Journal of Financial Econometrics, 2022, Vol. 20, No.

2, 219–252
https://doi.org/10.1093/jjfinec/nbab025
Advance Access Publication Date: 20 November 2021
Presidential Address

Realized Semi(co)variation: Signs That All

Volatilities are Not Created Equal*

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Tim Bollerslev
Duke University, NBER and CREATES

Address correspondence to Tim Bollerslev, Department of Economics, Duke University, Durham, NC

27708, USA, or e-mail: [email protected].
Received June 21, 2021; revised October 13, 2021; editorial decision October 15, 2021; accepted October 15, 2021

Abstract
I provide a selective review of recent developments in financial econometrics related
to measuring, modeling, forecasting, and pricing “good” and “bad” volatilities based
on realized variation type measures constructed from high-frequency intraday data. An
especially appealing feature of the different measures concerns the ease with which
they may be calculated empirically, merely involving cross-products of signed, or
thresholded, high-frequency returns. I begin by considering univariate semivariation
measures, followed by multivariate semicovariation and semibeta measures, before
briefly discussing even richer partial (co)variation measures. I focus my discussion on
practical uses of the measures emphasizing what I consider to be the most noteworthy
empirical findings to date pertaining to volatility forecasting and asset pricing.

Key words: cross-sectional return variation, downside risk, high-frequency data, jumps and co-
jumps, partial variation, realized variation, return predictability, semibeta, semi(co)variation, vola-
tility forecasting
JEL classification: C22, C51, C53, C58, G11, G12

It is difficult to pinpoint the origin of the research field that we now refer to as financial
econometrics. What is clear, however, is that the rapid growth of the field over the past
three decades has in no small part been fueled by the development of new procedures and

* Presidential address presented at the Thirteenth Annual Society for Financial Econometrics (SoFiE)
Conference, hosted by the University of California, San Diego, June 2021. I will like to thank Allan
Timmermann, Fabio Trojani, and two anonymous referees for their helpful comments. The paper
draws heavily from joint published and unpublished work with Jia Li, Andrew Patton, and Rogier
Quaedvlieg over the past several years. I am deeply indebted to each of them for our many discus-
sions and fun collaborations. I also draw on recent joint work with Sophia Li, Marcelo Medeiros,
Haozhe Zhang, and Bingzhi Zhao. In addition, I would like to thank Saketh Aleti and Haozhe Zhang
for their excellent research assistance.

C The Author(s) 2021. Published by Oxford University Press. All rights reserved.
V
For permissions, please email: [email protected]
220 Journal of Financial Econometrics

empirical findings related to the measurement, modeling, forecasting, and pricing of time-
varying financial market volatility. The importance of volatility for the field as a whole is
also underscored by the SoFiE logo, which prominently features the time series plot of a fi-
nancial asset price subject to volatility clustering. Looking at the price path featured in the
logo, it is evident that “large changes tend to be followed by large changes—of either
sign—and small changes tend to be followed by small changes” (Mandelbrot, 1963).
The initial research on time-varying volatility in financial markets, and the importance
thereof, were primarily based on parametric GARCH (Engle, 1982; Bollerslev, 1986) and

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

stochastic volatility type models (Taylor, 1982). However, the advent of high-frequency
intraday data for a host of different assets and instruments, starting in the early 2000s,
spurred somewhat of a paradigm shift, and many of the most influential developments over
the past two decades have involved so-called realized volatility measures constructed from
high-frequency intraday data (Andersen and Bollerslev, 1998; Andersen et al., 2001b;
Barndorff-Nielsen, and Shephard, 2002). Importantly, this approach also helped to embed
the econometric analyses of return volatility within the vast probability and statistics litera-
tures on Itô semimartingales and the theory of quadratic variation, thereby affording a
rigorous justification for the realized volatility concept based on no-arbitrage assumptions
and theoretical in-fill asymptotic arguments relying on the idea of ever finer sampled
returns over fixed time intervals.1 The quadratic variation estimated by the original realized
volatility measures is effectively blind to the signs of the underlying returns. On the other
hand, numerous studies, dating back to Roy (1952) and Markowitz (1959), have argued
that investors primarily care about negative returns and downside risks.2 Correspondingly,
the basic mean–variance tradeoff arguments that underlie many popular asset pricing mod-
els and predictions, the traditional CAPM included, should instead be based on the down-
side portion of the variation only (Hogan and Warren, 1972; Bawa and Lindenberg, 1977).
An extensive body of research in behavioral finance, supported by experimental evidence
and more formal theoretical arguments rooted in prospect theory and loss aversion
(Kahneman and Tversky, 1979), also suggest that up and downside risks are not treated the
same by investors.3 The ubiquitous Value-at-Risk (VaR) and Expected Shortfall measures
used for assessing the risk of an investment portfolio, and the Sortino ratio (Sortino and
van der Meer, 1991) sometimes employed in lieu of the traditional Sharpe ratio for port-
folio performance evaluation, further echo this asymmetric treatment of gains and losses.4

1 The introductory chapter to the collection of seminal volatility papers in Andersen and Bollerslev
(2018) provides a more in depth discussion of the key new concepts and ideas that have helped
shape these developments.
2 The Merriam-Webster dictionary also explicitly defines “risk” as the “possibility of loss or injury,”
as exemplified by “the chance that an investment (such as a stock or commodity) will lose value.”
3 Scenarios in which upside gains and downside losses are not symmetric arise in many other eco-
nomic situations and prediction problems; for additional discussion, see, for example,
Christoffersen and Diebold (1997), Patton and Timmermann (2007), and Babii et al. (2021).
4 A growing recent literature has also sought to relate these measures of downside risks to macro-
economic outcomes; see, for example, Giglio, Kelly, and Pruitt (2016), Adrian, Boyarchenko, and
Giannone (2019), and Carriero, Clark, and Marcellino (2020). Some of the ideas and new empirical
measures that I discus below could possibly be exploited in that context as well.
Bollerslev j Realized Semi(co)variation 221

Motivated by this line of reasoning and the idea that “good” and “bad” volatilities are
not necessarily created equal, Barndorff-Nielsen, Kinnebrock, and Shephard (2010) first
proposed decomposing the original realized volatility measure into separate up and down-
side realized semivariation measures based on the summation of the squared positive and
negative high-frequency returns, respectively. As I will discuss below, this simple decom-
position has in turn resulted in a number of new and interesting empirical findings pertain-
ing to both volatility forecasting and the differential pricing of the up and downside
realized semivariation measures. Anticipating some of the key findings, higher values of

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

downside (upside) realized semivariance for the aggregate market portfolio appear to be
associated with higher (lower) future aggregate market volatility, higher (lower) future ag-
gregate market return, while differences in the up minus downside semivariances for indi-
vidual stocks appear to be priced positively in the cross-section.
Of course, most issues in asset pricing finance are inherently multivariate in nature,
entailing non-diversifiable risks and the covariation among multiple assets and/or the co-
variation with specific systematic risk factor(s), or benchmark portfolios. Extending the
realized semivariance concept to a multivariate setting, Bollerslev et al. (2020) first pro-
posed an analogous decomposition of the standard realized covariance matrix into four
additive realized semicovariance components defined by the sum of the cross-products of
the signed pairs of high-frequency returns. In parallel to the findings for the realized semi-
variances, this “look inside” of the quadratic covariation has similarly been used in the con-
structions of improved covariance matrix forecasts.
The semicovariances have also been used in the definition of so-called realized semibetas
(Bollerslev, Patton, and Quaedvlieg, 2021a) and in turn more accurate cross-sectional asset
price predictions. Consistent with the implications from a downside CAPM and the results
based on separately estimated up and downside betas (Ang, Chen, and Xing, 2006), only
the two semibetas associated with downside market risk appear to be priced. However,
counter to the implications from a conventional downside CAPM model, the risk premiums
for two separate downside semibetas associated with “good” and “bad” asset-specific
covariations seemingly differ.
The choice of a zero threshold underlying the definitions of the realized univariate semi-
variation, multivariate semicovariation, and semibeta measures is often motivated by the
idea that investors value gains and losses asymmetrically, and therefore also price and pro-
cess “good” and “bad” volatility differently. However, other economic considerations nat-
urally dictate different definitions of “good” and “bad” returns based on some fixed non-
zero threshold, or some benchmark return. From a purely statistical perspective, the choice
of a zero threshold is also somewhat arbitrary, and a non-zero threshold and/or multiple
thresholds could in principle be used in the definition of more refined decompositions of
the quadratic variation. I will briefly discuss recent results based on this idea and corre-
sponding so-called realized partial (co)variation measures (Bollerslev et al., 2021).
Other measures of asymmetries and non-linear dependencies based on high-frequency
data, including measures of coskewness and cokurtosis (e.g., Neuberger, 2012; Amaya
et al., 2015), have, of course, been proposed and analyzed empirically in the literature.5

5 The SoFiE Presidential Addresses by Engle (2011) and Ghysles (2014) also both explicitly highlight
the importance of allowing for asymmetries and skewness in the calculation of VaR and other
downside risk measures.
222 Journal of Financial Econometrics

However, many of these measures involving higher order moments can be difficult to accur-
ately estimate in practice. By comparison, the realized semi(co)variation measures, con-
structed from sums of squares and cross-products of signed high-frequency returns, afford a
simple “look inside” the quadratic variation.
Realized bipower variation measures (Barndorff-Nielsen and Shephard, 2004b), specif-
ically designed to be robust to jumps, afford another such look. Along these lines, several
studies have argued for the importance of separately considering the quadratic variation
stemming from price discontinuities, or jumps and cojumps. I will not discuss this extensive

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

literature on estimating and testing for jumps and co-jumps based on high-frequency-data
at any great length here (see, e.g., Bollerslev, Law, and Tauchen, 2008; Lee and Mykland,
2008; Jacod and Todorov, 2009; Mancini and Gobbi, 2012; Aı̈t-Sahalia and Xiu, 2016; Li,
Todorov, and Tauchen, 2017, among others). However, as I will discuss below, appropri-
ately defined differences between semi(co)variances also consistently estimate jumps and
co-jumps.
Closely related to the above-mentioned studies on jumps, there is a recent and rapidly
growing literature on the pricing of downside tail, or crash risk (including, e.g., Bollerslev
and Todorov, 2011; Kelly and Jiang, 2014; Cremers, Halling, and Weinbaum, 2015;
Bollerslev, Li, and Todorov, 2016; Chabi-Yo, Ruenzi, and Weigert, 2018; Lu and Murray,
2019; Orlowski, Schneider, and Trojani, 2020, among others). However, in contrast to the
realized measures constructed solely from high-frequency data that I focus on here, all of
these studies rely on additional information gleaned from options prices for inferring risk-
neutral distributions, and assessing tail dependencies and the pricing thereof.
Directly paralleling the decompositions of realized volatilities into separate up and
downside variation measures, options implied volatilities may similarly be decomposed
into “good” and “bad” variation through the use of options with different strikes (see, e.g.,
Andersen, Bondarenko, and Gonzalez-Perez, 2015). Going one step further, the variance
risk premium, defined as the difference between the risk-neutral and the actual expected re-
turn variation, may also be separated into up and downside variance risk premiums.6
Consistent with the findings discussed below that most of the volatility persistence can be
traced to “bad” volatility, most of the return predictability inherent in the variance risk pre-
mium (as originally documented by Bollerslev, Tauchen, and Zhou, 2009) seemingly stems
from negative jumps and the downside portion of the premium (see, e.g., Andersen, Fusari,
and Todorov, 2015; Bollerslev, Todorov, and Xu, 2015; Feunou, Jahan-Parvar, and Okou,
2018; Kilic and Shaliastovich, 2019).
To help focus the paper, I will not discuss any of these results based on options-implied
up and downside variation measures any further here. To be clear, however, I think there is
much to be learned from comparing and contrasting the high-frequency-based realized
semi(co)variation measures to the corresponding “good” and “bad” variation measures
implied from options prices. Not only in terms of return predictability, but also in terms of
changes in market-wide perceptions of risks and risk aversion and the underlying economic
mechanisms at work (see, e.g., Bekaert and Engstrom, 2017; Bekaert, Engstrom, and Xu,
2021; Feunou et al., 2020).

6 Further extending this idea, Chabi-Yo and Loundis (2021) provides an options-based approach for
calculating conditional up and downside risk premiums for arbitrary moments.
Bollerslev j Realized Semi(co)variation 223

The plan for the rest of the paper is as follows. I begin with a discussion of the original
univariate realized semivariation measures in Section 1, followed by multivariate semico-
variation measures in Section 2, before briefly considering newly proposed partial (co)vari-
ation measures in Section 3. I will concentrate my discussion on intuition and what I
consider to be the most important new empirical insights to date pertaining to volatility
forecasting and asset pricing, leaving more technical details and formal theoretical argu-
ments aside.

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

1 Semivariation Measures
To formally set out the basic ideas, let ps denote the time-s logarithmic price for some asset,
with the underlying price process originating at time 0. Consistent with the absence of arbi-
trage, assume that the price process may be described by an Itô semimartingale of the form,
ðs ðs
ps ¼ ls ds þ rs dWs þ Js ; s  0; (1)
0 0

where ls corresponds to the drift, rs defines the diffusive volatility processes, Ws is a stand-
ard Brownian motion, and Js denotes a finite activity pure jump process. For concreteness, I
will refer to the unit time interval ½t; t þ 1 as a day, with the realized measures being
defined at the daily frequency.7 For ease of notation, I will assume that high-frequency
intraday prices pt ; ptþ1=K ; . . . ; ptþ1 are observed at K þ 1 equidistant times over the day,
with the corresponding logarithmic discrete-time return over the kth time interval denoted
by rt;k  ptþk=K  ptþðk1Þ=K . It follows then from the theory of quadratic variation for semi-
martingales that for K ! 1 (see, e.g., the general discussion in Aı̈t-Sahalia and Jacod,
2014),8

X
K ðt X
P
RVt  r2t;k ! r2s ds þ ðDJs Þ2 ; (2)
k¼1 t1 t1  s  t

where DJs captures the jump in the price if a jump occurred at time s, and otherwise equals
zero. In other words, for increasingly finer sampled intraday prices, the realized daily vari-
ance, defined as the summation of the within-day high-frequency squared returns, con-
verges uniformly in probability to the sum of the “continuous” price variation plus the
squared “jump” price increments over the day.9

7 Most of the empirical results in the literature, the illustrations highlighted below included, have
also been based on daily realized measures constructed from high-frequency intraday data.
However, the same measures could, of course, be defined over other non-trivial time intervals,
such as a week or a month.
8 The seminal insight underlying this result and the approximation of the quadratic variation process
of a semimartingale by its approximate quadratic variation, aka the realized volatility in the present
context, is according to Jacod and Protter (2012) attributable to Meyer (1967).
9 The assumption of equidistant prices and intraday returns spanning the same 1=K time interval is
not critical for this result, as long as the span of the longest intraday return-interval converges to
zero; see, for example, the references and discussion of alternative sampling schemes in Hansen
and Lunde (2006).
224 Journal of Financial Econometrics

Data limitations and market microstructure complications that disrupt the martingale
property of the price path over very fine time intervals invariably put an upper bound on
the practical choice of K, and thus render the continuous limit unattainable in practice. The
“volatility signature plot” (Andersen et al., 2000b) provides an oft-used informal diagnostic
tool to help gauge the sampling frequency at which market microstructure “noise” and
scarcity of observations start to overwhelm the signal, and in turn bias the realized volatility
estimates.10 I will not discuss this deliberate approach for choosing K in practice any fur-
ther here, nor will I discuss any of the other more advanced estimators and adjustment pro-

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

cedures that have been proposed in the literature to more efficiently estimate the quadratic
variation in the presence of “noise.” Suffice it to say, that the same practical complications
pertain to the semivariation measures that I will discuss next, and that some of these same
“noise robust” procedures might fruitfully be applied in that context as well.11 Meanwhile,
the comprehensive empirical analysis in Liu, Patton, and Sheppard (2015) comparing more
than 400 alternative estimators across numerous asset classes suggests that in practice it is
difficult to beat a simple sub-sampled RV estimator based on 5-min returns.
The realized variation measure in Equation (2) does not differentiate between “good”
and “bad” volatility. In an effort to do so, the realized up and down semivariance measures
first proposed by Barndorff-Nielsen, Kinnebrock, and Shephard (2010) (BNKS henceforth)
separate the total realized variation into two components associated with the positive and
12
negative high-frequency returns, say rþ 
t;k and rt;k , respectively,

X
K
2
X
K
2
RVþ
t  ðrþ
t;k Þ ; RV
t  ðr
t;k Þ : (3)
k¼1 k¼1

It follows trivially that RVt ¼ RVþ 

t þ RVt for all values of K. Meanwhile, maintaining
the basic Itô semimartingale setup and assumptions in Equation (1), in which the order of
the drift is dominated by the diffusive and jump components in the in-fill asymptotic limits,
BNKS show that the separately defined positive and negative semivariance measures con-
verge uniformly in probability to one-half of the integrated variation plus the sum of the
squared positive and negative price jumps, respectively. Hence, to a first-order asymptotic

10 Corsi et al. (2001) and Oomen (2003) have independently proposed a similar procedure for deter-
mining an appropriate sampling frequency.
11 The most prominent ways in which to accommodate “noisy” and scarce high-frequency prices,
without resorting to coarser sampling frequencies and discarding potentially valuable information,
include the two-scale approach of Zhang, Mykland, and Aı̈t-Sahalia (2005), the kernel-based ap-
proach of Barndorff-Nielsen et al. (2008), and the pre-averaging techniques of Jacod et al. (2009).
12 While the basic realized volatility concept, the semivariation measures included, rely on the notion
of ever finer sampled returns over a fixed time interval and asymptotically vanishing drifts, as typ-
ically implemented empirically in the form of daily measures constructed from intraday data, simi-
larly defined measures based on more coarsely sampled data, say monthly measures constructed
from daily returns, are often employed in the finance literature. In this situation, the impact of
price drifts may not be as immaterial. Accordingly the in-fill asymptotic arguments discussed
below may also provide a poorer guide. To help alleviate those concerns, one could rely on mean-
adjusted positive and negative returns.
Bollerslev j Realized Semi(co)variation 225

approximation, the difference between the semivariances is entirely determined by the dif-
ference in the signed squared jumps,
X
 P
SJt  RVþ t  RVt ! ðDJsþ Þ2  ðDJs Þ2 ; (4)
t1  s  t

where DJsþ (DJs ) denotes any positive (negative) price jump occurring at time s. A more
refined second-order asymptotic theory would allow for further differentiation between
RVþ 
t and RVt arising from a non-zero price drift and/or a non-zero correlation between

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

the innovations to the price and stochastic volatility processes, also commonly referred to
as a “leverage effect.” However, this second-order asymptotic theory involves bias terms
that are difficult to quantify, rendering it of limited practical use (for additional details see
BNKS, Kinnebrock and Podolskij, 2008).13 I will briefly return to this issue in my discus-
sion of the realized semicovariance measures in Section 2.
In addition to “looking inside” the semivariation measures to try and identify where dif-
ferences in RVþ 
t and RVt might be coming from, it would also be interesting to investigate
potential time-of-day effects. For instance, it is possible that two different intraday price
paths that manifest in the same daily RVþ 
t and RVt , say one up-down and one down-up,
may have different implications for investor’s risk perception and end-of-day investment
and portfolio rebalancing decisions (see, e.g., the experimental evidence and related discus-
sion in Grosshans and Zeisberger, 2018). As such, semivariation measured over shorter
interday time intervals may afford additional useful information.
Meanwhile, to empirically illustrate the daily semivariation measures, Figure 1 shows
the logarithmic prices for the S&P 500 SPY ETF at 5-min intervals on four different days in
2020. The first panel shows March 3. At 10:00 am that day the Federal Open Market
Committee (FOMC) issued a press release (outside its regularly scheduled announcement
cycle) intended to install confidence and ensure that it was closely monitoring the evolving
pandemic14:

The fundamentals of the U.S. economy remain strong. However, the coronavirus poses evolving
risks to economic activity. In light of these risks and in support of achieving its maximum em-
ployment and price stability goals, the Federal Open Market Committee decided today to lower
the target range for the federal funds rate by 1/2 percentage point, to 1 to 1-1/4 percent. The
Committee is closely monitoring developments and their implications for the economic outlook
and will use its tools and act as appropriate to support the economy.

As the figure shows, the market initially interpreted the statement by the FOMC
very positively, resulting in a 2% jump in the value of the index. However, prices
gradually drifted down over most of the remaining part of the trading day, and as a
result RVþ 
t and RVt ended up fairly close for the day (52.85% and 45.71%, respectively,

13 The Online Appendix to Bollerslev et al. (2020) also provides a more general second-order multi-
variate convergence result, which encompasses the convergence of the realized semivariances
as a special “diagonal” case.
14 The effect of FOMC announcements on asset prices has been the subject of a growing recent lit-
erature; see, for example, Lucca and Moench (2015); Bollerslev, Li, and Xue (2018); and Cieslak,
Morse, and Vissing-Jorgensen (2019), and the many references therein. I will not discuss this lit-
erature here.
226 Journal of Financial Econometrics

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Figure 1 Intraday S&P 500 index prices. The figure shows the logarithmic prices for the S&P 500 SPY
ETF at 5-min intervals for four different days in 2020.

in annualized volatility unit, implying a total daily realized volatility of

69.88% ¼ (52.85%2 þ 45.71%2)1=2 ).15 In contrast, on June 15, as shown in the second top
panel, the S&P 500 steadily rose over most of the trading day. The FED’s long-planned pro-
gram to help facilitate lending to smaller businesses, which finally launched on that day,
may in part account for this increase. At 2:00 pm on that same day, the FED further
announced that it would begin purchasing individual corporate bonds to inject liquidity
into the economy, resulting in an apparent jump in the price at that exact time.
Correspondingly, RVþ 
t far exceeded RVt for the day (equaling 24.28% and 15.01%, re-
spectively). By comparison, on July 9, as shown in the first panel in the bottom row, RVþ t
was less than RV t (equaling 11.63% and 16.08%, respectively). That day began with
reports of sharply increasing coronavirus cases in many parts of the United States, while a
subsequent report of falling unemployment insurance claims may have helped alleviate
some of the worst fears about the adverse economic consequences of the pandemic. The
final March 17 panel shows another day with exceptionally high overall RVt (78.18% for
the day), yet almost identical RVþ 
t and RVt (equal to 54.99% and 55.58%, respectively).
The previous day, March 16, was one of the historically worst days for the market, with
the S&P 500 falling by more than 12% (and also one of the highest daily RVt’s on record at
83.71%). The strong positive pre-noon trend on March 17 obviously helped recover some
of those losses.

15 Guided by the aforementioned findings in Liu, Patton, and Sheppard (2015), the realized measures
reported throughout are based on 5-min returns sub-sampled and averaged at a 1-min frequency.
Bollerslev j Realized Semi(co)variation 227

Looking closer at the price paths for the four deliberately chosen days in Figure 1, the
prolonged intraday return persistence evident for all of the days arguably goes beyond the
notion of gradual jumps put forth by Barndorff-Nielsen et al. (2009) (see also Li et al.,
2017). Nor can the identical signed successive price changes be explained by short-lived
bursts in volatility, as put forth by Christensen, Oomen, and Podolskij (2014) (see also
Bajgrowicz, Scaillet, and Treccani, 2015). Instead, the observed price paths point to more
sustained violations of the basic Itô semimartingale assumption and short-lived price drifts,
possibly associated with more difficult to interpret “soft” news, as also recently discussed

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

by Bollerslev et al. (2020).16 The occasional occurrence of such periods of extreme return
persistence, or drift-bursts, and the importance thereof for realized volatility estimation,
have also been emphasized by Laurent and Shi (2020) and Christensen, Oomen, and Renò
(2021). Further along these lines, Andersen et al. (2021c) and Laurent, Renò, and Shi
(2021) have both proposed a new family of realized volatility estimators for the integrated
volatility in Equation (2) based on a difference-in-difference type approach explicitly
designed to negate the impact of temporary price drifts. It would be interesting to formally
extend and apply these ideas to the “robust” estimation of semivariation type measures.
For less liquid assets, price staleness, or zero returns, as recently emphasized by Bandi et al.
(2020), will also need to be properly accounted for in the estimation.

1.1 Semivariance-Based Volatility Forecasting

The vast empirical literature on GARCH and other parametric stochastic volatility models
suggests that equity return volatility tends to increase more following negative return
shocks than equally sized positive return shocks. Following Black (1976) and Christie
(1982), this return-volatility asymmetry is commonly referred to as the “leverage effect.”
This common terminology notwithstanding, financial leverage has long since been discred-
ited as the primary explanation for the observed asymmetries, as the effect is too large em-
pirically to be explained solely by changes in leverage. Additionally, the asymmetries also
tend to be much stronger for aggregate equity index portfolio returns than for individual
stock returns, further discrediting an all leveraged-based explanation.17
Expanding on this theme, BNKS in their original empirical investigations of the semivar-
iance measures report that the inclusion of RV t in a daily asymmetric GJR-GARCH model
(Glosten, Jagannathan, and Runkle, 1993),18

htþ1 ¼ x þ ar2t þ bht þ dr2t Iðrt < 0Þ þ cRV

t ; (5)

typically renders d and the traditional daily leverage effect term insignificant. The downside
realized semivariance RV t is generally also more informative and drives out the signifi-
cance of the total realized variance RVt when both are included in the conditional variance

16 Relatedly, the recent parametric error correction type model developed by Andersen et al. (2021a)
explicitly seeks to link the arrival of new information and price discovery to short-lived price dri
fts and local mispricing.
17 For additional discussion and more recent estimates of the leverage effect based on high-
frequency data see Bollerslev, Litvinova, and Tauchen (2006); Aı̈t-Sahalia, Fan, and Li (2013); Corsi
and Renò (2012); and Kalnina and Xiu (2017).
18 Following Zakoı̈an (1994), the GJR-GARCH models are also sometimes referred to as a threshold
TGARCH model.
228 Journal of Financial Econometrics

equation. Multiplicative Error Models (Engle, 2002), high-frequency-based volatility

models (Shephard and Sheppard, 2010), or augmented realized GARCH models (Hansen,
Huang, and Shek, 2012) explicitly characterizing the dynamic dependencies in the realized
semivariaton measures, would allow for a more thorough exploration of these features, and
the idea that investors process and interpret “good” versus “bad” news differently (see also
the “good” versus “bad” environment GARCH type of models proposed by Bekaert,
Engstrom, and Ermolov (2015).
Instead of the GARCH-based approach in Equation (5), building on the unified frame-

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

work of Andersen et al. (2003), most realized volatility-based forecasting procedures now
tend to rely on simple-to-implement reduced form time series models estimated directly on
the realized measures. The heterogeneous autoregressive (HAR) model of Corsi (2009), in
which the future realized volatility depends linearly on past realized volatilities over differ-
ent horizons, has arguably emerged as the most popular such model. This particular formu-
lation is now also routinely used as the benchmark model for volatility forecast
comparisons (see, e.g., Bollerslev et al., 2018).
The HAR model also provides an especially convenient framework for incorporating
the realized semivariation measures (and other explanatory variables) into the construction
of volatility forecasts. In particular, following Patton and Sheppard (2015), the basic HAR
model for forecasting the daily realized volatility as a function of the lagged daily, weekly,
and monthly realized volatilities, is readily extended to a semivariance HAR (SHAR)
model,

RVtþ1 ¼ /0 þ /þ þ  
D RVt þ /D RVt þ /W RVt:t4 þ /M RVt:t22 þ etþ1 ; (6)

where RVt:t4 and RVt:t22 refer to the weekly and monthly realized volatilities defined by
the summation of the daily volatilities over the past 5 and 22 days, respectively. For
/þ 
D ¼ /D , the model obviously reduces to a conventional symmetric HAR model. To help
alleviate concerns of heteroscedasticity, the HAR and SHAR models are also sometimes
estimated in logarithmic form.
Meanwhile, estimating the SHAR model in Equation (6) with daily realized volatilities
for the S&P 500 SPY ETF from January 2, 2002, to December 31, 2020, for a total of 4532
observations, the estimate for / D equals 1.127, with a heteroscedasticity robust standard
error of 0.360, while the estimate for /þD equals 0.320, with a standard error of 0.373. In
other words, short-run changes in aggregate market volatility are primarily driven by “bad”
volatility. As such, differentiating between “good” and “bad” volatility also results in more
accurate volatility forecasts, with the R2 from the SHAR model equal to 0.624, compared
with an R2 of 0.594 for the basic HAR model.19 In addition to corroborating Patton and
Sheppard (2015), these results also align with the earlier empirical findings of Chen and
Ghysels (2011), and the more recent analyses in Audrino and Hu (2016) and Baillie et al.
(2019). These same qualitative findings carry over to individual stock return volatilities, al-
though consistent with the “leverage effect” manifesting more strongly at the aggregate

19 Further restricting /þ 
D ¼ /D , corresponding to a HAR model with the signed jump variation in
place of RVt, the estimated coefficient for SJt equals 0.483, with a standard error of 0.331. Along
these lines, the recent text-based analysis in Baker et al. (2021) also suggests that stock market
jumps triggered by monetary policy news, which tend to be positive on average, typically lead to
lower future volatility.
Bollerslev j Realized Semi(co)variation 229

market level, the differences in the estimates for / þ

D and /D tend to be more muted for indi-
20
vidual stocks.
It would be interesting to further explore the economic mechanisms underlying these
differences. It is possible that “bad” news tend to be released more slowly than “good”
news, thereby resulting in a stronger degree of persistence and predictability. Relatedly, it
may be that negative news simply receive more and prolonged coverage by the financial
news media than positive news. Another possibility is that investors, or market makers,
tend reduce their risk exposures after negative shocks, whether voluntarily or by decree,

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

which then through a gradual process of portfolio rebalancing induces higher volatility per-
sistence. In any event, regardless of the underlying mechanism, as I will discuss next, these
differences in the way in which “good” and “bad” volatilities impact future changes in total
volatility also naturally translate into differences in the way in which the semivariation
measures are priced, both in the time-series and cross-sectional dimensions.

1.2 Semivariance-Based Asset Pricing

The “leverage effect” and the apparent asymmetry in the relationship between aggregate
equity index return and volatility discussed above may alternatively be interpreted as indir-
ect evidence for a risk-based volatility feedback effect. As elucidated by Pindyck (1982) and
French, Schwert, and Stambaugh (1987),21 if expected returns and expected volatility are
indeed positively related, as in Merton (1973), such a relationship should in turn induce a
negative relation between realized returns and unexpected volatility; see also Campbell and
Hentschel (1992). Meanwhile, estimates based on GARCH-in-mean models, or simple
regressions of returns on lagged measures of volatility and/or surprises therein, often fail to
establish a significant risk-return tradeoff relationship, and sometimes even suggest that
expected returns and volatility are negatively related (see, e.g., Bekaert and Wu, 2000;
Ghysels, Santa-Clara, and Valkanov, 2005; Bollerslev, Litvinova, and Tauchen, 2006;
Rossi and Timmermann, 2015; Theodossiou and Savva, 2016; Hong and Linton, 2020, for
empirical evidence and reviews of this extensive literature). However, if investors rationally
price downside risk more dearly than upside potential, the risk-return tradeoff might natur-
ally manifest differently in “good” versus “bad” volatility measures. The empirical results
discussed in the previous section that improved volatility forecasts may be obtained by dif-
ferentiating between the influence of past “good” and “bad” volatilities also indirectly sup-
ports this conjecture.
To illustrate, Table 1 reports the results from a set of simple return predictability regres-
sions, in which I regress daily,
pffiffiffiffiffiffiffiffiffiffiweekly, and monthly S&P 500 returns on a constant and
pffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffi
RV, and a constant and RVþ and RV . For simplicity, and ease of comparisons, I de-
fine the weekly (monthly) returns as the sum of 5 (21) daily returns divided by 5 (21). The
sample spans the same 2002–2020 time period analyzed above, for a total of 4783 daily

20 As a case in point, the average estimates for / þ

D and /D over the identical 2002–2020 sample
period for the “bad covid” stocks and FAANG stocks discussed further in Section 2, equal to 0.621
and 0.008, respectively, with most of the /
D s being strongly statistically significant, and most of
the /þ
D s insignificant.
21 Robins and Smith (2021) provides a fresh look and empirical re-evaluation of the widely cited
French, Schwert, and Stambaugh (1987) paper with recent data and modern econometric
techniques.
230 Journal of Financial Econometrics

Table 1 S&P 500 return predictability regressions

Daily Weekly Monthly

pffiffiffiffiffiffiffi
RV 0.038 0.003 0.003
(0.056) (0.049) (0.030)
½0:675 ½0:055 ½0:111
pffiffiffiffiffiffiffiffiffiffi
RVþ 0.246 0.100 0.051
(0.236) (0.081) (0.036)

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

½1:041 ½1:227 ½1:439
pffiffiffiffiffiffiffiffiffiffi
RV 0.303 0.106 0.057
(0.248) (0.085) (0.027)
½1:223 ½1:246 ½2:145

The table reports the results from simple return predictability regressions of daily, weekly (5 days), and
monthly (21 days) S&P 500 SPY returns on a constant and RV, and a constant and RVþ and RV–. The
sample spans 2002–2020, for a total of 4783 daily observations. All of the regressions are estimated at a
daily frequency with overlapping returns. Newey–West standard errors accounting for the serial correlation
in the errors induced by the overlap are reported in parentheses. The square brackets report alternative
Hodrick (1992) t-statistics for testing the null of no predictability.

return observations. All of the regressions are estimated at a daily frequency with overlap-
ping weekly and monthly returns. In addition to Newey–West standard errors accounting
for the serial correlation induced by the overlap (given in parentheses), following Hodrick
(1992) I also report t-statistics (in square brackets) for testing the null hypothesis of no-
predictability based on the rearranged return regressions without any overlap.
Looking at the results, there is a clear pattern in the point estimates. In line with the ex-
pffiffiffiffiffiffiffi
tant empirical literature referred to above, the estimated regression coefficients for RV
are all small and insignificant, underscoring the difficulties in empirically establishing a
traditional risk-return tradeoff relationship. However, differentiating
pffiffiffiffiffiffiffiffiffiffi between “good” and
“bad” volatility, the estimated regression coefficients for RVþ are all negative, albeit not
pffiffiffiffiffiffiffiffiffiffi
statistically significant at conventional levels, while the coefficients associated with RV
are all positive, and also statistically significant at the longer monthly return horizon.22
These simple regression-based results also mirror those of the earlier unpublished
study by Breckenfelder and Tédongap (2012), and the time-series regressions reported
therein in which the up minus down realized semivariance measure for the aggregate
market portfolio negatively predicts future market returns. The results are also broadly
consistent with Feunou, Jahan-Parvar, and Tédongap (2013), and the empirically sig-
nificant risk-return tradeoff relationship for the model-based relative downside risk
measure established therein.23 The equilibrium asset pricing model in Farago and
Tédongap (2018), based on a representative investor with generalized disappointment

22 The degree of return predictability afforded by any one of the regressions is invariably very lim-
ited, with R2s close to zero. As such, the “economic significance” of the estimates should be inter-
preted with some caution.
23 The results also echo the earlier empirical findings of Bali, Demirtas, and Levy (2009), and a posi-
tive tradeoff between expected aggregate market returns and VaR, with VaR interpreted as a
measure of downside risk.
Bollerslev j Realized Semi(co)variation 231

aversion, provides a possible rational for this differential pricing of “good” versus
“bad” volatility at the aggregate market level.
Motivated in part by these empirical findings for the market portfolio, Bollerslev, Li,
and Zhao (2020) look instead at the role of “good” versus “bad” volatility at the indi-
vidual stock level. Focusing on the stock-specific differences in RVþ and RV—and the
signed jump variation defined in Equation (4) scaled by total RV, they find that stocks
with relatively higher (scaled) SJ earn systematically lower future returns than stocks
with relatively lower (scaled) SJ. Also, these cross-sectional differences cannot be

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

explained by any of the standard controls and/or systematic risk factors traditionally
used for explaining the variation in individual stock returns. Nor can the differences be
explained by high-frequency-based realized skewness and/or kurtosis measures (Amaya
et al., 2015), reinforcing that the semivariation measures are not simply capturing pre-
viously documented skewed or fat tailed distributional deviations from normality.
Further expanding on this, Yu, Mizrach, and Swanson (2020) find that SJ-based sorts
restricted to “smaller sized” jumps result in even larger and more significant cross-
sectional return differences.
It is possible that a systematic risk factor explicitly related to downside aggregate market
volatility, along the lines of the aforementioned study by Farago and Tédongap (2018),
could help explain why investors value individual stocks with relatively high SJ more dearly
than comparable stocks with low SJ, although that has yet to be established. An alternative,
and in my opinion more likely, explanation is that the cross-sectional differences in
future returns associated with differences in firm level “good” versus “bad” volatility
may be linked to behavioral biases and investors’ overreaction to “bad” news coupled with
limits to arbitrage, along the lines of Shleifer and Vishny (1997). Corroborating this thesis,
the differences in returns for stocks with high versus low relative SJ appear stronger for
smaller firm stocks, stocks with higher overall volatility, and less liquid stocks, all of which
arguably pose greater arbitrage risks. I will return to that same theme in my discussion of
the empirical results pertaining to the pricing of the semicovariation measures introduced
next.

2 Multivariate Semicovariation Measures

The univariate realized semivariation measures and empirical results discussed in the previ-
ous section were based on the decomposition of the total variation for a given asset into
“good” versus “bad” volatility for that particular asset. However, most empirical questions
and hypotheses in asset pricing finance are inherently multivariate in nature, entailing
measures of the covariation among multiple assets and considerations of systematic non-
diversifiable risks. Fortunately, the traditional realized volatility measure in Equation (2)
is readily extended to a realized covariance measure by considering the sum of cross-
products of vectors of high-frequency intraday returns (Andersen et al., 2003; Barndorff-
Nielsen and Shephard, 2004a). Following Bollerslev et al. (2020) (henceforth BLPQ) the
realized semivariances defined in Equation (3) may similarly be extended to a multivariate
setting and a decomposition of the realized covariance matrix into four unique additive
realized semicovariance components determined by the signs of the underlying high-
frequency returns.
232 Journal of Financial Econometrics

To fix ideas, consider the multivariate extension of the generic Itô semimartingale in
Equation (1) to a d-dimensional vector log-price process,
ðs ðs
ps ¼ p0 þ ls ds þ rs dWs þ J s ; s  0; (7)
0 0

where ls is a R -valued drift process, Ws is a d-dimensional standard Brownian motion, rs

d

is a d  d dimensional stochastic volatility matrix, and J s is a finitely active pure-jump pro-

cess.24 For notational simplicity, and in parallel to the discussion and definitions in Section

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

1, I will assume that intraday prices for all of the d assets are available at K equally spaced
times over the trading day ½t; t þ 1. Then, in a direct parallel to the seminal result in
Equation (2),

X
K ðt X
P
RCOVt  rt;k r0t;k ! rs r0s þ J s J 0s ; (8)
k¼1 t1 t1  s  t

where rt;k ¼ ptþk=K  ptþðk1Þ=K denotes the logarithmic discrete-time return vector for the
kth time interval on day t. Now, denote the corresponding vectors of signed positive and
25
negative high-frequency returns by rþ 
t;k and rt;k , respectively. Mirroring the decomposition
of the realized variance into two semivariances in Equation (3), the four realized semicovar-
iance matrices are then simply defined by,

X
K X
K X
K
Pt  rþ þ0
t;k rt;k ; Nt  r 0
t;k rt;k ; Mþ 0
t  Mt  rþ 0
t;k rt;k : (9)
k¼1 k¼1 k¼1

Note that by definition RCOVt  Pt þ Nt þ Mþ 

t þ Mt .
The two “concordant” realized semicovariance matrices (Pt and Nt ) comprise the posi-
tive and negative realized semivariances on their diagonals together with scalar realized
covariances constructed from identical-signed positive or negative high-frequency returns
on their off-diagonals. The “discordant” realized semicovariance matrices (Mþ 
t and Mt )
have zeros along their diagonals and scalar realized covariances constructed from
opposite-signed returns on their off-diagonals. Since Pt and Nt are defined as sums of vec-
tor outer-products, these matrices are both positive semidefinite, while Mþ 
t (and Mt ) is
indefinite. In situations where the ordering of the assets is arbitrary, the two discordant
matrices are naturally combined into a single “mixed” matrix Mt  Mþ 
t þ Mt . Note that
while the realized variance of a portfolio may be calculated from the realized covariance
matrix of the assets included in the portfolio based on the identity RVt  w0 RCOVt w,
where w refers to the vector of portfolio weights, the up and down semivariance measures
for a portfolio are not simply equal to portfolio weighted averages of the semicovariances.

24 This general setup, which also underlies the asymptotic theory in BLPQ, explicitly allows for multi-
variate “leverage effects” in the form of dependence between changes in the price and changes
in volatility, as well as stochastic volatility-of-volatility, volatility jumps and price-volatility co-
jumps.
þ 
25 Formally, rþ 
t ;k  rt ;k  It ;k and rt ;k  rt ;k  It;k , where  denotes the Hadamard (element-by-elem-
ent) product, and It ;k  ½1frt ;k;1 >0g ; . . . ; 1frt ;k;N >0g 0 and I
þ 0
t ;k  ½1frt ;k;1  0g ; . . . ; 1frt ;k;N  0g  ,
respectively.
Bollerslev j Realized Semi(co)variation 233

26
In particular, RVþ 0 þ  0
t 6¼ w ðPt þ Mt Þw and RVt 6¼ w ðNt þ Mt Þw.

Instead, the realized
semicovariance measures afford an alternative asset-specific “look inside” the quadratic
covariation.
To help further intuit the measures, consider the stylized setting in which the logarithmic
price process in Equation (7) follows a continuous semi-martingale with no drift (ls  0),
no jumps (J s  0), constant unit volatility for all of the assets, and instantaneous correlation
q among all of the different pairs of assets, in which case
P

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

RCOVt ! Id þ ðJd  Id Þq;

where Id denotes the d  d identity matrix and Jd is a d  d matrix of ones. In this situation,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P 1 1  q2 þ qarccosðqÞ
Pt ; Nt ! Id þ ðJd  Id Þ  ;
2 2p
while
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P qarccosq  1  q2
Mþ
t ; M
t ! ðJd  Id Þ  :
2p
As these limiting values make clear, the more strongly correlated the returns and the
larger the value of q, the larger (respectively, smaller) and closer to 1/2 (respectively, 0) are
the limiting values of the off-diagonal elements in the two concordant (resp. discordant)
semicovariances. Accordingly, one might naturally expect the relative importance of the
concordant semicovariances to increase during times of financial crises and “market
stress,” periods that are typically accompanied by higher overall asset return correlations.
The above stylized diffusive setting does not allow for any differences between Pt and
Nt (or Mþ 
t and Mt ). Nor does it provide for any distinction between the limiting values of
þ 
RVt and RVt for portfolios constructed from the d assets. Meanwhile, in line with the dis-
cussion pertaining to the realized semivariances in the previous section, and as illustrated
further below, empirically on days with important “directional news,” Pt and Nt can differ
quite dramatically. The more advanced limit theory developed in BLPQ identifies three dis-
tinct channels through which such differences can occur, namely directional “co-jumps,” a
type of “co-drifting,” and a specific form of “dynamic leverage effect.” In parallel to the
results for the semivariances in Equation (4), the first of these channels manifests directly in
the first-order asymptotics, with the difference between the probability limits of Pt and Nt
being entirely determined by identically signed co-jumps. That is,
P
X
P t  Nt ! DJ þ þ0  0
s DJ s  DJ s DJ s ; (10)
t1  s  t

where DJ þ
s (respectively, DJ 
s )
refers to the d-dimensional vector of time-s positive
(respectively, negative) jumps in each of the d assets. As shown in BLPQ, a feasible central
limit theorem further permits the formulation of a formal test (termed the JCSD test) for
significant differences in Pt and Nt due to co-jumps. By comparison, co-drifting and dynam-
ic leverage effects, both of which can cause the diffusive components of Pt and Nt to differ,

26 This also implies that the portfolio weights for the standard minimum variance portfolio will gener-
ally not minimize RVt . To minimize the “bad” volatility, one could rely on the approximate proce-
dures recently developed by Rigamonti et al. (2021).
234 Journal of Financial Econometrics

formally manifest in second-order bias terms in a non-central limit theorem.27 To aid in

interpreting the magnitude of these diffusive differences, BLPQ provides an additional in-
ference tool (termed the DCSD detection scheme), which under more restrictive regularity
conditions may be justified as an asymptotically valid test.
To empirically illustrate these features, Figure 2 shows the within-day 5-min logarithmic
prices (normalized to zero at the beginning of the day) on June 15, 2020, and July 9, 2020,
for two separate groups of stocks: “bad covid” stocks and the FAANG group of stocks.
The FAANG group of stocks comprises the five tech giants: Facebook (FB), Amazon

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

(AMZN), Apple (AAPL), Netflix (NFLX), and Alphabet (GOOG). These five stocks cur-
rently account for close to 20% of the total market capitalization of the S&P 500 index.
My definition of the “bad covid” group of stocks follows Bollerslev, Patton, and Zhang
(2021), who rely on hierarchical clustering methods together with a novel cross-validation
approach for determining the optimal number of clusters and cluster groupings for the S&P
100 individual stocks based on their daily realized semicorrelations. The “bad covid” clus-
ter of stocks that I consider here first arose on January 31, 2020, coincident with the World
Health Organization) first declared the coronavirus outbreak a health emergency of inter-
national concern.28 The cluster consists of: Boeing (BA), Occidental Petroleum (OXY),
Raytheon (RTX), Schlumberger (SLG), and Simon Property Group (SPG). These companies
obviously range quite widely in terms of their main lines of business, and do not line up
with conventional industry type classifications. Instead, the grouping reflects commonal-
ities in the way in which the prices of the different stocks respond to new information, as
seen through the lens of the semicovariation measures.
Looking first at June 15, the price paths for the “bad covid” stocks fairly closely mirror
the intraday price path for the S&P 500 SPY ETF on that same day shown in Figure 1.
There is a mostly monotonic increase in the prices for all of the stocks over the earlier part
of day, along with an apparent jump at 2:00 pm when the FED announced its intention to
inject additional liquidity into the economy. By comparison, the price paths for the
FAANG stocks on that same day, shown in the second top panel, evidence much more
muted appreciations. Looking at the bottom two panels for July 9, which as previously
noted saw a sharp rise in new reported coronavirus cases across the United States, the price
paths again show clear within-cluster similarities and across cluster differences. All of the
“bad covid” stocks performed very poorly losing more than 4% on that day, while the
FAANG stocks as a group ended up almost flat for the day.
In line with these observations, the relative importance of the different semicovariance
components also differ quite markedly within and across the two clusters of stocks on each
of the two different days. For instance, while the average Pij normalized by ðRVi RVj Þ1=2
for all of the (i, j) pairs of “bad covid” stocks equals 0.528 on June 15, the similarly nor-
malized Nij and Mij components only amount to 0.279 and 0.028, respectively, implying

27 These terms also set the analysis pertaining to realized semivariances and semicovariances apart
from most other high-frequency econometrics and related in-fill asymptotics, which typically on
standard central limit theorem type arguments; see, for example, Aı̈t-Sahalia and Jacod (2014).
28 In contrast, when clustering the S&P 100 stocks based on their standard realized correlations, a
similar “bad covid” cluster of stocks did not arise until March 18, 2020.
Bollerslev j Realized Semi(co)variation 235

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Figure 2 Intraday individual equity prices. The figure shows the intraday (normalized to zero at the be-
ginning of the day) logarithmic prices at 5-min intervals on June 15, 2020 (top panels) and July 9 (bot-
tom panels) for each of the “bad covid” stocks (left panels) and FAANG stocks (right panels), as
defined in the main text.

a total within-cluster average realized correlation of 0.779 for the day.29 By comparison,
the average normalized Nij for the cluster of “bad covid” stocks equals 0.442 on July 9,
compared with 0.235 and 0.097 for the normalized Pij and Mij components. The mixed
semicovariance components, of course, tend to play a relatively more important role for the
between cluster correlations. As a case in point, looking at all of the pairwise combinations
of “bad covid” stocks and FAANG stocks on July 9, the average normalized Mij equals
0.134, while the normalized Pij and Nij equal 0.197 and 0.365, respectively.
Putting the discussion pertaining to Figure 2 further into perspective, recall that under
the standard Itô semimartingale assumption, to a first-order asymptotic approximation dif-
ferences in the concordant (discordant) semicovariance components are entirely driven by
co-jumps. Empirically, many large-sized jumps tend to be readily associated with precisely
timed news announcements, or “sharp” news. However, aside from the fairly minor-sized
jumps for most of the stocks evident at 2:00 pm on June 15 in response to the FED’s an-
nouncement at that time, co-jumps do not seem to account for the differences in the real-
ized semicovariation measures for either of the 2 days depicted in Figure 2. Instead, echoing
the discussion pertaining to the price paths for the S&P 500 SPY ETF in Figure 1, both of

29 This particular normalization ensures that the so-defined three realized semicorrelations add up
to the standard realized correlation. Following Bollerslev, Patton, and Quaedvlieg (2020), other nor-
malizations based on the realized semivariances could be employed in the definition of alternative
realized semicorrelation type measures.
236 Journal of Financial Econometrics

the days shown in Figure 2 seem to be characterized by “soft” and more difficult to inter-
pret news, resulting in distinctly different intraday return persistence for the two different
clusters of stocks, and in turn varying importance of the three different semicovariance
components.30
Further highlighting these differences, Table 2 shows the average values of the previous-
ly defined daily realized semicorrelations for all of 2020 for the 10 unique pairs of stocks in
the “bad covid” and FAANG clusters, along with the 25 unique “bad covid” versus
FAANG stock pairs. As expected, the within-cluster averages of the concordant qP s and

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

qN s are very close, as are the corresponding cross-cluster averages. Meanwhile, underscor-
ing the different high-frequency dynamic features of the stocks in each of the two clusters,
the average value of the discordant qM s for the between cluster pairs of stocks is much
larger (in an absolute value sense) than the two averages for the within-cluster stock pairs. I
turn next to a discussion of how these differences in the relative importance of the semico-
variance components across stocks and through time may be used in the construction of
improved volatility forecasts.

2.1 Semicovariance-Based Volatility Forecasting

A number of different GARCH and stochastic volatility type models have been proposed in
the literature to account for return-volatility asymmetries in multivariate settings (e.g.,
Kroner and Ng, 1998; McAleer, Hoti, and Chan, 2009; Francq and Zakoı̈an, 2012). To il-
lustrate, let rþ 
t and rt denote the vector of signed daily positive and negative returns, re-
spectively. A straightforward multivariate generalization of the aforementioned univariate
GJR-GARCH model, in which the conditional covariance matrix Htþ1 responds asymmet-
rically to cross-products of lagged returns depending on the signs of the returns, may then
be expressed as,31

Htþ1 ¼ X þ aP rþ þ0  0 þ 0  þ0
t rt þ aN rt rt þ aM ðrt rt þ rt rt Þ þ bHt :

Directly paralleling the univariate models discussed in Section 1.1, the cross-products of
the daily lagged return vectors in this multivariate model may naturally be replaced by their
respective realized semicovariance components,

Htþ1 ¼ X þ aP Pt þ aN Nt þ aM Mt þ bHt : (11)

For aP ¼ aN ¼ aM this obviously collapses to a symmetric multivariate realized GARCH

model. However, by allowing aP, aN, and aM to differ, the realized semicovariance-based
model in Equation (11) allows for more refined and potentially more informative intraday
“continuous” as opposed to daily threshold-based multivariate “leverage effects.”
The estimation results in Bollerslev, Patton, and Quaedvlieg (2020) for a cross-section
of individual stocks support this idea, and point to significant improvements in overall
model fit by allowing the impact of the lagged semicovariance components to differ. In line

30 Relatedly, Jiang, Li, and Wang (2021) have recently documented the existence of pervasive under-
reaction to various types of firm-specific news and strong intraday individual stock price drifts fol-
lowing large (in an absolute sense) “news-driven” returns.
31 For illustrative purposes, I assume the a’s and b to be scalar, but richer non-scalar paramateriza-
tions, and models in which the impact of rþ 0  þ0
t rt and rt rt are not necessarily the same, could, of
course, be, and has been, entertained empirically.
Bollerslev j Realized Semi(co)variation 237

Table 2 Average daily realized semicorrelations

q qP qN qM

“Bad Covid” 0.422 0.268 0.290 0.139

FAANG 0.525 0.315 0.315 0.103
“Bad Covid” vs. FAANG 0.148 0.196 0.206 0.254

The table shows the average daily realized semicorrelations for 2020 for the 10 unique pairs of stocks in the
“Bad Covid” and FAANG clusters of stocks, respectively, along with the 25 unique pairs of cross-group corre-

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

lations. The three realized semicorrelations for stock pair (i, j) are defined as Pij ; Nij and Mij divided by
ðRVi RVj Þ1=2 .

with existing empirical evidence pertaining to more traditional multivariate asymmetric

GARCH models, the estimates for aN for the realized models in Equation (11) are typically
larger than the estimates for both aP and aM, implying that most of the co-persistence may
be traced to common “bad” intraday news. As such, the asymmetric realized
semicovariance-based models generally also produce more accurate covariance matrix fore-
casts compared with the forecasts from symmetric multivariate realized GARCH models
that restrict all of the a’s to be the same.
Multivariate GARCH models, their realized versions included, can be challenging to im-
plement empirically, especially in large dimensions. Alternatively, and in direct parallel to
the univariate semivariance-based HAR models discussed in Section 1.1, the realized semi-
covariance measures may similarly be used in the formulation of simple-to-implement
multivariate HAR type forecasting models. To illustrate the basic idea, consider the bivari-
ate case and the 3D HAR model originally estimated in BLPQ, in which each of the scalar
semicovariance components are allowed to depend on its own daily, weekly, and monthly
lags, as well as the lags of the other two components. Specifically, for asset pairs (i, j),
2 3 2 3 2 3 2 3 2 3 2 P 3
Ptþ1;ij /P Pt;ij Pt:t4;ij Pt:t21;ij tþ1;ij
6
4 Ntþ1;ij 5 ¼ 4 /N 5 þ UD 4 Nt;ij 5 þ UW 4 Nt:t4;ij 5 þ UM 4 Nt:t21;ij 5 þ 4 N 7
tþ1;ij 5;
Mtþ1;ij /M Mt;ij Mt:t4;ij Mt:t21;ij M
tþ1;ij

where UD ; UW , and UM are 3  3 parameter matrices. Restricting the U matrices to be sca-

lar, and forcing the intercepts to be the same (/P ¼ /N ¼ /M ), the above formulation obvi-
ously collapses to a standard univariate HAR model for RCOVij  Pij þ Nij þ Mij .
However, the estimates for a sample of individual stocks reported in BLPQ reveal highly
significant differences in the freely estimated parameters, with the dynamic dependencies in
both Pij and Nij driven almost exclusively by the lagged Nij terms, while the mixed semico-
variances Mij appear to be mostly driven by their own lags. Digging deeper, the
semicovariance-based models generally also assign greater weights to the daily lagged meas-
ures and thus respond faster to new information, compared with traditional multivariate
HAR models for RCOVij . Of course, to ensure that the forecasts for a full covariance ma-
trix based on these element-wise forecasts for the individual covariances are positive defin-
ite, additional parameter restrictions, and/or regularization will have to be imposed.
These gains from the use of the realized semicovariance measures for covariance matrix
forecasting extend to the forecasts of portfolio variances. Consider the decomposition of
238 Journal of Financial Econometrics

the realized variance of a portfolio with portfolio weights w into its portfolio specific semi-
covariance components,

RVt  w0 RCOVt w ¼ w0 Pt w þ w0 Nt w þ w0 Mt w:

As previously noted, these portfolio semicovariance measures differ from the up and
down semivariance measures defined in Equation (3), and in contrast to the latter, which
only require high-frequency returns on the portfolio itself, the semicovariance measures re-
quire high-frequency returns for all of the assets included in the portfolio. BLPQ again find

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

univariate HAR models for realized portfolio variances that exploit these portfolio specific
semicovariance measures to be both faster, in the sense of assigning larger weights to the
lagged daily realized measures, and more persistent, in the sense of shocks decaying at a
slower rate, than conventional HAR and SHAR models based on lagged portfolio realized
variances and semivariances only.
I will next discuss various findings which suggest that the separate semicovariation com-
ponents underlying this covariance matrix and portfolio variance forecasting results are not
priced the same either.

2.2 Semicovariance-Based Asset Pricing

The basic premise that investors only care about downside systematic risk(s) effectively
implies that only “bad” covolatility should be priced. In the case of aggregate market risk,
only the covariation with negative market returns ought to carry a risk premium, as in the
downside version of the CAPM (Ang, Chen, and Xing, 2006). Meanwhile, market frictions
and/or behavioral biases may cause assets with identical downside covariation to be priced
differently. In particular, consider two assets with the same total downside covariation. If
one of the two assets covaries more strongly with the market when the market is perform-
ing poorly, thereby exacerbating the systematic downside risk, it may naturally be expected
to carry a higher overall risk premium than the other asset, which covaries less strongly
with the market when the market is performing poorly.
To succinctly illustrate this idea, consider Figure 3 adapted from Hogan and Warren
(1974). If investors do not care about “good” volatility, the covariation stemming from the
two states where the market returns are positive (P and Mþ ) should not earn any risk pre-
mium. In contrast, any covariation associated with joint negative market and individual
asset returns (N) should be positively compensated, while the mixed covariation stemming
from negative market returns and positive individual asset returns (M ) ought to carry a
negative, and in an absolute value sense lower, risk premium. The decomposition of the
conventional CAPM beta into four realized semibetas proposed by Bollerslev, Patton, and
Quaedvlieg (2021a) is directly motivated by these considerations.
Specifically, omitting the time t subscript for notational convenience, and relying on the
same element-wise notation as above, the four realized semibetas for asset i with respect to
the market portfolio f are simply defined by,

Pfi Nfi þ Mþ

fi  M
fi
bPi  ; bN
i  ; bM
i  ; bM
i  ; (12)
RVf RVf RVf RVf

where RVf refers to the realized variance of the return on the market. The negative signs for
the two discordant semibetas ensure that all of the realized semibetas are non-negative by
Bollerslev j Realized Semi(co)variation 239

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Figure 3 Semicovariance pricing. The figure, adapted from Hogan and Warren (1974), shows the signs
of the risk premiums (k’s) associated with the four semicovariance components implied by a mean
semivariance pricing framework.

definition. If the four semibetas move proportionally with each other, as would be the case
in a Gaussian world, they would convey no additional information over and above the con-
þ 
ventional realized beta, b  bP þ bN  bM  bM (Barndorff-Nielsen and Shephard,
2004a; Andersen et al., 2006). As such, the four semibetas would also necessarily be priced
the same.
To illustrate, consider the same stylized diffusive setting with no drift, no jumps, and
constant volatility discussed in Section 2, for which RCOVt !P Id þ ðJd  Id Þq. Further
denote the probability limit of the standard realized beta by b. It follows that in this
situation,

P b  1 pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi 
bP ; bN ! q 1  q þ arccosðqÞ ;
2p
while

þ  P b  1 pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi 
bM ; bM ! q 1  q  arccosðqÞ :
2p
Hence, while the relative contribution of the concordant versus discordant semibetas to
the standard beta obviously depends on the strength of the correlation between the asset
and the market, each of the two concordant (discordant) semibetas consistently estimate
the same quantity. In a non-diffusive setting, or non-Gaussian world, however, the prob-
ability limits for all of the semibetas may formally differ. In accordance with the economic
intuition conveyed by Figure 3, they may also be priced differently.
The empirical results reported in Bollerslev, Patton, and Quaedvlieg (2021a) support

this conjecture. Only bN and bM appear to be priced in the cross-section of individual

stocks, in the sense that higher values of bN and bM both tend to be associated with
higher realized returns. Meanwhile, the estimated risk premium for bN across a variety of
specifications and different samples of stocks and time periods is typically around double

that of the estimated premium for bM , possibly related to a “betting against beta” type
story (Frazzini and Pedersen, 2014). The hypothesis that the two risk premiums are
240 Journal of Financial Econometrics

numerically the same is also easily rejected statistically.32 By comparison the estimated risk

premium for the traditional realized b is even less than the premium for bM .

On the face of it, the different risk premiums for bN and bM may seem puzzling. In a
frictionless financial market the sign of the covariation with the market can costless be
changed through short positions, so that in order to prevent arbitrage opportunities the two
risk premiums ought to be the same. The downside version of the CAPM also effectively
combines the “good” and “bad” downside semibetas into a single downside beta

bD  bN  bM , with a single risk premium. However, as argued by Pontiff (1996) and

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Shleifer and Vishny (1997), with legal constraints and charters impeding many institutional
investors from short-selling, and many individual investors simply reluctant to sell short,
this may result in limits-to-arbitrage and accompanying arbitrage risks (see also Hong and
Sraer, 2016). These arbitrage risks may in turn induce a wedge between the pricing of the
N and M semicovariation components. Consistent with this reasoning, Bollerslev,
Patton, and Quaedvlieg (2021a) further find that the hypothesis of identical risk premiums

for bN and bM is more strongly rejected for stocks with higher idiosyncratic volatility, a
commonly used proxy for greater impediment to price-correcting arbitrage (see, e.g.,
Stambaugh, Yu, and Yuan, 2015).
I will not pursue this line of reasoning and the pricing of the different semibetas any fur-
ther here. Instead, to merely illustrate the practical calculation of the semibetas, Table 3
reports the averages of the daily realized semibetas for all of 2020 for the same five “bad
covid” and five FAANG stocks discussed in Section 2. All of the betas are calculated with
respect to the S&P 500 market portfolio. As the table shows, the averages of the semibetas
for the “bad covid” stocks all exceed those for the FAANG stocks, indicative of more pro-
nounced non-normal dependencies. It would be interesting to further explore the economic
mechanisms behind these differences, and what explain the variation in semibetas across
stocks and time more generally. The model in Boloorforoosh et al. (2020), explicitly allow-
ing for time-varying betas and beta risk, may prove useful in thinking about these issues.33
In addition to the averages of the individual stocks semibetas, Table 3 also reports the
averages of the 2020 daily realized semibetas for the two equally weighted portfolios
comprise the five “bad covid” and five FAANG stocks, respectively. While the traditional
realized beta of a portfolio, and the up and downside portfolio betas, may be calculated as
the portfolio weighted averages of the respective realized betas for the individual stocks
included in the portfolio, the semibetas of a portfolio are not simply equal to the portfolio
weighted averages of the individual semibetas. Rather, consistent with the idea that port-
folio formation mute the impact of firm-specific jumps and other idiosyncratic risks, there-
by rendering the portfolio returns closer to the returns on the market portfolio compared
with the returns on the individual stocks included in the portfolio, the values of the two

32 The significant premium (respectively, discount) for lower-tail (respectively, upper-tail) asymmetric
dependence estimated by Alcock and Hatherley (2017) and Alcock and Sinagl (2020) also indirect-
ly supports these findings. The findings are also generally in line with Schneider, Wagner, and
Zechner (2020), and the tendency for assets with positive (respectively, negative) coskewness to
offer lower (respectively, higher) returns than predicted by the traditional CAPM.
33 Relatedly, the framework developed in Buraschi, Porchia, and Trojani (2010) for intertemporal port-
folio choice with stochastic second moments may be helpful in thinking about optimal asset allo-
cation decisions and how to hedge against time-varying semicovariances and semibetas.
Bollerslev j Realized Semi(co)variation 241

Table 3 Average daily realized semibetas

þ
b bP bN bM bM

Individual stocks
“Bad Covid” 1.230 0.863 0.874 0.271 0.237
FAANG 1.185 0.681 0.690 0.096 0.089
Portfolios
“Bad Covid” 1.233 0.740 0.769 0.146 0.130

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

FAANG 1.185 0.622 0.637 0.037 0.036

The top panel shows 2020 average daily realized semibetas with respect to the S&P 500 for the five stocks
included the “Bad Covid” and FAANG clusters of stocks, respectively. The bottom panel shows the average
daily realized semibetas for 2020 for equally weighted portfolios comprise the five stocks in each of the two
clusters.

portfolio discordant semibetas are both closer to zero than their individual stock averages.
Correspondingly, the two concordant portfolio semibetas are also both smaller than the
averages of the concordant semibetas for the individual stocks. This phenomenon, of
course, is not unique to the semibetas. Many other non-linear features, some of which have
previously been associated with cross-sectional differences in returns, are similarly dimin-
ished through the effects of portfolio diversification.
In addition to their use for more accurate ex-ante return predictions, the realized semi-
betas could also help shed new light on ex-post investment return performance. There is an
extensive literature, dating back to the early work by Treynor and Mazuy (1966) and
Merton and Henriksson (1981), devoted to the question of whether mutual funds and other
investment vehicles are able to “time” the market. This question is typically answered em-
pirically by comparing what effectively amounts to estimates of the up and downside betas
of a fund, with good timing ability manifest by bD < bU and/or changes in bU (bD) posi-
tively (negatively) correlated with the performance of the market. Unfortunately, the
returns for many funds are only available at relatively coarse monthly or quarterly frequen-
cies, hindering direct estimation of dynamically varying fund betas at the horizons over
which the funds might actively be changing their market exposures. Alternatively, the up
and downside betas of a fund may be estimated as the portfolio weighted averages of the up
and downside betas of the individual asset included in the fund’s portfolio, thereby allow-
ing for the calculation of time-varying beta estimates at the same frequency over which
fund holdings are available (see, e.g., Bodnaruk, Chokaev, and Simonov, 2019). Further
decomposing these estimates into separate “good” and “bad” up and downside fund betas,

X
d
þ þ X
d
 
bU
F ¼ wi ðbPi  bM P M
i Þ  bF  bF ; bD
F ¼ wi ðbN M N M
i  bi Þ  bF  bF ;
k¼1 k¼1

may afford additional insights into where the fund performance is coming from.34 In par-
M þ
ticular, it follows that if bPF > bN
F and bF > bM D U
F , then bF < bF , which is traditionally

34 Note, that while bUF and bDF are identical to the up and downside fund betas that would obtain
Mþ M
with high-frequency fund returns, if such returns were available, bPF ; bN
F ; bF , and bF do not
directly match the four fund semibetas that would be calculated directly with high-frequency fund
returns.
242 Journal of Financial Econometrics

considered to be indicative of good overall market timing. Appropriately timed variation in

a fund’s semibetas, may, of course, also be commensurate with superior overall market tim-
ing ability.35 Counter to this, however, Bandi and Renò (2021) report that many hedge
funds seemingly have larger betas when market returns are especially low, or what effect-
ively amounts to larger “betas in the tails.”
The semibetas discussed above are all rooted in the standard one-factor CAPM.
However, the same basic idea readily extends to multi-factor pricing models based on fac-
tors for which high-frequency returns are available, whether in the form of ETFs or other

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

actively traded financial instruments, or in the form of brute force constructed high-
frequency factor returns.36 With a single factor, or approximately uncorrelated factors,
realized semiloadings may naturally be defined by the same expressions as in Equation
(12), with the relevant factor in place of the market factor f. With multiple non-trivially
correlated factors different normalizations may be called for. In parallel to the semibeta
pricing results discussed above, the law-of-one-price coupled with limits-to-arbitrage may
again impose certain restrictions and/or bounds on the values of the risk premiums associ-
ated with the relevant semiloadings.
Along these lines, Aı̈t-Sahalia, Jacod, and Xiu (2021) reports that allowing for separate
risk premiums for the continuous and jump components of the Fama–French risk factors
significantly enhances the explanatory power of second-stage Fama–MacBeth cross-
sectional return regressions compared with the fit afforded by conventional models that
price the two components the same.37 Relatedly, Massacci, Sarno, and Trapani (2021) find
that allowing for different factor structures in endogenously determined up and downside
regimes afford a superior fit compared with traditional factor models that do not condition
on the state of the economy. It would be interesting to further explore the interplay between
these findings and the semiloading idea proposed here, both empirically and theoretically.
The framework developed by Engle and Mistry (2014) for analyzing asymmetric volatility
and skewness in factor returns within the context of an intertemporal capital asset pricing
model may prove useful in guiding such investigations.

3 Partial (Co)Variation Measures

The zero-threshold that underlies the realized semi(co)variation measures, and related vola-
tility forecasting and pricing results, discussed above is firmly rooted in the idea that invest-
ors process and price up and downside risks differently. From a purely statistical

35 Further expanding on this theme and the previous analysis in Artavanis, Eksi, and Kadlec (2019),
“good” and “bad” up and downside fund betas could potentially also help explain mutual fund
flows.
36 Intraday high-frequency versions of the Fama–French size and value factors were first con-
structed by Bollerslev and Zhang (2003), while high-frequency versions of all the five Fama–
French factors and the Carhart momentum factor have been put together by Aı̈t-Sahalia, Kalnina,
and Xiu (2020). High-frequency versions of the more than one hundred factors defined in Jensen,
Kelly, and Pedersen (2021) have also recently been explored by Aleti (2021).
37 The earlier study by Bollerslev, Li, and Todorov (2016) similarly finds that continuous and jump
CAPM betas are not priced the same. The much more extensive empirical analysis of the full
“factor zoo” in Aleti (2021) also further corroborates this differential pricing.
Bollerslev j Realized Semi(co)variation 243

perspective, however, the choice of a zero threshold may seem somewhat arbitrary. More
elaborate GARCH and other parametric volatility forecasting models involving multiple
non-zero thresholds have also been entertained in the literature (see, e.g., Medeiros and
Veiga, 2009; Cai and Stander, 2019, and the many references therein). Following Bollerslev
et al. (2021), the high-frequency-based realized semi(co)variation measures may similarly
be extended to so-called partial (co)variation measures, by allowing for a non-zero thresh-
old and/or multiple thresholds.
To fix ideas, let fg ð Þ, for g ¼ 1; 2; . . . ; G, denote a partition of the real line into G non-

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

overlapping intervals, so that the kth intraday return may be expressed as
rt;k  f1 ðrt;k Þ þ f2 ðrt;k Þ þ þ fG ðrt;k Þ. The basic realized variation measure in Equation (2)
may then alternatively be expressed as the sum of the corresponding G partial variation
ðgÞ
measures PVt implicitly defined by,

X
K X
G
ðgÞ
RVt ¼ f1 ðrt;k Þ2 þ þ fG ðrt;k Þ2  PVt : (13)
k¼1 g¼1

The realized semivariation measures in Equation (3) obviously obtained as special cases
by setting G ¼ 2, and f1 ðrt;k Þ  rt;k Iðrt;k < 0Þ and f2 ðrt;k Þ  rt;k Iðrt;k > 0Þ, respectively.
But, other thresholds may be used in the definition of more refined partial realized variation
measures.
Multivariate realized partial covariation measures may be defined analogously by parti-
tioning the vectors of high-frequency returns. Specifically, let rt;k ¼ f1 ðrt;k Þ þ f2 ðrt;k Þ þ þ
fG ðrt;k Þ denote an exact decomposition of the kth intradaily return vector into G compo-
nents based on the partition functions fg ðxÞ ¼ x 1fcg < x  cgþ1 g, where the thresholds
are monotonically increasing cg1  cg , with c1 ¼ 1 and cGþ1 ¼ 1. The resulting G2
ðg;g0Þ
realized partial covariation measures PCOVt are then implicitly defined by,

X
K
RCOVt  rt;k r0t;k
k¼1
X
K
¼ f1 ðrt;k Þf1 ðrt;k Þ0 þ f1 ðrt;k Þf2 ðrt;k Þ0 þ þ fG ðrt;k ÞfG ðrt;k Þ0 (14)
k¼1
XG XG
ðg;g0Þ
¼ PCOVt :
g¼1 g0 ¼1

The realized semicovariance measures defined in Equation (9) again obtain as special
cases for a single threshold at zero. In situations when the ordering of the assets is arbitrary,
mirroring the combination of the two discordant semicovariance matrices into a single
“mixed” matrix (Mt  Mþ 
t þ Mt ), all of the matched pairs of “mixed” partial covariance
ðg;g0Þ ðg0;gÞ
matrices may similarly be combined (i.e., PCOVt þ PCOVt for g 6¼ g0 ), resulting in
2
“only” GðG þ 1Þ=2, as opposed to G , partial covariance matrices in total.
Rather than relying on time-invariant thresholds in the definition of the partition func-
tions fg ð Þ, one might also naturally consider time-varying thresholds based on the
volatility-standardized intraday returns and the quantiles of said distributions. Doing so
will help avoid certain partitions becoming especially thinly or densely populated during
extended time periods of high or low volatility. The conventional zero threshold, of course,
is typically very close to the median of both the raw and the standardized high-frequency
244 Journal of Financial Econometrics

return distributions, and as such the semi(co)variation measures still obtain as special cases
of the so-defined partial (co)variation measures.

3.1 Partial (Co)variance-Based Volatility Forecasting

The freedom to choose the number and location of the thresholds in the definition of the
realized partial (co)variance measures affords a great deal of added flexibility compared to
the semi(co)variance measures discussed in Sections 1 and 2.38 Meanwhile, faced with the
oft-observed empirical tradeoff between better in-sample fits of more complicated models

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

versus better out-of-sample forecast performance of simpler models, it is not clear whether
the use of the richer partial (co)variation measures will necessarily result in superior volatil-
ity forecasting models compared to the semi(co)variance-based forecasting models dis-
cussed in Sections 1.1 and 2.1.
The empirical analyses in Bollerslev et al. (2021) shed a first light on this question.
Using the same sample of stocks and same univariate and multivariate HAR structures ana-
lyzed in BLPQ, the results suggest that it is difficult, although not impossible, to improve
upon the fixed threshold at zero. At the same time, however, when considering only a single
threshold, or G ¼ 2, zero clearly emerges as the “hero.” Allowing for multiple thresholds,
G ¼ 3 and partial (co)variance-based models with one threshold close to zero and another
threshold in the left tail of the standardized intraday return distributions typically emerge
as the best performing forecasting models. In other words, “good” and “bad” (co)volatility
and negative (co)jumps all manifest differently in terms of their dynamic dependencies.
The partial covariances defined in Equation (14) rely on simple threshold-based cutoffs
and rectangular-shaped partitions of the total covariation. Alternative partitions based on
ellipsoids, or other geometric shapes, possibly centered at non-zero coordinates, could be
employed in the definition of alternative classes of partial covariation measures. It is pos-
sible that some of these alternative decompositions may be used in the construction of even
better multivariate volatility forecasting models. Given the vast set of decompositions and
related models to potentially consider, I would envision ideas and techniques adapted from
machine learning to be very helpful in terms of disciplining or regularizing the estimation of
such models and further exploring this question.

3.2 Partial (Co)variance-Based Asset Pricing

There is a rapidly growing recent literature on the use of machine learning techniques in
economics and finance. A common finding for many of these studies in the area of asset
pricing finance concerns the importance of allowing for non-linearities and interactions
among predictor variables (e.g., Freyberger, Neuhierl, and Weber, 2020; Gu, Kelly, and
Xiu, 2020). Relatedly, and in parallel to the use of the semicovariances for asset pricing in
the form of the semibetas or semi-factor-loadings discussed in Section 2.2, the partial cova-
riances may similarly be used in the definition of partial betas or partial-factor-loadings
capturing different parts of the systematic risk exposures. This again is reminiscent of the
aforementioned studies by Bollerslev, Li, and Todorov (2016); Aı̈t-Sahalia, Jacod, and Xiu
(2021); and Aleti (2021), and the idea that the continuous and jump components of the sys-
tematic risk factors may be priced differently. However, other partitions of the systematic

38 The freedom to choose the number and location of the thresholds also pose formidable theoretical
challenges in establishing the in-fill asymptotic distributions of the partial (co)variation measures.
Bollerslev j Realized Semi(co)variation 245

risks different from continuous versus jump variation, or “good” versus “bad” covolatility,
may possibly result in even better asset price predictions. I am currently exploring this idea
in joint ongoing work (Bollerslev, Patton, and Quaedvlieg, 2021b), in which we seek to esti-
mate a full “risk premium surface.”39 Following the discussion above, the law-of-one-price
together with considerations of arbitrage again impose certain restrictions on the surface.
The random field regressions combined with sieve approximations that we rely on for the
estimation affords an especially convenient framework for incorporating such restrictions.

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

4 Conclusion
Financial markets are inherently forward-looking. The dynamics of financial asset prices
thus encodes potentially valuable information about investors’ expectations and beliefs
about future state variables, as well as preferences and attitudes toward different types of
risks. In particular, as I have argued here, there are both sound economic reasons and ample
empirical evidence to support the thesis that “good” and “bad” volatilities are not created
equal. “Looking inside” the quadratic return variation through the lens of newly developed
simple-to-implement high-frequency-based realized semi(co)variation measures, it is clear
that “bad” (co)volatility is both more informative about future (co)volatility and priced
more dearly by investors than “good” (co)volatility. In addition to these conclusions,
gleaned from the uses of realized semi(co)variance measures in simple reduced form volatil-
ity forecasting models and regression-based return predictions, there are many other intri-
guing questions still left to be explored in regards to the theoretical properties and wider
empirical applications of the new semi and partial (co)variation measures. I look forward
to seeing the fruits from continued financial econometrics research efforts devoted to these
ideas.

Supplementary Data
Supplementary data are available at https://www.datahostingsite.com.

References
Adrian, T., N. Boyarchenko, and D. Giannone. 2019. Vulnerable Growth. American Economic
Review 109: 1263–1289.
Aı̈t-Sahalia, Y., J. Fan, and Y. Li. 2013. The Leverage Effect Puzzle: Disentangling Sources of Bias
at High Frequency. Journal of Financial Economics 109: 224–249.
Aı̈t-Sahalia, Y., and J. Jacod. 2014. High-Frequency Financial Econometrics. Princeton, NJ:
Princeton University Press.
Aı̈t-Sahalia, Y., J. Jacod, and D. Xiu. 2021. “Inference on Risk Premia in Continuous-Time Asset
Pricing Models.” Working Paper. Chicago, IL: University of Chicago.
Aı̈t-Sahalia, Y., I. Kalnina, and D. Xiu. 2020. High-Frequency Factor Models and Regressions.
Journal of Econometrics 216: 86–105.
Aı̈t-Sahalia, Y., and D. Xiu. 2016. Increased Correlation among Asset Classes: Are Volatility or
Jumps to Blame, or Both. Journal of Econometrics 194: 205–219.

39 Our partition of the factor space underlying this estimation bears some resemblance to the asset
pricing trees and related set of test assets developed by Bryzgalova, Pelger, and Zhu (2020).
246 Journal of Financial Econometrics

Alcock, J., and A. Hatherley. 2017. Characterizing the Asymmetric Dependence Premium. Review
of Finance 21: 1701–1737.
Alcock, J., and P. Sinagl. 2020. “The Price of Asymmetric Dependence: International Evidence.”
Working Paper. Iowa City, IA: University of Iowa.
Aleti. S. 2021. “The High-Frequency Factor Zoo.” Working Paper. Durham, NC: Duke
University.
Amaya, D., P. Christoffersen, K. Jacobs, and A. Vasquez. 2015. Does Realized Skewness Predict
the Cross-Section of Equity Returns. Journal of Financial Economics 118: 135–167.
Andersen, T. G., I. Archakov, G. Cebiroglu, and N. Hautsch. 2021a. Local Mispricing and

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Microstructural Noise: A Parametric Perspectove. Journal of Econometrics, forthcoming.
Andersen, T. G., and T. Bollerslev. 1998. Answering the Skeptics: Yes, Standard Volatility Models
Do Provide Accurate Forecasts. International Economic Review 39: 885–905.
Andersen, T. G., and T. Bollerslev. 2018. “Volatility: Introduction.” In T. G. Andersen and T.
Bollerslev (eds.), Volatility, pp. xiiv–xliii. Cheltenham: Edward Elgar Publishing.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and P. Labys. 2000. Great Realizations. Risk 13:
105–108.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and P. Labys. 2001b. The Distribution of Realized
Exchange Rate Volatility. Journal of the American Statistical Association 96: 42–55.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and P. Labys. 2003. Modeling and Forecasting
Realized Volatility. Econometrica 71: 579–625.
Andersen, T. G., T. Bollerslev, F. X. Diebold, and G. Wu. 2006. Realized Beta: Persistence and
Predictability. Advances in Econometrics 20: 1–40.
Andersen, T. G., O. Bondarenko, and M. T. Gonzalez-Perez. 2015. Exploring Return Dynamics
via Corridor Implied Volatility. Review of Financial Studies 28: 2902–2945.
Andersen, T. G., N. Fusari, and V. Todorov. 2015. The Risk Premia Embedded in Index Options.
Journal of Financial Economics 117: 558–584.
Andersen, T. G., Y. Li, V. Todorov, and B. Zhou. 2021c. Volatility Measurement with Pockets of
Extreme Return Persistence. Journal of Econometrics, forthcoming.
Ang, A., J. Chen, and Y. Xing. 2006. Downside Risk. Review of Financial Studies 19: 1191–1239.
Artavanis, N., A. Eksi, and G. Kadlec. 2019. “Downside Risk and Mutual Fund Flows.” Working
Paper. Blacksburg, VA: Virginia Tech University.
Audrino, F., and Y. Hu. 2016. Volatility Forecasting: Downside Risk, Jumps and Leverage Effect.
Econometrics 4: 1–24.
Babii, A., X. Chen, E. Ghysels, and R. Kumar. 2021. “Binary Choice with Asymmetric Loss in a
Data-Rich Environment: Theory and an Application to Racial Justice.” Working Paper. Chapel
Hill, NC: University of North Carolina.
Baillie, R. T., F. Calonaci, D. Cho, and S. Rho. 2019. Long Memory, Realized Volatility and HAR
Models. Journal of Time Series Analysis 40: 609–628.
Bajgrowicz, P., O. Scaillet, and A. Treccani. 2015. Jumps in High-Frequency Data: Spurious
Detections, Dynamics, and News. Management Science 62: 2198–2217.
Baker, S. R., N. Bloom, S. J. Davis, and M. Sammon. 2021. “What Triggers Stock Market
Jumps.” Working Paper. Chicago, IL: University of Chicago.
Bali, T. G., K. O. Demirtas, and H. Levy. 2009. Is There an Intertemporal Relation between
Downside Risk and Expected Returns. Journal of Financial and Quantitative Analysis 44:
883–909.
Bandi, F., and R. Renò. 2021. b in the Tails. Journal of Econometrics, forthcoming.
Bandi, F. M., A. Kolokolov, D. Pirino, and R. Renò. 2020. Zeros. Management Science 66:
3466–3479.
Barndorff-Nielsen, O., S. Kinnebrock, and N. Shephard. 2010. “Measuring Downside
Risk-Realised Semivariance.” In T. Bollerslev, J. R. Russell, and M. W. Watson (eds.), Volatility
Bollerslev j Realized Semi(co)variation 247

and Time Series Econometrics: Essays in Honor of Robert F. Engle, pp. 117–36. Oxford:
Oxford University Press.
Barndorff-Nielsen, O. E., P. R. Hansen, A. Lunde, and N. Shephard. 2008. Designing Realized
Kernels to Measure the Ex-Post Variation of Equity Prices in the Presence of Noise.
Economerrica 76: 1481–1536.
Barndorff-Nielsen, O. E., P. R. Hansen, A. Lunde, and N. Shephard. 2009. Realized Kernels in
Practice: Trades and Quotes. The Econometrics Journal 12: C1–C32.
Barndorff-Nielsen, O. E., and N. Shephard. 2002. Econometric Analysis of Realised Volatility and
Its Use in Estimating Stochastic Volatility Models. Journal of the Royal Statistical Society: Series

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

B (Statistical Methodology) 64: 253–280.
Barndorff-Nielsen, O. E., and N. Shephard. 2004a. Econometric Analysis of Realized
Covariation: High-Frequency Based Covariance, Regression, and Correlation in Financial
Economics. Econometrica 72: 885–925.
Barndorff-Nielsen, O. E., and N. Shephard. 2004b. Power and Bipower Variation with Stochastic
Volatility and Jumps. Journal of Financial Econometrics 2: 1–37.
Bawa, V. S., and E. B. Lindenberg. 1977. Capital Market Equilibrium in a Mean-Lower Partial
Moment Framework. Journal of Financial Economics 5: 189–200.
Bekaert, G., and E. Engstrom. 2017. Asset Return Dynamics under Habits and Bad
Environment—Good Environment Fundamentals. Journal of Political Economy 125: 713–760.
Bekaert, G., E. Engstrom, and A. Ermolov. 2015. Bad Environments, Good Environments: A
Non-Gaussian Asymmetric Volatility Model. Journal of Econometrics 186: 258–275.
Bekaert, G., E. Engstrom, and N. Xu. 2021. The Time Varation in Risk Appetite and Uncertainty.
Management Science, forthcoming.
Bekaert, G., and G. Wu. 2000. Asymmetric Volatility and Risk in Equity Markets. Review of
Financial Studies 13: 1–42.
Black. F. 1976. “Studies of Stock Price Volatility Changes.” Proceedings of the 1976 Meetings of
the American Statistical Association, pp. 177–181.
Bodnaruk, A., B. Chokaev, and A. Simonov. 2019. Downside Risk Timing by Mutual Funds. The
Review of Asset Pricing Studies 9: 171–196.
Bollerslev, T. 1986. Generalized Autoregressive Conditional Heteroskedasticity. Journal of
Econometrics 31: 307–327.
Bollerslev, T., B. Hood, J. Huss, and L. H. Pedersen. 2018. Risk Everywhere: Modeling and
Managing Volatility. The Review of Financial Studies 31: 2729–2773.
Bollerslev, T., T. Law, and G. Tauchen. 2008. Risk, Jumps and Diversification. Journal of
Econometrics 144: 234–256.
Bollerslev, T., J. Li, A. J. Patton, and R. Quaedvlieg. 2020. Realized SemiCovariances.
Econometrica 88: 1515–1551.
Bollerslev, T., J. Li, and Y. Xue. 2018. Volume, Volatility, and Public News Announcements. The
Review of Economic Studies 85: 2005–2041.
Bollerslev, T., S. Z. Li, and V. Todorov. 2016. Roughing up Beta: Continuous vs. Discontinuous
Betas, and the Cross-Section of Expected Stock Returns. Journal of Financial Economics 120:
464–490.
Bollerslev, T., S. Z. Li, and B. Zhao. 2020. Good Volatility, Bad Volatility and the Cross-Section
of Stock Returns. Journal of Financial and Quantitative Analysis 55: 751–781.
Bollerslev, T., J. Litvinova, and G. Tauchen. 2006. Leverage and Volatility Feedback Effects in
High-Frequency Data. Journal of Financial Econometrics 4: 353–384.
Bollerslev, T., M. Medeiros, A. J. Patton, and R. Quaedvlieg. 2021. From Zero to Hero: Realized
Partial (Co)Variances. Journal of Econometrics, forthcoming.
Bollerslev, T., A. J. Patton, and R. Quaedvlieg. 2020. Multivariate Leverage Effects and Realized
Semicovariance GARCH Models. Journal of Econometrics 217: 411–430.
248 Journal of Financial Econometrics

Bollerslev, T., A. J. Patton, and R. Quaedvlieg. 2021a. Realized Semibetas: Disentangling ‘Good’
and ‘Bad’ Downside Risks. Journal of Financial Economics, forthcoming.
Bollerslev, T., A. J. Patton, and R. Quaedvlieg. 2021b. “The Risk Premium Surface.” Work in
Progress. Durham, NC: Duke University.
Bollerslev, T., A. J. Patton, and H. Zhang. 2021. “Equity Clusters through the Lens of Realized
Semicorrelations.” Working Paper. Durham, NC: Duke University.
Bollerslev, T., G. Tauchen, and H. Zhou. 2009. Expected Stock Returns and Variance Risk
Premia. Review of Financial Studies 22: 4463–4492.
Bollerslev, T., and V. Todorov. 2011. Tails, Fears, and Risk Premia. The Journal of Finance 66:

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

2165–2211.
Bollerslev, T., V. Todorov, and L. Xu. 2015. Tail Risk Premia and Return Predictability. Journal
of Financial Economics 118: 113–134.
Bollerslev, T., and B. Y. B. Zhang. 2003. Measuring and Modeling Systematic Risk in Factor
Pricing Models Using High-Frequency Data. Journal of Empirical Finance 10: 533–558.
Boloorforoosh, A., P. Christoffersen, M. Fournier, and C. Gouriéroux. 2020. Beta Risk in the
Cross-Section of Equities. The Review of Financial Studies 33: 4318–4366.
Breckenfelder, J., and R. Tédongap. 2012. “Asymmetry Matters: A High-frequency risk-return
trade-off.” Working Paper. Stockholm, Sweden: Stockholm School of Economics.
Bryzgalova, S., M. Pelger, and J. Zhu. 2020. “Forest Through the Trees: Building Cross-section of
Stock Returns.” Working Paper. London: London Business School.
Buraschi, A., P. Porchia, and F. Trojani. 2010. Correlation Risk and Optimal Portfolio Choice.
The Journal of Finance 65: 393–420.
Cai, Y., and J. Stander. 2019. The Threshold GARCH Model: Estimation and Density Forecasting
for Financial Returns. Journal of Financial Econometrics 18: 395–424.
Campbell, J. Y., and L. Hentschel. 1992. No News is Good News. Journal of Financial Economics
31: 281–318.
Carriero, A., T. E. Clark, and M. Marcellino. 2020. “Capturing Macroeconomic Tail Risk with
Bayesian Vector Autoregressions.” Working Paper. Cleveland, OH: Federal Reserve Bank of
Cleveland.
Chabi-Yo, F., and J. Loundis. 2021. “A Decomposition of Conditional Risk Premia and
Implications for Representative Agent Models.” Working Paper. Amherst, MA: University of
Massachusetts, Amhurst.
Chabi-Yo, F., S. Ruenzi, and F. Weigert. 2018. Crash Sensitivity and the Cross Section of
Expected Stock Returns. Journal of Financial and Quantitative Analysis 53: 1059–1100.
Chen, X., and E. Ghysels. 2011. News—Good or Bad—and Its Impact on Volatility Predictions
over Multiple Horizons. Review of Financial Studies 24: 46–81.
Christensen, K., R. C. Oomen, and M. Podolskij. 2014. Fact or Friction: Jumps at Ultra High
Frequency. Journal of Financial Economics 114: 576–599.
Christensen, K., R. C. Oomen, and R. Renò. 2021. The Drift Burst Hypothesis. Journal of
Econometrics, forthcoming.
Christie, A. A. 1982. The Stochastic Behavior of Common Stock Variances: Value, Leverage and
Interest Rate Effects. Journal of Financial Economics 10: 407–432.
Christoffersen, P. F., and F. X. Diebold. 1997. Optimal Prediction under Asymmetric Loss.
Econometric Theory 13: 808–817.
Cieslak, A., A. Morse, and A. Vissing-Jorgensen. 2019. Stock Returns over the FOMC Cycle. The
Journal of Finance 74: 2201–2248.
Corsi, F. 2009. A Simple Approximate Long-Memory Model of Realized Volatility. Journal of
Financial Econometrics 7: 174–196.
Bollerslev j Realized Semi(co)variation 249

Corsi, F., and R. Renò. 2012. Discrete-Time Volatility Forecasting with Persistent Leverage Effect
and the Link with Continuous-Time Volatility Modeling. Journal of Business & Economic
Statistics 30: 368–380.
Corsi, F., G. Zumbach, U. A. Müller, and M. M. Dacorogna. 2001. Consistent High-Precision
Volatility from High-Frequency Data. Economic Notes by Notes 30: 183–204.
Cremers, M., M. Halling, and D. Weinbaum.2015. Agregate Jump and Volatility Risk in the
Cross-Section of Stock Returns. The Journal of Finance 70: 577–614.
Engle, R. F. 1982. Autoregressive Conditional Heteroscedasticity with Estimates of the Variance
of United Kingdom Inflation. Econometrica 50: 1987–1007.

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Engle, R. F. 2002. New Frontiers for ARCH Models. Journal of Applied Econometrics 17:
425–446.
Engle, R. F. 2011. Long-Term Skewness and Systemic Risk. Journal of Financial Econometrics 9:
437–468.
Engle, R. F., and A. Mistry. 2014. Priced Risk and Asymmetric Volatility in the Cross Section of
Skewness. Journal of Econometrics 182: 135–144.
Farago, A., and R. Tédongap. 2018. Downside Risks and the Cross-Section of Asset Returns.
Journal of Financial Economics 129: 69–86.
Feunou, B., R. R. Aliouchkin, R. Tédongap, and L. Xu. 2020. The Term Structure of Expected
Loss and Gain Uncertainty. Journal of Financial Econometrics 18: 473–501.
Feunou, B., M. R. Jahan-Parvar, and C. Okou. 2018. Downside Variance Risk Premium. Journal
of Financial Econometrics 16: 341–383.
Feunou, B., M. R. Jahan-Parvar, and R. Tédongap. 2013. Modeling Market Downside Volatility.
Review of Finance 17: 443–481.
Francq, C., and J. M. Zakoı̈an. 2012. QML Estimation of a Class of Multivariate Asymmetric
GARCH Models. Econometric Theory 28: 179–206.
Frazzini, A., and L. H. Pedersen. 2014. Betting against Beta. Journal of Financial Economics 111:
1–25.
French, K. R., G. W. Schwert, and R. F. Stambaugh. 1987. Expected Stock Returns and Volatility.
Journal of Financial Economics 19: 3–29.
Freyberger, J., A. Neuhierl, and M. Weber. 2020. Dissecting Characteristics Nonparametrically.
The Review of Financial Studies 33: 2326–2377.
Ghysels, E., P. Santa-Clara, and R. Valkanov. 2005. There is a Risk-Return Tradeoff after All.
Journal of Financial Economics 76: 509–548.
Ghysles, E. 2014. Conditional Skewness with Quantile Regression Models: SoFiE Presidential
Address and a Tribute to Hal White. Journal of Financial Econometrics 12: 620–644.
Giglio, S., B. Kelly, and S. Pruitt. 2016. Systemic Risk and the Macroeconomy: An Empirical
Evaluation. Journal of Financial Economics 119: 457–471.
Glosten, L. R., R. Jagannathan, and D. E. Runkle. 1993. On the Relation between the Expected
Value and the Volatility of the Nominal Excess Return on Stocks. The Journal of Finance 48:
1779–1801.
Grosshans, D., and S. Zeisberger. 2018. All’s Well That Ends Well? On the Importnace of How
Returns Are Achieved. Journal of Banking & Finance 87: 397–410.
Gu, S., B. Kelly, and D. Xiu. 2020. Empirical Asset Pricing via Machine Learning. The Review of
Financial Studies 33: 2223–2273.
Hansen, P. R., Z. Huang, and H. H. Shek. 2012. Realized GARCH: A Joint Model for Returns
and Realized Measures of Volatility. Journal of Applied Econometrics 27: 877–906.
Hansen, P. R., and A. Lunde. 2006. Realized Variance and Market Microstructure Noise. Journal
of Business & Economic Statistics 24: 127–161.
Hodrick, R. J. 1992. Dividend Yields and Expected Stock Returns: Alternative Procedures for
Inference and Measurement. Review of Financial Studies 5: 357–386.
250 Journal of Financial Econometrics

Hogan, W. W., and J. M. Warren. 1972. Computation of the Efficient Boundary in the ES
Portfolio Selection Model. The Journal of Financial and Quantitative Analysis 7: 1881–1896.
Hogan, W. W., and J. M. Warren. 1974. Toward the Development of an Equilibrium
Capital-Market Model Based on Semivariance. The Journal of Financial and Quantitative
Analysis 9: 1–11.
Hong, H., and D. A. Sraer. 2016. Speculative Betas. The Journal of Finance 71: 2095–2144.
Hong, S. Y., and O. Linton. 2020. Nonparametric Estimation of Infinite Order Regression and Its
Application to the Risk-Return Tradeoff. Journal of Econometrics 219: 389–424.
Jacod, J., Y. Li, P. A. Mykland, M. Podolskij, and M. Vetter.2009. Microstructure Noise in the

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Continuous Case: The Pre-Averaging Approach. Stochastic Processes and Their Applications
119: 2249–2276.
Jacod, J., and P. E. Protter. 2012. Discretization of Processes. Heidelberg, Germany: Springer.
Jacod, J., and V. Todorov. 2009. Testing for Common Arrivals of Jumps for Discretely Observed
Multidimensional Processes. Annals of Statistics 37: 1792–1838.
Jensen, T. I., B. Kelly, and L. H. Pedersen. 2021. “Is There a Replication Crisis in Finance.”
Working Paper. Frederiksberg, Denmark: Copenhagen Business School.
Jiang, H., S. Z. Li, and H. Wang. 2021. Pervasive Underreaction: Evidence from High-Frequency
Data. Journal of Financial Economics 141: 573–599.
Kahneman, D., and A. Tversky. 1979. Prospect Theory: An Analysis of Decision under Risk.
Econometrica 47: 263–291.
Kalnina, I., and D. Xiu. 2017. Nonparametric Estimation of the Leverage Effect: A Trade-off be-
tween Robustness and Efficiency. Journal of the American Statistical Association 112: 384–396.
Kelly, B., and H. Jiang. 2014. Tail Risk and Asset Prices. Review of Financial Studies 27:
2841–2871.
Kilic, M., and I. Shaliastovich. 2019. Good and Bad Variance Premia and Expected Returns.
Management Science 65: 2522–2544.
Kinnebrock, S., and M. Podolskij. 2008. A Note on the Central Limit Theorem for Bipower
Variation of General Functions. Stochastic Processes and Their Applications 118: 1056–1070.
Kroner, K. F., and V. K. Ng. 1998. Modeling Asymmetric Comovements of Asset Returns. Review
of Financial Studies 11: 817–844.
Laurent, S., R. Renò, and S. Shi. 2021. “Realized Drift.” Working Paper. Verona VR, Italy:
University of Verona.
Laurent, S., and S. Shi. 2020. Volatility Estimation and Jump Detection for Drift–Diffusion
Processes. Journal of Econometrics 217: 259–299.
Lee, S. S., and P. A. Mykland. 2008. Jumps in Financial Markets: A New Nonparametric Test and
Jump Dynamics. Review of Financial Studies 21: 2535–2563.
Li, J., V. Todorov, and G. Tauchen. 2017. Jump Regressions. Econometrica 85: 173–195.
Li, J., V. Todorov, G. Tauchen, and R. Chen. 2017. Mixed-Scale Jump Regessions with Bootstrap
Inference. Journal of Econometrics 202: 417–432.
Liu, L. Y., A. J. Patton, and K. Sheppard. 2015. Does Anything Beat 5-Minute RV? A Comparison
of Realized Measures across Multiple Asset Classes. Journal of Econometrics 187: 293–311.
Lu, Z., and S. Murray. 2019. Bear Beta. Journal of Financial Economics 131: 736–760.
Lucca, D. O., and E. Moench. 2015. The pre-FOMC Announcement Drift. The Journal of Finance
70: 329–371.
Mancini, C., and F. Gobbi. 2012. Identifying the Brownian Covariation from the Co-Jumps Given
Discrete Observations. Econometric Theory 28: 249–273.
Mandelbrot, B. 1963. The Variation of Certain Speculative Prices. The Journal of Business 36:
394–419.
Markowitz, H. 1959. Portfolio Selection: Efficent Diversification of Investments. London: Yale
University Press.
Bollerslev j Realized Semi(co)variation 251

Massacci, D., L. Sarno, and L. Trapani. 2021. “Factor Models with Downside Risk.” Working
Paper. London: King’s College London.
McAleer, M., S. Hoti, and F. Chan. 2009. Structure and Asymptotic Theory for Multivariate
Asymmetric Volatility. Econometric Reviews 28: 422–440.
Medeiros, M. C., and A. Veiga. 2009. Modeling Multiple Regimes in Finncial Volatiloty with a
Flexible Coefficient GARCH(1,1) Model. Econometric Theory 25: 117–161.
Merton, R. C. 1973. An Intertemporal Capital Asset Pricing Model. Econometrica 41: 867–887.
Merton, R. C., and R. D. Henriksson. 1981. On Market Timing and Investment Performance. The
Journal of Business 54: 513–534.

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022

Meyer, P. A. 1967. “Intégral Stochastiques—1.” In:Séminaire de Probabilités I, Lecture Notes in
Mathematics, Vol. 39, pp. 72–94. Heidelberg, Germany: Springer Verlag.
Neuberger, A. 2012. Realized Skewness. Review of Financial Studies 25: 3423–3455.
Oomen, R. C. 2003. Three Essays on the Econometric Analysis of High Frequency Financial Data.
Ph.D. Thesis, European University Institute.
Orlowski, P., P. Schneider, and F. Trojani. 2020. “On the Nature of Jump Risk Premia.” Working
Paper. Lugano, Switzerland: University of Lugano.
Patton, A. J., and K. Sheppard. 2015. Good Volatility, Bad Volatility: Signed Jumps and the
Persistence of Volatility. Review of Economics and Statistics 97: 683–697.
Patton, A. J., and A. Timmermann. 2007. Properties of Optimal Forecasts under Asymmetric
Loss. Journal of Econometrics 140: 884–918.
Pindyck, R. S. 1982. Risk, Inflation, and the Stock Market. American Economic Review 74:
335–351.
Pontiff, J. 1996. Costly Arbitrage: Evidence from Closed-End Funds. The Quarterly Journal of
Economics 111: 1135–1151.
Rigamonti, A., D. Ferrari, A. Weissensteiner, and S. Paterlini. 2021. “Smoothed Semicovarinace
Estimation.” Working Paper. Liechtenstein: University of Liechtenstein.
Robins, R. P., and G. P. Smith. 2021. A New Look at Expected Stock Returns and Volatility.
Critical Finance Review, forthcoming.
Rossi, A. G., and A. Timmermann. 2015. Modeling Covariance Risk in Merton’s ICAPM. The
Review of Financial Studies 28: 1428–1461.
Roy, A. D. 1952. Safety First and the Holding of Assets. Econometrica 20: 431–449.
Schneider, P., C. Wagner, and J. Zechner. 2020. Low-Risk Anomalies? The Journal of Finance 75:
2673–2718.
Shephard, N., and K. Sheppard. 2010. Realising the Future: Forecasting with
High-Frequency-Based Volatility (HEAVY) Models. Journal of Applied Econometrics 25:
197–231.
Shleifer, A., and R. W. Vishny.1997. The Limits of Arbitrage. The Journal of Finance 52: 35–55.
Sortino, F. A., and R. van der Meer. 1991. Downside Risk. The Journal of Portfolio Management
17: 27–31.
Stambaugh, R. F., J. Yu, and Y. Yuan. 2015. Arbitrage Asymmetry and the Idiosyncratic
Volatility Puzzle. The Journal of Finance 70: 1903–1948.
Taylor, S. J. 1982. “Financial Returns Modelled by the Product of Two Stochastic Processes –
a Study of Daily Sugar Prices, 1961–1979.” In: O. D. Anderson (ed.), Time Series Analysis:
Theory and Practice, pp. 203–26. Amsterdam, The Netherlands: North-Holland
Publishing.
Theodossiou, P., and C. S. Savva.2016. Skewness and the Relation between Risk and Return.
Management Science 62: 1598–1609.
Treynor, J., and K. Mazuy. 1966. Can Mutual Funds Outguess the Market? Harvard Business
Review 44: 131–136.
252 Journal of Financial Econometrics

Yu, B., B. Mizrach, and N. R. Swanson. 2020. New Evidence of the Marginal Predictive Content
of Small and Large Jumps in the Cross-Section. Econometrics 8: 1–52.
Zakoı̈an, J. M. 1994. Threshold Heteroskedastic Models. Journal of Economic Dynamics and
Control 18: 931–955.
Zhang, L., P. A. Mykland, and Y. Aı̈t-Sahalia. 2005. A Tale of Two Time Scales: Determining
Intergated Volatility with Noisy High-Frequency Data. Journal of the American Statistical
Association 100: 1394–1411.

Downloaded from https://academic.oup.com/jfec/article/20/2/219/6432504 by guest on 28 March 2022