You are on page 1of 17

J Econ Finan (2012) 36:712727

DOI 10.1007/s12197-010-9147-6

Sudden equity price declines and the flight-to-safety


phenomenon: additional evidence using daily data
Joe Brocato & Kenneth L. Smith

Published online: 10 August 2010


# Springer Science+Business Media, LLC 2010

Abstract Using daily data we show sudden, extreme declines in the U.S. stock
market for crash dates to lead to a capital preserving (as opposed to strategic or
tactical) reallocation to government debt securities. In most cases we find flightinduced reallocation reverses direction within one day of a crash. However, for the
1987 world crash we find increased and persistent return volatility in both equity and
bond returns lasting up to five days following this dramatic decline in world equity
prices. Like previous research in this area, we find equity crashes alter long-run
stock/bond return correlations and lead to increased stock and bond return volatility.
Finally, we describe the somewhat unique stock and bond correlation adjustments
triggered by the 9/11 attack and the impact this event had on the behavior of U.S.
equity investors flight-to-safety reaction.
Keywords Flight-to-Safety . Stock and Bond Correlations . Return Volatility .
Risk Analysis
JEL Classification G01 . G11 . G12

1 Introduction
Typically the total returns of stocks and bonds are positively correlated. To some extent
this positive return association results from common discounted cash flow characteristics that drive both stock and bond prices during periods of macroeconomic stability.

J. Brocato (*)
Department of Accounting, Finance, and Economics, College of Business Administration,
Tarleton State University, Stephenville, TX 76402, USA
e-mail: jbrocat@tarleton.edu
K. L. Smith
Department of Economics, Finance, and Banking, Kelce College of Business,
Pittsburg State University, Pittsburg, KS 66762, USA

J Econ Finan (2012) 36:712727

713

While positive stock/bond return correlations are the norm, a casual reading of the
literature presents a wide array of numerically different return correlations (differences
in magnitude as well as in sign) for both the U.S. and other developed economies. These
varying correlations arise from differences in the time period under investigation and
different economic environments existing when the stock and bond price data are
generated, the frequency of the observations used in the correlation computation,
differences in the nature of the index from which the return calculations are obtained,
and many other factors.
For example, using annual U.S. stock and U.S. long-term government bond data
from Ibbotson and Associatesnow published by Morningstar (2007)we find a
long-run return correlation of 0.162 when the time period is 1927 through 2001.
Extending the interval from 1927 through 2006 increases the coefficient to 0.190.
Excluding the Great Depression and WWII from the Morningstar data series
produces a coefficient of 0.231. More recently from 1970 through the end of 1980
Dopfel (2003) reports a U.S. correlation of 0.40.
Scruggs and Glabadanidis (2003) find that the conditional correlation between
U.S. stock and bond returns varies considerably over the post-WWII period, being
negative in the late 1950s and early 1960s but positive since the mid-1960s. Engle
(2002) using various ARCH models finds that the correlations in the 1990s to be
mostly positive. However, in earlier periods Ilmanen (2003) identifies three episodes
of negative U.S. stock/bond correlations (19291932, 19561965, and 19982001).
He relates these to unique time-dependent economic conditions.
Brocato and Steed (1998) in a study of asset allocation over the business cycle
from 1972 through 1993 find the monthly stock/bond total return correlation to be
0.253 over expansions, rising to 0.524 over recessions, and stabilizing to 0.325 when
expansion and recession data are pooled.
Schwert (1990) finds considerable temporal instability of the stock/bond return
correlations in the U.S. economy particularly at turning points in the business cycle.
A related research finding is reported by Guidolin and Timmermann (2005) for the
U.K. They report stock and bond correlations to be positive and significant during
economic expansions but negative for economic downturns.
Changing the frequency to daily data and using the S&P 500 index and long-term
U.S. government bond data compiled by Thomson DataStream we obtain in this
paper a correlation coefficient of 0.095 for the March 1984October 2006 interval.
This positive but materially lower correlation coefficient is due to the use of daily
frequency data which contain a large random noise component.
The chief conclusion to be drawn from this brief review of the evidence is that
while historical stock and bond return correlation coefficients exhibit considerable
instability over time depending upon economic and expected factors, observation
frequency, and index construction, they are usually positive over longer intervals
when macroeconomic stability characterizes the financial environment.
For this paper, we formally define the flight-to-safety phenomenon as a sudden
and unexpected investor-driven re-allocation from risky equity securities to the
safety of default-free U.S. government Treasury bonds. At any point in time, an
observed stock/bond correlation coefficient and flight-to-safety behavior bear a close
relationship. The purpose here is to quantify the economic and statistical significance
of this relationship using daily data.

714

J Econ Finan (2012) 36:712727

From a traditional asset-pricing perspective, an abrupt decrease in the value of stocks


can be explained in terms of sudden changes in perceived capital loss probabilities and
in the equilibrium required risk premium for bearing this possible loss. As will be
statistically quantified, a daily flight-to-safety reaction in fully functioning securities
markets has a predictable impact on long-run stock/bond total return correlations.
The term flight-to-safety as used in this discussion should be distinguished from
asset re-allocation motivated by strategic versus tactical motives. Pure flight-to-safety
behavior is characterized by market participants desire to abruptly decrease their
portfolios exposure to securities bearing increased risk of capital loss. Conversely,
strategic asset re-allocation decisions are driven by deliberate longer-run portfolio
adjustments aimed at a target mix of securities for a given time line, risk profile, return
objective, and expected capital market outlook. Tactical re-allocation decisions are
geared to receiving (or avoiding) a short-run abnormal return (or loss) due to a
perceived alteration in relative security prices that is expected to be temporary.
In distinction to strategic and tactical motives, pure flight behavior is motivated by a
fear of sudden and permanent capital loss without regard for any long-run strategic
portfolio imbalance that may result or to a perceived short-run tactical excess return gain
or loss. Whereas strategic and tactical decisions are based upon some measure of
information (either current or expected), a flight-to-safety play is predicated upon a
sudden information void regarding what future equilibrium prices and required returns
will be. Flight-to-safety actions are examples of very short-run, purely idiopathic
investor behavior regarding subsequent relative stock and bond prices. While one can
tacitly relate ex post unexpected macroeconomic or political shocks to a sudden move
from equity securities as we do in this paper, such behavior is initially motivated by the
ex ante (and unknown) consequences of not re-allocating quickly.
In an innovative paper, Gulko (2002) presents convincing evidence that rapid and
unexpected declines in U.S. equity prices shifts the stock/bond correlation from its
normal positive sign to a statistically significant negative sign. His methodology is
as follows. Using daily data from 1987 through 2000 Gulko identifies six stock
market crashes related to macroeconomic or political events. Gulko defines a crash
as one where equity prices as measured by the S&P 500 index decline by five
percent or more over a one-day (or weekend) trading period. Using an event studytype model he aggregates the data into three sub-groups, a pre- and post-crash set of
observations covering ten days prior and subsequent to a crash day. The event period
which is approximately two weeks in length is nested between the pre- and postcrash periods. Regressions estimated separately on the three spliced and pooled time
series over the 1987 through 2000 period show the stock/bond regression slopes to
be positive over the pre- and post-crash periods while the event period regression
produces a statistically significant negative slope coefficient. Gulko cites this sign
change as evidence of a market-decline-induced flight-to-safety. More formally,
Gulko uses the term decoupling in the sense that the stock/bond return correlation
temporarily disconnects from its normal positive relationship. The sign change can
readily be attributed to perceived increases in the equity risk premium associated
with sharp and unexpected declines in equity prices and the desire by investors to
preserve capital value by rapid purchases of risk-free government debt securities.
In this paper we extend and articulate Gulkos work on the flight-to-safety
phenomenon, decoupling, and the temporary shift in the normally positive stock/

J Econ Finan (2012) 36:712727

715

bond return correlation. In particular, we investigate four additional avenues of


market-induced shifts from stocks to bonds: (1) We do not aggregate the data.
Gulkos methodology pools daily data by splicing pre- and post-crash intervals. Such
aggregation masks potential differences between various equity market crashes.
Using non-pooled daily data our method brings into sharper relief the stock/bond
market interaction both before and after a given crash date. This allows us to relate
sudden economic disruptions to specific calendar days and any subsequent daily
reversion activity; (2) Preceding our regression results is percentage change analysis.
Percentage change analysis gives both the sign and magnitude of stock and bond
prices (stated as daily percentages) for crash days and for the subsequent trading
days. From these results we obtain indirect evidence on the flow of funds between
stock and bond markets as these flows relate to crash events; (3) We carry out preand post-event correlation and variance analysis so as to gauge the extent of postevent persistence caused by sudden equity market declines. In this context,
correlation persistence has been identified in the literature as a by-product of market
shocks (Solnik et al. 1996); (4) We extend the time interval so as to allow an analysis
of the flight-to-safety behavior resulting from the September 11, 2001 attack. As we
will explain, the September 11 flight response is a delayed temporal analysis since
the stock market was closed the week of the attack.
The paper is organized as follows. Section 2 describes the data, related summary
statistics, and raw stock/bond return correlations. Section 3 identifies the crash dates
and associated macroeconomic or political event activity, the daily percentage declines
occurring on these dates for both stocks and bonds, and the next-day rate of return
responses in these markets. Section 4 describes the regression model and the
methodology used. Section 5 presents the regression results. Section 6 compares preand post-event return correlations and volatility. Here we relate the crash events to postevent turbulence in the U.S. stock and bond markets. Section 7 concludes the paper.

2 Data and descriptive statistics


2.1 Data
The complete interval data set extends from March 3, 1984 through October 17, 2006.
We commence the time series on this date because many consider early 1984 as the
beginning of the great bull market that culminated in the worldwide crash of 1987 (see
selected papers in Kamphuis et al. (1988)). From March 3, 1984 through October 17,
2006 provides a time series of 5,882 daily NYSE close-to-close observations. Since
the U.S. stock market was closed for four days subsequent to the September 11th
attack (Tuesday September 11 at 11:48 am EST through Friday September 14) we also
use an abbreviated time series of 5,877 observations to allow for an analysis of the
U.S. bond market which remained open during the week of the attack.
Our flight-to-safety analysis does not continue through 20072009 for the
following reasons: First, we do not find any large declines in daily stock prices
in 2007. The largest one-day decline in stock prices for 2007 occurs on February
27 (3.53%). The second reason has to do with the sudden and unprecedented
bank liquidity constraints and resultant Federal Reserve actions that characterized

716

J Econ Finan (2012) 36:712727

the 2008 and 2009 U.S. financial freeze-up. The events leading up to and
exacerbating the largest U.S. financial market meltdown since the Great
Depression are well known and need not be repeated here. It is sufficient to
point out that the 1984 to 2006 period of U.S. financial history was marked by
overall confidence in the payments ability of the financial system, adequate
liquidity necessary for an orderly trading and pricing of securities, and the
absence of severe, persistent, and defensive Federal Reserve intervention in
financial markets. As a result, we can safely attribute daily changes in bond
prices over the full time series interval to normal investor behavior, behavior
which can reasonably include temporary bond price spikes as they relate to
flight-to-safety behavior and the movement from stocks to bonds.
This assumption does not hold for the 20072009 period. In early 2008 and
continuing into 2009 the Federal Reserve engaged in large and continuing Open
Market purchases. This was accompanied by Fed loan and security purchases
financed by creating new reserves for the banking system. From these two actions
alone bank reserves grew from $11 billion in mid-year 2008 to approximately $900
billion by mid-year 2009an 8,081% increase. This dramatic increase in reserves
lowered interest rates to near zero levels and is responsible for the 25.87 percent
increase in long-term government bond prices in 2008.
Given this environment of massive liquidity injections into the financial system,
we conclude that our statistical model that uses daily bond price data cannot be
relied upon to discriminate between private sector flight behavior and the larger
workings of the Federal Reserve and Treasury debt market over the 20082009
period. For this reason we terminate our time series near the end of 2006.
Equity data for the U.S. uses the S&P 500 index. This index is widely recognized
as the best single indicator of the U.S. equities market and includes a representative
sample of 500 leading companies in all industrial sectors of the domestic economy
with a minimum market capitalization of $5 billion. Daily government bond indices
for U.S. Treasuries are obtained from Thomson Datastream. The Datastream
benchmark bond index uses a common maturity bond portfolio of 10+ years and
is standardized by yield, liquidity, issue size, coupon, and duration.
While equity market shocks can persist in the data beyond one trading day, we are
concerned about whether or not intra-market adjustments can be completed within
the normal operating hours of the NYSE and the major U.S. government bond
dealers in New York. Trading hours for the NYSE are 8:30 am to 4:00 pm EST; for
government bond dealers the New York trading hours extend from 7:30 am to
5:00 pm EST. The fact that NYSE trading hours are closely nested within those of
the bond markets gives ample time for equity shocks to be registered in bond prices
within the same calendar day. We thus minimize any trading anomalies that might
result from daily non-synchronous trading hours. This fact provides more reliability
for the estimated model coefficients while still allowing for persistence (or price
reversion) to continue into the next trading day.
2.2 Descriptive statistics
Table 1 gives the basic descriptive statistics of the stock and bond return data stated
as daily percentages.

J Econ Finan (2012) 36:712727


Table 1 Summary descriptive
statistics for the return data in
daily percentages (March 3,
1984 October 17, 2006)

717

Stocks

Bonds

Arithmetic average

0.37

0.31

Geometric average

0.31

0.30

Maximum value

0.80

0.40

Minimum value

0.22

0.28

Median

0.46

0.24

Standard deviation

0.10

stocks/bonds

0.40
0.096

As expected, the range of minimum and maximum daily returns for stocks is
much greater than for bonds. Likewise, the daily standard deviation for stocks is
higher. The daily correlation between stock and bond returns, shown in the last row
of the table, is quite low at 0.096. This is the result of the large random noise
component which dominates daily data. To some extent this feature is beneficial in
that it allows the flight-to-safety metrics of the regression model to be more
prominent than would be the case with observations of lower frequency.

3 Chronology of crash dates and daily percentage change analysis


Table 2 presents a chronology of crash dates and related daily percentage rates of
change in the stock and bond indexes. For the identified crash dates Table 2 also lists
the relevant economic or political event that can reasonably be related to flight
behavior. Following Gulko (2000) we define a stock market crash as one
characterized by a decline in the daily close-to-close S&P 500 index of five percent
or more. Searching the S&P 500 index time series from March 3, 1984 through
October 17, 2006 we find six days where a decline of this magnitude occurs (from
day t -1 to t).1 Between September 10, 2001 and the resumption of NYSE trading on
September 17, the percentage decline in the S&P 500 index was 4.92%. While this
four-day delayed response is slightly less than our five percent daily decline floor we
choose to include this interim four-day trading hiatus given the global importance of
the September 11th attack both financially and politically. We also report the
percentage return calculated from the close of Monday September 10 through the
halt in trading by the NYSE at 11:48 am EST on September 11. While this return is
only 0.05%, it surely would have reached the negative five percent floor had the
NYSE not terminated trading early.
An interesting comparison here is provided by the bond market. While this
market was closed on September 11 and 12, it registered a relatively large +1.55%
gain upon reopening on September 13. Clearly, this price increase reflects a flight-to1

Prior to regression estimation using these pre-identified event dates we apply the Inclan-Tsiao (1994)
variance change point method to the stock and bond return time series over the 19842006 period. This
cumulative sum-of-squares method is based upon an iterative two-step approach where potential return
breakpoints are first identified and then confirmed. We therefore let the data in conjunction with the
Inclan-Tiao algorithm determine the crash dates prior to using them in the regression model. See Smith
and Brocato (2010) and Smith and Bracker (2003) for applications of the Inclan-Tsiao technique.

+4.02%

Bonds

Sept 12Wed

NYSE closed

0.05%

Closed

Stocks

Bonds

Closed

t+1

Sept 11Tues

+0.50%

+1.66%

Jan 11Mon

Day

+5.19%

Stocks

2001 Terrorist attack

Oct 20Tues

t+1

+1.55%

NYSE closed

Sept 13Thur

+0.67%

NYSE closed

Sept 14Fri

t+3

0.53%

0.87%

t+2

+4.98%

Oct 28Tues

+0.75%

7.13%

Oct 27Mon

Asian currency crisis

1997

+2.72%

Oct 16Mon

6.31%
+1.51%

7.00%

Bonds

1.22%

22.90%

+1.11%

Stocks

Oct 13Fri

Jan 8Fri

Leveraged buyout of UAL

1989

Oct 19Mon

Crash of global markets

Day t

1988

1987 global crash aftershock

1987

+1.92%

4.92%

Sept 17Mon

t+6

% is from Sept 10
through Sept 17

0.13%

+3.79%

Sept 1Tues

+0.34%

7.04%

Aug 31Mon

Russian debt, LTCM crisis

1998

Table 2 Five percent or greater stock market decline dates, event activity, and daily percentage index changes, days t (the event day) and t+1

0.99%

0.52%

Sept 18Tues

t+7

% is from Sept 17
through Sept 18

0.82%

+3.25%

Apr 17Mon

+0.44%

6.01%

Apr 14Fri

Tech stock decline

2000

718
J Econ Finan (2012) 36:712727

J Econ Finan (2012) 36:712727

719

safety even though the liquidity could not come from the sale of NYSE stocks. For
the remaining three days of trading following the attack and the day of resumed
stock market trading (September 17th) the bond market appears to have acted as a
safe harbor with gains on September 13th and September 17th.
To record any persistence or reversion in the returns following an event day, t,
Table 2 also presents the percentage returns from the close of the event day, t, to the
close of the subsequent trading date, t +1, for both stocks and bonds.
Table 2 shows all stock market crashes over the March 3, 1984October 17, 2006
period occur either on Monday or Friday. This finding is consistent with the
empirical finding that information contagion spawned during weekends when
markets are closed can suddenly be reflected in global stock prices as markets reopen chronologically carrying new information (see Schwert 1990 and Roll 1988).
Flight-to-safety behavior is also demonstrated. Specifically, Table 2 shows, with
one exception (Friday January 8), negative returns on day t in stocks are matched
with positive returns in bonds on the same day. While the bond returns are more
muted, the positive signs support flight behavior as investors flee from equities to the
safety of government Treasuries. With the data used in this study we are not able to
fully explain why the percentage values of stocks and bonds seen in Table 2 are not
closer in absolute value. A number of explanations might be offered as they relate to
investments in alternative instruments. Perhaps the most obvious explanation is that
a sudden flight from stocks is immediately side-lined in cash before the funds are
used to purchase government debt securities.2
Table 2 also reveals another interesting feature of investor behavior. We see that
for all event days t from 1987 through 2000 stock returns rebound with a positive
but muted return on the subsequent trading day t+1. We can offer no apparent
theoretical reason for this result. However, it is consistent with a flight-to-safety
mentality that adjusts to equity shocks after new information is absorbed while the
stock market is still closed. Wall Street Journal writer, John Parry, describes this
behavior in more casual terms, When stocks are expected to show weakness,
investment funds often flow to the perceived haven of the bond market, with that
shift usually going into reverse when, (as yesterday), equities start to strengthen.
(Wall Street Journal, August 1, 1001, page C1). Table 2 also shows some (mild)
negative returns in the bond market following the day of a stock market crash, t + 1.
These return data suggest funds flow fairly quickly out of bonds and back into
equities after a stock market decline.3

2
Neither U.S. corporate grade debt nor foreign government sovereign debt are included in the model
estimated in this paper. Informal statistical tests suggest that these alternative assets do not provide a
temporary safe haven for equity investors during the crashes studied. These tests are available from the
authors upon request.
3
Since Table 2 only reports data for days t and t+1 it omits three equity decline dates in what was a
reverberating sequence of crashes in October 1987. On Friday October 16, the S&P 500 index declined by
5.29 percent, while the U.S. government bond market index increased by 0.11 percent. On Monday,
October 26, U.S. equities declined by 8.64 percent, while government bonds increased by 1.58 percent
(see Table 3 for supporting regression evidence). Finally, on Tuesday, October 27 U.S. equities increased
by 2.39 percent while bonds decreased by 1.19 percent. Checking the daily data through Monday
November 2, we find no other significant positive or negative percentage changes in either the stock or
bond indexes.

720

J Econ Finan (2012) 36:712727

In general, the crash dates seen in Table 2 show that daily U.S. stock market
movements are quite sensitive to global economic events. For example, the 1987
world crash continued to have contagion and over-reaction effects in the U.S.
equities market lasting well into 1988 (Telser 1988; Roll 1988). The failure of the
United Airlines $6.75 billion leveraged buyout on October 13, 1989 is the probable
cause the S&P index decline of 6.31%. The tremendous surge in capital inflows to
the Asian economies in 19951997, much of it from U.S. investors, along with
excessive risk-taking, failed government policies, and the collapse of the Thai baht
in July 1997, can be cited as chief causes of Asias financial crisis. These factors
spilled over to the U.S. equities market in 1997 (Sacks and Radelet 1998 and
Moreno et al. 1998). In 1998, the U.S. equity market was rocked in close succession
by the Russian bond default in August of 1998 and then the Federal Reservefacilitated recapitalization of the Long Term Capital Management hedge fund in
September. Dungey et al. (2006) show that the close proximity of these two events
induced an unusual period of volatility in international bond markets that quickly
spread to global equity markets. They point out that the U.S. equity market was
spared somewhat from these combined events since the U.S. had minimal exposure
to Russian debt in combination with the fact that the Federal Reserve acted
aggressively in easing monetary policy following the LTCM recapitalization
announcement. While the September 11th attack was a strictly U.S.-based shock,
the contagion literature shows that the aftershocks continued to impact all global
markets including the U.S. equity market (Mun 2005; Nikkinen et al. 2008).

4 Model and test methodology


To more formally quantify crash day flights-to-safety within a crash window we use
an altered version of dummy variable regression developed by Dufour (1980). As
Dufour explains this approach to binary variable OLS provides a computationally
convenient method of coefficient estimation when interest is on one observation only
as is the case here. In a regression framework Dufour shows that to test for a
structural break in the slope coefficient with the null hypothesis H0: 1 =2 = is
equivalent to a test of the intercept H0: ij =0 in a model that uses dummies for
selected individual observations.4 This allows us to target specific days surrounding
a particular crash episode thereby bringing into sharper relief daily bond market
flight behavior before, on, and after a given crash date. Such investor behavior is
masked by Gulkos data pooling event methodology.
The total return series for stock and bond returns is constructed by computing for
each market the time series equation;
lnRt  lnRt1

While other methodologies are available in the empirical finance literature to test for time-varying
correlations (e.g., ARCH, DCC, copulas, Markow switching, etc.), we choose the Dufour dummy variable
technique because the estimated intercept shift estimates allow for straightforward economic and statistical
interpretation of investor crash response we are interested in.

J Econ Finan (2012) 36:712727

721

where Rt and Rt-1 represent the daily index values for either the stock or bond market
on day t and t 1, respectively. Return calculations are based upon daily close-toclose index values.
The crash day is period t as described in Table 2. Our crash days and associated
event windows are defined as follows. Each of the crash days is centered between
the five preceding and five subsequent trading days. This produces an eleven day
trading sequence for each window, t - 5, t - 4,... t - 1, t, t +1, t +2,... t+5. As Table 2
shows there are seven windows over which the flight-to-safety phenomenon is
tested.
The model is;
Yt b Xt lij Dijt "t

where Yt =the daily rate of return for the government bond index for day t, Xt =the
daily rate of return for the S&P 500 stock index for day t, and t =an error term. Tests
on t show it to follow the classical properties, [N(0,2)], with a standard deviation
of 0.004. The long-run slope coefficient between bonds and stocks over the 1984
through 2006 period is given by the estimate . A priori we expect the sign of this
estimate to be positive.
The coefficient ij is intended to capture the extent (sign and magnitude) of flightto-safety behavior for a given day j within an event window i. Since we are
interested in whether flight behavior is significantly different from zero we estimate
Eq. 2 without an intercept term.
We define the dummy variables, Dijt, as follows:
i=1... 8, where i represents a particular crash window over the full time series
from March 3, 1984 through October 17, 2006.
j=a particular day within each crash window such that j=t - 5, t - 4,... t...t+4, t+5.
t=a time subscript extending over the full time period for each observation.
For example, for a given day we have:
Dijt 1 if the day including the event day falls within window i
0 if the day is not within window i
Substituting Eqs. 3a and 3b into Eq. 1 and taking expected values gives

E Yt jDijt 0 b Xt

E Yt jDijt 1 b Xt lij

3a
3b
4a
4b

where ij, is our numerical flight-to-safety estimate for a given day j within a given
event window i.
For this paper the null hypothesis is that there is no flight from stocks to bonds for
a given crash day, i.e., H0: ij =0; the alternative hypothesis is H1: ij >0. Using a
two-tailed test we examine this hypothesis in the usual manner by determining if, on

722

J Econ Finan (2012) 36:712727

the basis of a t test, an estimated ij coefficient is positive and statistically


significant.5 A rejection of the null hypothesis (i.e., ij >0) provides evidence of a
move from stocks to bonds on a particular day and is our statistical proxy for
Gulkos OLS finding that the slope coefficients in his pooled event study regressions
change from positive to negative (i.e., decouples) during periods of crash behavior
(2002, 61). However, our interpretation varies from Gulkos in that we identify flight
behavior on specific crash days over the event intervals rather than draw inferences
from aggregated data.

5 Results
Estimation of Eq. 2 from March 3, 1984 through October 17, 2006 (excluding the
four observations from September 11thSeptember 14th) gives the following general
results:

b 0:0579t  value 9:659


Adjusted R2 0:0404
D:W: 1:90
The slope coefficient is positive and statistically significant showing that bond
and stock returns are directly related over the long-run, a finding that is consistent
with most empirical research. However, the coefficient is quite small. This is
expected given the fact that a large component of the variation in daily stock and
bond returns is made up of random noise. The low adjusted R2 value also reflects the
high random noise component in the data. The null hypothesis of no autocorrelation
in the residuals is supported by the Durbin-Watson statistic at the five percent level.
In order to determine the statistical importance of the model with and without the
crash day dummies we test with a restricted F-test the null hypothesis that the daily
dummy variable coefficients, ij, in Eq. 2 are jointly equal to zero. This test produces
F82,5794 =4.15 which is significant at the one percent level. Thus we can reject the
null hypothesis that the crash day dummies as a group do not add explanatory power
to the model as specified in Eq. 2.
Table 3 provides evidence on the intercept coefficients, ij, for each of the seven
crash windows and the five days preceding and the five days subsequent to each of
the crash dates. Probable macroeconomic and/or political events are repeated from
Table 2. The intercept coefficients, ij, for the five percent or greater crash day
declines are given at the column event date headings from t 5 through t+5 and are
the same as those listed in Table 2. The M, T, W, R, F indicators in each cell of
Table 3 refer to Monday, Tuesday, Wednesday, etc. In each cell, t - statistics are
reported directly below each dummy variable coefficient. Cells containing
statistically significant shift dummy coefficients are highlighted by bold shading.
The term na in the 2001 row indicates the days that the NYSE was closed.

5
The model is estimated without slope dummies. Tests for changes in the slope coefficient using dummy
variables produced statistically insignificant results.

J Econ Finan (2012) 36:712727

723

In general Table 3 shows we can reject the null hypothesis of no flight-to-safety


with one exception the crash days show positive and statistically significant shift
dummies indicating a material movement into government Treasury bonds on
October 19, 1987, October 13, 1989, October 27, 1997, August 31, 1998, April 14,
2000, and after the reopening of the NYSE on September 17, 2001. Also
noteworthy is the reverse movement out of bonds following a crash day, t+1.
This reversion pattern is absent on October 20, 1987 suggesting investors required
more than one trading day to reduce their equity holdings. In general, however,
our findings support the previous observation that crash-generated funds typically
flow quickly from bonds back into stocks as new information is processed after
a crash.
Of the seven crash events, clearly the 1987 global stock market crash shows the
most persistent bond movement activity with continued statistically significant

Table 3 Crash window intercept shift estimates (ij) from Eq. 2. These dates correspond to those described
in Table 2. The first row shows the five days prior to the event date, t-5, the event date t, and the five days
after the event date, t+5. The M, T, W, R, F in each cell refers to Monday, Tuesday, etc. The na in the 2001
row indicates that the NYSE was closed the week of the terrorist attack. Statistically significant dummy
coefficients are highlighted by shading. The last two cells show the intercept coefficients when the times
series excludes Sep 11 (T) through Sept 14 (F). We include shift dummy coefficients for the re-opening of
the NYSE (Sept 17) and the subsequent trading day (Sept 18). *** indicates significant at the 0.01 level, **
indicates significant at the 0.05 level, * indicates significant at the 0.10 level
event dates

t-5

t-4

t-3

t-2

t-1

t+1

t+2

t+3

t+4

t+5

1987
worldwide
crash

Oct
12/M

Oct
13/T

Oct
14/W

Oct
15/R

Oct
16/F

Oct
19/M

Oct
20/T

Oct
21/W

Oct
22/R

Oct
23/F

Oct
26/M

-0.0045
(-1.039)

0.0066
(1.539)

0.0022
(0.515)

0.0255***
(5.894)

-0.0009
(-0.207)

0.0210***
(4.848)

Dec
31/R

Jan
4/M

Jan
5/T

Jan
6/W

Jan
7/R

Jan
8/F

Jan
11/M

Jan
12/T

Jan
13/W

Jan
14/R

Jan
15/F

-0.6000
(-1.399)

0.0007
(0.160)

0.0045
(0.733)

-0.0084
(-1.507)

0.0024
(0.670)

-0.0072
(-1.297)

0.0052
(0.915)

0.0022
(0.393)

0.0047
(0.847)

0.0039
(0.559)

0.0196***
(2.787)

Oct
6/F

Oct
9/M

Oct
10/T

Oct
11/W

Oct
12/R

Oct
13/F

Oct
16/M

Oct
17/T

Oct
18/W

Oct
19/R

Oct
20/F

0.0003
(0.0807)

0.008**
(1.956)

-0.0021
(-0.497)

-0.0025
(-0.587)

0.0020
(0.471)

0.0188*** -0.0103** -0.0026


(4.369)
(-2.394) (-0.597)

0.0007
(0.168)

0.0033
(0.765)

-0.0012
(-0.280)

1997
Asian Crisis

Oct
20/M
0.0007
(0.161)

Oct
21/T
-0.0005
(-.126)

Oct
22/W
0.0016
(0.377)

Oct
23/R
0.0077*
(1.780)

Oct
24/F
0.0032
(0.741)

Oct
27/M
0.0127***
(2.715)

Oct
28/T
-0.0083*
(-1.927)

Oct
29/W
0.0052
(1.206)

Oct
30/R
0.0060
(1.393)

Oct
31/F
0.0007
(0.158)

Nov
3/M
-0.0065
(-1.513)

1998
Russian debt
default/LTCM

Aug
24/M

Aug
25/T

Aug
26/W

Aug
27/R

Aug
28/F

Aug
31/M

Sep
1/T

Sep
2/W

Sep
3/R

Sep
4/F

Sep
7/M

0.0019
(0.435)

0.0038
(0.878)

0.0039
(0.903)

0.0087**
(2.026)

0.0034
(0.784)

0.0075*
(1.760)

-0.0026
(-0.619)

-0.0023
(-0.530)

0.0048
(1.103)

0.0021
(0.489)

-0.0014
(-0.314)

2000
Tech decline

Apr
7/F
0.0044
(1.013)

Apr
10/M
0.0048
(1.103)

Apr
11/T
-0.0059
(-1.363)

Apr 12
W
-0.0050
(-1.167)

Apr
13/R
0.0036
(0.840)

Apr
14/F
0.0079*
(1.844)

Apr
17/M
-0.010***
(-2.344)

Apr
18/T
-0.0063
(-1.453)

Apr
19/W
0.0070
(1.627)

Apr
20/R
-0.0004
(-0.096)

Apr
21/F
0.0007
(0.151)

2001
September 11
attack

Sep
4/T

Sep
5/W

Sep
6/R

Sep
7/F

Sep
10/M

Sep
11/T

Sep
12/W

Sep
13/R

Sep
14/F

Sep
17/M

Sep
18/M

-0.013***
(-2.916)

0.0011
(0.262)

0.0100**
(2.324)

0.0061
(1.426)

-0.0032
(-0.733)

na

na

na

na

1988
worldwide
aftershock

1989
UAL
buyout failure

-0.014*** -0.0038
(-3.190) (-0.877)

0.0042
(0.970)

0.0248*** 0.0371***
(5.4207) (8.584)

0.0221*** -0.0095**
(5.118)
(-2.215)

724

J Econ Finan (2012) 36:712727

upward shifts in the regression intercept lasting for most of the five days following
the initial crash on October 19th. This evidence of continued flight-to-safety
behavior reflects the severity of the impact of this worldwide equity decline and the
contagious nature of the information aftershocks that accompanied this event. Most
of the research in this area attributes the 1987 crash to the bursting of a speculative
bubble (Kamphuis et al. 1988; Chowdhury and Lin 1993); that the crash was a
rational response to the slow build-up of irrational world assets prices that continued
to move ahead of fundamentals and where the post-crash process of information
revision continued to elicit ongoing global equity market volatility (Telser 1988;
Poterba and Summers 1998; Roll 1988).
Table 3 does show some anomalous results. For example, given the 7.00% decline
in the S&P 500 on Friday January 8, 1988 (see Table 2), we cannot explain why the
shift dummy does not register a positive and statistically significant movement into
bonds. Table 2 also shows a large percentage decline in the bond index on this date.
Perhaps both markets were rattled by the severity of the decline in global equity
prices such that investors sought refuge in crash.6 Table 3 also shows six of the cells
preceding the crash days (the so-called clean days) having statistically significant
movement into or out of bonds. Saha et al. (2009) report a robust finding that
extreme movements in stock prices are often preceded by large average daily price
movements during the prior three days. Our results are consistent with these
findings.
Tables 2 and 3 allow for additional comparisons. While the numerical magnitudes
in these two tables are not directly comparable, the algebraic signs are comparable
and should provide some measure of concurrence. This is the case for most dates.
For example, we see negative percentage changes in stocks in period t (Table 2)
reflected in positive intercept coefficient shifts in period t (Table 3), with the
exception of January 8, 1988. Further, for period t+1 we see a concurrence of signs
with positive (negative) percentage changes in bond returns the day after a crash
(Table 2) reflected in positive (negative) intercept shifts (Table 3) for period t +1.
This feature supports our previous finding on flight-to-safety behavior, one that
shows a tendency for investors to quickly revert back to equities the day after a
crash.

6 Pre- and post-measures of stock and bond return correlations and volatility
Existing research strongly indicates that financial shocks alter asset return
correlations and volatility from their pre-shock values (Campbell et al. 2002; Solnik
et al. 1996; Schwert 1989; Hamao et al. 1990). Applying this general belief to our
data, Tables 4 and 5 present pre-and post-crash stock and bond return correlations
and standard deviations, respectively. Pre-crash data are obtained by splicing the
daily close-to-close stock and bond returns from days t - 5 through t - 1 for each

6
We point out that the previously-cited Inclan-Tsiao procedure does identify a shift breakpoint with a
statistically significant increase in the variance of equity returns on this date.

J Econ Finan (2012) 36:712727

725

Table 4 Pre- and post-crash stock and bond return correlations


Stocks PRE

Bonds PRE

Bonds PRE

0.1443

Stocks POST

0.1629

0.5562

Bonds POST

0.0934

0.5155a

Stocks POST

0.1976a

Stocks PRE/Bonds PRE correlation is significantly different from the Stocks POST/Bonds POST
correlation at the 5% level under the null hypothesis that both correlations are equal. Likewise, the Bonds
PRE/Bonds POST correlation is significantly different from the Stocks PRE/Stocks POST correlation at the
5% level. The statistic used to test these hypotheses is computed using the Fisher r-to-z transformation.
For a description of this statistic see Hays and Winkler (1971, 653)

of the crash windows from 1987 through 2006. Post-crash returns are obtained in
the same manner for days t+1 through t+5. This provides a spliced sequence of
35 stock and bond return observations prior to and subsequent to crashes (for both
stocks and bonds we omit the days following the September 11th attack since the
NYSE was closed). In the tables the acronyms PRE and POST represent
observations prior to and subsequent to crash dates, respectively.
Table 4 shows that prior to crash dates stock and bond returns share a positive
correlation (0.1443). However, following crash dates the sign become negative
(0.1976) and the difference in correlations is statistically significant at the five
percent level. In a related fashion we see pre-crash stock returns negatively
correlated with post-crash bond returns (0.0934). These negative correlations
suggest that portfolio diversification is enhanced during crash episodes. We also see
that stock returns become more highly (positively) correlated after crash days while
the correlation between pre-and post-bond returns is significantly different with a
value of 0.5155. Collectively, these findings support our flight hypothesis.
Table 5 shows pre- and post-crash stock and bond standard deviations and
associated F-values. As expected the volatility of stock returns rises during postcrash days. This is a statistically significant volatility increase (the five percent
critical F-value is F34/34 =1.84). The volatility for bond returns rises after crashes and
is statistically significant at the ten percent level. In general, Tables 4 and 5 support
the conclusion that extreme and sudden declines in equity values result in altered
return correlations that are different in both magnitude and sign. Additionally, return
volatility increases as well.

Table 5 Pre- and post-crash


stock and bond return standard
deviations and F-values

stocksPRE

bondsPRE

0.0143

0.0056

stocksPOST

bondsPOST

the critical F34/34 =1.84 at the


5% level

0.0359

0.0091

F34/34 value

F34/34 value

2.51a

1.63b

the critical F34/34 =1.61 at the


10% level

726

J Econ Finan (2012) 36:712727

7 Summary and conclusion


Using an altered version of the Gulko (2002) decoupling model we test the flight-tosafety hypothesis between March 3, 1984 and October 17, 2006. Unlike Gulko we
use a dummy variable regression model suggested by Dufour (1980) which allows
us to test individual stock/bond shifts for specific equity market crash dates. The
preponderance of evidence for the seven crash episodes in addition to percentage
change analysis allows us to reject the hypothesis of no daily flight-to-safety.
Additionally, with the exception of the world crash in October 1987, our tests show
flight episodes are short-lived with statistically significant movement out of bonds
and back into stocks the day after a crash. Pre-and post- correlation analysis supports
our findings that extreme and sudden stock market declines alter return correlations
and volatility. Specifically, we find that while stock and bond returns are positively
correlated for the five days prior to a crash, they turn statistically negative for the
five days subsequent to a crash. Likewise we find increased post-crash volatility in
both stock and bond returns lasting at least five days after an equity market decline.
Logically the reader might ask why the authors have not included data subsequent
to 2006 so as to investigate flight behavior over this period of unprecedented
financial market turbulence in U.S. macroeconomic history. As pointed out earlier,
we do not feel that financial market activity in 20082009 is directly comparable
such that flight-to-safety estimates would be meaningful relative to the 19842006
period. A study of these unique 20082009 financial market disruptions and how
investor flight-to-safety behavior was impacted by uncertainty and lack of liquidity
provides a rich topic for further research.

References
Brocato J, Steed S (1998) Optimal asset allocation over the business cycle. Financ Rev 33:129148
Campbell R, Koedijk K, Kofman P (2002) Increased correlation in bear markets. Financ Anal J 58:8794
Chowdhury M, Lin J (1993) Fads and the crash of 87. Financ Rev 28:385401
Dopfel F (2003) Asset allocation in a lower stock-bond correlation environment. J Portf Manage 2528
Dufour J (1980) Dummy variables and predictive tests for structural change. Econom Lett 6:241247
Dungey M, Fry R, Gonzalez-Hermosillo B, Martin V (2006) Contagion in international bond markets
during the Russian and the LTCM crises. J Financ Stab 2:127
Engle R (2002) Dynamic conditional correlation: a simple class of multivariate generalized autoregressive
conditional heteroskedasticity models. J Bus Econ Stat 20:339350
Guidolin M, Timmermann A (2005) Economic implications of bull and bear regimes in UK stock and
bond returns. Econ J 115:111143
Gulko L (2002) Decoupling. J Portf Manage 5966
Hamao Y, Masulis R, Ng V (1990) Correlation in price changes and volatility across international stock
markets. Rev Financ Stud 3:281307
Hays W, Winkler R (1971) Statistics: probability, inference, and decision. Holt, Rinehart & Winston, New
York, p 653
Ilmanen A (2003) Stock-bond correlations. J Fixed Income 5566
Inclan C, Tsiao G (1994) Use of cumulative sum of squares for retrospective detections of changing
variance. J Am Stat Assoc 89:913923
Kamphuis R, et al. (ed) (1988) Black Monday and the future of financial markets. Dow Jones-Irwin Inc
Mun C (2005) Contagion and impulse response of international stock markets around the 911 terrorist
attacks. Glob Financ J 16:4868
Moreno R, Pasadilla G, Remolona E (1998) Asias financial crisis: lessons and policy responses. In: Asia:
responding to crisis. Asian Development Bank Institute

J Econ Finan (2012) 36:712727

727

Morningstar, Inc (2007) SBBI Yearbook


Nikkinen J, Omran M, Sahlstrom P, Aifo J (2008) Stock returns and volatility following the September 11
attacks: evidence from 53 equity markets. Int Rev Financ Analy 17:2746
Poterba J, Summers M (1998) Mean reversion in stock prices: evidence and implications. J Financ Econ 2759
Roll R (1988) The international crash of October 1987. In: Kamphuis R et al. (ed) Black Monday and the
future of financial markets. Dow Jones Irwin Inc
Sacks J, Radelet S (1998) The onset of the east Asian financial crisis. In: Harvard Institute for
International Development (working paper)
Saha A, Malkiel B, Grecu A (2009) The clustering of extreme movements: stock prices and the weather. J
Invest Manag 7:115122
Schwert G (1989) Why does stock market volatility change over time? J Financ 44:11151153
Schwert G (1990) Stock volatility and the crash of 87. Rev Financ Stud 3:77102
Scruggs J, Glabadanidis P (2003) Risk premia and the dynamic covariance between stock and bond
returns. J Financ Quant Anal 38:295316
Smith KL, Bracker K (2003) Forecasting changes in copper futures volatility with GARCH models using
an iterative algorithm. Rev Quant Financ Account 20:245265
Smith KL, Brocato J (2010) Applying the Inclan-Tsiao breakpoint algorithm in the search for the flight-tosafety phenomenon. Appl Financ Econ 20(5):371380
Solnik B, Boucrelle C, LeFur Y (1996) International market correlation and volatility. Financ Anal J 1734
Telser L (1988) October 1987 and the structures of financial markets: An exorcism of demons. In:
Kamphuis R, et al. Black Monday and the future of financial markets. Dow Jones-Irwin Inc

Copyright of Journal of Economics & Finance is the property of Springer Science & Business Media B.V. and
its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's
express written permission. However, users may print, download, or email articles for individual use.

You might also like