Journal of Financial Markets 25 (2015) 3351

Contents lists available at ScienceDirect

Journal of Financial Markets

journal homepage: www.elsevier.com/locate/finmar

Equity volatility as a determinant of future

term-structure volatility$
Naresh Bansal a,n, Robert A. Connolly b,1, Chris Stivers c,2
a

John Cook School of Business, Saint Louis University, St. Louis, MO, United States
UNC Kenan-Flagler Business School, University of North Carolina Chapel Hill, Chapel Hill, NC, United States
c
College of Business, University of Louisville, Louisville, KY, United States
b

a r t i c l e in f o

a b s t r a c t

Article history:
Received 28 September 2013
Received in revised form
8 May 2015
Accepted 13 May 2015
Available online 22 May 2015

We show that equity volatility serves as a determinant of future

Treasury term-structure volatility over the recent October 1997 to
June 2013 period. We nd that equity volatility contains incrementally reliable information for the subsequent volatility of: (1)
10-year and 30-year bond futures returns, (2) the term-structure's
level, and (3) the term-structure's slope. We present additional
evidence that suggests a ight-to-quality/ight-from-quality pricing avenue is a likely contributor to the volatility linkages, where
time-varying economic uncertainty can generate both a large
positive serial correlation in stock volatility and a time-variation in
the precautionary savings motive and diversication benets of
holding bonds.
& 2015 Elsevier B.V. All rights reserved.

JEL classication:
G12
G14
Keywords:
Equity risk
Term structure
Bond volatility

1. Introduction
Understanding term-structure volatility is a fundamental issue in nancial economics with both
theoretical and practical importance. In this paper, we show that realized equity volatility can serve as
an important determinant of future term-structure volatility. By term-structure volatility, we refer to

We thank Andrew Lim and seminar participants at the 2013 Financial Management Association meetings, the 2013
Midwest Finance Association meetings, and the 2013 Missouri Economics Conference for helpful comments. We also gratefully
acknowledge an anonymous referee who provided thoughtful comments that improved the paper.
n
Corresponding author. Tel.: 1 314 977 7204.
E-mail addresses: nbansal@slu.edu (N. Bansal), Robert_Connolly@unc.edu (R.A. Connolly),
chris.stivers@louisville.edu (C. Stivers).
1
Tel.: 1 919 962 0053.
2
Tel.: 1 502 852 4829.

http://dx.doi.org/10.1016/j.nmar.2015.05.002
1386-4181/& 2015 Elsevier B.V. All rights reserved.

34

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

the volatility of Treasury bond futures returns, and the volatility of both the level and slope of the
Treasury term-structure. By determinant, we refer to an intertemporal relation between the lagged
realized equity volatility and the subsequent bond-market volatility that holds in a multivariate
framework, when also controlling for the past term-structure volatility and other term-structure state
variables.
Researchers have offered both theory and empirical evidence that suggest important linkages
between equity risk and the Treasury bond market. For example, Bekaert, Engstrom, and Xing (2009)
nd that higher economic uncertainty can lead to both higher equity volatility and an increased
motive for precautionary savings that can depress interest rates. Fleming, Kirby, and Ostdiek (1998)
and Kodres and Pritsker (2002) suggest cross-asset-class effects tied to hedging and portfolio
rebalancing. Pricing effects linking the stock and bond markets have been attributed to ight-toquality/ight-from-quality (FTQ/FFQ), where some investors (presumably) switch between riskier
stocks and safer Treasuries as risk perceptions change (Connolly, Stivers, and Sun, 2005, 2007;
Underwood, 2009; Baele, Bekaert, and Inghelbrecht, 2010; BenRaphael, Kandel, and Wohl, 2012;
Jubinski and Lipton, 2012; Bansal, Connolly, and Stivers, 2014).3 Chordia, Sarkar, and Subrahmanyam
(2005) nd that innovations to stock volatility forecast an increase in bond bidask spreads.
Why might the realized equity volatility contain important incremental information for the
subsequent bond volatility? First, consider a FTQ/FFQ avenue, as motivated by the literature cited
above. With linkages between the economic state and stock volatility, a higher stock volatility this
month is likely to be associated both with more extreme stock price movements over the next month
(volatility clustering), and with higher economic uncertainty and volatility in that uncertainty (stock
volatility tending to be higher in stressful economic times with greater economic-state uncertainty). If
a higher stock-return volatility and a higher time series variability in economic uncertainty are likely
following months with a high realized stock volatility, then the likelihood of FTQ/FFQ pricing
inuences over the subsequent month is presumably much greater.4 Second, the return volatility of
both equities and bonds may be responding to some omitted factor or news that bears on the
volatility of each asset class, in the sense of Fama and French (1993). If there is volatility clustering in
that common factor, then equity volatility may be providing an additional signal about the underlying
volatility environment for subsequent bond returns.5
Our empirical investigation is also motivated by Andersen and Benzoni's (2010) ndings. Under
standard afne term structure models, they note that the instantaneous yield volatility should be
spanned by the cross-section of yields. They nd evidence inconsistent with this prediction and
conclude that a broad class of afne diffusive, quadratic Gaussian, and afne jump-diffusive models
cannot accommodate the observed yield volatility dynamics, (p. 603). Their ndings suggest that
factors outside the bond market are likely to be important for understanding yield volatility. In this
paper, we examine the role of equity volatility as one potential factor.
We focus on the October 1997 to June 2013 period since the literature indicates a clear change in
the joint distribution of stock and bond returns around October 1997. Fig. 1 in Baele, Bekaert, and
Inghelbrecht (2010, p. 2376) depicts the shift in the stock-bond correlation from sizably positive to
predominantly negative in the latter part of 1997. Bansal, Connolly, and Stivers (2014) argue that the
equity-risk dynamics and ight-to-quality pricing inuences may be particularly important for
understanding bond market dynamics over the post-1997 period since this period features a
predominantly negative stock-bond-return correlation, a low ination-risk environment, and several
episodes of high and volatile equity risk. We also briey examine an earlier period in the mid-1990's

3
Some authors use the phrase ight-to-safety, rather than ight-to-quality. For the purposes of our study, we consider
these terms as interchangeable.
4
We focus on volatility measures over the monthly horizon, but also evaluate the quarterly horizon.
5
See Fleming, Kirby, and Ostdiek (1998), for an alternate discussion on the intuition behind these two avenues for a stockbond volatility linkage. In their model, two distinct sources of linkages arise. One is common information, such as news about
ination, which simultaneously affects investor expectations in multiple markets. The second source is due to cross-market
hedging. When information alters expectations in one market, traders adjust their holdings across markets, producing an
information spillover, (p. 135). In our view, their cross-market hedging and our FTQ/FFQ capture a similar perspective and their
common information is similar to our omitted common factor perspective.

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

35

to evaluate an alternate period with low equity risk and a sizably positive stock-bond return
correlation.6
We analyze the volatility of returns on long-term (30-year) and medium-term (10-year) Treasury
futures contracts, the volatility in the change in term structure's level, and the volatility in the
change in term structure's slope. The volatility of long-term and medium-term bond returns may be
attributed to changes in the level of yields or the slope of the yield curve. Because principal
component representations of the yield curve are orthogonal by construction, we can identify the
separate effects of equity volatility on the principal components. Following standard practice in the
term structure literature, we measure the change in the term structure's level as the change in the rst
principal component (PC1) of the term structure, and the change in the term structure's slope as the
change in the second principal component (PC2) of the term structure.7 Our term-structure volatility
measures are based on either daily returns (for the 30-year and 10-year futures contracts) or daily
changes (for the principal components) over rolling one-month and one-quarter periods. For the
lagged equity volatility, we use the lagged realized stock volatility over a one-month period, as
calculated from either past daily S&P 500 futures returns or past 5-minute returns on the SPY (S&P
500) Exchange Traded Fund.
To summarize our primary empirical results, we nd that lagged equity volatility is a substantial,
reliable determinant of the subsequent T-bond and T-note futures return volatility and of the
subsequent volatility of the level and slope of the term-structure. The information content of equity
volatility is incremental in nature, in the sense that we control for the volatility information contained
in the lagged realized term-structure volatility and other term-structure state variables (e.g., Andersen
and Benzoni, 2010; Cochrane and Piazzesi, 2005). The intertemporal aspect of our ndings supports
the notion that equity risk can help us to understand movements in the term-structure, beyond an
approach that only looks at the bond market in isolation.
We also provide additional evidence to probe the underlying mechanisms behind the
intertemporal stock-to-bond volatility (ISBV) relation. We nd evidence consistent with FTQ/FFQ
dynamics being a key contributor to the ISBV relation. For example, we nd that the ISBV relation is
linked to the economic state, with a much stronger ISBV relation in stressful uncertain economic times
(such as around recessions). Further, we nd that the partial ISBV relation remains strong when
controlling for the lagged volatility of economic variables such as ination and the default yield
spread, variables that seem likely to be more linked to bond volatility but might also be embedded in
equity volatility.
The paper proceeds as follows. Section 2 describes our data and sample selection. Section 3
presents our main empirical results. Sections 4 and 5 present additional evidence that bears on
understanding the underlying mechanisms behind our ISBV ndings. Section 6 discusses our ndings
in relation to earlier empirical studies on stock-bond volatility linkages, and Section 7 concludes.

2. Data description and sample selection

2.1. Data description
We investigate the relation between lagged equity volatility and three dimensions of the realized
term-structure volatility over the next month. First, we investigate the volatility of Treasury bond
returns, measured by the daily returns on Treasury futures contracts for two different maturities, the
10-year T-note futures contract (medium-term bond) and the 30-year T-bond futures contract (longterm bond). By futures returns, we refer to the daily price change (close-to-close, based on the daily
6
Fleming, Kirby, and Ostdiek (1998) and Chordia, Sarkar, and Subrahmanyam (2005) consider stock and bond volatility
linkages, but with data only through 1995 and 1998, respectively, and with substantially different empirical approaches. In
Section 6, we discuss our ndings in relation to these earlier studies.
7
Researchers have shown that the term-structure's rst three principal components are closely related to its level, slope,
and curvature, respectively, and capture almost all of the variation in the yields. Diebold, Piazzesi, and Rudebusch (2005) nd
that the rst two principal components alone account for almost all (99%) of the variation in the yields.

36

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

mark-to-market contract price) divided by the preceding day's closing price, as obtained from
DataStream's continuous futures series. Second and third, we investigate the volatility of the termstructure's level and the slope, based on the daily changes in the term-structure's rst principal
component (PC1) and second principal component (PC2), respectively. In our principal component
analysis, we compute the rst three principal components for every trading day from the 10 zerocoupon-bond yields from year-one to year-ten, using the instantaneous continuously-compounded
forward-rate yields as described in Gurkaynak, Sack, and Wright (GSW) (2007).
Following from Andersen and Benzoni (2010), we also use the lagged values of the three principal
components (from day t  1) as term-structure state variables when modeling the subsequent termstructure volatility over trading days t to t j. In robustness testing, we also use the lagged GSW
instantaneous continuously-compounded forward-rate yields at 1 year, 2 years, 3 years, 5 years, 7 years,
and 10 years out as alternate term-structure state variables, following from Cochrane and Piazzesi (2005).
For the stock volatility, we examine two alternate measures. First, to treat the stock volatility and
term-structure volatility symmetrically, we calculate the realized stock volatility from daily returns
from S&P 500 futures contracts. Second, as an alternate measure of realized monthly stock volatility,
we use the standard deviation of 5-minute returns of the SPY (S&P 500) Exchange Traded Fund (ETF).
By calculating the realized volatility from 5-minute returns, our volatility measure captures more
information so we should end up with a higher quality volatility estimate over the respective time.8
Note, in both our measures of lagged stock volatility, the lagged realized stock volatility is calculated
from the past stock returns over trading days t  1 to t  22, relative to the subsequent term-structure
volatility that is measured over trading days t to t 21. In an Online Appendix, we provide details on
constructing the realized volatility from the 5-minute SPY returns.
Aspects of our empirical investigation also use: (1) the Chicago Board Option Exchange's Volatility
Index (VIX), dened as the implied volatility from S&P 500 equity-index options standardized for a
one-month expiration; and (2) ination compensation data, based on yield differences between 10year TIPS and 10-year nominal Treasuries per GSW (2010).
2.2. Sample selection
In our introduction, we explained why our primary sample period is over October 1997 to June 2013,
with the start date coinciding with the shift from a sizably positive to a predominantly negative stock-bond
correlation. Further, in contrast to the relatively high ination risk of the 1970s and early 1980s, Campbell,
Sunderam, and Viceira (2013) and David and Veronesi (2013) present evidence that, over the 19972013
period, bond investors faced relatively lower ination risk and Treasury bonds likely became more of a
hedge instrument. We note that the 19972013 period also largely postdates the earlier work on stockbond volatility linkage in Fleming, Kirby, and Ostdiek (1998) and Chordia, Sarkar, and Subrahmanyam
(2005), whose samples end in 1995 and 1998, respectively. To expand our analysis and assist in
interpretation, we also estimate our volatility models over the 1993-1996 period to provide a lower-stress
stock-market period to contrast with the higher-stress periods from our primary sample period.
2.3. Summary statistics
Table 1 reports the mean, median, standard deviation, skewness, and excess kurtosis for the four
different term-structure volatility measures (rows 14), for the volatility of daily stock futures returns
(row 5), and for the stock volatility from 5-minute SPY returns (row 6). The four term-structure volatility
measures and the daily-stock-return volatility are monthly measures, as computed from daily
observations over the rolling 22-trading-day period. The 5-minute-stock-return volatility is also computed
in similar fashion over the same 22-trading-day periods, but with returns at ve-minute intervals.
For the different volatility measures in Table 1, we present the statistics both for the raw variable
(row a) and for the logarithmic transformation of the raw variable (row b). In our empirical regression
8
Andersen and Bollerslev (1998) and Andersen, Bollerslev, Diebold, and Ebens (2001) argue that such a high-frequency
volatility estimate should be less noisy than a comparable estimate from daily returns.

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

37

Table 1
Summary data statistics.
This table reports the summary statistics for the key volatility variables. We report the mean, median, standard deviation,
skewness, and excess kurtosis for each volatility measure. The annualized volatility measures for the variables in rows 1 to 5 are
computed from the square-root of the sum of 22 squared daily returns for the futures or daily changes for the principal
components over the rolling 22-trading-day period, from t to t 21. Rows 1 and 2 report the 30-year T-bond futures return
volatility and 10-year T-note future return volatility, respectively. Rows 3 and 4 report the volatility of the rst and second
principal component, respectively. Row 5 reports the volatility of the S&P 500 futures returns. Finally, row 6 reports the
volatility calculated from 5-minute returns of the SPY ETF over the same 22-trading-day period. In all rows, the rst row
(labeled a) presents the statistics for the raw variable, and the second row (labeled b) presents the statistics for logarithmic
transformation of the variable. The sample period is from October 1997 to June 2013.
Row

Variable

Mean

Median

Std. Dev.

Skewness

Kurtosis

1a.
1b.
2a.
2b.
3a.
3b.
4a.
4b.
5a.
5b.
6a.
6b.

TB
ln(TB)
TN
ln(TN)
PC1
ln(PC1)
PC2
ln(PC2)
StFt
ln(StFt)
SPY5 minute
ln SPY5 minute

9.56
2.20
5.91
1.71
1.80
0.52
0.58
 0.64
18.32
2.78
19.22
2.84

8.96
2.19
5.47
1.70
1.69
0.53
0.53
 0.64
15.87
2.76
16.93
2.83

3.26
0.33
2.24
0.36
0.66
0.36
0.26
0.42
10.85
0.48
10.18
0.46

1.01
0.17
1.30
0.12
1.07
 0.03
1.29
0.13
2.79
0.46
2.05
0.45

0.87
 0.24
2.77
 0.02
1.91
 0.11
1.68
 0.08
12.36
0.38
6.68
 0.13

models, we use the log transformation to transform the raw volatility variable into a series that is
closer to normally distributed.
Fig. 1 exhibits the time series of the volatility of the 30-year T-bond-futures return (Panel A) and
the volatility of the S&P 500 futures return (Panel B). The graphs show a sizable relation between the
equity volatility and the bond volatility series, which is the subject of our empirical investigation.

3. Main empirical results for the ISBV relation

To investigate how equity volatility is related to the future term-structure volatility, we regress the
realized monthly volatility of our term-structure variables on our measures of lagged equity volatility,
while controlling for other relevant variables from the literature. We estimate variations of the
following regression for each of our four term-structure volatility measures:
TmSt
ST
TmSt
t;t 21 0 1 t  1;t  22 2 t  1;t  22

3
X
j1

j PrCompj;t  1 t ;

where the dependent variable, TmSt

t;t 21 , is the logarithmic transformation of one of the four termstructure volatility measures over trading days t to t21, calculated as the log of the square-root of the
sum of 22 squared daily values (daily returns for the T-bond and T-note futures, and daily changes for
the principal components) over the rolling 22-trading-day period. The explanatory variables are: (1)
ST
TmSt
t  22;t  1 , the rst lag of the dependent variable (to address volatility clustering); (2) t  1;t  22 , the log of
volatility for the S&P 500 futures returns over trading days t 22 to t  1; and (3) PrCompj;t  1 are the
three principal components at the end of day t 1. We use the three principal component at time t 1
as term-structure state variables, since the principal components are well known to represent the level,
slope, and curvature in the term structure.9 The s and s are coefcients to be estimated. We also
report results for an alternate estimation where the 2 stock-volatility term is based on high-frequency
9
Our principal components usage here follows from Andersen and Benzoni (2010). In robustness checks, we replace the
lagged principal components with multiple lagged forward rates [following Cochrane and Piazzesi, 2005], and nd that the
coefcients of interest remain qualitatively similar.

38

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

Panel A: Ln(Volatility of 30-year T-bond Futures Returns)

3.5

Ln(Volatility)

3.0
2.5
2.0
1.5
1.0

Oct-2012

Oct-2011

Oct-2010

Oct-2008

Oct-2009

Oct-2007

Oct-2006

Oct-2005

Oct-2004

Oct-2003

Oct-2002

Oct-2001

Oct-2000

Oct-1999

Oct-1998

0.0

Oct-1997

0.5

Panel B: Ln(Volatility of Stock Futures Returns)

5.0
4.5

Ln(Volatility)

4.0
3.5
3.0
2.5
2.0
1.5
1.0

Oct-2012

Oct-2011

Oct-2010

Oct-2009

Oct-2008

Oct-2007

Oct-2006

Oct-2005

Oct-2004

Oct-2003

Oct-2002

Oct-2001

Oct-2000

Oct-1999

Oct-1998

0.0

Oct-1997

0.5

Fig. 1. Time series of rolling volatility measures. This gure displays the time series of the rolling monthly volatilities of the
long-term T-bond returns and the stock-market returns, where the rolling volatilities are constructed from the 22 daily
observations within the 22-trading-day period. Panel A reports the log of the volatility of 30-year T-bond futures returns, and
Panel B reports the log of the volatility of S&P 500 futures returns. The units are the natural log of the annualized sample
standard deviation, with the returns in percentage terms. For the x-axis, the value for day t is for the volatility over trading days
t to t 21. The sample period is October 1997 to June 2013.

(5-minute) SPY returns, to evaluate robustness of the ISBV relation and to investigate whether the ISBV
relation is stronger with a presumably higher-quality stock volatility measure. Test statistics are
calculated based on heteroscedastic- and autocorrelation-consistent standard errors, with the number
of lags for the autocorrelation structure set to 22 since we use 22-trading-day overlapping variables.
We focus on rolling monthly periods since the monthly horizon is common in nance studies and
this approach should mitigate some of the noise that would be present in a daily-volatility analysis.
We evaluate rolling periods because this approach is likely to better use the available data to capture
the return dynamics (Richardson and Smith, 1991).
The next four subsections report estimation results for each of our four term-structure volatility
measures, respectively. In this analysis, we report results for separate estimations over our complete
primary sample period (1997:102013:06), an approximate rst-half subperiod (1997:102005.06),
and an approximate second-half subperiod (2005.072013:06).
3.1. Long-term T-bond return volatility
Table 2 reports on the volatility of 30-year T-bond-futures return as the term-structure volatility
measure (i.e., TmSt
TB
i;j
i;j ) in Eq. (1). Our primary coefcient of interest here is 2, the coefcient on the
lagged stock return volatility. We report the results from estimating three variations of Eq. (1).

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

39

Table 2
Volatility in 30-year T-bond futures returns.
This table reports results for how the equity risk is related to the subsequent 30-year T-bond return volatility. Panel A reports
three variations of the following regression, denoted as models (a) to (c) in the table:
TB
ST
TB
t;t 21 0 1 t  1;t  22 2 t  1;t  22

3
X
j1

j PrCompj;t  1 t ;

ST
where TB
i;j i;j is the logarithmic transformation of volatility for the 30-year T-bond (S&P 500) futures contract over trading

days i to j, calculated as the log of the square-root of the sum of 22 squared daily futures returns over the rolling 22-trading-day
period; PrCompj;t  1 are the three principal components from day t  1; and the s and s are coefcients to be estimated. Panel B
reports on a similar model, except the log of the standard deviation of realized 5-minute S&P 500 ETF returns over t  1 to t  22
replaces the realized stock volatility from daily returns for the 2 term. We report on the October 1997 to June 2013 sample
period, along with two subperiods (1997:102005:06 and 2005:072013:06). T-statistics are in parentheses, calculated with
heteroscedastic and autocorrelation consistent standard errors. An F-test (3) test statistic is in brackets, which jointly tests the
coefcients on the three lagged principal components. nnn, nn, and n indicate 1%, 5%, and 10% p-values.
Period

Model

[Pr Comp]

Panel A: Realized stock volatility from daily returns over t  1 to t  22 as the 2 Term
a.
0.347 (9.96)nnn
b.
0.594 (11.82)nnn
0.132 (3.93)nnn
c.
0.318 (5.51)nnn
0.189 (5.92)nnn
[17.62]nnn
First Subperiod
a.
0.155 (2.55)nn
1997:102005.06
b.
0.485 (6.52)nnn
0.118 (2.22)nn
c.
0.201 (2.77)nnn
0.199 (3.90)nnn
[17.07]nnn
Second Subperiod
a.
0.466 (12.44)nnn
2005.072013.06
b.
0.661 (8.60)nnn
0.118 (2.25)nn
c.
0.456 (5.39)nnn
0.099 (1.92)n
[9.20]nn
Full Period
1997:102013:06

Panel B: Realized stock volatility from 5-minute returns over t  1 to t  22 as the 2 Term
Full Period
0.286 (4.89)nnn
0.230 (7.01)nnn
[19.92]nnn
1997:102013:06
First Subperiod
0.173 (2.35)nn
0.233 (4.33)nnn
[18.58]nnn
1997:102005.06
Second Subperiod
0.397 (4.88)nnn
0.174 (3.10)nnn
[7.62]nnn
2005.072013.06

R2

26.4%
51.5%
58.8%
5.1%
28.3%
43.1%
46.6%
64.3%
69.1%
60.3%
44.4%
70.1%

The rst model has equity-risk as the only explanatory term. In the second model, we add the lagged
value of the dependent variable as an additional explanatory term. The third model is the full
specication given by Eq. (1).
In Panel A of Table 2 with the daily-return-based stock volatility, we nd that the estimates of 2 in
all three models are positive and statistically signicant over all three estimation periods. We also nd
that both the rst lag of the historical volatility (the 1 term) and the three lagged principal
components are reliably related to the subsequent volatility.
In Panel B of Table 2, where the lagged 5-minute-return-based stock volatility replaces the dailyreturn-based stock volatility, we nd that the ISBV relation remains reliably evident. For Panel B, the
statistical signicance of the estimated 2 and the R2 value is higher than the comparable model (c) in
Panel A for all three periods. This supports both the robustness of our ISBV ndings and the notion
that the ISBV relation is more reliably evident when using a higher quality stock-volatility measure.

3.2. Medium-term T-bond return volatility

Next, in Table 3, we repeat a comparable analysis for the volatility for the 10-year T-note-futures
returns (i.e., TmSt
TN
i;j
i;j ). For all three estimation periods, we again nd that the 2 estimates in all the
regressions are positive and statistically signicant. In Panel B, we again nd that the statistical
signicance of the estimated 2 and the R2 value is higher than the comparable model (c) in Panel A for
all three periods. For the other explanatory terms, the results are comparable to those for the longerterm bond.

40

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

Table 3
Volatility in 10-year T-note futures returns.
This table reports results for how the equity risk is related to the subsequent 10-year T-note return volatility. Panel A reports
three variations of the following regression, denoted as models (a) to (c) in the table:
TN
ST
TN
t;t 21 0 1 t  1;t  22 2 t  1;t  22

3
X
j1

j PrCompj;t  1 t ;

ST
where TN
i;j i;j is the logarithmic transformation of volatility for the 10-year T-note (S&P 500) futures contract over trading

days i to j, calculated as the log of the square-root of the sum of 22 squared daily futures returns over the rolling 22-trading-day
period; PrCompj;t  1 are the three principal components from day t  1; and the s and s are coefcients to be estimated. Panel B
reports on a similar model, except the log of the standard deviation of realized 5-minute S&P 500 ETF returns over t  1 to t  22
replaces the realized stock volatility from daily returns for the 2 term. We report on the October 1997 to June 2013 sample
period, along with two subperiods (1997:102005:06 and 2005:072013:06). T-statistics are in parentheses, calculated with
heteroscedastic and autocorrelation consistent standard errors. An F-test (3) test statistic is in brackets, which jointly tests the
coefcients on the three lagged principal components. nnn, nn, and n indicate 1%, 5%, and 10% p-values.
Period

Model

[Pr Comp]

Panel A: Realized stock volatility from daily returns over t  1 to t  22 as the 2 Term
a.
0.361 (9.54)nnn
b.
0.575 (10.36)nnn
0.237 (3.22)nnn
c.
0.382 (6.14)nnn
0.154 (4.05)nnn
[10.67]nnn
First Subperiod
a.
0.191 (2.63)nn
1997:102005.06
b.
0.495 (5.85)nnn
0.116 (1.85)n
c.
0.256 (2.97)nnn
0.205 (3.25)nnn
[10.63]nnn
Second Subperiod
a.
0.453 (10.37)nnn
2005.072013.06
b.
0.655 (9.85)nnn
0.099 (2.10)nn
c.
0.444 (5.59)nnn
0.093 (1.73)n
[8.21]nn
Full Period
1997:102013:06

Panel B: Realized stock volatility from 5-minute returns over t  1 to t  22 as the 2 Term
Full Period
0.365 (5.99)nnn
0.184 (4.62)nnn
[11.38]nnn
1997:102013:06
First Subperiod
0.236 (2.80)nnn
0.236 (3.49)nnn
[11.50]nnn
1997:102005.06
Second Subperiod
0.390 (4.94)nnn
0.169 (2.92)nnn
[7.12]nnn
2005.072013.06

R2

23.6%
46.9%
51.5%
5.7%
29.3%
39.4%
38.6%
58.0%
63.2%
52.3%
40.2%
64.1%

The evidence in Tables 2 and 3 indicates that the equity volatility contains substantial and reliable
forward-looking information about the subsequent bond return volatility. The information is
incremental, in that the partial ISBV relation remains strong when controlling for the information
contained in the lagged bond volatility and the three lagged principal components.

3.3. Volatility of the term-structure's level

We also investigate the term-structure's level as proxied by the rst principal component of the
term structure. Note the term-structure's rst principal component is by construction orthogonal to
the second and third principal component (which are representative of the slope and curvature,
respectively).
Table 4 reports on an estimation of Eq. (1) with the realized volatility of the change in the termstructure's rst principal component as the term-structure volatility measure (i.e., TmSt
PC1
i;j
i;j ). In
Panel A, we nd that the estimates of 2 are positive in all cases, and are strongly statistically
signicant for all cases except regressions (b) and (c) for the second-half subperiod. The positive 2
indicates that the volatility of the rst principal component tends to be larger following higher lagged
equity volatility. We note that the coefcient on the equity risk term is largely similar for both models
(b) and (c) (with or without the three lagged principal components), which indicates that the
information from the equity risk term is largely distinct from the information in the lagged crosssection of yields. In Panel B of Table 4, we nd that the 2 estimates are larger in both magnitude and
statistical signicance for all three estimation periods, as compared to regression (c) in Panel A.

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

41

Table 4
Volatility of change in the term-structure's level (rst principal component). This table reports results for how the equity risk is
related to the subsequent volatility of the term-structure's rst principal component. Panel A reports three variations of the
following regression, denoted as models (a) to (c) in the table:
PC1
ST
PC1
t;t 21 0 1 t  1;t  22 2 t  1;t  22

3
X
j1

j PrCompj;t  1 t ;

ST
where PC1
i;j ( i;j ) is the logarithmic transformation of volatility of the change in the rst principal component (S&P 500 futures

contract) over trading days i to j, calculated as the log of the square-root of the sum of 22 squared daily changes in the rst
principal component (daily futures returns) over the rolling 22-trading-day period; PrCompj;t  1 are the three principal
components from day t  1; and the s and s are coefcients to be estimated. Panel B reports on a similar model, except the log
of the standard deviation of realized 5-minute S&P 500 ETF returns over t  1 to t  22 replaces the realized stock volatility from
daily returns for the 2 term. We report on the October 1997 to June 2013 sample period, along with two subperiods (1997:10
2005:06 and 2005:072013:06). T-statistics are in parenthesis, calculated with heteroscedastic and autocorrelation consistent
standard errors. An F-test (3) test statistic is in brackets, which jointly tests the coefcients on the three lagged principal
components. nnn, nn, and n indicate 1%, 5%, and 10% p-values.
Period

Model

[Pr Comp]

Panel A: Realized stock volatility from daily returns over t  1 to t  22 as the 2 Term
a.
0.342 (9.41)nnn
b.
0.191 (1.66)n
0.264 (4.99)nnn
c.
0.097 (1.26)
0.222 (5.99)nnn
[30.05]nnn
First Subperiod
a.
0.204 (3.01)nnn
1997:102005.06
b.
0.058 (0.90)
0.194 (3.01)nnn
c.
0.013 (0.50)
0.271 (4.76)nnn
[17.16]nnn
Second Subperiod
a.
0.404 (9.07)nnn
2005.072013.06
b.
0.693 (11.23)nnn
0.051 (1.06)
c.
0.441 (5.52)nnn
0.065 (1.13)
[10.35]nn

Full Period
1997:102013:06

Panel B: Realized stock volatility from 5-minute returns over t  1 to t  22 as the 2 Term
Full Period
0.086 (1.12)
0.256 (6.37)nnn
[31.49]nnn
1997:102013:06
First Subperiod
 0.002 (  0.08)
0.326 (5.20)nnn
[20.38]nnn
1997:102005.06
Second Subperiod
0.395 (4.96)nnn
0.131 (2.11)nn
[9.58]nnn
2005.072013.06

R2

21.0%
28.0%
42.0%
7.3%
8.8%
35.4%
29.3%
54.8%
61.4%
43.1%
38.4%
61.9%

3.4. Volatility of the term-structure's slope

In Table 5, we report on an examination of the volatility of the term-structure's slope, as proxied
PC2
in Eq. (1)). In Panel A, the
for by the term-structure's second principal component (i.e., TmSt
i;j
i;j
estimates of 2 are positive in all cases, and are statistically signicant for all cases except regression
(c) for the second subperiod. The positive 2 indicates that the volatility of the second principal
component tends to be larger following higher lagged equity volatility. For model (a) with only the
equity risk as an explanatory term, the R2 values are again appreciable, at 24.8% for our entire sample
period and 38.4% for the second-half subperiod. Again, we note that the coefcient on the equity risk
term is largely similar for both models (b) and (c) (with or without the three lagged principal
components), which indicates that the information from the equity risk term is largely distinct from
the information in the lagged cross-section of yields. Once again, the 2 estimates in Panel B of Table 5
are larger in both magnitude and statistical signicance for all three estimation periods, as compared
to regression (c) in Panel A.
3.5. Summary of main results and robustness checks
The evidence in Tables 25 indicates that the lagged equity volatility contain substantial and
reliable forward-looking information about the subsequent term-structure volatility, beyond the
information contained in the lagged own volatility and the lagged principal components (as termstructure state variables). In all four tables, the evidence is consistent for both the subperiods, both in

42

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

Table 5
Volatility of change in the term-structure's slope (second principal component). This table reports results for how the equity
risk is related to the subsequent volatility of the term-structure's second principal component. Panel A reports three variations
of the following regression, denoted as models (a) to (c) in the table:
PC2
ST
PC2
t;t 21 0 1 t  1;t  22 2 t  1;t  22

3
X
j1

j PrCompj;t  1 t ;

( ST
where PC2
i;j
i;j ) is the logarithmic transformation of volatility of the change in the second principal component (S&P 500
futures contract) over trading days i to j, calculated as the log of the square-root of the sum of 22 squared daily changes in the
second principal component (daily futures returns) over the rolling 22-trading-day period; PrCompj;t  1 are the three principal
components from day t  1; and the s and s are coefcients to be estimated. Panel B reports on a similar model, except the log
of the standard deviation of realized 5-minute S&P 500 ETF returns over t  1 to t  22 replaces the realized stock volatility from
daily returns for the 2 term. We report on the October 1997 to June 2013 period, along with approximate one-half subperiods
(1997:102005:06 and 2005:072013:06). T-statistics are in parenthesis, calculated with heteroskedastic and autocorrelation
consistent standard errors. An F-test (3) test statistic is in brackets which jointly tests the coefcients on the three lagged
principal components. nnn, nn, and n indicate 1%, 5%, and 10% p-values.
Period

Model

[Pr Comp]

Panel A: Realized stock volatility from daily returns over t  1 to t  22 as the 2 Term
a.
0.534 (8.82)nnn
b.
0.188 (1.39)
0.351 (5.14)nnn
c.
0.093 (1.29)
0.352 (8.36)nnn
[51.52]nnn
First Subperiod
a.
0.247 (3.87)nnn
1997:102005.06
b.
 0.002 (  0.05)
0.247 (3.88)nnn
c.
0.004 (0.12)
0.298 (5.36)nnn
[5.51]nnn
Second Subperiod
a.
0.404 (10.57)nnn
2005.072013.06
b.
0.777 (11.32)nnn
0.092 (1.65)n
c.
0.459 (5.26)nnn
0.089 (1.48)
[17.46]nn
Full Period
1997:102013:06

Panel B: Realized stock volatility from 5-minute returns over t  1 to t  22 as the 2 Term
Full Period
0.063 (1.00)
0.433 (10.31)nnn
[67.07]nnn
1997:102013:06
First Subperiod
 0.010 (  0.34)
0.361 (6.32)nnn
[7.56]nnn
1997:102005.06
Second Subperiod
0.422 (5.05)nnn
0.156 (2.33)nn
[16.87]nnn
2005.072013.06

R2

24.8%
32.1%
55.5%
10.8%
10.8%
21.7%
38.4%
71.7%
77.5%
59.4%
25.7%
78.0%

terms of the statistical signicance and the magnitude of the coefcients on the equity-risk terms. In
all four tables, the estimated coefcients on the lagged equity volatility changes only modestly when
adding the three lagged principal components (comparing models (b) to (c) in each table). Finally, in
all four tables, the ISBV relation is stronger in Panel B with the 5-minute-return-based stock volatility,
which supports the robustness of our ISBV ndings and indicates that the ISBV relation is more
reliably evident when using a higher-quality stock volatility measure.
We next probe the robustness of these ndings. First, we re-estimate the primary regressions from
Tables 25, but with six lagged forward rates (the 1-year, 2-year, 3-year, 5-year, 7-year, and 10-year
GSW instantaneous forward rates) replacing the lagged three principal components as term-structure
state variables. This approach follows from Cochrane and Piazzesi (2005), who nd that the termstructure of forward rates has strong explanatory power for one-year bond excess returns. We nd
qualitatively very similar results to those depicted in Tables 25. For all four term-structure
volatilities, the estimated 2 on the lagged stock volatility remains reliably positive, with a p-value of
1% or better.
Second, a natural question is whether our results are evident when analyzing longer time horizons
(rather than the rolling monthly horizon analyzed in the tables) and for alternate specications. In an
Online Appendix, we evaluate the subsequent quarterly term-structure volatility (from 66 trading-day
observations) with an alternate specication that allows for additional information from the prior
realized term-structure volatility by including multiple lags as explanatory terms. We nd that the
lagged stock volatility remains a reliable incrementally informative determinant for the subsequent

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

43

term-structure volatility. Taken together, these additional ndings support the robustness of our
primary ndings.

4. A ight-to-quality/ight-from-quality avenue?
We next discuss and present evidence regarding FTQ (and FFQ) as a potential underlying economic
mechanism that might be an important contributor behind the documented ISBV relation.
Theoretically, we rst appeal to the framework in Bekaert, Engstrom, and Xing (BEX) (2009). BEX
consider the joint pricing of stocks and bonds in a market where both economic uncertainty and risk
aversion may change over time. One result from BEX is that the volatility of economic fundamentals
(or economic uncertainty) is very highly correlated with expected stock-market volatility, where
fundamentals refers to dividend growth. BEX also nd that stock volatility is systematically higher in
bad economic times such as recessions, which indicates positive serial correlation in stock volatility or
volatility clustering.10 Their model also features a classic FTQ avenue where bond prices are likely to
appreciate with heightened economic uncertainty due to a precautionary savings effect.
The theoretical framework of Veronesi (1999) also provides a rationale for volatility clustering. In
his model, time-varying volatility is tied to uncertainty about the economic state, and the price impact
of news is higher when uncertainty about the underlying economic state is higher.11 Bollerslev, Chou,
and Kroner (1992) provide a survey of the empirical evidence on volatility clustering, along with a
discussion on the theoretical underpinnings of volatility clustering.
Thus, if a relatively high stock-return volatility and high time series variability in economic
uncertainty are likely following months with a relatively high realized stock volatility, then the
likelihood of FTQ pricing inuences over the subsequent month is presumably much greater. With this
economic intuition in mind, we next report ve additional evaluations that bear on the plausibility of
a FTQ avenue.
4.1. Variation in the ISBV relation with the market state
The basic premise of FTQ is that the phenomenon would be largely episodic around times with
higher stock-market stress or economic uncertainty. Accordingly, we expect that the ISBV relation
would tend to be stronger around recessions. Further, under a FTQ avenue, the ISBV relation would
presumably be largely non-existent over low stock-market stress periods (periods with a low and
stable stock volatility and with no prominent economic or international crises). In this subsection, we
explore these predictions.
For our investigation, we choose two high-stress and two low-stress stock-market subperiods and
estimate the ISBV relation over each of these four subperiods separately. While such market-state
classications are admittedly somewhat subjective, we rely on the National Bureau of Economic Research
(NBER) recession classication and the VIX behavior to provide objectivity in our classications.
We categorize two subperiods as having relatively higher stock market stress: March 2001 to
November 2002 and December 2007 to June 2010. The beginning month of these higher-stress states
is the rst month of a formal NBER recession. The nal month of the higher-stress states is 12 months
following the last month of each NBER recession. Our choice of a one-year after recession terminal
month recognizes that uncertainty and market stress typically remain past the formal end of a
recession as the market learns of the recovery. We note that the formal NBER announcement of the
10
The BEX measure of economic uncertainty is the volatility of fundamentals. In their Table 5 (p. 71), they report on
simulated moments from their model using estimated parameters from actual data for the 1927 to 2004 period. In this exercise,
they estimate that the volatility of fundamentals is highly correlated to the expected stock market volatility, with a correlation
coefcient of 0.88. Further, they estimate an autocorrelation coefcient of 0.98 for the stock-market's conditional variance, or
strong volatility clustering.
11
The notion of economic uncertainty is different in Veronesi (1999) versus BEX (2009). In Veronesi, the uncertainty refers
to the notion that the true underlying economic state is unobservable and unknown by investors. In certain market states, the
uncertainty about the market state is relatively higher.

44

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

end of the respective recession occurred after the November 2002 and June 2010 end-months for the
higher-stress states, with the NBER announcements occurring in July 2003 and September 2010 for
the earlier and later recession, respectively. We also feel that this one-year post choice is a good t
with the VIX time series behavior, as discussed further below.
We also categorize two subperiods as having relatively lower stock market stress: January 1993 to
November 1996 and April 2004 to June 2007.12 These choices rely heavily on the VIX behavior, as
follows. The beginning month has the following two criteria: (1) occurs after the end of the preceding
recession has been formally announced by the NBER, and (2) has the closing daily VIX o20% for that
entire month and for the next 11 calendar months (for one consecutive year with the closing daily
VIX o20%). The nal month for the lower-stress states is selected as the rst month that: (1) occurs
after the beginning-month criteria is met, and (2) precedes two consecutive months that have
episodes where the VIX exceeds 20%.
Our choice of the two high and low market-stress subperiods is supported by Table 6, Panel A,
which reports VIX summary statistics for each of the four subperiods. For our two high-stress
subperiods, the daily closing VIX value is above the full-sample median of 19.03% for 95.2% and 89.7%
of the days for our rst and second high-stress subperiods, respectively. For our two low-stress
subperiods, the daily closing VIX value is above the full-sample median for less than 1.8% of the days
for both low-stress subperiods.
We estimate variations of the following regression separately for these four subperiods:
TB
ST
TB
t;t 21 0 1 t  1;t  22 2 t  1;t  22 t ;

where the terms are as dened for Eq. (1) and Table 2, Panel A.
In Panel B of Table 6, we report the results for the T-bond futures return volatility. The row-1
specication includes only the lagged stock-futures volatility as the explanatory term. Note that the
estimated 2 coefcients on the stock volatility term are about twice as large for the high-stress
subperiods as compared to the same coefcients for the low-stress subperiods. Further, the R2 values
are strikingly different, at an average of 28.9% for the high-stress periods versus 3.6% for the low-stress
periods.
In the row-2 specication, we evaluate a variation that includes only the own lagged T-bond
volatility as the explanatory term. When comparing the row-1 model (with stock-volatility as the sole
explanatory term) to the row-2 model (with the T-bond volatility as the sole explanatory term), we
nd that the stock volatility contains more forward-looking information about the subsequent T-bond
volatility than does the own-lagged T-bond volatility for the high-stress state (in terms of the R2
values).
Finally, the row-3 specication includes both the lagged T-bond volatility and lagged stock
volatility as explanatory terms. The results again indicate that the lagged stock volatility is the more
important explanatory term for the high-stress periods, whereas the own-lagged T-bond return
volatility is the more important explanatory term for the low-stress subperiods. In untabulated
results, we nd similar results when estimating a comparable model where the T-bond futures return
volatility is replaced by the 10-year T-note futures return volatility.
Overall, the results in Table 6 support the premise that the ISBV relation is stronger with higher
stress/volatility in the stock market. These ndings indicate an episodic nature of the ISBV relation in a
manner that one would expect through a FTQ avenue.
4.2. Evidence of stock market volatility clustering
With the intertemporal nature of our primary ndings, a FTQ avenue would require that a
relatively high realized stock volatility last month must be reliably associated with a higher
subsequent stock volatility over the next month. As previously discussed, Veronesi (1999) and BEX
(2009) provide theory and evidence about stock volatility clustering. Consistent with their ndings,
12
For this exercise, we extend our sample earlier back to 1993 in order to capture a second low-stress market state for
evaluation.

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

45

Table 6
The intertemporal stock-to-bond volatility relation for four key subperiods. This table reports how equity volatility is related to
the subsequent volatility of the T-bond-futures returns for four different key subperiods. Panel A reports statistics for the
CBOE's implied volatility index (VIX) from S&P 500 index options for the overall January 1993 to June 2013 period and
separately for the four subperiods to highlight subperiods differences in stock market stress. Panel B reports estimation
variations of the volatility model from Table 2, with separate regression results for the two separate lower market stress and
uncertainty subperiods (rows 13 and 79) and the two separate higher market stress and uncertainty subperiods (rows 46
and 1012). The volatilities here are calculated from daily return observations over a 22-trading-day period. For the estimated
coefcients, T-statistics are in parentheses, calculated with heteroscedastic and autocorrelation consistent standard errors. nnn,
nn
, and nn indicate 1%, 5%, and 10% p-values.
Panel A: VIX Subperiod Statistics
Subperiod
Full Sample
I. Low Stress
II. High Stress
III. Low Stress
IV. High Stress

Dates

Mean

Median

Low

Max

Std. Dev.

% 4 19:03

1993:012013:06
1993:011996:11
2001:032002:11
2004:042007:06
2007:122010:06

20.52
13.75
26.66
13.39
30.09

19.03
13.23
24.46
13.04
25.41

9.31
9.31
17.40
9.89
15.58

80.86
23.87
45.08
23.81
80.86

8.46
2.17
6.29
2.13
12.61

50%
1.8%
95.2%
1.1%
89.7%

Panel B: Subperiod Regression Results for the 30-year T-bond Futures Volatility
Subperiod

Dep. Var.

R2

1 ( TB
t  1;t  22 )

2 ( ST
t  1;t  22 )

n/a

0.167 (2.33)nn

4.8%

0.460 (4.54)nnn

n/a

21.6%

I. Low Stress

1. TB
t;t 21

1993:01 to

2. TB
t;t 21

1996:11

3. TB
t;t 21

0.447 (4.11)nnn

0.024 (0.33)

21.7%

II. High Stress

4. TB
t;t 21
5. TB
t;t 21
6. TB
t;t 21
7. TB
t;t 21
8. TB
t;t 21
9. TB
t;t 21
10. TB
t;t 21
11. TB
t;t 21
12. TB
t;t 21

n/a

0.326 (4.17)nnn

21.6%

0.236 (1.63)

n/a

5.5%

0.141 (0.82)

0.304 (3.69)nnn

23.5%

2001:03 to
2002:11
III. Low Stress
2004:04 to
2007:06
IV. High Stress
2007:12 to
2010:06

n/a

0.134 (0.95)

2.4%

0.574 (6.72)nnn

n/a

36.4%

0.579 (5.79)nnn

 0.016 (  0.14)

36.4%

n/a

0.307 (6.31)nnn

36.2%

0.596 (5.29)nnn

n/a

35.7%

0.325 (1.52)

0.176 (1.94)n

40.3%

clustering in stock volatility is also evident over our sample. The simple correlation between the 22trading-day stock-futures volatility over days t to t 21 and the same volatility over trading-days
t 22 to t  1 is 0.69 for the simple standard deviation and 0.70 for the log of the standard deviation.
To illustrate further the tendency of stock volatility to cluster, we perform the following sorting
exercise to analyze only extreme volatility episodes. With the realized stock volatility over trading
days t  22 to t  1 as the sorting variable, we sort ve variables that are constructed from daily
observations over trading days t to t21: (1) the S&P 500 futures volatility, (2) the T-bond futures
volatility, (3) the 10-year T-note futures volatility, (4) the correlation between the T-bond and S&P 500
futures returns, and (5) the correlation between the T-note and S&P 500 futures returns. For the
lagged volatility in this sorting exercise, we use the more precise volatility measure from 5-minute
returns on the SPY ETF.
In Table 7, Panel A, we report the summary statistics for the volatilities and correlations that follow
the largest 10% of the realized stock volatilities. The results indicate that the subsequent stock
volatility is appreciably larger than average for months when the prior realized stock volatility was
high. For this high volatility state, the row-1 results show that the mean/median of this conditional
subsequent stock volatility was 33.3%/28.1% (versus 18.3/15.9% over our primary 1997:10 to 2013:06
sample), with 96.4% of these conditional stock-volatilities being above the full-sample median.
Conversely, in Table 7, Panel B, we present comparable statistics for observations that follow the
smallest 10% of the prior realized stock volatilities. The results indicate that the subsequent stock
volatility is appreciably smaller than average for months when the prior realized stock volatility was

46

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

Table 7
Return volatilities and correlations that follow an extreme realized stock volatility. This table reports subset statistics for
volatilities and correlations that follow an extreme realized stock volatility over the prior month. For every rolling 22-tradingday period, the monthly volatility and correlation are calculated from the 22 daily observations. Panel A (B) reports statistics for
the subset of volatilities and correlations over trading days t to t 21 that follow the largest (smallest) 10% of the realized 5minute S&P 500 ETF return volatilities over trading days t  22 to t  1. Rows 13 report on the S&P 500 futures return volatility,
the T-bond futures return volatility, and the 10-year T-note futures return volatility, respectively, in annualized standarddeviation percentage units. Row 4 reports on the VIX level on day t, in percentage units. Row 5 reports on the realized volatility
of the daily VIX changes over t to t 21 in daily-change VIX units. Rows 6 and 7 report the correlations between the stock
futures returns and the T-bond and T-note futures returns, respectively. For each subset, we report the mean, median, 25th
percentile, and 75th percentile. The sample period is October 1997 to June 2013.
Panel A: Observations following the Largest 10% of Realized Stock Volatility
Variable
StFt

1.
2. TB
3. TN
4. VIX
5. VIX
6. StFt;TB
7. StFt;TN

Mean

Median

25th Pctl

75th Pctl

33.29
13.43
8.41
38.89
2.66
 0.33
 0.34

28.13
13.30
8.00
36.22
2.11
 0.32
 0.35

21.94
10.54
6.50
31.19
1.58
 0.51
 0.51

37.11
16.62
9.83
43.67
2.91
 0.12
 0.14

Panel B: Observations following the Smallest 10% of Realized Stock Volatility

Variable
StFt

1.
2. TB
3. TN
4. VIX
5. VIX
6. StFt;TB
7. StFt;TN

Mean

Median

25th Pctl

75th Pctl

9.92
6.80
4.17
12.51
0.77
 0.04
 0.03

9.35
6.49
4.10
12.08
0.60
 0.07
 0.04

7.81
5.94
3.54
11.19
0.48
 0.33
 0.33

11.35
7.47
4.57
13.34
0.91
0.27
0.29

low. For this low volatility state, the row-1 results show that the mean/median of this conditional
subsequent stock volatility was 9.9%/9.4%, with only 5.1% of these conditional stock volatilities being
above the full-sample median.
Finally, rows 2 and 3 of Table 7 show that the conditional volatility of the T-bond and T-note
futures returns are also strikingly different, depending upon whether the prior month's stock
volatility was extremely high or low. The T-bond and T-note return volatilities that follow a high stock
volatility (Panel A) have a mean and median that are about twice the comparable values for the
observations that follow a low stock volatility (Panel B).
4.3. VIX characteristics following an extreme stock market volatility
Under a FTQ avenue that is linked to periods of stock market stress and time-varying economic
uncertainty, we would expect that periods following a high realized stock volatility would also be
periods with both a relatively high level and high variability in the option-derived implied stockmarket volatility. We examine this proposition using VIX data. In addition to VIX's basic interpretation
as a measure of expected stock volatility, VIX has also been interpreted as a fear index that reects
economic uncertainty and, perhaps, risk aversion (e.g., Bollerslev, Tauchen, and Zhou, 2009).
Using the same sorting exercise as in Section 4.2, we nd that the VIX is strikingly higher and more
variable for observations that follow a high realized stock volatility. In Table 7, Panels A and B, rows
4 and 5 report these conditional VIX statistics, with the VIX variability dened as the square root of
the sum of squared daily VIX-changes over t to t 21. For the days that follow a high (low) realized
stock volatility in the top (bottom) decile, the conditional mean/median of VIX is 38.89/36.22 (12.51/

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

47

12.08). For the periods that follows a high (low) realized stock volatility in the top (bottom) decile, the
conditional mean/median of the VIX variability is 2.66/2.11 (0.77/0.60).
4.4. The stock-bond return correlation and high stock market volatility
Under a FTQ avenue for understanding the ISBV relation, we would expect that a high realized
stock volatility this month would be associated with a relatively more negative stock-bond return
correlation over the next month. Again using the same sorting approach as in Section 4.2, we nd that
the stock-bond return correlations are appreciably more negative for observations that follow a high
realized stock volatility. For example, in Table 7, we nd that the median of the 22-trading-day stockbond return correlations (between daily S&P 500 futures returns and 10-year T-note futures returns)
is 0.35 (  0.04) for observations that follow a high (low) realized stock volatility.13
4.5. T-bond diversication benets with large stock market declines
Finally, under a FTQ avenue, one would expect that: (1) T-bonds would have actually served as a
good diversication instrument against bad stock market outcomes over our sample, and (2) that such
diversication benets would have been relatively stronger following periods with a higher realized
stock volatility.14 We compute conditional average returns for 30-year T-bond futures and 10-year
T-note futures for those observations when there was a concurrent extreme stock return, with
separate evaluations for those observations that follow a relatively low stock volatility and for those
observations that follow a relatively high stock volatility (based on the lagged volatility being below or
above its median value).
Table 8 reports these conditional average returns both at the daily horizon (Panel A) and at the
weekly horizon (Panel B), where a week is dened as ve consecutive trading days with overlapping
weekly observations. Column (2) reports on the stock-return threshold for both extremely negative
returns ( o5th percentile) and extremely positive returns ( 4 95th percentile), based on the realized
stock returns over the October 1997 to June 2013 sample period. Column (3) reports if the stock
volatility over day t  1 to t  22 is below or above its median value. Column (4) reports the number of
observations for each stock-return threshold. Columns (5)(7) report the average returns for the 30year T-bond futures, the 10-year T-note futures, and the S&P 500 futures, respectively, for the
observations when the realized stock return falls in the threshold listed in column (2).
As expected, the extreme stock returns are much more likely when the lagged stock volatility is
above its median value. In all the sub-panels of the table, the number of observations (column 4) is
dramatically lower for rows when the lagged stock volatility was below its median (rows 1 and 3) as
compared to that for rows when the lagged stock volatility was above its median (rows 2 and 4).
When the stock return is extreme, we also nd a sizably opposite average T-bond return. For each row,
we nd that the conditional average of the T-bond-futures returns (columns 5 and 6) are appreciable
and of opposite sign, as compared to the average of the extreme S&P 500 futures returns (column 7).
Together, these observations provide support for the notion that the diversication benet of T-bonds
looks to be appreciably greater following periods of higher realized stock volatility.
4.6. Summary discussion
This section's evidence supports the plausibility of a FTQ mechanism being an important
contributor to the ISBV relation. Perhaps most importantly, we found that the ISBV relation is stronger
during periods of stock-market stress and weaker during periods of relative calm. Further, we found
that a relatively higher stock-market volatility this month is associated with: (1) a much higher stock13
These ndings are consistent with related ndings in Connolly, Stivers, and Sun (2005) and Baele, Bekaert, and
Ingehelbrecht (2010) that a higher VIX is associated with a subsequently more negative stock-bond correlation.
14
This intuition also ts with ndings in Campbell, Sunderam, and Viceira (2013), which indicate more of a hedge role for
T-bonds since around 2000.

48

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

Table 8
Average T-bond returns when stock returns are extreme. This table examines T-bond and 10-year T-note futures returns over
days and weeks that experienced extreme stock returns. We report the conditional averages of futures returns for four separate
subsets of observations using the following double-sort criteria: Returns coincident with either extremely negative/positive
stock returns (rst sort), but then separate subsets depending upon whether the prior month had a relatively low/high realized
stock volatility (second sort). An extreme stock return (column 2) is if the observation is either in the top or bottom vigintile. A
prior low/high realized stock volatility (column 3) is whether the stock volatility over the prior month was above or below its
median value. Panels A and B report on the daily and weekly horizon, respectively, where a week is a rolling 5-trading-day
period. The stock returns are S&P 500 futures returns, and the realized stock volatility is computed from the daily S&P 500
futures returns over the rolling 22-trading-day period. The sample period is October 1997 to June 2013.
Panel A: Daily Returns
If coincident

If StFt
t  1;t  22

of

Avg Ret

Avg Ret

Avg Ret

(1)

stock return is:

(2)

was:
(3)

Obs.
(4)

30-yr T-bond
(5)

10-yr T-note
(6)

S&P500
(7)

1.
2.
3.
4.

o 5th pctl
o 5th pctl
495th pctl
495th pctl

31
167
18
180

0.59
0.61
 0.49
 0.46

0.37
0.41
 0.20
 0.27

 2.75
 3.22
2.15
3.20

Row

Below
Above
Below
Above

Median
Median
Median
Median

Panel B: Weekly Returns

Row

If coincident

(1)

stock return is:

(2)

1.
2.
3.
4.

o 5th pctl
o 5th pctl
4 95th pctl
4 95th pctl

If

StFt
t  1;t  22
was:
(3)

Below
Above
Below
Above

Median
Median
Median
Median

of

Avg Ret

Avg Ret

Avg Ret

Obs.
(4)

30-yr T-bond
(5)

10-yr T-note
(6)

S&P500
(7)

20
178
10
188

0.99
1.27
 0.32
 0.70

0.63
0.90
 0.12
 0.31

 5.07
 6.54
4.56
5.91

market volatility next month; (2) both a higher VIX level and higher VIX variability next month;
indicating higher risk perceptions and greater time series variation in risk perceptions; (3) a more
negative stock-bond return correlation next month; and (4) an increased diversication benet for
holding T-bonds in a stock-bond portfolio.

5. An omitted factor perspective?

In this section, we present evidence regarding a potential omitted common-factor avenue that
may bear on understanding our ISBV ndings. Under this channel, presumably there might be
common factors that are important determinants of both stock and bond volatility, with the commonfactor's volatility exhibiting substantial positive serial correlation. If so, then the lagged stock volatility
might provide information about the underlying common-factor-news volatility, which could
contribute to an empirical ISBV link.
We consider two potential common-factor candidates. First, Fama and French (1993) found that
the monthly returns of both stock portfolios and government bond portfolios responded similarly to
their default yield spread (DYS) variable.15 We investigate whether or not the volatility of a DYS
variable, dened as the yield difference between Moody's Baa- and Aaa-rated bonds, is an important
omitted variable that may be relevant for understanding our primary results. We add the lagged
volatility of DYS, based on daily changes, as an additional explanatory variable in our primary
regressions in Tables 2 and 3. In an Online Appendix, we report that: (1) the lagged stock volatility
15
Chen, Roll, and Ross (1986), Fama and French (1993), and Jagannathan and Wang (1996), among others, relate yield
spreads to expected stock returns.

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

49

remains a reliable, incremental explanatory term, and (2) the lagged DYS volatility is not an important
incremental explanatory term in our setting.
Second, while ination may be more directly relevant to the valuation of nominal bonds,
researchers have proposed that ination news may drive both stocks and bond returns, sometimes in
opposite directions (e.g., Campbell and Ammer, 1993; David and Veronesi, 2013). Next, using the
method from Gurkaynak, Sack, and Wright (2010), we form a daily ination compensation variable
based on the difference between the 10-year TIPS yield and the yield of the 10-year nominal T-note.
We add the lagged ination-compensation volatility (based on daily changes) as an additional
explanatory variable in our primary regressions in Tables 2 and 3. In the Online Appendix, we show
that the lagged stock volatility remains a reliable, incremental explanatory term, and the lagged
ination-compensation volatility is not an important incremental explanatory term in our setting. To
conclude, while our limited evidence in this section is inconclusive (in that other factors or
approaches could be evaluated), our ndings lend no support to the notion that this omitted
common-factor avenue is of rst-order importance for understanding our ISBV ndings.

6. Other empirical studies on stock-bond volatility linkages

In this section, we briey discuss our ndings in the context of two earlier empirical studies that
also evaluated volatility linkages between the stock and Treasury bond markets.
6.1. Relation to Fleming, Kirby, and Ostdiek (1998)
Fleming, Kirby, and Ostdiek (FKO) (1998) evaluate daily return data from S&P 500, T-bond, and Tbill futures contracts for the January 1983 to August 1995 period. They modeled information ows and
evaluated how information inuences all three markets through both a direct effect and an
information-spillover effect tied to cross-market hedging. Their results show a greater volatility
linkage across the markets than is indicated by the modest correlations in daily returns and daily
absolute returns. Their nding of strong cross-market volatility linkages is consistent with the
premise of our main ndings.
However, FKO's empirical work is much different than ours. First, they analyze the volatility of
daily returns using a stochastic volatility model with an AR(1) process. Our focus is on monthly and
quarterly realized volatilities, estimated from daily or high frequency intraday returns. Second, their
notion of volatility linkages is based on the correlation of conditional daily variances of S&P 500 and Tbond futures returns. Our investigation is broader in the sense that we examine the intertemporal
linkages between lagged stock volatility and four different measures of the subsequent term-structure
volatility in a multivariate setting that controls for the lagged term-structure volatility and other
term-structure state variables. Finally, their sample (which predates our sample) has a stock-bond
return correlation of 0.35; the comparable correlation for our sample is 0.30. Such striking
correlation differences suggest differences in the relative importance of a FTQ/FFQ avenue between
our two samples.
6.2. Relation to Chordia, Sarkar, and Subrahmanyam (2005)
Chordia, Sarkar, and Subrahmanyam (CSS) (2005) evaluate the liquidity, returns, and volatility of
stock and bond markets for the June 1991 to December 1998 period. Their focus is on liquidity, but
they also evaluate cross-market volatility linkages. Their measure of daily volatility is the absolute
residual of a regression of daily returns against various calendar-related dummy variables.
In contrast to our overall ndings, they do not nd that stock volatility Granger-causes bond
volatility in a two-lag VAR model. While their empirical approach is substantially different than ours,
we believe that sample-period differences are central to understanding the differences between their
volatility ndings and ours. Only the last 15 months of their sample (October 1997 to December 1998)
overlap with our sample. Over June 1991 to September 1997 (the rst 6 years, 4 months of their

50

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

sample), the mean (median) VIX was 15.2% (14.7%), with a maximum of 25.99%, and the correlation
between daily S&P 500 and 10-year T-note futures returns was 0.42. Over our October 1997 to June
2013 sample, the mean (median) VIX was 22.2% (20.9%), with 947 different days having VIX values
greater than 25.99%. Also, in our Table 6, recall that we reported the results of our volatility model
over the low stock stress period from 1993 to 1996 (which is a subset of their sample) and did not nd
a reliable ISBV relation.
During times of predominantly low stock-market stress, the collective ndings in CSS (2005) and
our study indicate that the ISBV relation is likely to be weak. However, during times of high stockmarket stress, our ndings indicate that the ISBV relation is likely to be sizable.

7. Conclusions
Over the October 1997 to June 2013 period, we nd an economically substantial and statistically
reliable intertemporal relation between realized stock volatility and the subsequent realized volatility
of important Treasury market variables. Further, we nd that the intertemporal stock-to-bond
volatility (ISBV) relations remain substantial and reliable, even while controlling for: (1) the own
term-structure's lagged realized volatility, (2) various term-structure state variables proposed in the
literature, and (3) other potential omitted factors that might plausibly subsume the estimated ISBV
relations. Specically, we study the volatility of returns from both T-bond and 10-year T-note futures
contracts and the volatility of changes in the term-structure's rst and second principal component.
Our investigation focuses on rolling estimates of monthly realized volatilities, constructed from daily
observations over rolling 22-trading-day periods.
The intertemporal aspect of our ndings supports the notion that equity volatility can help
understand volatility behavior in bond markets, beyond an approach that only looks at the bond
market in isolation. This notion builds from the results in Andersen and Benzoni (2010), who nd that
the cross-section of yields does not span yield volatility and who suggest that linking term structure
dynamics to the general economic environment might prove productive.
While the ISBV relations are substantial and reliable over our full sample, we nd that the ISBV
relations are substantially stronger during times with notable stock-market stress (such as around
recessions) and appreciably weaker in calm times (periods with a sustained low VIX and no
prominent economic or international crisis). This suggests that a ight-to-quality/ight-from-quality
(FTQ/FFQ) avenue may be important for understanding our ndings. Consistent with this premise, we
nd that a high realized stock volatility this month is associated with the following over the next
month: (1) a much higher subsequent realized stock volatility, (2) much higher day-to-day variability
in the stock-market's option-derived implied volatility (VIX), (3) a more negative stock-bond return
correlation, and (4) an appreciably greater diversication benet to holding T-bonds in a stock-bond
portfolio.
We argued that the combined intuition from Bekaert, Engstrom, and Xing (2009) and Veronesi
(1999) is also consistent with our evidence favoring a FTQ/FFQ interpretation of our ISBV ndings.
Higher economic uncertainty is associated with a persistently higher stock volatility, more variability
in the perceived economic uncertainty, and a greater precautionary savings motive. So, a higher stock
volatility this month is likely to be followed by more extreme stock returns next month and with more
variability in market uncertainty measures (or fear measures), such as VIX. If so, then the likelihood of
FTQ/FFQ pricing inuences over the subsequent month should be greater, and this, we believe,
contributes to our ISBV ndings.
At rst glance, our ndings seem at odds with evidence in Chordia, Sarkar, and Subrahmanyam
(CSS) (2005), who nd that stock volatility does not Granger-cause bond volatility. However, their
sample is the largely low-risk period from June 1991 to December 1998. So, their ndings seem to t
with our evidence that the ISBV relation is weaker in calmer stock-market states (see our Table 6).
Thus, an additional contribution of our study is to show that the intertemporal stock-to-bond
volatility ndings in CSS (2005) lack some generality.
Finally, in the sense that equity risk is related to macroeconomic uncertainty and investor
sentiment, our intertemporal ndings seem consistent with recent evidence on bond-return

N. Bansal et al. / Journal of Financial Markets 25 (2015) 3351

51

predictability. For example, Cooper and Priestley (2009) and Ludvigson and Ng (2009) nd that
macroeconomic variables play a role in understanding bond risk premia. Laborda and Olmo (2014)
show that investor sentiment variables have predictive power for bond risk premia.
Appendix A. Supplementary data
Supplementary data associated with this article can be found in the online version at http://dx.doi.
org/10.1016/j.nmar.2015.05.002.

References
Andersen, T., Bollerslev, T., 1998. Answering the skeptics: yes, standard volatility models do provide accurate forecasts. Int.
Econ. Rev. 39, 885905.
Andersen, T., Bollerslev, T., Diebold, F., Ebens, H., 2001. The distribution of realized stock return volatility. J. Financ. Econ. 61,
4376.
Andersen, T., Benzoni, L., 2010. Do bonds span volatility risk in the U.S. treasury market? a specication test for afne term
structure models. J. Finance 65, 603653.
Baele, L., Bekaert, G., Inghelbrecht, K., 2010. The determinants of stock and bond return comovements. Rev. Financ. Stud. 23,
23742428.
Bansal, N., Connolly, R., Stivers, C., 2014. The stock-bond return relation, the term-structure's slope, and asset-class risk
dynamics. J. Financ. Quant. Anal. 49, 699724.
Bekaert, G., Engstrom, E., Xing, Y., 2009. Risk, uncertainty, and asset prices. J. Financ. Econ. 91, 5982.
Ben-Rephael, A., Kandel, S., Wohl, A., 2012. Measuring investor sentiment with mutual fund ows. J. Financ. Econ. 104, 363382.
Bollerslev, T., Chou, R., Kroner, K., 1992. ARCH modeling in nance: a review of the theory and empirical evidence. J. Econom. 52,
559.
Bollerslev, T., Tauchen, G., Zhou, H., 2009. Expected stock returns and variance risk premia. Rev. Financ. Stud. 22, 44634492.
Campbell, J., Ammer, J., 1993. What moves the stock and bond markets? a variance decomposition for long-term asset returns.
J. Finance 48, 337.
Campbell, J., Sunderam, A., Viceira, L., 2013. Ination bets or deation hedges? the changing risks of nominal bonds. Working
paper 09-088, Harvard Business School.
Chen, N., Roll, R., Ross, S., 1986. Economic forces and the stock market. J. Bus. 59, 383403.
Chordia, T., Sarkar, A., Subrahmanyam, A., 2005. An empirical analysis of stock and bond market liquidity. Rev. Financ. Stud. 18,
85129.
Cochrane, J., Piazzesi, M., 2005. Bond risk premia. Am. Econ. Rev. 95, 138160.
Cooper, I., Priestley, R., 2009. Time-varying risk premiums and the output gap. Rev. Financ. Stud. 22, 28012833.
Connolly, R., Stivers, C., Sun, L., 2005. Stock market uncertainty and the stock-bond return relation. J. Financ. Quant. Anal. 40,
161194.
Connolly, R., Stivers, C., Sun, L., 2007. Commonality in the time-variation of stock-stock and stock-bond return comovements.
J. Financ. Mark. 10, 192218.
David, A., Veronesi, P., 2013. What ties return volatilities to price valuations and fundamentals? J. Polit. Econ. 121 682746.
Diebold, F., Piazzesi, M., Rudebusch, G., 2005. Modeling bond yields in nance and macroeconomics. Am. Econ. Rev. 95,
415420.
Fama, E., French, K., 1993. Common risk factors in the returns on stocks and bonds. J. Financ. Econ. 33, 356.
Fleming, J., Kirby, C., Ostdiek, B., 1998. Information and volatility linkages in the stock, bond, and money markets. J. Financ. Econ.
49, 111137.
Gurkaynak, R., Sack, B., Wright, J., 2007. The U.S. treasury yield curve: 1961 to the present. J. Monet. Econ. 54, 22912304.
Gurkaynak, R., Sack, B., Wright, B., 2010. The TIPS yield curve and ination compensation. Am. Econ. J.: Macroecon. 2, 7092.
Jagannathan, R., Wang, Z., 1996. The conditional CAPM and the cross section of expected returns. J. Finance 51, 353.
Jubinski, D., Lipton, A., 2012. Equity volatility, bond yields, and yield spreads. J. Futures Mark. 32, 480503.
Kodres, L., Pritsker, M., 2002. A rational expectations model of nancial contagion. J. Finance 57, 769799.
Laborda, R., Olmo, J., 2014. Investor sentiment and bond risk premia. J. Financ. Mark. 18, 206233.
Ludvigson, S., Ng, S., 2009. Macro factors in bond risk premia. Rev. Financ. Stud. 22, 50275067.
Richardson, M., Smith, T., 1991. Tests of nancial models in the presence of overlapping observations. Rev. Financ. Stud. 4,
227254.
Underwood, S., 2009. The cross-market information content of stock and bond order ow. J. Financ. Mark. 12, 268289.
Veronesi, P., 1999. Stock market overreaction to bad news in good times: a rational expectations equilibrium model. Rev. Financ.
Stud. 12, 9751007.