1993) How markets process information: News releases andvolatility, Journal of Finance 48, 1161–1191.
Engle, R.F. and Rosenberg, J. (2002) Empirical pricing kernels, Journal of FinancialEconomics 64, 341–372.
Fleming, M.J. and Remolona, E.M. (1997) What moves the bond market?, Federal ReserveBank of New York Economic Policy Review, 31–50.
Fleming, M.J. and Remolona, E.M. (1999) Price formation and liquidity in the U.S. Treasurymarket: The response to public information, Journal of Finance 54, 1901–1915.
Fleming, M.J. and Remolona, E.M. (1999b) The term structure of announcement effects,Working Paper, Bank for International Settlements.
Gordon, S. and St-Amour, P. (2004) Asset returns and state-dependent risk preferences,Journal of Business and Economic Statistics 22, 241–252.
24
-
Hong, H., Torous, W., and Valkanov, R. (2007) Do industries lead stock markets?, Journalof Financial Economics 83, 367–396.
Jones, C.M., Lamont, O., and Lumsdaine, R.L. (1998) Macroeconomic news and bondmarket volatility, Journal of Financial Economics 47, 315–337.
Maheu, J.M. and McCurdy, T.H. (2004) News arrival, jump dynamics and volatilitycomponents for individual stock returns, Journal of Finance 59, 755–793.
Neely, C.J. (1999) Target zones and conditional volatility: The role of realignments, Journalof Empirical Finance 6, 177–192.
Stock, J. and Watson, M. (1989) New indexes of coincident and leading economic indicators,in Olivier, J., Fisher, S. (ed.): NBER Macroeconomics Annual 1989 (Cambridge, MA).
Veronesi, P. (1999) Stock market overreaction to bad news in good times: A rationalexpectations equilibrium model, Review of Financial Studies 12, 975-1007.
25
-
Table 1: Macroeconomic News and Futures Returns Summary Statistics
This table shows summary statistics for macroeconomic news and for daily returns on the 30-yearTreasury bond futures (US), the 10-year Treasury note futures (TY), the 5-year Treasury notefutures (FV), the 2-year Treasury note futures (TU), and the 90-day Treasury bill futures (TB).Panel A shows the announcements, their abbreviations, the reported units of the variables, the timesat which the announcements are normally released, and the number of times two announcementsare concurrent (same date and time). Panel B shows statistics on daily continuously compoundedreturns for the full sample that is available for each futures contract. Panel C shows statistics ondaily returns for the longest sample common to all the futures contracts.
Panel AConcurrent
AnnouncementsAnnouncement Abbrev. Units Time CPI CUR NFP PPI
Consumer Price Index CPI % Change 8:30 ET 282 0 0 0Civilian Unemployment CUR % Level 8:30 ET 0 282 282 13Nonfarm Payrolls NFP Thousands 8:30 ET 0 282 282 13Producer Price Index PPI % Change 8:30 ET 0 13 13 282
Panel B
US TY FV TU TB
First Obs Feb-1980 Jun-1982 May-1988 Jun-1990 Feb-1980Last Obs Dec-2003 Dec-2003 Dec-2003 Dec-2003 Sep-2003Obs 6054 5469 3968 3526 5965Mean+ 0.0399 0.0449 0.0238 0.0079 0.0079StDev+ 0.1189 0.0718 0.0425 0.0211 0.0214Skewness -0.0692 -0.0357 -0.4764 -1.0419 0.3349Kurtosis 5.1034 5.9126 6.4016 10.8327 14.5480Min -0.0324 -0.0286 -0.0222 -0.0103 -0.0071Max 0.0394 0.0342 0.0118 0.0065 0.0140
Panel C
US TY FV TU TB
First Obs Jun-1990 Jun-1990 Jun-1990 Jun-1990 Jun-1990Last Obs Sep-2003 Sep-2003 Sep-2003 Sep-2003 Sep-2003Obs 3353 3353 3353 3353 3353Mean+ 0.0420 0.0358 0.0276 0.0084 0.0077StDev+ 0.0921 0.0615 0.0421 0.0210 0.0074Skewness -0.4209 -0.5271 -0.5872 -1.0220 0.7597Kurtosis 4.5016 5.5767 6.5843 10.8010 12.1037Min -0.0298 -0.0286 -0.0222 -0.0103 -0.0036Max 0.0246 0.0142 0.0114 0.0065 0.0043
+ Daily mean return and standard deviation are annualized.
26
-
Table 2: Effect of Macroeconomic News on Bond Returns
This table shows parameter estimates for the following regression:
rt,t−1 = αk + βexpk (1−XRIt) Sk + βreckXRItSk + ek,
where rt,t−1 represents the day-to-day return for each of the Treasury bond futures, XRIt is theprobability of being in a recession phase at time t, and S denotes the standardized announcementsurprise. The day-to-day return is computed on the 30-year Treasury bond futures (US), the 10-year Treasury note futures (TY), the 5-year Treasury note futures (FV), the 2-year Treasury notefutures (TU), and the 90-day Treasury bill futures (TB).
US TY FV TU TB
CPI α 0.00028 0.00020 0.00018 0.00018 0.00011βexp -0.00123∗ -0.00085∗∗ -0.00091∗∗ -0.00043∗∗ -0.00017∗
βrec -0.00244 -0.00211∗∗ -0.00057 -0.00003 -0.00035R2 0.045 0.052 0.055 0.047 0.037
CUR α 0.00060 0.00021 0.00014 0.00007 0.00019∗
βexp 0.00091 0.00096∗∗ 0.00083∗∗ 0.00042∗∗ 0.00020∗
βrec 0.00297 -0.00001 0.00012 0.00027 0.00097∗∗
NFP βexp -0.00446∗∗∗ -0.00320∗∗∗ -0.00240∗∗∗ -0.00101∗∗∗ -0.00061∗∗∗
βrec -0.00303 -0.00211∗ -0.00170∗∗ -0.00102∗∗∗ -0.00059∗∗∗
R2 0.181 0.216 0.294 0.308 0.171
PPI α 0.00028 0.00075 0.00047 0.00015 0.00001βexp -0.00220∗∗∗ -0.00096∗∗ -0.000212 0.00001 -0.00043∗∗∗
βrec 0.00135 0.00111 0.00097∗ 0.00035 -0.00011R2 0.038 0.017 0.015 0.012 0.048
∗ ∗ ∗, ∗∗, and ∗ denote statistical significance at the one, five, and 10 percent levels,respectively.
27
-
Table 3: Effect of Good and Bad News on Bond Returns
This table shows parameter estimates for the following regression:
rt,t−1 = αk + βGexpk (1−XRIt) SktGkt + βBexpk (1−XRIt) SkBkt+ βGreckXRItSktGkt + βBreckXRItSkBkt + ekt,
where rt,t−1 represents the day-to-day return for each of the Treasury bond futures, XRIt is theprobability of being in a recession phase at time t, S denotes the standardized announcementsurprise, and Gkt = 1 and Bkt = 0 if the information released in announcement k at time t is goodnews (for the Treasury market) and Gkt = 0 and Bkt = 1 otherwise. The day-to-day return iscomputed on the 30-year Treasury bond futures (US), the 10-year Treasury note futures (TY), the5-year Treasury note futures (FV), the 2-year Treasury note futures (TU), and the 90-day Treasurybill futures (TB).
US TY FV TU TB
CPI α 0.00073 0.00109∗∗∗ 0.00092∗∗∗ 0.00047 0.00016βGexp 0.00080 0.00109 0.00043 0.00010 0.00015βBexp -0.00286∗∗∗ -0.00266∗∗∗ -0.00211∗∗∗ -0.00101∗∗∗ -0.00043∗∗∗
βGrec -0.00665∗∗∗ -0.00357∗∗ 0.00028 0.00023 -0.00116∗∗
βBrec -0.00019 -0.00148∗ -0.00117 -0.00016 0.00016R2 0.095 0.100 0.099 0.084 0.103
CUR α 0.00089 0.00014 0.00001 -0.00001 0.00017βGexp 0.00039 -0.00031 -0.00041 -0.00005 0.00014βBexp 0.00119 0.00185∗∗∗ 0.00248∗∗∗ 0.00108∗∗∗ 0.00026βGrec 0.00167 0.00338∗ 0.00271 0.00065 -0.00011βBrec 0.00303 -0.00082 -0.00138 0.00006 0.00108∗∗
NFP βGexp -0.00339∗∗∗ -0.00253∗∗∗ -0.00156∗∗∗ -0.00068∗∗∗ -0.00052∗∗∗
βBexp -0.00634∗∗∗ -0.00452∗∗∗ -0.00390∗∗∗ -0.00161∗∗∗ -0.00076∗∗∗
βGrec -0.00319 -0.00275∗∗ -0.00239∗∗∗ -0.00117∗∗∗ -0.00056∗∗
βBrec 0.00035 0.00110 0.00014 -0.00097 -0.00054R2 0.189 0.233 0.334 0.339 0.178
PPI α 0.00054 0.00046 0.00044 0.00002 -0.00001βGexp -0.00266∗∗ -0.00131∗∗ -0.00030 -0.00021 -0.00046∗∗
βBexp -0.00166 -0.00032 -0.00008 0.00033 -0.00042βGrec 0.00081 0.00102 0.00114∗∗ 0.00042∗∗ 0.00005βBrec 0.00217 0.00111 0.00072 0.00015 0.00022R2 0.040 0.019 0.016 0.023 0.048
∗ ∗ ∗, ∗∗, and ∗ denote statistical significance at the one, five, and 10 percent levels,respectively.
28
-
Table 4: Estimates of the State-Dependent Jump Model
This table shows estimates for daily returns on the 30-year Treasury bond futures (US), the 10-year Treasury note futures (TY), the 5-year Treasury note futures (FV), the 2-year Treasury notefutures (TU), and the 90-day Treasury bill futures (TB) of the following model:
²1,t+1 = σt+1zt+1, zt+1 ∼ N[0, 1], ²2,t+1 =nt+1∑
k=1
Yt+1,k − θt+1λt+1
σ2t+1 = ω +(α + αjE
[nt
∣∣Φt])
²2t + βσ2t , µt+1 = µ
expt+1(1−XRIt+1) + µrect+1XRIt+1
λt+1 = λ0 +∑
k=cpi,er,ppi
λkDk,t+1
θt+1 =(1−XRIt+1
)(θexp0 +
∑
k=cpi,cur,nfp,ppi
θexpk Sk,t+1
)+XRIt+1
(θrec0 +
∑
k=cpi,cur,nfp,ppi
θreck Sk,t+1
)
US TY FV TU TB
ω × 106 0.9810∗∗∗ 4.4483∗∗∗ 0.0021∗∗∗ 0.5273∗∗∗ 14.6029∗∗∗α 0.0737∗∗∗ 0.0878∗∗∗ 0.0701∗∗∗ 0.0411∗∗∗ 0.1387∗∗∗
αj -0.0561∗∗∗ -0.0640∗∗∗ -0.0531∗∗∗ -0.0352∗∗∗ -0.1239∗∗∗
β 0.9370∗∗∗ 0.9188∗∗∗ 0.9321∗∗∗ 0.9571∗∗∗ 0.8756∗∗∗
µexp × 104 1.5124∗ 0.8729∗ 0.4063 0.2004 0.0325µrec × 104 5.2735∗∗ 4.7239∗∗∗ 3.1109∗∗∗ 1.8189∗∗∗ 0.9623∗∗∗
λ0 0.3429∗∗∗ 0.3161∗∗∗ 0.3213∗∗∗ 0.1641∗∗∗ 0.1320∗∗∗
λcpi 0.3501∗∗∗ 0.3957∗∗∗ 0.4650∗∗∗ 0.1989∗∗∗ 0.0635λer 1.0574∗∗∗ 1.2604∗∗∗ 1.0151∗∗∗ 0.5447∗∗∗ 0.9802∗∗∗
λppi 0.6823∗∗∗ 0.6944∗∗∗ 0.5370∗∗∗ 0.1565∗∗∗ 0.1548∗∗∗
θexp0 × 104 -5.3039∗ -4.7729∗∗∗ -2.4921∗ -4.4213∗∗∗ -0.42493θexpcpi × 104 -29.2192∗∗∗ 0.5324 8.4604 9.6003∗∗ 3.2602∗∗θexpcur × 104 4.9841 -3.7622 -0.9755 -3.5334 -0.1045θexpnfp × 104 -33.1023∗∗∗ -21.3446∗∗∗ -22.2312∗∗∗ -4.1504 -0.8882∗θexpppi × 104 15.1112 4.8902 1.4131 5.1144 1.7234θrec0 × 104 -29.8123∗∗∗ -24.3031∗∗∗ -14.1431∗∗∗ -3.1804 4.4326∗∗∗θreccpi × 104 57.6348∗∗ 34.6237∗ 15.5484 8.1934 5.3218θreccur × 104 -30.5386 -26.1379 28.8390∗∗ 4.1004 3.7019∗θrecnfp × 104 43.4328∗∗ 21.1036∗∗ 19.7344∗∗ 18.4301∗∗∗ 6.8400∗∗∗θrecppi × 104 -22.4381 -18.2351 -18.7093∗∗ -14.7223∗∗∗ -2.7254δ 0.0055∗∗∗ 0.0039∗∗∗ 0.0026∗∗∗ 0.0022∗∗∗ 0.0010∗∗∗
logL 21,766.60 22,306.00 18,195.90 18,870.40 35,506.50Obs 6054 5469 3968 3526 5965
∗ ∗ ∗, ∗∗, and ∗ denote statistical significance at the one, five, and 10 percent levels,respectively.
29
-
Table
5:
Model-Im
plied
Jum
pand
Vola
tility
Dynam
ics
Thi
sta
ble
show
ske
yfe
atur
esim
plie
dby
the
esti
mat
edju
mp
mod
elfo
rda
ilyre
turn
son
the
30-y
ear
Tre
asur
ybo
ndfu
ture
s(U
S),
the
10-y
ear
Tre
asur
yno
tefu
ture
s(T
Y),
the
5-ye
arTre
asur
yno
tefu
ture
s(F
V),
the
2-ye
arTre
asur
yno
tefu
ture
s(T
U),
and
the
90-d
ayTre
asur
ybi
llfu
ture
s(T
B).
Pan
elA
show
sth
eex
pect
ednu
mbe
rof
jum
ps(λ
t),th
em
ean
ofth
eju
mp
size
(θt)
,an
dth
epr
opor
tion
ofth
eco
ndit
iona
lvar
ianc
eth
atis
expl
aine
dby
the
jum
pco
mpo
nent
(i.e
.,va
r[ ²2,t+
1
∣ ∣ Φt]/
var[ r
t+1
∣ ∣ Φt] )
.T
here
port
edva
lues
are
sam
ple
med
ians
abov
ean
dm
ean
abso
lute
devi
atio
nsfr
omth
em
edia
nsbe
low
.Pan
elB
show
sth
ein
ferr
ednu
mbe
rof
jum
psfo
rth
efu
llsa
mpl
e,th
eav
erag
enu
mbe
rof
jum
pspe
rye
ar,th
enu
mbe
rof
anno
unce
men
tda
ysin
the
sam
ple,
the
infe
rred
num
ber
ofju
mps
occu
rrin
gon
anno
unce
men
tda
ys,t
hepr
opor
tion
ofju
mps
occu
rrin
gon
anno
unce
men
tda
ys,a
ndth
epr
opor
tion
ofan
noun
cem
ent
days
that
resu
ltin
jum
ps.
Pan
elA
US
TY
FV
TU
TB
all
exp
rec
all
exp
rec
all
exp
rec
all
exp
rec
all
exp
rec
λt
0.34
290.
3429
0.34
290.
3161
0.31
610.
3161
0.32
130.
3213
0.32
130.
1641
0.16
410.
1641
0.13
200.
1320
0.13
200.
0970
0.09
730.
0955
0.10
940.
1094
0.10
990.
0933
0.09
330.
0930
0.04
060.
0406
0.04
090.
0530
0.05
310.
0520
θ t-6
.04e
-4-5
.79e
-4-2
.64e
-3-5
.16e
-4-5
.16e
-4-2
.16e
-3-2
.84e
-4-2
.72e
-4-1
.07e
-3-4
.29e
-4-4
.38e
-4-1
.02e
-4-2
.80e
-5-3
.30e
-53.
75e-
45.
38e-
42.
64e-
46.
52e-
43.
10e-
41.
40e-
44.
08e-
42.
24e-
41.
27e-
43.
17e-
41.
02e-
45.
50e-
51.
80e-
48.
60e-
53.
30e-
58.
80e-
5ju
mp-
indu
ced
0.31
870.
3264
0.26
090.
3697
0.36
970.
3727
0.43
060.
4332
0.41
580.
5551
0.56
120.
5255
0.41
960.
4597
0.17
52va
rian
ce0.
1261
0.12
350.
1342
0.13
110.
1317
0.12
610.
1190
0.12
150.
1000
0.10
460.
1060
0.09
170.
2458
0.23
760.
2232
30
-
Pan
elB
US
TY
FV
TU
TB
all
exp
rec
all
exp
rec
all
exp
rec
all
exp
rec
all
exp
rec
N.ju
mps
+10
0286
913
392
482
698
682
607
7532
228
735
473
418
55N
.ju
mps
per
year
++
41.3
842
.25
36.4
242
.24
42.7
438
.34
42.9
743
.41
39.6
422
.83
23.3
819
.10
19.8
220
.69
15.0
3N
.an
noun
cem
ent
days
839
718
121
759
670
8954
648
165
473
411
6283
171
012
1N
.ju
mps
onan
noun
cem
ent
days
612
534
7853
948
554
348
314
3411
010
010
245
209
36ju
mps
onan
n.da
ysov
eral
lju
mps
0.61
0.61
0.59
0.58
0.59
0.55
0.51
0.52
0.45
0.34
0.35
0.29
0.52
0.50
0.65
jum
pson
ann.
days
over
allan
n.da
ys0.
730.
740.
640.
710.
720.
610.
640.
650.
520.
230.
240.
160.
290.
290.
30ju
mps
onC
PI
ann.
days
9685
1179
736
7970
922
193
1412
2ju
mps
onC
PI
over
jum
pson
ann.
days
0.16
0.16
0.14
0.15
0.15
0.11
0.23
0.22
0.26