Electronic copy available at: http://ssrn.com/abstract=2527213 New Evidence on State-Dependent Dynamics of Risk, Inflation, and Asset Valuation 1 Robert Connolly Finance Area Kenan-Flagler School of Business University of North Carolina Chapel Hill, NC David Dubofsky Department of Finance College of Business University of Louisville Louisville, KY Chris Stivers Department of Finance College of Business University of Louisville Louisville, KY This version: March 23, 2015 1 We thank Ric Colacito, Cami Kuhnen, Ahn Le, Cipriana Prepeliuc, and seminar participants at UNC-Chapel Hill and the University of Louisville for helpful comments. Please address comments to Robert Connolly (email: Robert [email protected]; phone: (919) 962-0053); David Dubofsky (email: [email protected]; phone: (502) 852-3016); or Chris Stivers (e-mail: [email protected]; phone: (502) 852-4829).
71
Embed
New Evidence on State-Dependent Dynamics of … copy available at : http ://ssrn.com /abstract = 2527213 New Evidence on State-Dependent Dynamics of Risk, In ation, and Asset Valuation…
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Electronic copy available at: http://ssrn.com/abstract=2527213
New Evidence on State-Dependent Dynamics of
Risk, Inflation, and Asset Valuation1
Robert Connolly
Finance Area
Kenan-Flagler School of Business
University of North Carolina
Chapel Hill, NC
David Dubofsky
Department of Finance
College of Business
University of Louisville
Louisville, KY
Chris Stivers
Department of Finance
College of Business
University of Louisville
Louisville, KY
This version: March 23, 2015
1We thank Ric Colacito, Cami Kuhnen, Ahn Le, Cipriana Prepeliuc, and seminar participants at
UNC-Chapel Hill and the University of Louisville for helpful comments. Please address comments to
Robert Connolly (email: Robert [email protected]; phone: (919) 962-0053); David Dubofsky (email:
Electronic copy available at: http://ssrn.com/abstract=2527213
New Evidence on State-Dependent Dynamics of
Risk, Inflation, and Asset Valuation
Abstract
Over 1997 to 2013, we find striking economic-state differences in how bond and stock values
are linked to innovations in risk and inflation. We identify two prolonged periods (October 2001
- April 2004 and January 2009 - December 2013) that we characterize as ‘recessionary/post-
recessionary’ (RPR) states, where: (1) increases in perceived equity (bond) risk are strongly and
very reliably linked to higher (lower) T-bond returns and a decreased (increased) term-structure
slope, with regression R2 values up to 50% at the monthly horizon; and (2) inflation is positively
linked to stock returns. These RPR periods are also distinguished by an elevated stock-market
variance risk premium, lower economic growth, lower inflation, a close proximity to sizable stock
market declines, a relatively larger term yield spread as compared to very low T-bill yields,
and presumably higher risk aversion. Over the non-RPR growth periods in our sample, the
comparable risk-to-return connection in longer-term Treasuries is much weaker or non-existent
and the link between inflation and stock returns is negative or statistically insignificant. Our
findings fit with key implications from recent theory that emphasize time-varying risk aversion
and the signaling role of inflation.
JEL Classification: G12
Keywords: Treasury-bond and Stock futures returns, Forward interest rates, Equity and Bond
risk, Inflation news, Economic states
1. Introduction
Asset-class risk perceptions and inflation are of fundamental importance in asset pricing, port-
folio construction, and risk management. This paper uncovers striking economic-state differences
over 1997 to 2013 in how bond and stock values are linked to innovations in asset-class risk and
inflation. By ‘asset-class risk’, we refer to risk perceptions for T-bonds and equities, as measured
by the implied volatilities from 10-year T-Note futures options and S&P 500 options, respectively.
Our economic states are also distinguished by notable differences in the stock market’s variance
risk premium, economic growth, inflation, proximity to sizable stock market declines, the bond
market’s term yield spread relative to T-bill yields, and presumably risk aversion.
Specifically, we identify two prolonged periods (October 2001 - April 2004 and January 2009
- December 2013) that we characterize as ‘recessionary/post-recessionary’ (RPR) states, where:
(1) increases in perceived equity (bond) risk are strongly and reliably linked to higher (lower)
T-bond returns and a decreased (increased) term-structure slope; and (2) inflation is positively
related to stock market returns. Conversely, over the remaining non-RPR periods, the comparable
risk-return connection in longer-term Treasuries is much weaker or non-existent and the relation
between stock returns and inflation news is either negative or statistically insignificant over the
expansionary economic months. These stylized differences across economic states fit with key
implications from recent theory that emphasize time-varying risk aversion and the signaling role
of inflation; see especially Bekaert, Engstrom, and Xing (2009), Bekaert and Engstrom (2013),
and David and Veronesi (2013).
Our study commences in October 1997 due to data limitations, combined with evidence sug-
gesting a regime-shift around 1997 in stock-bond return dynamics and Treasury bond risk premia.
Baele, Bekaert, and Inghelbrecht (2010) show that the stock-bond return correlation shifted from
almost exclusively positive over the 1970’s through mid-1997 to predominantly negative around
October 1997.1 Analysis in Campbell, Sunderam, and Viceira (2013) and Campbell, Pflueger,
and Viceira (2014) also suggest a prominent regime shift about this time; since then Treasury
bonds appear to serve as more of a hedge instrument against the risk of a stock market decline
1Aslandis and Christiansen (2012) find a similar correlation shift using high-frequency intraday stock and T-
bond futures returns.
1
and/or a weak macroeconomy. In contrast, over the 1970s and 1980s, Treasury bonds added to
an investor’s macroeconomic risk exposure.
While perhaps no theoretical model is rich enough to encompass all these economic state dif-
ferences in a tractable fashion, our empirical investigation draws much from models in Bekaert,
Engstrom, and Xing (2009) (BEX), Bekaert and Engstrom (2013) (BE), and David and Veronesi
(2013) (DV). In the BEX framework, both economic uncertainty and risk aversion vary counter-
cyclically over time, where economic uncertainty refers to the volatility of economic fundamentals
such as the dividend growth rate. Risk aversion increases as consumption moves toward a sub-
sistence or habit level. ‘Flight-to-quality’ (FTQ) episodes arise endogenously in their model, as
increases in fundamental volatility lead to higher bond prices and may lead to lower equity prices
in some economic states. This occurs when heightened economic uncertainty acts to depress
interest rates due to a precautionary savings motive, and to depress stock values due to a risk-
premium feedback. Their model calibration indicates a very high correlation between expected
equity volatility and economic uncertainty, which suggests equity implied-volatility changes may
be a good proxy for changes in economic uncertainty. Periods with higher risk aversion also have
a stronger precautionary savings motive. This suggests a stronger positive link between changes
in economic uncertainty and bond values in weaker economic times. Thus, the risk-to-return
connections in Treasuries may vary with the economic state and risk aversion.
BE (2013) introduce non-Gaussian fundamentals where consumption dynamics are different
for good and bad economic times. During bad times with higher risk aversion, their framework
suggests: (1) a stronger precautionary saving effect, because a sizable negative shock to con-
sumption is more likely, and (2) an elevated equity variance risk premium. This perspective also
suggests important economic-state differences in risk-return dynamics.
In DV (2013), economic states vary along the two dimensions of economic growth and infla-
tion, but they assume constant relative risk aversion. In a low-growth/low-inflation state, news of
higher inflation is bad news for nominal bonds but good news for stocks because higher inflation
may signal that the economy is not slipping into a deflationary state with very low growth or
contraction. Thus, the reaction of stock prices to inflation news varies with the economic state,
and inflation news can generate a negative stock-bond return correlation.2
2Guidolin and Timmermann (2006) analyze monthly stock and bond returns over 1954 to 1999, and their results
2
Our empirical investigation identifies two prolonged RPR periods that are distinguished by
a noteworthy set of economically-important commonalities. In addition to the dramatically
stronger risk-return connection in longer-term Treasuries, both RPR periods: (1) exhibit a strong
and reliably positive link between inflation news and stock returns; (2) exhibit a higher variance-
risk premium in the stock market and a greater difference between the stock-market’s implied
volatility and subsequent realized volatility; (3) commence in the later part of a formal NBER
economic recession, after the Federal Reserve has largely or totally concluded its monetary easing
in the targeted Fed Funds rate; (4) have a relatively much larger term yield spread, as compared
to the very low T-bill yields; (5) can be characterized as a lower-growth and lower-inflation
economic state, especially relative to the preceding expansion; and (6) commence following a
sizable downturn in the stock market, with a decline exceeding 35% from the preceding stock
market peak.3 Figure 1 depicts this variation in the term-yield spreads (relative to the T-bill
yield) and the equity variance risk premium. Collectively, the higher variance risk premium, the
weak economic state, and the preceding stock-market decline suggest that our two RPR periods
have relatively higher aggregate risk aversion.4
To illustrate the economic-state differences, consider a regression of monthly T-Bond-futures
returns as the dependent variable against the concurrent and lagged monthly changes in the
asset-class risk as explanatory terms. For our two RPR periods (encompassing 91 months),
this regression has an average R2 value of 41.7%. For the other non-RPR periods in sample
(encompassing 104 months), the comparable average R2 value is 2.3%. For weekly returns, the
indicate that economic states are important for understanding the joint stock-bond return distribution.3To identify these two RPR periods, we primarily utilize the ‘structural break’ statistical methods of Bai and
Perron (1998) and (2003), but we also consider variations in the other market variables and characteristics described
here. Our Section 2.5 and Appendix A presents details for the structural-break analysis and the other economic
differences summarized in this paragraph.4See Bollerslev, Tauchen, and Zhou (2009), Bollerslev, Gibson, and Zhou (2011), and Bekaert, Hoerova, and Lo
Duca (2013) for discussion that links the variance risk premium to movements in aggregate risk aversion. Fama and
French (1993), Campbell and Cochrane (1999), and BEX (2009) present evidence that aggregate risk aversion is
countercyclical. Guiso, Sapienza, and Zingales (2014) present both survey and experimental evidence that suggests
higher risk aversion is likely following dramatic stock market declines. In controlled experiments with financial
professionals, Cohn, Engelmann, Fehr, and Marechal (2014) find convincing evidence in favor of the countercyclical
risk aversion hypothesis.
3
comparable average R2 value is 29.1% for our RPR periods versus 4.2% for the non-RPR periods.
Our ‘risk-to-Treasury-values’ findings fit with key empirical implications in BEX (2009) and
BE (2013). Consistent with BEX, we show empirically that bond values tend to increase with
heightened economic uncertainty, and there is a much stronger risk-return relation in longer-term
Treasuries over weak economic times with presumably higher risk aversion. Consistent with BE,
our findings suggest that bad economic times are likely to have both a stronger precautionary
savings motive and an elevated equity variance risk premium.
Consistent with Campbell, Sunderam, and Viceira (2013) (CSV), our findings suggest that
longer-term Treasuries became more of a hedge instrument against economic uncertainty and/or
equity risk since around 2000. Stronger FTQ linkages in the longer-term Treasury market over
our RPR periods (especially 2009-2013) seem intuitive because of the relatively low inflation risk,
the relatively high term yield spread, and the near-zero money market yields.
Our findings regarding economic-state differences in the stock-inflation connection fit with the
key empirical implication in DV (2013) that inflation news provides a signal about the underlying
economic state. Over our two RPR periods (which have low-growth and low inflation), we
find that the relation between inflation news and stock returns is appreciably and statistically-
reliably positive, coincident with a more negative stock-bond return correlation.5 Over the
non-RPR growth periods in our sample, the relation between inflation news and stock returns
is either reliably negative or statistically insignificant and the stock-bond return correlation is
either positive or appreciably less negative.
When evaluated jointly, we find that both equity-risk innovations and inflation news are
important in understanding the stock-bond return correlation over our sample. Thus, our evi-
dence suggests that both the economic-uncertainty perspective in BEX (2009) and the inflation
perspective in DV (2013) are valuable in understanding asset return dynamics.
To combat the severe economic downturn of 2008-09, the Federal Reserve undertook unprece-
dented large-scale purchases in the longer-term fixed income markets, commencing in early 2009.
There is evidence (e.g., Krishnamurthy and Vissing-Jorgensen (2011) and Jarrow and Li (2013))
that the intensity of Fed purchases affected Treasury yields; which suggests a possible link to our
5Inflation news is measured primarily by deduction from movements in TIPS and nominal Treasury yields, per
Gurkaynak, Sack, and Wright (2010). Inflation news based on CPI and PPI releases yield consistent results.
4
principal findings regarding the strong ‘risk-to-Treasury value’ connection over 2009-2013. We
present evidence in Section 6 that is inconsistent with a ‘Federal Reserve intensity’ explanation
of the strong risk-return connection over our 2009-2013 RPR period.
In sum, we document striking economic-state differences in both risk-return and inflation-
return connections. Our work is related to the existing literature that studies risk and return
in the Treasury markets, but unlike much of the term-structure literature we jointly explore the
impact of both bond and equity risk on Treasury values.6 Further, our work also looks across
markets by showing that the economic-state divisions that bear on understanding the risk-to-
Treasury-return connection are also the same economic-state divisions that bear on understanding
the inflation-to-stock-return connection.
The remainder of our study is organized as follows. Section 2 describes the data. Section 3
establishes our main empirical findings regarding the connection between asset-class risk percep-
tions and returns/yields. Section 4 expands our empirical investigation to also consider changes
in inflation expectations. Section 5 describes other properties of our two RPR periods. Finally,
Section 6 provides additional analysis concerning an alternative “Fed Intensity” hypothesis, and
Section 7 concludes.
2. Data Description and Sample Selection
2.1. Asset-class Risk from the Implied Volatility of Options
We use equity and T-Note implied volatility as observable, high-quality, and dynamic measures
of risk expectations (or risk perceptions) for each asset class. For equity risk, we use the Chicago
Board Options Exchange (CBOE) VIX measure, derived from options on the S&P 500. For the
risk of longer-term Treasuries, we use the implied volatility from 10-year T-Note futures options
from Bloomberg. Both implied volatilities are standardized to provide an annualized volatility
6For affine-quadratic models, see Ahn, Dittmar, Gao, and Gallant (2003), Leippold and Wu (2002), and Ahn,
Dittmar, and Gallant (2002). For regime-switching models, see Bansal and Zhou (2002), Bansal, Tauchen, and
Zhou (2004), Ang and Bekaert (2002), and Dai, Singleton, and Yang (2007). For nonlinear models, see Ahn and
Gao (1999) and Feldhutter, Heyerdahl-Larsen, and Illeditsch (2013). See Ghysels, Le, Park, and Zhu (2014) for
a discrete-time no-arbitrage term structure model that combines the tractability of affine term structure models
with the ability of GARCH models to deliver an accurate measure of yield volatility.
5
estimate over the subsequent one month.7 The VIX is available from 1990 and the T-Note-futures
implied volatility (TIV) from mid-1993, which limits our sample choices.
We also note that both the VIX and TIV are sizably positively correlated, with a correlation of
0.66 between their levels, 0.344 for the one-month changes, and 0.227 for the one-week changes.
Figure 2 depicts this time-series behavior for both the VIX and TIV values and their weekly
changes. This positive correlation makes economic sense; because, to some extent, stocks and
bonds may share exposure to the same economic risk factors (see, e.g., Fama and French (1993)).
This positive correlation also indicates the importance of including both stock and bond risk
when trying to isolate a partial risk-return relation linked to a specific asset-class risk.
Our empirical analysis relies upon the assumption that VIX and TIV are good proxies for the
forward-looking risk, or expected return volatility, of each respective asset-class return series. In
Appendix B, we report evidence that supports the use of VIX and TIV as comparable forward-
looking asset-class risk measures for equities and T-bonds.
Table 1 reports univariate summary statistics for the VIX and TIV percentage changes over
rolling one-week periods. Our empirical work focuses on these variables to measure implied-
volatility dynamics; denoted ∆log(IVt/IVt−j) as a ‘continuous percentage change’ variable with
IV indicating the implied volatility, either V IX or TIV in annualized percentage units, and
where j equals 5 (22) trading days for the rolling weekly (monthly) analysis. With this method,
the volatility of the VIX-changes and TIV-changes are more comparable and their volatility is
less variable across subperiods; see subperiod statistics in Table 1. Later, in robustness checks,
we find similar results when using the simple IV changes in place of the log changes.
If the VIX changes have much more volatility in our RPR periods, then this might help
explain why the regression R2 values are much greater when relating equity risk to longer-term
Treasuries over our RPR periods (see Forbes and Rigobon (2002)). Accordingly, we compare
the VIX variability during our RPR periods to the VIX variability during our non-RPR periods.
Table 1 reports the standard deviation of our primary ∆log(V IXt,t−5) weekly-change variable
7See Whaley (2000) and Blair, Poon, and Taylor (2001) for more background information on VIX, including
additional supportive evidence about the forward-looking volatility information in VIX. For the TIV series, we
smoothed out a few extreme data points that were inexplicably much different than the prior and subsequent day’s
value (about 0.5% of the observations), by replacing the inconsistent daily values with the prior day’s value.
6
for our RPR and non-RPR subperiods. The standard deviations are remarkably similar across
subperiods, all in the 0.10 to 0.13 range. The similarity in the VIX variability across our RPR
and non-RPR subperiods can also be seen in Figure 2, Panel B. Further, the average absolute
∆log(V IXt,t−5) is not statistically reliably different across our RPR and non-RPR periods. We
conclude there is not a pervasive or pronounced difference in the VIX variability across our RPR
and non-RPR periods.
2.2. Futures Contract Returns
To capture the dynamic behavior of stock and Treasury bond prices, we use the returns implied
by the prices of 30-year T-Bond, 10-year T-Note, and S&P 500 futures contracts. These fu-
tures contracts are all very widely traded, so stale prices and other liquidity concerns should be
minimized; see, e.g., Ahn, Boudoukh, Richardson, and Whitelaw (2002).
To calculate the implied returns, we use the continuous futures price series from Datastream
for the three contracts, where the futures prices are from the daily settlement prices, with rolling
contracts to provide a continuous series. These series from DataStream International (with codes
CUSCS04, CTY CS04, and ISPCS04) are constructed so that the returns derived from the price
series may be interpreted as excess returns. Table 1 reports univariate summary statistics for
these futures returns at the weekly horizon over our subperiods of interest.
2.3. Treasury Yields and Forward Interest Rates
For our Treasury yield data, we primarily rely on the data set as described in Gurkaynak, Sack,
and Wright (2007). Our study uses their zero-coupon bond yields and forward interest rates. For
our calculation of the term-structure’s principal components, we use their 10 zero-coupon bond
yields at years one to 10. We also use their instantaneous continuously-compounded forward-
rate (FR) yields to evaluate the dynamics of marginal interest rates at specific points in the term
structure; see Figure 3 for the time-series of the FR yields over our sample. Table 1 reports
univariate summary statistics for weekly changes in these forward rates over our subperiods of
interest. Our investigation also uses ‘inflation compensation’ data, based on yield differences
between 10-year TIPS and 10-year nominal Treasuries, using the method from Gurkaynak, Sack,
7
and Wright (2010). Finally, we use the 10-year and 6-month Treasury Constant Maturity (TCM)
yields from the Federal Reserve as an estimate of the term yield spread.
2.4. Realized Stock Market Volatility
To estimate the equity ‘variance risk premium’, we construct a rolling 22-trading-day realized
volatility measure from 5-minute returns on the widely traded SPY S&P 500 ETF. Our equity
‘variance risk premium’ equals the difference between VIX at the close of day t and the realized
volatility from 5-minute returns over trading days t−21 through day t. Bollerslev, Tauchen, and
Zhou (2009) and Bollerslev, Gibson, and Zhou (BGZ) (2011) refer to this implied-minus-realized
volatility difference as the ‘variance risk premium’ and BGZ note that it “is sometimes used by
market participants as a measure for the market-implied risk aversion” (page 239). Appendix C
provides details on the data and how we calculated this realized volatility.
2.5. Selection of our Two ‘Recessionary/Post-Recessionary’ (RPR) Periods
In Appendix A, we provide details for our rationale and procedures for selecting the RPR periods
that are featured in our paper. The first-order criteria is a ‘structural break’ analysis for the
risk-return relation in longer-term Treasuries, using the statistical methods of Bai and Perron
(1998, 2003). Additionally, we also consider several other economic properties that reinforce the
selection of our two RPR periods. Details are relegated to Appendix A for brevity in our main
text, with Appendix A.8 elaborating on related evidence in CSV (2013).
As we noted in our introduction, our RPR periods commence later in a recession after the
Federal Reserve has totally or largely completed their easing in the targeted Fed Funds rate
(FFR). Further, our first RPR period ended a few months before the Fed finally began increasing
the targeted FFR; and our second RPR period was ongoing at the end of our sample period with
the Fed maintaining a near-zero targeted FFR. In Appendix A.4, we discuss our RPR periods in
the context of these Fed actions.
In Appendix A.4, we discuss why the early parts of the recessions are not included in our
RPR periods. We present evidence there that suggests that Federal Reserve actions are likely to
have obscured the risk-return Treasury connections around the onset of recessions, as both risk
8
perceptions and Treasury values may have been responding to Fed actions rather than typical
free-market forces. The Fed lowered the targeted Fed Funds rate 11 times in 2001 and 7 times in
2008. Given these observations, our empirical investigation contrasts our RPR periods to both:
(1) the full non-RPR periods (1997:10 - 2001:09 and 2004:05 - 2008:12) that contain economic-
growth periods and the onset of the recession; and (2) subsets of the non-RPR periods that
include months with economic growth only, or non-RPR-growth periods (1997:10 - 2001:02 and
2004:05 - 2007:11). With this approach, we investigate whether the results for the non-RPR
periods are appreciably different when omitting the onset of the recessions.
3. The Risk-to-Return Connection and Economic States
In this section, we present our primary empirical results analyzing the connection between changes
in asset-class risk perceptions and three aspects of the Treasury market: (1) T-Bond and 10-year
T-Note futures returns; (2) changes in Treasury forward rates at the 1-, 5-, and 10-year points
in the term structure; and (3) changes in the term-structure’s slope. We also present evidence
regarding the relation between stock returns and changes in asset-class risk perceptions.
We investigate the Treasury forward rates as a complement to analysis of the longer-term
Treasury-futures return because the instantaneous forward rates indicate the marginal interest
rate at specific points on the yield curve. Thus, changes in forward rates should better capture
pricing influences at different specific points in the yield curve. We investigate the instantaneous
forward interest rates from Gurkaynak, Sack, Wright (GSW, 2007) that are 1-year, 5-years, and
10-years out to capture short-, mid-, and longer-horizon points in the term structure.8
We report on two measures of the change in the Treasury term-structure’s slope. The first is
the change in the term-structure’s second principal component, where the principal components
are estimated from the 10 zero-coupon bond yields with maturities from 1-year to 10-years using
8Additionally, the GSW forward-rate data is constructed from spot quotes on Treasury bonds, so our analysis is
extended beyond the futures market. GSW’s intent is to estimate a yield curve that is suitable for “understanding
its fundamental determinants such as ... perceived risk, and investors’ risk preferences” (page 2295). Their
parametric yield curve allows for very rich shapes of the forward curve while largely ruling out variation resulting
from anomalous prices of a small number of securities.
9
the GSW data.9 The second is the change in the term yield spread, where the term yield spread
is defined as the difference between the 10-year and 6-month Treasury Constant Maturity Yield
from the Federal Reserve.
We choose to investigate the sizable change horizons of one week and one month for two
reasons. First, the longer horizons should eliminate some of the noise that would be evident in a
daily or intra-day analysis, such as non-synchronous market closings that introduce measurement
errors in daily data. Second, our view is that relations at these longer horizons are more likely
to reflect fundamental underlying economic forces. Shorter-horizon changes (such as daily or
intra-day changes) are more likely to be distorted by market microstructure influences.
In this section, we motivate and interpret our empirical investigation primarily from the
theoretical framework of BEX (2009). Accordingly, we begin by providing further background
on their model.
3.1. Implications from BEX’s (2009) Theoretical Framework
The BEX (2009) theoretical framework features time-variation in both economic uncertainty and
risk aversion. Economic uncertainty refers to the conditional variance of fundamentals, such as
dividend growth. Risk aversion is countercyclical and varies with the difference between current
consumption and an ‘external habit level,’ following from the preference structure in Campbell
and Cochrane (1999).10 BEX evaluate the relative importance of uncertainty and risk aversion
in understanding equity risk premia, equity volatility, and the term structure of interest rates.
They calibrate their model to actual U.S. economic data over 1927 to 2004.
There are several prominent dimensions of BEX that are central to our empirical investiga-
tion. First, because their measure of economic uncertainty is highly correlated with conditional
stock market variance (ρ = 0.88), we posit that time-variation in the stock-market’s conditional
variance is a reasonable proxy for movements in their concept of economic uncertainty (at least as
9Previous research has shown that the term-structure’s first three principal components are closely related to
its level, slope, and curvature, respectively. See, e.g., Diebold, Piazzesi, and Rudebusch (2005), who find that the
first two principal components account for almost all (99 percent) of the variation in the yields.10Bekaert, Engstrom, and Grenadier (2010) also use this preference structure and note that it suggests “moody
investors”, since risk aversion can increase dramatically as consumption falls back to near the habit level.
10
a first-order interpretation).11 Thus, when interpreting our empirical investigation from BEX’s
perspective, we assume that VIX movements are a reasonable proxy for short-horizon movements
in economic uncertainty and that TIV movements are more related to bond-specific risk.12
Second, in BEX, a higher economic uncertainty is associated with a higher precautionary
savings motive. This feature can induce a FTQ effect on bonds, where increased economic
uncertainty can drive up bond prices.
Third, because our RPR periods commence later in recessions and are ongoing through the
beginning of an uncertain economic recovery, consumption in our RPR periods should be rela-
tively closer to the habit/subsistence consumption level. If so, under BEX, risk aversion would
be both higher during our RPR periods and more sensitive to small consumption changes. In
Section 3.6, we argue that this risk-aversion property suggests that FTQ influences, where in-
creased economic uncertainty acts to drive down interest rates, is likely to be stronger over our
RPR periods. The intuition is that weak economic times have higher risk aversion and risk aver-
sion is more sensitive to consumption changes, so when uncertainty increases there is a relatively
stronger precautionary savings effect in these states.
Fourth, the effect of time-varying economic uncertainty on stock prices is ambiguous and
can vary with the economic state, because of opposing economic forces that affect stock prices.
To begin with, greater economic uncertainty lowers interest rates; which, ceteris paribus, would
promote lower discount rates for stock valuation and higher stock prices. On the other hand, the
11In BEX (2009), the stock-market’s conditional variance is also modestly correlated with the inverse of the
surplus ratio (ρ=0.38), a key driver in risk aversion movements. This implies that movements in the stock-
market’s conditional variance are likely to be associated with movements in risk aversion to a lesser degree. BEX
argue that risk aversion is less variable than economic uncertainty, so presumably our weekly and monthly VIX
changes should largely reflect variation in uncertainty.12An alternative to using VIX-changes directly would be to decompose the VIX into an economic-uncertainty
component and a risk-aversion component along the lines of Bekaert, Hoerova, and Lo Duca (2013) (BHL), and then
evaluate changes in the uncertainty component (rather than VIX). With that approach, the uncertainty component
is the projected conditional variance from a regression of the realized variance against the lagged VIX and lagged
realized variance. The risk-aversion component is the difference between VIX and the projected variance. We
evaluated this approach using equation (1) in BHL (page 774). Over our sample, the realized variance loaded more
heavily on the VIX. Thus, changes in VIX (as used as an explanatory variable in equation (1)) and comparable
changes in the projected conditional variance had extremely high positive correlations (ρ > 0.98).
11
increased economic uncertainty also increases stock risk with a higher conditional stock variance,
which would promote higher equity risk premia and lower stock prices. With the two opposing
effects, the net effect may vary with the economic state. From BEX, “there may be instances
where our model will generate a classic ‘flight to quality’ effect with uncertainty lowering interest
rates, driving up bond prices, and depressing equity prices” (page 72).13
Finally, the results from the BEX calibration over 1927 to 2004 suggests that increases in
uncertainty should be associated with a steeper real term structure. However, there are opposing
effects with complex interactions between uncertainty and risk aversion, so this prediction may
not hold in all economic states.
The flexibility and complexity of the BEX framework suggest it is an interesting empirical
question to evaluate how the risk-return connections vary with the economic state. In this section,
we investigate four empirical questions. First, is there a positive relation between economic-
uncertainty changes (as proxied by VIX movements) and Treasury values; and, if so, is this
positive risk-return connection reliably stronger over our RPR periods with presumably higher
risk aversion? Second, under a premise of higher risk aversion, is the relation between bond
returns and own-bond risk stronger over our RPR periods? Third, does the relation between the
term-structure’s slope and economic uncertainty vary with the economic states? Finally, under
a premise of higher risk aversion, is the relation between stock returns and economic-uncertainty
changes stronger over our RPR periods?
3.2. Estimation with State-dependent Risk-to-Treasuries Relations
To investigate the first two questions, we begin by estimating variations of the regression:
TrFtRtt−j,t = α0 + (λ1 + λ2D0104t + λ3D
0913t )∆log(V IXt−j,t)+ (1)
(γ1 + γ2D0104t + γ3D
0913t )∆log(TIVt−j,t) + ϵt−j,t
13The BEX (2009) model is complex, with five state variables, 19 parameters, and 34 moment conditions in
their estimation. “Our model involves more state variables and parameters than much of the existing literature,
making it difficult to trace pricing effects back to any single parameter’s value,” (BEX, 2009, page 62). As such,
there are complex interactions such that movement in state variables can sometimes have different influences on
asset-pricing issues for different economic conditions.
12
where TrFtRt indicates the percentage change in the Treasury futures contract’s price or the
‘futures return’ over the period from t−j to t; V IX and TIV are the equity and T-bond implied
volatilities; the ∆log and subscripts t − j, t indicate the difference in the log of each variable
between trading days t− j and t; D0104t is a dummy variable that equals one over our first RPR
period over 2001:10 - 2004:04; D0913t is a dummy variable that equals one over our second RPR
period over 2009:01 - 2013:12; ϵt−j,t is the residual; and the α, λ’s and γ’s are coefficients to
be estimated. The dependent variable is either the 30-year T-Bond futures return (TB) or the
10-year T-Note futures return (TN), and j is either 5 or 22, indicating rolling 5-trading-day
and 22-trading-day returns and change horizons. We also report on comparable estimations, but
where the dependent variables are changes-in-forward rates for 10-year, 5-year, and 1-year out
forward rates (FR10, FR05 and FR01). We later report on a similar specification that includes
time-varying volatility of ϵt−j,t.
Table 2, Panel A, reports the results for Treasury futures returns. The regressions in rows 1,
3, 5, and 7 report the results on the unconditional relation (without the dummy variables). These
unconditional relations indicate that the VIX-changes (TIV-changes) are positively (negatively)
related to the futures returns in all four cases (T-Bond-futures and T-Note-futures, for both
weekly and monthly returns). Thus, estimates for our full sample indicate a classic FTQ dynamic,
as suggested in BEX (2009), where longer-term Treasury values increase reliably with increased
equity risk perceptions. The relation to the TIV-changes suggests a risk-premium-feedback effect,
where increases in own-asset-class risk are associated with declining prices.
Estimates from the full model are reported in Table 2, Panel A, rows 2, 4, 6, and 8. In all
four cases, we find that the risk-return connection is much stronger for our two RPR periods. For
example, in row 2, the estimated relation between the monthly VIX-changes and the monthly
T-bond-futures return is 7.26 for our first RPR period (λ1 + λ2) and 8.85 for our second RPR
period (λ1 + λ3), versus 0.35 for the remainder of our sample (λ1 only). In row 4, the estimated
relation between the monthly TIV-changes and the monthly T-note-futures return is -5.19 for
our first RPR period (γ1+γ2) and -4.53 for our second RPR period (γ1+γ3), versus 1.13 for the
remainder of our sample (γ1 only). The estimated λ2, λ3, γ2, and γ3 coefficients are all highly
statistically significant. For the monthly horizon, there is no reliable relation between the risk
13
changes and the Treasury-futures return for the non-RPR periods (λ1 and γ1).
Table 2, Panel B, reports the results for the change in the forward rates (FR). The results for
the 5-year and 10-year FRs depict the same qualitative relation as those in Panel A: (1) overall,
unconditionally, VIX-increases (TIV-increases) are associated with declining (increasing) FRs;
and (2) the conditional relations are much stronger over our RPR periods. For the monthly
horizon for the 5-year and 10-year FRs, note that the ∆V IX-∆FR relation has a different
algebraic sign for the RPR versus non-RPR periods.
The results for the 1-year FR, FR01, are somewhat different with the ∆V IX-∆FR relation
being stronger for our first RPR period but weaker for our second RPR period. By December
2008, the Federal Reserve had essentially locked T-bill yields to near zero for the remainder of
our sample. This suggests little variability in short rates over this period, so we would expect
smaller variability in the 1-year FR over our later RPR period. Figure 3 displays the time-
series of FR levels over our sample, and shows that the 1-year FR fell below 1% in early 2009
where it remained for the majority of our second RPR period. By mid 2011, the one-year FR
became especially low, with little variability. Given these observations, it is not surprising that
the connection between the 1-year FR and the risk changes is weaker over our second RPR
period, with the weakening for the ‘1-year FR’-risk connection corresponding with a substantial
strengthening for the ‘10-year FR’-risk connection.
3.3. Separate Risk-to-Treasuries Estimations for RPR & Non-RPR Subperiods
Next, we expand our analysis along two dimensions. First, we estimate the risk-to-Treasuries
connection separately for key subperiods so we can compare each RPR period with our non-RPR
periods. Second, we also add the lagged risk-change terms as additional explanatory terms to
evaluate the intertemporal risk relation. If there is any reliable relation between the dependent
variables and the lagged risk changes, then the evidence may prove useful in interpreting the
concurrent relation. For example, if the T-bond return’s relation to the concurrent ∆V IX term
and the lagged ∆V IX term have different algebraic signs, then this would indicate a reversal of
the concurrent relation. A reversal would suggest that the concurrent relation is partially due
to a premium earned by liquidity providers to absorb any concurrent demand and supply shocks
where TrV art−j,t indicates one of five different Treasury-related dependent variables over t − j
to t as in Section 3.2 and Table 2; the subscripts t−2j,t−j on a variable indicate the first-order
lags of the respective term; and the other terms are as defined for equation (1). We report on
rolling monthly and weekly horizons, where j equals 22 or 5 trading days.
Table 3, Panel A, reports on these estimations over the RPR subperiods. For the longer-term
Treasury-futures returns and the 5- and 10-year FR changes, we note that the R2 values are all
quite sizable. For example, for the T-Bond and T-Note futures returns at the monthly horizon,
the average R2 value is over 45% across Panels A.1 to A.4. The estimated λ1 coefficients (γ1
coefficients) on the concurrent ∆V IX (∆TIV ) term are sizable and highly statistically significant
for all eight cases (rows 1 to 4 and 6 to 9); indicating that VIX increases (TIV increases) are
associated with increasing (decreasing) bond prices and decreasing (increasing) distant forward
rates. The strong risk-return connections over the 2011:07 - 2013:12 subperiod (Panel A.4)
indicate that the strong connection over 2009-2013 is not simply due to extreme market dynamics
over the 2008-09 recession and financial crisis, since this subperiod begins over two years after
this recession formally ended in June 2009.14
Regarding the intertemporal relation between the Treasury variables and the lagged risk-
change terms for the RPR periods, we find that the coefficients on the lagged risk-change terms
have the same algebraic sign as the comparable coefficients on the concurrent risk-change terms
for all but three of the 64 estimated coefficients for the TB, TN, FR05, and FR10 variables (Panels
A.1 to A.4). A majority of the estimated λ2 and γ2 coefficients are statistically significant. This
continuation in the risk-return relations for our RPR periods casts doubt on a ‘return to liquidity
provider’ explanation for the concurrent relations and suggests a striking risk-return linkage that
14We evaluate one-half subperiods for our second RPR period because the second-half of this RPR period is
substantially removed from the economic crisis of 2008-09 and a 30-month one-half RPR subperiod evaluation
seemed of adequate length. The first RPR is only 31 months so we did not evaluate one-half subperiods for it.
15
extends beyond the concurrent relation.
The FR01 results are similar. However, as compared to the risk-FR relations for FR05 and
FR10, the relations between FR01 changes and the risk changes tend to be stronger over our first
RPR period (2001:10 - 2004:04), but weaker over our second RPR period (2009:01 - 2013:12), in
terms of the size of the estimated coefficients and the R2 values. Our discussion in Section 3.2
discusses likely influences behind these differences.
Table 3, Panel B, reports coefficient estimates for our non-RPR subperiods, with separate
estimations for both the full non-RPR period and the portion of the non-RPR periods that is
classified as being in an economic expansion by the NBER (or ‘non-RPR-Growth’ subperiods, see
Section 2.5 and Appendix A). We evaluate the growth portion of the non-RPR period separately
to ensure our results are not being driven by the early recession months. For the longer-term
Treasury-futures returns and the 5- and 10-year FR changes, we note that the risk-return connec-
tions are much weaker, as compared to Panel A. The estimated λ1 coefficients on the concurrent
VIX-change term are not statistically significant for the 10-year FR, at either time horizon for
either non-RPR period. The estimated γ1 coefficients on the concurrent TIV-change terms are
mixed. For FR05 and FR10, only one of the estimated coefficients on the lagged risk-change
terms is statistically significant at a 5% p-value.
The contrast between the Panel A and Panel B results is striking. The average R2 of rows 1-4
and 6-9 for the four RPR subperiods (Panel A) is 36.7% with a range of 16.9% to 61.5% versus a
comparable average R2 for the four non-RPR periods (Panel B) of 3.6% with a range of 0.5% to
11.1%. The estimated coefficients (λ1 and γ1) also tend to have a much larger magnitude for our
RPR periods. For example, for the 5-year FR, the average estimated λ1’s on the ∆V IX term is
-1.09 (-0.68) for the monthly (weekly) horizon for the four RPR periods in Panel A versus -0.08
(-0.11) for the monthly (weekly) horizon for the four non-RPR periods in Panel B.
The results are somewhat different for the 1-year FR in Panel B. In all four subpanels in Panel
B, the connection between ∆V IX and the 1-year FR is negative and statistically significant for
the weekly change horizon. The R2 values are somewhat higher than for FR05 and FR10, ranging
from 4 to 26%. This finding is suggestive that FTQ pricing influences over the non-RPR periods
(while seemingly less important as compared to our RPR periods) were likely more tied to linkages
16
with the money market. Recall that the T-bill yields were appreciably higher over these periods,
and the term-yield-spread appreciably lower, which suggests that the money market might have
been the more favored safe-haven lower-risk asset.
To evaluate further the efficacy of our RPR vs. non-RPR economic-state divisions, we es-
timate the same risk-return Treasury connections for the weekly horizon over rolling one-year
estimation periods of 251 trading days each. We then retain the R2 values for each of the rolling
regressions for the Treasury-futures returns and plot the time-series of R2 values in Figure 4. The
results indicate much higher R2 values since about January 2009, corresponding to our second
RPR period. The figure also indicate a second region where the R2 values tend to be consistently
elevated, roughly corresponding to our first RPR period over October 2001 to April 2004. The
consistency of these elevated R2 values across these two RPR periods reinforces the notion of
prolonged periods with different risk-return dynamics, rather than limited results linked to a few
isolated extreme cases.
3.4. The Risk-to-‘Term Structure Slope’ Connection in Treasuries
Next, we evaluate the connection between the term-structure’s slope and changes in asset-class
risk perceptions, which addresses our third empirical question. We estimate the following model:
∆TSSt−j,t = α0 + (λ1 + λ2D0104t + λ3D
0913t )∆log(V IXt−j,t)+ (3)
(γ1 + γ2D0104t + γ3D
0913t )∆log(TIVt−j,t) + ϵt−j,t
where ∆TSS indicates the change in the Treasury term-structure’s slope over the period from
t−j to t; and the other terms are as defined for equation (1). We use two measures of the change
in the term-structure’s slope: (1) the change in the term-structure’s second principal component
(∆PC2), and (2) the change in the term yield spread(∆TY S). The principal components are
estimated from the 10 zero-coupon bond yield with maturities from 1-year to 10-years, and the
term yield spread is defined as the difference between the 10-year and 6-month Treasury Constant
Maturity Yield. As before, we report on the monthly and weekly change horizons.
Table 4 reports the results. We note two primary findings. First, the estimated relation
between ∆TSS and ∆V IX is positive, but statistically insignificant, over the non-RPR periods
(λ1 in the row-2,-4,-6, and -8 models). Second, the ∆TSS-∆V IX relation is negative and much
17
lower for our two RPR periods (λ2 and λ3 in the row-2,-4,-6, and -8 models). The findings reflect
the relatively stronger relation between equity-risk innovations and longer-term Treasuries over
our RPR periods (relative to short-term Treasuries), which suggest that longer-term Treasuries
became relatively more of a favored safe-haven asset over our RPR periods (in a FTQ sense).
We also note that the estimated relation between ∆TSS and ∆TIV tends to become more
positive over our RPR periods, but only reliably so for our second RPR period (2009:01 - 2013:12).
This suggests that an increased TIV induces relatively higher term risk premia over our RPR
periods, presumably through an own-asset risk feedback mechanism. If so, this seems consistent
with the notion of higher risk aversion over our RPR periods. Collectively, the Table 4 evidence
again indicates state-dependent dynamics between risk and term-structure behavior.
3.5. Robustness with Alternative Specifications
In Appendix D, we investigate robustness of results in Tables 2 through 4 with a variety of
alternative empirical approaches, including: (1) jointly estimating the conditional mean and
conditional volatility in a specification that also models time-variation in volatility; (2) a specifi-
cation that includes additional term-structure state variables and the lagged dependent variable
as additional explanatory terms; (3) specifications that consider alternative functional forms for
the VIX and TIV changes. To summarize, our primary results depicted in Tables 2 through 4
remain reliably evident.
3.6. Risk and Return in Treasuries and BEX (2009) and BE (2013)
Our findings bear on empirical implications from BEX (2009) and BE (2013). First, the BEX
model generates a FTQ effect where higher economic uncertainty is associated with lower interest
rates due to a precautionary savings effect. This dynamic is evident in our estimates. The
unconditional relations depicted in Table 2 (models in rows 1, 3, 5, 7, 9 and 11) support this
implication.15
Second, we find that the risk-return connections are dramatically stronger for the longer-term
Treasuries in our RPR state, as compared to the much weaker or even non-existent connections
15Bansal, Connolly, and Stivers (2014) find similar unconditional results in comparable regressions on monthly
returns that also include the concurrent stock return as an additional explanatory variable.
18
for our non-RPR state. Qualitatively, we argue that these state-dependent differences are also
suggested by the economic forces inherent in the BEX framework. Such an interpretation requires
a more detailed review of their model, which we provide next.16
In BEX (2009), risk aversion increases as current consumption declines toward a habit or
subsistence consumption level. Specifically, the coefficient of relative risk aversion is equal to
γCt
Ct−Ht; where γ is a risk-aversion parameter, Ct is current consumption, and Ht is a subsistence
or habit level of consumption. They define Qt =Ct
Ct−Htas a preference shock. In BEX, “increased
volatility (referring to higher economic uncertainty) unambiguously drives up bond prices” (page
63), due to a precautionary saving motive. The response of interest rates to movements in
economic uncertainty is shown in their equation (28) (BEX page 70). There, the relation between
interest rates and economic uncertainty movements is more negative when either the positive risk
aversion parameter γ is larger or when the negative σqc parameter is more negative, where σqc is
a correlation parameter between consumption growth and qt (qt = ln(Qt)).17
We expect that σqc would become relatively more negative in our RPR state. Recall that
our RPR periods commence later in recessions (after most or all of the Fed easing in the Fed
Funds rate has occurred) and continue through the early, uncertain stages of recovery. Thus,
over our RPR periods, current consumption should be relatively closer to the habit/subsistence
consumption level. If so, qt is relatively high, driving up risk aversion, and changes in qt will
also be more sensitive to small consumption changes. Thus, one would expect a relatively more
negative σqc in our RPR state. In addition, higher risk aversion in our RPR periods could also be
linked to a relatively larger γ parameter, presumably reflecting investor psychology in the RPR
state with a weak economy that followed a dramatic stock market decline.
For these reasons, we argue that the much stronger relation between ∆V IX and the Treasury
values fits a ‘conditional version’ of the BEX framework, in an intuitive ‘comparative statics’ sense
16Our discussion in this subsection appeals to intuition and qualitative deduction from the BEX and BE frame-
works, rather than exact quantitative results based on their formal calibrations or model. The BEX calibrations
focus on unconditional parameters, while our focus is on economic-state-based differences.17We note that the BEX equation (28) applies strictly to real interest rates, but their results relating uncertainty
to nominal rates are similar. Additionally, with generally modest and stable inflation over our sample period,
movement in nominal rates should be largely aligned with movements in real rates over our sample, especially for
modest one-week and one-month horizons.
19
where our RPR state seems likely to have a more negative σqc and/or larger γ. The opposing
‘comparative static’ state is a more typical growth state where one would expect lower risk
aversion with a less negative σqc. The economic intuition is that bad economic times have higher
risk aversion that is also more sensitive to consumption changes. Thus, positive uncertainty
where the terms are as defined for Table 2, except that here we use non-overlapping Friday-to-Friday weekly
returns. When allowing for three structural breaks with a minimum subperiod length of 10% of the sample,
the estimation identifies three breaks with the following four subperiods for both the weekly T-Bond-
futures returns and the weekly changes in the 10-year forward rates: 10/3/1997-10/26/2001, 11/02/2001-
4/23/2004, 4/30/2004-1/9/2009, and 1/16/2009-12/27/2013. The 2001:10 - 2004:04 and 2009:01 - 2013:12
subperiods here have strong ‘risk-return’ and ‘risk-FR’ connections. For the 30-year T-Bond-futures
(change in 10-year forward rates), the λ1’s average 5.68 (-0.53) for these two high-connection subperiods
vs. an average λ1 of 1.65 (-0.02) over the remaining 1997-2001 and 2004-2008 subperiods. When allowing
for a maximum of five structural breakpoints, the indicated strong-connection subperiods are similar.
We favor the approach with three structural breaks, because of the links to formal recessions and other
economic-state distinctions (see A.2 through A.7 in this appendix). Thus, we select the periods over
2001:10-2004:04 and 2009:01-2013:12 as our ‘recessionary/post-recessionary’ (RPR) periods that have
relatively strong risk-return connections in longer-term Treasuries. The remaining two subperiods in our
sample, 1997:10 - 2001:09 and 2004:05 - 2008:12, are referred to as non-RPR periods.
We also analyze weekly 10-year T-Note futures returns using the same structural-break methods. For
the weekly 10-year T-Note futures return when allowing for three structural breakpoints, we estimate that
the high risk-return subperiods are 12/15/2000 to 4/23/04 and 12/28/07 to 12/27/13. These subperiods
51
are similar to our RPR periods above, but the beginning dates here are somewhat earlier in December
2000 and December 2007. These December 2000 and December 2007 dates are around the beginning of
the nearby recession (March 2001 and December 2007 are the recession-start months).
Thus, the results for the 10-year T-Note futures returns suggest that we also evaluate the 1997:10 -
2001:02 and 2004:05 - 2007:11 as alternative non-RPR periods that only contain NBER expansion months,
which we refer to as non-RPR-growth periods. Accordingly, in Tables 3, 5, 6, and 7, we evaluate the non-
RPR-growth periods as a comparison to the full non-RPR periods proposed above. If our findings are
similar for each non-RPR period and the corresponding inclusive expansionary non-RPR period, then this
would imply that our results for the full non-RPR periods are not being dominated by high volatility
times associated with the onset of the corresponding regression. Contrasting the expansionary non-
RPR periods with our RPR periods also fits with theoretical implications in David and Veronesi (2013),
regarding economic-state differences in the relation between stock returns and inflation news. Also, see
our supporting discussion in Section A.3, regarding Federal Reserve action around the onset of recessions.
A.2. Economic-State Differences in the Relation between Stock Returns and Inflation
News. This is the focus of Section 4 in our main text, so we refer readers to that section.
A.3. Bai-Perron Analysis of the Variance-Risk Premium and the Bias in VIX for Fore-
casting the Subsequent Realized Volatility. The stock-market’s variance risk premium (VRP) in-
dicates when option values are high relative to the recent realized stock volatility, measured in our case
by the difference between the implied stock-market volatility (VIX) at the close of day t and the realized
volatility of the SPY ETF using 5-minute stock returns over trading days t− 21 to t. Movements in the
VRP are likely due to variations in the premium linked to high uncertainty about the underlying volatility
and, perhaps, a higher aggregate risk aversion in the market (see Bollerslev, Tauchen, and Zhou (2009);
Bollerslev, Gibson, and Zhou (2011); and Bekaert, Hoerova, and Lo Duca (2013)). When we estimate
a Bai-Perron structural break model on the average VRP over October 1997 - December 2013, we find
high-VRP subperiods over 2001:10 - 2004:11 and 2009:01 - 2013:12. These high-VRP subperiods overlap
closely with the RPR periods proposed in A.1 above. We describe this VRP behavior in Table 7 and
Figure 1 in the main text. Similarly, Appendix Table A1 shows that the bias in VIX is reliably higher for
our RPR periods.
A.4. Relation to Economic Recessions and Federal Reserve Action. The onset of our RPR
periods occurs after NBER recessions have been ongoing for some time, with the timing appearing to have
ties to Federal Reserve actions. Consider Federal Reserve actions over the 2008-2009 economic crisis. Over
2008, the Federal Reserve lowered the targeted Fed Funds rate (FFR) seven times from a beginning-of-the
year value of 4.25% down to an end-of-the year value of ‘0 to 0.25%’ with the final change on 12/16/2008.
The targeted FFR remained at essentially zero through our sample end-date of December 2013. This final
reduction in the FFR was just two weeks prior to the onset of our second RPR period in January 2009.
Additionally, the Federal Reserve announced their Quantitative Easing I on November 25, 2008, when
it announced the intention of purchasing $800 billion in bank debt, MBS, and Treasury Notes. Sizable
purchases began in early 2009. Thus, by January 2009, the Federal Reserve had essentially completed its
available easing in the short-end of the yield curve and announced a major initiative in the longer-end of
52
the yield curve. This suggests, by January 2009, that the Federal Reserve had limited additional options
to influence prices with traditional monetary policy tools. Thus, by then, weekly and monthly changes in
market-based risk perceptions might have become relatively more important for understanding Treasury
values, rather than direct Federal Reserve actions or announcements. We also note that January 2009, the
first month of this RPR period, is also in the later stages of a recession. The NBER dated the so-called
Great Recession as occurring over December 2007 to June 2009, with the the official announcement of the
beginning month in December 2008.22
Next, recall that our first RPR period commenced in October 2001. October 2001 followed a major
civil crisis, the 9/11 terrorist attack, and was in the later stages of the recession that occurred over March
to November 2001. This recession’s beginning was retroactively announced in November 2001, within one
month of the beginning of our first RPR period; this recession’s end was retroactively declared on July
17, 2003. Over 2001, the Federal Reserve lowered the targeted FFR 11 times from 6.5% at the beginning-
of-the year to 1.75% at the end-of-the year, with the last decrease on 12/11/2001. After 2001, the Federal
Reserve lowered the targeted FFR only twice more (to 1.25% on 11/6/2002, and to 1.0% on 6/25/2003)
before reversing and raising the targeted FFR to 1.25% on June 30, 2004. Note that this increase in June
2004 was within a couple months of the ending month for our first RPR period (April 2004).
We have argued that our RPR periods were likely to have had greater aggregate risk aversion, especially
relative to the higher growth expansionary periods in our sample. However, there is a sizable proportion
of the early stages of each corresponding recession that is not included in our RPR periods.
We were initially puzzled by the fact that earlier recession months are not in our RPR periods. We
offer the Fed’s strong fast responses to the recession as a potential explanation. To probe this issue, we
analyze the calendar year 2008 and examine whether the relation between VIX changes and bond returns
is different around days on which the Federal Reserve announced a lowering of the targeted FFR. Since
the announcements come late in the day (2:15 PM Eastern time), we analyze two-day returns and two-day
VIX changes to cover the announcement day and subsequent day. Over 2008, for the non-announcement
periods, we find that the typical positive relation between VIX changes and T-Note returns is reliably
evident. However, for the announcement periods, we find that the relation between VIX changes and the
T-Note return is negative and reliably lower.
These findings are largely tied to the episodes over October 8-9 and December 16-17, 2008. On October
8, the targeted FFR was lowered from 2% to 1.5%; and over October 8-9, the VIX increased 10.2%, the
S&P 500 futures return was -9.1%, and the 10-year T-Note futures return was -3.1%. This large negative
T-Note futures return with such a large VIX increase is unusual. For this FFR announcement, the
market appeared to be either disappointed that the FFR reduction was not larger, or interpreted the
FFR reduction as confirming a collapsing economy which spurred a flight to the safest T-Bills.23 Next, on
December 16, 2008, the targeted FFR was lowered from 1% to ‘0 to 0.25%’; and over December 16-17, the
22On September 20, 2010, the NBER announced an end date of June 2009 for the 2008-09 recession.23Kontonikas, MacDonald, and Saggu (2013) argue that over the 2008 financial crisis that the stock market
appeared to interpret some FFR reductions as signals of worsening future economic conditions, with a corresponding
stock decline.
53
VIX decreased by 6.9%, the S&P 500 futures return was +3.6%, and the 10-year T-Note futures return
was +1.7%. This positive T-Note futures return with such a large VIX decrease is also unusual. So, for
this case, the market seemed to be reassured by the FFR announcement and both stock and longer-term
Treasury investors benefitted. In both of these instances, stocks and bonds moved together (with the same
algebraic signs), with the relation between the T-Note-futures returns and VIX changes being opposite
to its more typical ‘weak economy’ relation. By comparison, over all of 2008, the two-day returns of the
S&P 500 futures and 10-year T-Note futures return had opposite signs over 71% of the time.
In 2001, the targeted FFR was lowered 11 times. While there is not a statistically reliably difference
in the relation between the VIX changes and the Treasury-futures return for the FFR announcement
periods over 2001, we note that there was both a VIX decrease and positive returns for both the stock
and Treasury futures for three of the two-day announcement periods.
The above observations are consistent with the premise that the relation between VIX changes and
T-bond values early in these recessions is likely obscured by strong Federal Reserve responses, at least to
some degree. This is one of the reasons that we evaluate the expansionary non-RPR periods separately
(1997:10 - 2001:02 and 2004:05 - 2007:11), as discussed in Section A.1 above.
A.5. Differences in the Term Yield Spread and T-bill Yields. The Treasury term-yield-spread
(TYS, defined as the difference between the Treasury’s 10-year and 6-month Constant Maturity yields)
and the Treasury short-rate (defined as the 6-month Treasury Constant Maturity yield) also display
considerable state-based variation, with a much higher TYS and lower T-bill yields during our RPR
periods. See Section 5.2, Table 7, and Figure 1 for discussion and evidence.
A.6. Relation to Economic Growth and Inflation. We examine the quarterly growth in real U.S.
GDP. Over 1960 through the third quarter 1997 (preceding our sample period), the average annualized
quarterly GDP growth was 3.45%. Over October 1997 to December 2000 (our first expansionary non-RPR
period), the average GDP growth rate was a relatively high 4.09%. For our subsequent RPR period over
October 2001 to March 2004, the average GDP growth rate was 2.85% over the full subperiod and 1.81%
for the first half of the subperiod.24 Next, over April 2004 to September 2007 (our second expansionary
non-RPR period), the average GDP growth was 2.68%. For our subsequent RPR period over January
2009 to December 2013, the average GDP growth rate was 1.80% over the full subperiod and 1.19% for
the first half of the subperiod. Thus, relative to the preceding period back to 1960 and relative to the
preceding economic expansion, both our RPR subperiods had low GDP growth.
We also evaluate inflation over our RPR periods, using the the monthly percentage change in the CPI
index (CPI for All Urban Consumers: All Items). Over 1960 through the third quarter 1997 (preceding our
sample period), the average annualized monthly inflation rate was 4.52%, with an average of 6.07% over
the 1970’s and 80’s. Over October 1997 to February 2001 (our first expansionary non-RPR subperiod),
the average inflation rate was 2.58%. For our subsequent RPR period over October 2001 to March 2004,
the average inflation rate was 1.98%. Next, over April 2004 to November 2007 (our second expansionary
non-RPR period), the average inflation rate was 3.30%. For our subsequent RPR period over January
24Since the GDP numbers are reported quarterly, we select the quarters to best match our subperiods, which
are delineated by calendar months.
54
2009 to December 2013, the average inflation rate was 2.09%. Thus, relative to the preceding period back
to 1960 (and especially relative to the 1970’s and 80’s) and relative to the preceding economic expansion,
both of our RPR subperiods had lower inflation. Overall, the GDP and CPI statistics here support the
premise that our RPR state can be considered as having relatively low economic growth and low inflation.
A.7. Proximity to Stock Market Declines. Both of our RPR periods follow substantial stock
market declines. For our RPR period commencing in October 2001, the stock market had declined by
-35.4% on October 1, 2001 from its peak on March 24, 2000 (based on the CRSP value-weighted stock
index). For our RPR period commencing in January 2009, the stock market had declined by -39.8% on
January 2, 2009 from its peak on October 9, 2007. Evidence in Guiso, Sapienza and Zingales (2013)
suggests that aggregate risk aversion is likely to be higher following such sizable stock market declines.
A.8. Relation to Economic-State Information in Campbell, Sunderam, and Viceira (CSV)
(2013). After completing our analysis in A.1 through A.7 above and the initial draft of our study, we
noted a strong relation between the onset of our RPR periods and movements in a key CSV state variable,
ψ1. In CSV, ψ1 is an important state variable that governs time variation in the volatility of the real
interest rate and its covariation with their stochastic discount factor. CSV note that their ψ1 variable is
identified primarily through the covariance of stock and bond returns and the volatility of bond returns.
We direct readers to Figure 8 in CSV, which depicts the time-series of their ψ1 state variable. First, in
regard to our full October 1997 to December 2013 period, we note that the ψ1 falls precipitously around
1997 until it turns negative in the early 2000’s, and it has remained negative since then (CSV’s graph
stops in late 2009). Recall that the stock-bond return correlation shifted to predominantly negative in the
fall of 1997. Analysis in Campbell, Pflueger, and Viceira (2014) also suggest a regime shift in 1997. These
observations support the notion of a key regime shift around the fall of 1997, which we believe supports
our rationale for starting our analysis then.
We also note that their ψ1 state variable moved to local minimums following the recession of 2001
and in the fall of 2008 (see CSV’s Section 4.3), roughly corresponding to the onset of our RPR periods.
In terms of our RPR classification, we find this observation to be reassuring: CSV’s analysis supports the
notion of economic-state transitions at about the same times as does our empirical-based analysis.
Additionally, CSV’s analysis suggests that longer-term Treasury bonds have become more of a hedge
asset in recent times. Their Figure 10 suggests that the expected excess return of 10-year nominal T-
bonds was negative over much of our two RPR periods, especially around 2009. This is consistent with our
evidence that suggests longer-term Treasuries became more of a favored safe-haven asset over our RPR
periods, especially over our later RPR period from 2009-2013.
While these links to the CSV study are reassuring, we primarily explain our findings in our main text
in relation to BEX (2009), BE (2013), and DV(2013). A key difference between CSV versus BEX and
BE is that time-varying risk aversion is important in BEX and BE, while CSV assume a constant price
of risk. CSV note that they have estimated an extension of their model with time-varying risk aversion,
but their primary paper assumes a constant variance for the stochastic discount factor. Accordingly, the
BEX model seems a better fit for our purposes, especially in view of our evidence that strongly suggests
time-varying risk aversion.
55
Appendix A - Table A1:
VIX and the Future Realized Stock Volatility: Bias Variation Linked to our RPR Periods
This table reports how the option-derived implied stock volatility, VIX, is related to the subsequent
realized stock return volatility. We estimate variations of the following model:
σStt,t+21 = (γ1 + γ2Dum
0104t + γ3Dum
0913t )V IXt−1 + εt,t+21
where, σSTt,t+21 is the annualized standard deviation that is estimated from the volatility of the 5-minute SPY
returns over trading days t to t+21 (see Appendix C); V IXt−1 is the closing V IX from trading day t− 1,
Dum0104t is a dummy variable that equals one if the lagged VIX is in our first RPR period over 2001:10 to
2004:04, Dum0913t is a dummy variable that equals one if the lagged VIX is in our second RPR period over
2009:01 to 2013:12), and the γ’s are coefficients to be estimated. For model variation b. through e., we
report on the model estimated over the denoted subperiod, so the dummy variables are not applicable. The
overall sample period is 1997:10 - 2013:12. T-statistics are in parentheses, calculated with heteroskedastic
and autocorrelation consistent standard errors.
Period Model γ1 γ2 γ3 R2
Full Sample a. 0.946 (23.92) -0.142 (-2.91) -0.168 (-4.15) 66.5%
1997:10-2013:12
Non-RPR I b. 0.958 (31.84) n/a n/a 24.3%
1997:10-2001.09
RPR I c. 0.802 (23.10) n/a n/a 62.7%
2001:10 - 2004:04
Non-RPR II d. 0.954 (14.49) n/a n/a 71.8%
2004:05-2008.12
RPR II e. 0.778 (39.97) n/a n/a 71.5%
2009:01 - 2013:12
56
Appendix B - Table B1:
Implied Volatility and the Future Realized Volatility of Stock and Bond Returns
This table reports how our option-derived implied volatilities, VIX and TIV, are related to the subse-
quent realized stock and T-bond return volatility, respectively. We estimate the following two models for
both the subsequent stock return volatility and the T-bond return volatility.
(a) σzt,t+21 = γ0 + γ1IVt−1 + εt,t+21
(b) log(σzt,t+21) = γ0 + γ1log(IVt−1) + εt,t+21
where, for equation (a), σzt,t+21 is the annualized standard deviation for the returns of asset-class z over
trading-days t to t + 21, calculated as the square-root of the sum of 22 squared daily returns over the
rolling 22-trading-day period; IVt−1 is either the lagged V IX or TIV for the stock-volatility model and
Treasury-volatility model, respectively, where V IX and TIV are the equity and T-note implied volatilities
as explained in Section 2.1, and the γs are coefficients to be estimated. For equation (b), the model is the
same except that we take the log of the realized standard deviation and of the implied volatilities in order
to transform the variables to be closer to normally distributed. Panel A reports on stock-market volatility,
where z indicates stock market returns, using the daily S&P 500 futures returns. Panel B reports on T-bond
volatility, where z indicates 10-year T-Notes, using daily 10-year T-Note futures returns. The sample period
is 1997:10 to 2013:12. T-statistics are in parentheses, calculated with heteroskedastic and autocorrelation
consistent standard errors. The superscript 1 indicates a 1% p-value.
Panel A: Realized Stock Return Volatility over t to t+ 21
as the dependent variable, V IXt−1 as the explanatory variable
Period Model γ1 R2
1997:10-2013:12 a. 0.911 (10.07)1 55.2%
b. 1.050 (21.31)1 60.1%
Panel B: Realized 10-yr T-Note Return Volatility over t to t+ 21
as the dependent variable, TIVt−1 as the explanatory variable
Period Model γ1 R2
1997:10-2013:12 a. 0.872 (16.09)1 52.5%
b. 0.963 (20.31)1 55.5%
57
Appendix C:
Calculation of the Stock Market’s Realized Volatility from High-Frequency Returns
Trading records for the SPY S&P 500 ETF were downloaded from the TAQ dataset on WRDS. We
deleted any trading records with a negative price or trading volume, or where the correction indicator
showed a trade had been corrected or cancelled. We also eliminated any trading record where the sale
condition met any of these criteria: cond =“O”, “Z”, “B”, “T”, “L”, “G”, “W”, “J”, or “K”. These
screens mimic typical filter rules applied in empirical microstructure studies. We also dropped any record
with a timestamp before 9:30 a.m. or after 4:00 p.m.
Using data cleaned in this way, we identify the first trade of the day, and then our algorithm identifies
the first trade after 300 seconds have elapsed, and then the first trade after the next 300-second interval,
and so forth through the end of the trading day. In the early years of the sample, the volume of trading
was sufficiently low on some days such that the interval between trades was larger than 300 seconds, but
this was relatively rare. In later years of the sample when there are multiple trades per second, we use
the first trade after the 300-second interval since trades within the second are arranged in the order of
execution. See Holden and Jacobsen (2014) for details on this timing issue.
We compute five-minute squared returns (r2) from this sequence of 5-minute prices, and with that
sequence, we compute the following estimate of the realized volatility (RV, in standard deviation units):
RVt =√12 ×
√√√√ 21∑i=0
r2t−i (8)
where the summation and i subscript indicates all the five-minute return shocks, calculated in this way,
over trading days t back through t− 21. Thus, the RV denoted on day t captures a rolling 22-trading-day
period to generate a daily estimate of a rolling one-month RV in S&P500 returns. Our RV data is also
computed over 1997:10 - 2013:12.
58
Appendix D:
Robustness of the Risk-Return Connection in Longer-term Treasuries
In this appendix, we probe robustness of our risk-return findings for longer-term Treasuries by report-
ing on different variations of the models that we estimate in Sections 3.2 and 3.4.
D.1. Allowing for Time-variation in Volatility. Time-variation in risk is a fundamental premise
behind our empirical investigation, so it would seem natural to estimate a specification that also allows for
time-varying volatility in our models’ residuals. Accordingly, to probe robustness for our results in Tables
2 through 4, we estimate a system where equation (1) or equation (3) is the conditional mean equation
and the following equation is the conditional variance equation:
vt−j,t = α0 + α1TIVt−j (9)
where vt−j,t is the conditional variance of the residual, ϵt−j,t, either from equation (1) or equation (3);
TIVt−j is the implied volatility from the 10-year T-Note futures options at the close of day t− j; the α’s
are additional coefficients to be estimated, with j=5 for weekly change horizons and j=22 for monthly
change horizons. Results in Appendix B indicate that the TIV provides substantial information about the
subsequent volatility of longer-term Treasury prices, so we use TIV to parsimoniously capture the bond-
market’s volatility environment in this specification.25 We use maximum likelihood estimation that jointly
estimates the conditional mean and conditional variance equations, assuming a conditional normal density.
For the conditional mean equation, we find results that are quite similar to those depicted in Tables 2 and
4. For the conditional variance equation, the estimated α1 coefficient is positive and highly statistically
significant in all cases, which supports the efficacy of using the TIV as a risk measure. With our extensive
evaluation and comparison of subperiods, we elect to present OLS results (with heteroskedastic-consistent
standard errors) in our main tables, due to the simplicity of the OLS model and its intuitive R2 measure.26
D.2. The Risk-Return Connection with Lagged Term-Structure State Variables. Next, to
further probe the relation between T-bond returns and changes in risk perceptions, we extend our prior
models to also control for well-known term-structure state variables and for the lagged dependent variable.
Cochrane and Piazzesi (2005) and others have shown that the term-structure of current yields, or forward
rates, can serve as state variables and provide information about the subsequent Treasury bond returns.
We add the lagged dependent variable to control for and evaluate autoregressive behavior.
25We also evaluated specifications for the conditional variance that include the lagged squared residual, in a
traditional ARCH sense. We find that the lagged residuals do not add additional reliable volatility information
beyond the lagged TIV term.26Tabular results for the specification with time-varying volatility are available from the authors upon request.
59
Accordingly, we next report on variations of the following model: