23.03.29 SW the-Estimation-Of-Leverage-Effect-With-High-Frequency-Data (2014, Christina D. Wang, Leverage Influences Skewness and Volatility)

This article was downloaded by: [134.117.10.
200]
On: 03 July 2014, At: 22:43
Publisher: Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House,
37-41 Mortimer Street, London W1T 3JH, UK
Journal of the American Statistical Association

Publication details, including instructions for authors and subscription information:
http://amstat.tandfonline.com/loi/uasa20
The Estimation of Leverage Effect With High-Frequency

Data
Christina D. Wang & Per A. Mykland
Accepted author version posted online: 03 Dec 2013.Published online: 19 Mar 2014.
To cite this article: Christina D. Wang & Per A. Mykland (2014) The Estimation of Leverage Effect With High-Frequency Data,
Journal of the American Statistical Association, 109:505, 197-215, DOI: 10.1080/01621459.2013.864189
To link to this article: http://dx.doi.org/10.1080/01621459.2013.864189
PLEASE SCROLL DOWN FOR ARTICLE
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of the
Content. Any opinions and views expressed in this publication are the opinions and views of the authors, and
are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied upon and
should be independently verified with primary sources of information. Taylor and Francis shall not be liable for
any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities whatsoever
or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of the use of
the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
amstat.tandfonline.com/page/terms-and-conditions
The Estimation of Leverage Effect With
High-Frequency Data
Christina D. WANG and Per A. MYKLAND
The leverage effect has become an extensively studied phenomenon that describes the (usually) negative relation between stock returns
and their volatility. Although this characteristic of stock returns is well acknowledged, most studies of the phenomenon are based on
cross-sectional calibration with parametric models. On the statistical side, most previous works are conducted over daily or longer return
horizons, and few of them have carefully studied its estimation, especially with high-frequency data. However, estimation of the leverage
effect is important because sensible inference is possible only when the leverage effect is estimated reliably. In this article, we provide
nonparametric estimation for a class of stochastic measures of leverage effect. To construct estimators with good statistical properties, we
introduce a new stochastic leverage effect parameter. The estimators and their statistical properties are provided in cases both with and
without microstructure noise, under the stochastic volatility model. In asymptotics, the consistency and limiting distribution of the estimators
are derived and corroborated by simulation results. For consistency, a previously unknown bias correction factor is added to the estimators.
Applications of the estimators are also explored. This estimator provides the opportunity to study high-frequency regression, which leads to
the prediction of volatility using not only previous volatility but also the leverage effect. The estimator also reveals a theoretical connection
between skewness and the leverage effect, which further leads to the prediction of skewness. Furthermore, adopting the ideas similar to the
estimation of the leverage effect, it is easy to extend the methods to study other important aspects of stock returns, such as volatility of
volatility.
KEY WORDS: Consistency; Discrete observation; Efficiency; Itô process; Microstructure noise; Realized volatility; Skewness; Stable
Downloaded by [134.117.10.200] at 22:43 03 July 2014
convergence.
1. INTRODUCTION based on the “leverage effect” hypothesis: A drop in the value

of the stock (negative return) increases the financial leverage
The leverage effect has become an extensively studied empir-
(debt-to-equity ratio), which makes the stock riskier and in-
ical phenomenon in the form of the (usually negative) correla-
creases its volatility. Since then, leverage effect has been taken
tion between (current) returns and (current and future) volatil-
to be synonymous with asymmetric volatility. Financial lever-
ity (Engle and Ng 1993; Zakoian 1994; Wu and Xiao 2002,
age itself, however, seems not enough to explain either the large
etc.). It is one of the many stylized facts of the security re-
magnitude of the effect of declines in current price on future
turn distribution, along with the well-known fat tails, skew-
volatility (Figlewski and Wang 2001), or the phenomenon that
ness, excess kurtosis, and heteroscedasticity. The discovery of
the asymmetry of market index returns is generally larger than
leverage effect closely relates to the study of stochastic volatil-
that of individual stocks (Kim and Kon 1994; Tauchen and
ity. Although for very low frequency data, such as monthly or
Zhang 1996; Andersen et al. 2001). In another point of view,
yearly asset returns, the assumption of homogeneity seems not
the asymmetric nature of volatility to return shocks simply re-
to be entirely unreasonable (Mandelbrot 1963; Fama 1965; Of-
flects time-varying risk premium (Pindyck 1984; French et al.
ficer 1973), the increasing frequency of observed data in stud-
1987; Campbell and Hentschel 1991). This explanation is of-
ies suggests heterogeneity in volatility, in other words, time-
ten referred to as the “volatility feedback effect”: If volatility is
varying volatility (Engle 1982; Bollerslev 1986; Andersen and
priced, an anticipated increase in volatility raises the required
Bollerslev 1998; Engle 2000; Andersen et al. 2001). This finding
return on equity, leading to an immediate stock price decline.
has had profound implications in both the theory and practice
Many later works either compare the two effects or seek
of financial economics and econometrics. It has inspired new
to argue that they can both be at work (Nelson 1991; Engle
model building, such as the emergence of ARCH models and the
and Ng 1993; Glosten et al. 1993; Bekaert and Wu 1997; Wu
later stochastic volatility models. Modeling volatility as a sepa-
2001; Hasanhodzic and Lo 2011). While there is little agreement
rate process allows the study of its relation with the associated re-
concerning the fundamental causes behind the leverage effect,
turn process, which leads to the discovery of asymmetric volatil-
that is not the focus of this article.
ity. Time varying volatility is also of substantial importance in
As most early studies are conducted over daily or longer
modeling for options pricing, as in Hull and White (1987), Stein
time horizons, it is worthwhile to examine this phenomenon
and Stein (1991), Heston (1993), and Ball and Roma (1994).
with high frequency data, which provides the opportunity to
Black (1976) and Christie (1982) were among the first to
explore more closely the relation between stock price and its own
document the volatility asymmetry, and gave an explanation
volatility. Some recent work has demonstrated that the volatility
asymmetry still appears over fairly small time intervals. But
Christina D. Wang is Post Doctoral Research Fellow, Department of Oper- some new aspects are added as both very good and bad news
ations Research and Financial Engineering and Bendheim Center for Finance,
Princeton University, Princeton, NJ 08544 (E-mail: danw@princeton.edu). Per increase volatility, with the latter having a more severe effect
Mykland is Professor, Department of Statistics and the College, University of
Chicago, Chicago, IL 60637 (E-mail: mykland@galton.uchicago.edu). Finan-
cial support From the National Science Foundation Under grants DMS 06- © 2014 American Statistical Association
04758, SES 06-31605, SES 11-24 526 and from the Oxford-Man Institute is Journal of the American Statistical Association
gratefully acknowledged. The authors thank Neil Shephard, Kevin Sheppard, March 2014, Vol. 109, No. 505, Theory and Methods
and Lan Zhang for their helpful comments and suggestions. DOI: 10.1080/01621459.2013.864189
197
198 Journal of the American Statistical Association, March 2014
(Barndorff-Nielsen et al. 2008b; Chen and Ghysels 2011). Also, (2011). Another carefully chosen function of volatility can sim-
the leverage effect decays exponentially as the time lag between plify the estimation of high-frequency regression coefficients.
return and volatility increases. In the literature, the peak effect This leads to an interesting discovery in volatility prediction.
is obtained at the instantaneous correlation between return and The empirical study with Microsoft stock data (2008–2011)
volatility (Bouchaud et al. 2001; Bollerslev et al. 2006).1 This shows strong predictive power of a term containing the leverage
corresponds to our definition of the leverage effect as being effect on the next period volatility. The power is comparable to
instantaneous. For further discussion of this effect; see the end that of current period volatility which is believed to be the most
of Section 2.2. significant term in volatility prediction.
Although many papers deal with the source or new properties The main results of this article will be given in Sections 2 and
of the leverage effect, few have tried to rigorously estimate it, 4. The data-generating mechanism and model setting can be
which is critical for supporting any conclusive claims. Simple found in Section 2.1. The (stochastic) parameter of the leverage
correlation estimators may be applied to the estimation of the effect is defined in Section 2.2. Based on this, for the case
leverage effect with caution. Those estimators lose consistency without microstructure noise, the estimator and limit theorems
with high-frequency data (Aı̈t-Sahalia et al. 2013). It is the can be found in Sections 2.3 and 2.4. Simulation results are
purpose of this article to construct nonparametric estimators provided in Section 3. Results that corroborate the theorems can
of leverage effects in the stochastic volatility model. To study be found in Section 3.1. Section 4 studies the case where market
the estimation, we define a new leverage effect parameter as microstructure noise is present in the data. The estimator and
the covariance (a covariation, to be precise) between the stock limit theorems for this case are provided in Sections 4.1 and
return and a function of its volatility. To construct the estimator 4.2. Simulation results for this case are provided in Section 5.
for the new parameter, we first study the classical equi-distant The extension to irregularly spaced data can be found in Section
sampling case without microstructure noise as the foundation 6. The relation between leverage effect and skewness is shown
for the study (later in the article) of more complicated cases, in Section 7. An application of the leverage effect in high-
such as the case with microstructure noise. As is emphasized frequency regression is implemented in Section 8. The details
in several studies (Mykland and Zhang 2006; Barndorff-Nielsen of empirical studies are in Section 9. The conclusion is provided
et al. 2008a; Renault and Werker 2011), it is more natural to work in Section 10. Proofs are in the Appendix.
with irregularly spaced data in practice. Based on the results
in equidistant cases, the extension of estimators to irregularly 2. MAIN RESULTS
spaced data can be constructed in a similar way, with some
2.1 Data-Generating Mechanism
adaptations.
Even with equidistant observations without microstructure In general, we shall work with a broad class of continuous
noise, we discover a previously unknown bias correction factor. semimartingales, namely Itô processes. In econometrics and
This bias correction factor is critical to obtaining consistency. financial mathematics studies, this is the most popular model for
The factor may have a substantial impact on the estimated value log price processes due to nonarbitrage considerations (Delbaen
since it functions as a magnifier, especially when estimates are and Schachermayer 1994, 1995, 1998).
close to zero. The bias correction factor may play an even more
important role in the estimation, when the situation becomes Definition 1. A process Xt is called an Itô process provided
more complicated as market microstructure noise is present. In- it satisfies
deed, in the case with microstructure noise, the bias correction dXt = μt dt + σt dWt , X0 = x0 , (1)
factor is found to be bigger than that in the case with uncontam-
inated continuous price paths. where μt and σt are adapted càdl àg locally bounded random
Statistical properties such as consistency and asymptotic dis- processes, and Wt is a Wiener process. The underlying filtration
tribution are carefully studied in different settings. The theo- will be called (Ft ). The probability measure will be called P.
retical findings of these statistical properties are corroborated The integrated variance process is given as
by the simulation results. These asymptotic properties have ap- t
plications to hypothesis testing (e.g., for model checking) and X, Xt = σu2 du. (2)
constructing confidence intervals. 0
There are many ways to apply the estimators of the leverage The process (2) is also known as the quadratic variation of X.
effect depending on the practical purpose. One way to explore We shall sometimes also use the term “integrated volatility.”
the potential application of the estimators is embedded in the
definition of the stochastic parameter of the leverage effect. Ac- We further assume that σt is also an Itô process (see the next
cording to the definition, one specific choice of the function section for discussion of this)
imposed on volatility gives rise to a unique relation between the dσt = at dt + ft dWt + gt dBt , (3)
leverage effect and skewness, which will help to estimate skew-
ness consistently. This relation may introduce further applica- where Bt is another Wiener process independent of Wt , and at ,
tions in hedging strategy or new product design; see Neuberger ft , and gt are all assumed to be Itô processes.
Clearly, in this stochastic volatility (SV) model, Xt corre-
1The
sponds to the log price process and σt is its own volatility pro-
relative change between time lag and lagged leverage effect should be
maintained, even if the consistency may be an concern, since the consistency cess. Both processes have a common driving Wiener process
can be achieved by a bias correction multiplier from our study shown later. Wt , which accommodates the leverage effect.
Wang and Mykland: The Estimation of Leverage Effect With High-Frequency Data 199
To summarize the technical requirements, we specify exact the options pricing side (Hull and White 1987; Stein and Stein
conditions as follows: 1991; Heston 1993; Ball and Roma 1994), and on the econ-
metric side (Barndorff-Nielsen and Shephard 2002; Barndorff-
Assumption 1. The system satisfies (1) and (3), where X and
Nielsen et al. 2006; Jacod 2008; Barndorff-Nielsen and Veraart
σ are continuous processes (the continuous modification). We
2009, Aı̈t-Shalia and Jacod 2009; Mykland and Zhang 2011b).
assume that all the coefficients ft , gt , at , μt are locally bounded
A parallel development can be carried out under assumptions of
in absolute value. We also assume that σt is locally bounded
fractional Brownian motion (Comte and Renault 1998; Gloter
away from zero.2
and Hoffmann 2004; Brockwell and Marquardt 2005; Nualart
2.2 The Parameter: A Definition of Leverage Effect 2006; Comte et al. 2010).
The above is, of course, a set of theoretical considerations.
As we have seen in the Introduction, the literature offers var- We finally appeal to the results in Section 9 to show that our
ious perspectives on how to specify a parameter for this effect; current definition of volatility asymmetry does find something
see also the discussion below in this section. We concentrate on empirically relevant; we substantially improve the prediction of
the estimation of the following stochastic parameter: next-period volatility using the current-period leverage effect.
Definition 2. The stochastic parameter of the contempora-
2.3 Estimation in the Absence of Microstructure Noise
neous leverage effect is defined as the quadratic covariation
between Xt and F (σt2 ) As the first step, we shall work with the equally spaced
T case for the process (Xt ); specifically it is observed every

X, F (σ 2 )T = 2F σt2 σt2 ft dt, (4) tn,i+1 = t = Tn units of time, at times 0 = tn,0 < tn,1 <
0 tn,2 < · · · < tn,n = T . Furthermore, we divide
√ observed values
where we suppose that into Kn blocks, with block size Mn = [c n] (except possi-
bly for the first and last block, which does not matter for the
Assumption 2. x → F (x) is twice continuously differen-
asymptotics), for some constant c. The boundary points are
tiable, monotone on (0, ∞).
on the grid H = {0 < τn,1 < τn,2 < · · · < τn,Kn −1 ≤ T }, where
The incorporation of the function F allows more flexibil- Kn = [ Mnn ].
ity and a wider range of applications when different forms of Define3
volatility are of interest, such as log-volatility processes which n −2

K

tend to be more stationary over time as implied by many empir- X,
F (σ 2 )T = 2 Xτn,i+1 − Xτn,i F σ̂τ2n,i+1 − F σ̂τ2n,i ,
ical studies. The actual choice of F will depend on the practical i=0
purpose and empirical evidence. The inclusion of this func- (5)
tion can also help reveal some interesting connections between
leverage effects and other statistics. Further interpretation of this and
specification will be seen in Sections 7 and 8. 1 2
σ̂τ2n,i = Xtn,j +1 − Xtn,j .
We stop here for a moment to reflect on the definitions. First Mn × t tn,j ∈(τn,i ,τn,i+1 ]
of all, we work with continuous processes. The interface with
jump processes remains to be explored. The latter permits ad- The factor 2 in the first part of Equation (5) might look unnat-
ditional concepts of asymmetry, in particular the semivariance ural to be included. However, it is crucial for the consistency of
of Barndorff-Nielsen, Kinnebrock, and Shephard (2008b). The the estimator (see Remark 3 for a discussion of this previously
connection between semivariance and the leverage effect (and unknown factor).
skewness, see Section 7) in this article is an important question
Theorem 1. Under Assumptions 1–2, as n → ∞ and T fixed,
which we leave for future investigation. This is necessarily a
complex matter, as it involves a different model of the price n1/4 (X,
F (σ 2 ) − X, F (σ 2 )T )
process (continuous paths vs. jumps). TT
L 16 2 2 6
Once one works with continuous paths, the assumption that →Z F σt σt dt
the leverage effect is instantaneous is both natural in a semi- c 0
T 1/2
martingale model for σt , and is empirically supported by the 2 2 4 44 2 32 2
+ cT F σt σt f + gt dt , (6)
finding in Bollerslev, Litvinova, and Tauchen (2006), where it 0 3 t 3
was shown that the connection between return and volatility is
most significant when the time lag is 0. This does not contra- stably in law,4 where Z is a standard normal random variable
dict the fact that the effect can appear at a greater time lag, as and independent of FT .
documented by Chen and Ghysels (2011).
As far as the Itô process (continuous semimartingale) assump- 3One can also consider a kernel estimator of the spot volatility in (5), by applying
tion is concerned, this assumption appears frequently both on the methods in Kristensen (2010), with some adaptation. A detailed study is
beyond the scope of this article.
4Suppose that all relevant processes (X , σ , etc.) are adapted to the filtration
t t
2To get from local boundedness to results that cover the whole time interval, use (Ft ). Let Zn be a sequence of Ft -measurable random variables. We say that
arguments as in chap. 2.4.5 (p. 160-161) of Mykland and Zhang (2012). |σt | is Zn converges stably in law to Z as n → ∞ if Z is measurable with respect to
locally bounded from above by continuity. The assumptions guarantee that the an extension of FT so that for all A ∈ FT and for all bounded continuous g,
equivalent martingale measure for X exists locally. This is used in the proofs; EIA g(Zn ) → EIA g(Z) as n → ∞. The same definition applies to triangular
see the beginning of Section A.1. arrays.
Another natural estimator analogous to X,

F (σ 2 )T is converge to zero but to one half of the leverage effect. To see
Xt2
n −2
why it is one half, note that each increment t j term is roughly

K
2 2
X,
F (σ 2 )T = 2 Xτn,i+1 − Xτn,i F στn,i+1 − F
στn,i , an (inconsistent) estimator of σt2j . Thus the cross product gives an
i=0 average of leverage effects over (τi , t1 ], (τi , t2 ], . . . , (τi , τi+1 ]. If
1 X, σ 2 t is considered to be constant over (τi , τi+1 ], that average

στ2n,i = of those leverage effects will give a value of about half of the
(Mn − 1) × t t ∈(τ ,τ ]
n,j n,i n,i+1 leverage effect over the entire interval. Hence, we have reduced
2
× Xtn,j +1 − Xτn,i+1 , (7) the estimation of leverage effect by half. An adjustment factor
of 2 therefore needs to be added to achieve consistency.
and
1 1 2.4 Estimation of Asymptotic Variance
Xτn,i+1 = Xtn,j +1 = Xτn,i+1 − Xτn,i .
Mn tn,j ∈(τn,i ,τn,i+1 ]
Mn Let

Kn −2
Noticing the relation between the two estimators when 1 2 2 2
G1n = 2n 2 (Xτn,i+1 − Xτn,i )2 F
στn,i+1 − F
στn,i ,
F (x) = x: X, σ 2 T = MMn −1
n
X, σ 2 − i Mn (Mn2−1)t Xτn,i+1
i=0
(Xτn,i+2 )2 + i Mn (Mn2−1)t (Xτn,i+1 )3 , the following theorem and (10)
can be easily derived:

Kn −2
1 2 2 2
Theorem 2. Under Assumptions 1–2 as in Theorem 1, as G2n = 2n 2 Mn t στ2n,i F
στn,i+1 − F
στn,i .
n → ∞ and T fixed, i=0
By the same methods as in the proof of Theorem 1, we have

n1/4 (X,
F (σ 2 ) − X, F (σ 2 )T )
TT the following convergences in probability:
L 16 2 2 6 T
→Z F σt σt dt 2 2 6
c 0 1 p 8
Gn → F σt σt dt
T 1/2 c 0
2 2 4 44 2 32 2 T
+ cT F σt σt f + gt dt , (8) 2 2 4 28 2 16 2
0 3 t 3 + cT F σt σt f + gt dt, (11)
0 3 t 3
stably in law,5 where Z is a standard normal random variable T
p 8 2 2 6
and independent of FT . G2n → F σt σt dt
c 0
T
Remark 1. From the limit theorems, it is not hard to see that 2 2 4 16 2
by properly choosing c, one can minimize the limit variance. + cT F σt σt ft + gt2 dt, (12)
0 3
The optimal value is

T 2 and
16 0 F σt2 σt6 dt
16 T 2 2 6
c =
T 2
2
. (9) G1n +
p
G2n → F σt σt dt
T 0 F σt2 σt4 44 f 2
+ 32 2
g dt c 0
3 t 3 t

T 2 2 4 44 2 32 2
See Section 2.4 for an estimator6 of c. + cT F σt σt f + gt dt.
0 3 t 3
Remark 2. The two estimators have the same asymptotic (13)
properties. Even though the centered version gives slightly more
symmetric results, it does not behave very differently from the Equation (13) gives the estimation of the asymptotic vari-
noncentered version. In practice, the noncentered version can be ance. With this estimation, a feasible version of the central limit
applied with less programming effort. Therefore, our later sim- distribution can be derived.
ulation mainly adopts the noncentered version of estimators. Theorem 3. Under Assumptions 1–2, as n → ∞ and T fixed,
Remark 3. The origin of the factor 2 in the estimator can be n1/4 (X,
F (σ 2 )T − X, F (σ 2 )T ) L
found in the proof of Theorem 1. For intuition, however, we →Z1 ,
give here a verbal explanation of the source of this adjustment G1n + G2n
constant. Let us consider the case where F (x) = x. Since σ̂t2 is a and (14)
consistent estimator of σt2 , then the first multiplication (Xτi+1 −
Xτi )(σ̂τ2i+1 − στ2i ) already gives a consistent (though infeasible) n1/4 (X,
F (σ 2 )T − X, F (σ 2 )T ) L
→Z1
estimator of the leverage effect in the interval (τi , τi+1 ]. Then G1n + G2n
one may expect the remainder term (Xτi+1 − Xτi )(σ̂τ2i − στ2i ) to
stably in law,7 where Z1 is a standard normal random variable
have mean zero. However, since σ̂τ2i employs data in the time
and independent of FT .
interval (τi , τi+1 ], as does (Xτi+1 − Xτi ), the product does not
Notice that the limiting distribution Z1 is the same in both
limits. In other words, the difference between the two statistics
5See Footnote 4.
6Here and in the continuation of Remark 1, we assume that the denominator in
(9) is nonzero. 7See Footnote 4.
Table 1. The summary statistics do exhibit the target normality
MSE Mean Median Q1 Q3
n = 390, T = 1/250 infeasible 1.097112 −0.02108 −0.005707 −0.676 0.6496

n = 390, T = 1/250 feasible 1.051725 0.006831 −0.004835 −0.754 0.7399
n = 23400, T = 1/250 infeasible 1.009285 −0.006430 0.003727 −0.6982 0.6796
n = 23400, T = 1/250 feasible 1.002125 0.002964 0.004267 −0.6964 0.6917
n → ∞, fixed T (asymptotic value) 1 0 0 −0.674 0.674
NOTE: This corroborates the theorems and shows that the asymptotics can predict small sample behavior. For sample size 390, both the mean and median are very close to 0. The MSE
is close to 1 and the quartiles are close to the theoretical values from N(0, 1). As sample size increases, the MSE decreases further closer to 1.
converges to zero in probability. With this feasible CLT, one can are studied. So if the asymptotics correctly predict small sample
conduct hypothesis testing and construct confidence interval for behavior, the distributions should be close to the standard normal
the leverage effect parameter. distribution.
The Heston model used in the simulation is of the form:
Remark 1 (continued). The result (13) opens paths to es-
timating the tuning parameter c in (9). We here outline two dXt = σt dWt ,

approaches. dσt2 = κ θ − σt2 dt + γ σt (ρ dWt + 1 − ρ 2 dBt ),
where Wt ⊥⊥ Bt . (15)
Method 1: The conceptually simplest possibility is to pick
c = arg min{G1n + G2n } over a suitable grid of c’s. If the grid

is nested and becomes dense as n → ∞, this automatically 3.1 Normality Demonstration
provides a consistent estimator of c. In the simulation, the true log price is simulated from the He-
Method 2: Since Method 1 is computationally heavy, we
ston model with broadly √ realistic parameter values: κ = 5, θ =
0.04, γ = 0.5, ρ = − 0.5 over 1 trading day. Two different
here also propose an alternative two-step method. Fix an ini-
sampling frequencies are studied to examine the small sample
tial value c0 , and compute (G1n + G2n )(c0 ). On the other hand,

T behavior. The first is when the data are observed at 1-min fre-
we can reduce estimation of γ 2 = 0 (F (σt2 ))2 σt6 dt to the lo-
quency, which corresponds to sample size 390. The second is
cal estimation of volatility by the methods in Section 4.1 in
when the data are observed at every second, which corresponds
Mykland and Zhang (2009). Call this latter estimate γ̂ 2 . We thus
to sample size 23,400. The results are given in Table 1.
obtain that c0−1 (G1n + G2n )(c0 ) − 16γ̂ 2 /c02 consistently estimates

T
T 0 (F (σt2 ))2 σt4 ( 44
3 t
f 2 + 32
3 t
g 2 ) dt. A consistent estimate of c2 4. ESTIMATION WITH MICROSTRUCTURE NOISE
is thus given from (9) as
It is well known that markets are not so ideal that log price pro-
−1
ĉ2 = 16γ̂ 2 c0−1 G1n + G2n (c0 ) − 16γ̂ 2 /c02 . cesses can be simply represented by pure semimartingales. This
has long been thought about as “microstructure” see, e.g., Roll
The two methods can be used together, with the second provid- (1984), O’Hara (1995), Harris (1990), and Hasbrouck (1996).
ing a starting point for seaching for a minimum in Method 1. In the context of high-frequency data, such microstructure was
originally observed through the so-called signature plot (intro-
3. SIMULATION RESULTS duced by Andersen et al. (2000); see also the discussion in
Mykland and Zhang (2005)). This led researchers to investigate
All simulation results are based on 10,000 sample paths while
a model where the efficient price is latent, and one actually
varying the sample size n, function F, and optimal choice of
observes
c (path dependent). In the simulation, the properties of the es-
timator are studied with the Heston model (Heston 1993). To Yt = Xt + t . (16)
examine the theoretical limit distribution, the distribution of the
statistics in Theorem 1 Several approaches8 seek to deal with microstructure noise
while estimating integrated volatility, and they shed light on how
n1/4 (X,
F (σ 2 )T − X, F (σ 2 )T ) to proceed in the estimation of leverage effects in the similar
16
T 2 2 6
T 2 1/2 , situation. Among these approaches, we have focused on preav-
σt dt + c 0 F σt2 σt4 44 f2 + 32 2
cT 0 F σt 3 t
g
3 t
dt
eraging. The preaveraging method (Jacod et al. 2009; Podolskij
and the statistics in Theorem 3, and Vetter 2009a; Mykland and Zhang 2011a) provides a plausi-
ble way to solve the problem with microstructure. Therefore, all
n1/4 (X,
F (σ 2 )T − X, F (σ 2 )T ) of the following discussion will be in the framework of preav-

G1n + G2n eraging and the blocking method will be adjusted as follows:
and
n1/4 (X,
8Such
F (σ 2 )T − X, F (σ 2 )T ) as Zhang, Mykland, and Aı̈t-Sahalia (2005), Zhang (2006), Barndorff-
, Nielsen et al. (2008a), Reiss (2010), and Xiu (2010), as well as the preaveraging
G1n + G2n papers cited in the text.
The contaminated log price process (Yt ) is observed every Note that the factor 2 in the previous proposed estimator in
tn,i = Tn units of time, at times 0 = tn,0 < tn,1 < tn,2 < · · · < Equation (5) is now changed to 3 instead. This change is due
tn,n = T . to the preaveraging method we adopted first to asymptotically
eliminate the impact of noise on the estimation. The change is
Assumption 3.
consistent with the adjustment to the realized volatility estimated
Yt = Xt + t , where t ’s are iid. N (0, a 2 ) and by preaveraging; see Jacod et al. (2009).
t ⊥⊥ the W and B processes, for all t. (17)
We also assume that t ’s have finite fourth moment, and are
independent of both return and volatility processes. n1/8 (X,F (σ 2 )T − X, F (σ 2 )T )
⎛
T
We can also relax these assumptions; see the development in L √ 2 2 4 44 2 32 2
Mykland and Zhang (2011a). →Z ⎝c c1 T F σt σt ft + gt dt
0 3 3
Here two nested levels of blocks will be required. The first √ T
level of blocks defines the range of preaveraging and the sec- 16 c1 2 2 6
+ F σt σt dt
ond one implements a blocking idea similar to that in the case c 0
T
without noise in Section 2.3. 96a 2 2
Blocks are defined on a much less dense grid of τn,i , also + 3/2 F σt2 σt4 dt
cc1 T 0
spanning [0, T ], so that ⎞1/2
T
block # i = {tn,j : τn,i ≤ tn,j < τn,i+1 } (18) 216a 4 2 2 2
+ 7/2 F σt σt dt ⎠ , (21)
(the last block, however, includes T). We define the block size cc1 T 2 0
by
stably in law,9 where Z is a standard normal random variable
Mn,i = #{j : τn,i ≤ tn,j < τn,i+1 }. (19) and independent of FT .
In principle, the block size Mn,i can vary across the trading
The optimal c and c1 that minimize the asymptotic variance
period [0, T ], but for this development we take Mn,i = Mn ; it
are derived as follows:
depends on the sample size n, but not on the block index i.
We then use as an estimated value of the efficient price in the √
−C 2 + 12AD + C C 2 + 12AD
time period [τn,i , τn,i+1 ): c= , (22)
9BD
1
X̂τn,i = Yt . and
Mn t ∈[τ ,τ ) n,j
n,j n,i n,i+1
√
Treating the estimated efficient price as a new data frame, we C+ C 2 + 12AD
c1 = , (23)
proceed as in Section 2.3 but with Xt replaced by X̂t , n by n = 2A
n/Mn (up to rounding), and tn,i by τn,i . Furthermore, we divide
√
T
T
X̂t values into Kn blocks, with block size L = Ln = [c n ] where A = 16 0 (F (σt2 ))2 σt6 dt, B = T 0 (F (σt2 ))2 σt4
2
T 2 2 4 216a 4
(except possibly for the first and last block, which does not f 2 + 32 g 2 ) dt, C = 96a
( 44

3T t 23 2 t 2 T 0 (F (σt )) σt dt, and D = T 2
matter for the asymptotics), for some constant c. The boundary
points are on the grid G = {0 < λn,1 < λn,2 < · · · < λn,Kn −1 ≤ 0 (F (σt )) σt dt.
In practice, c and c1 can be estimated by minimizing G1n + G2n
T } ⊂ H.
defined in the next section, over a suitable grid of c’s and c1 ’s.
4.1 The Case With Microstructure Noise If the grid is nested and becomes dense as n → ∞, this auto-
matically provides a consistent estimator of c and c1 .
In the case with microstructure noise, the data blocking
mechanism will be similar to
√ that just stated, but less compli- 4.2 Estimation of Asymptotic Variance
cated where M = Mn = [c1 n], τn,i = iMn Tn , and L = Ln =
1/4 Let
[ cn
√ ]. The interval between successive observations is now
c1
Kn −2
t = tn = tn,j +1 − tn,j = T /n. 9 1 2 2
Define G1n = n 4 X̂λn,i+1 − X̂λn,i F σ̂λ2n,i+1 − F σ̂λ2n,i ,
2 i=0

Kn −2

X,
F (σ 2 )T = 3 X̂λn,i+1 − X̂λn,i F σ̂λ2n,i+1 − F σ̂λ2n,i , and (24)
i=0
1 9 1
K n −2
2 2 2
X̂τn,i = Ytn,j , (20) G2n = n 4 Ln Mn t σλ2n,i+1 F
σλn,i+1 − F
σλn,i+1 .
M tn,j ∈[τn,i ,τn,i+1 )
2 i=0
and
1 2
σ̂λ2n,i = X̂τn,j +1 − X̂τn,j .
L × M × t τn,j +1 ∈(λn,i ,λn,i+1 ]
9See Footnote 4.
Table 2. Because of the much slower convergence rate, the simulation results are not as good as in the case without microstructure noise
MSE Mean Median Q1 Q3
n = 5 days, T = 1/50 infeasible 1.315581 −0.05502 −0.01236 −0.71910 0.66330

n = 5 days, T = 1/50 feasible 1.142911 0.02566 −0.02703 −0.79680 0.80940
n = 20 days, T = 2/25 infeasible 1.193074 −0.03032 −0.003359 −0.6793 0.6578
n = 20 days, T = 2/25 feasible 1.125859 0.02167 −0.05247 −0.77740 0.76390
n → ∞, fixed T (asymptotic value) 1 0 0 −0.674 0.674
NOTE: Even so, with reasonably large sample size, the mean and median are still close to 0. The MSE is not very far from 1, and the quartiles are reasonably close to the theoretical
values from the standard normal distribution.
By the same methods as in the proof of Theorem 1, we have the 5. SIMULATION FOR THE CASE WITH
following convergences in probability MICROSTRUCTURE NOISE
√ T
p 8 c1 2 2 6 Similarly to the case without microstructure noise, the small
G1n → F σt σt dt sample behavior of the asymptotic normality can be demon-
c 0
T strated by simulating the statistics
√ 2 2 4 28 2 16 2
+ c c1 F σt σt ft + gt dt
0 3 3
2 T
48a
+ 3/2 F σt2 σt4 dt
2 n1/8 (X, F (σ 2 )T − X, F (σ 2 )T )
√
cc1 T 0 16 c1
T 2 2 6 √
T 2 2
T
108a 4 2 2 2 c 0 (F (σt )) σt dt + c c1
T 0 (F (σt ))
+ 7/2 F σt σt dt, (24) σ 4 ( 44 f 2 + 32 g 2 ) dt + 96a 2 T
(F (σt2 ))2 σt4 dt
t 3 t 3 t 3/2
cc1 T 0
cc1 T 0 2

4 T 2 2 2
√ T + 216a 0 (F (σt )) σt dt
2 2 6
7/2 2
p 8 c1 cc1 T
G2n → F σt σt dt
cT 0
T and
√ 2 2 4 16 2
+ c c1 F σt σt ft + gt2 dt
3
0
T n1/8 (X,
F (σ 2 )T − X, F (σ 2 )T )
48a 2 2 2 4 .
+ 3/2 F σt σt dt G1n + G2n
cc1 T 2 0
T
108a 4 2 2 2 The Heston model is once again adopted in the simulation.
+ 7/2 F σt σt dt, (25)
cc1 T 0 3 The parameterization
√ is the same with κ = 5, θ = 0.04, γ =
0.5, ρ = − 0.5. The true log-price process is latent. It is con-
and taminated by market microstructure as in Equation (17). The
√
p 16 c1 T 2 2 6 standard deviation of noise is set to be a = 0.0005. This is also
G1n + G2n → F σt σt dt a realistic value in practice. Since the first step of preaveraging
cT 0
T consumes part of the data and reduces the sample size for the
√ 2 2 4 44 2 32 2
+ c c1 F σt σt ft + gt dt second step of estimation, the choices of n are bigger than those
0 3 3
T in the case without noise. The frequency is chosen as 1 sec,
96a 2 2 2 4 which produces 23, 400 observations in each trading day. The
+ 3/2 F σt σt dt
cc1 T 2 0 results corroborate the theorem and are demonstrated in Table 2.
T Even though only the simulations with F (x) = x are pre-
216a 4 2 2 2
+ 7/2 F σt σt dt. (26) sented, the results with other functions satisfying the condition
cc1 T 03
in definition (2) such as F (x) = log x have been investigated.
Equation (27) gives the estimation of the asymptotic vari- The results look very similar and the tables are omitted for the
ance. With this estimation, a feasible version of the central limit reasons of space.
distribution can be derived.
6. IRREGULARLY SPACED DATA
So far our analysis in cases both with and without microstruc-
n1/8 (X,
F (σ 2 )T − X, F (σ 2 )T ) L ture noise has been based on measuring prices in regularly
→Z1 , (28)
G1n + G2n spaced intervals. In some ways it is more natural to work with
prices measured in tick time and so it would be desirable to ex-
stably in law,10 where Z1 is a standard normal random variable tend the above theory to cover irregularly spaced data. This
and independent of FT . is emphasized by Zhang, Mykland, and Aı̈t-Sahalia (2005),
Barndorff-Nielsen et al. (2008a), and Renault and Werker (2011)
in their studies. We here use the framework from Barndorff-
10See Footnote 4. Nielsen et al. (2008a).
Assumption 4. The observation times (tn,i ) satisfy the follows:

condition

Kn −2
iT
T n Z,
F (s 2 )T = 3 Ẑλn,i+1 − Ẑλn,i F ŝλ2n,i+1 − F ŝλn,i ,
tn,i = G i = G (s) ds, i = 0, 1, . . . , n, i=0
n 0
1
Ẑτn,j = (Y ◦G(tn,p+1 ) − Y ◦G(tn,p )),
where G : [0, T ] → [0, T ] is a strictly increasing, twice differ- M
tn,p+1 ∈
entiable function with G(0) = 0, G(T ) = T . G (s) is locally (j Mt, (j + 1)Mt]
bounded away from 0, and G is bounded.
and
With the change of time under the Assumption 4, the stochas-
tic volatility model can be written as: 1
ŝλ2n,i = (Ẑτn,j +1 − Ẑτn,j )2 . (30)
λn,i+1
dZt = dX ◦ G(t) = μG(t) G (t) dt + σG(t) G (t) dWte , τn,j +1 ∈
(iLMt, (i + 1)LMt]
and (29) We emphasize that this estimator is feasible (observable), since

both contaminated price Yt and observation times tn,i and λn,i
dst = dσ ◦ G(t) = aG(t) G (t) dt + fG(t) G (t) dWte
are directly observable from the market.
+ gG(t) G (t) dBte , The CLT, estimator of asymptotic variance and feasible CLT
can be derived in a similar manner as when observations are
where W e and B e are independent Wiener processes.
regularly spaced.
Proposition 1. The leverage effect in Definition 2 satisfies
X, F (σ 2 )T = Z, F (s 2 )T . 7. LEVERAGE EFFECT AND SKEWNESS
From sec. 2 of Mykland and Zhang (2009), leverage
Proof:
effect(F (x) = x) and skewness have a close relationship. For
T equidistant data, the skewness of returns in high-frequency data
Z, F (s 2 )T = F (s 2 )ds 2 , Zt satisfies (as n → ∞)
0

T 2 2 n
= F σG(t) 2σG(t) fG(t) G (t) dt lim Xt3n,i+1
0 T ≤T
t n,i+1
T
T

2
= F 2
σG(t) 2σG(t) fG(t) dG(t) 3 L
→ σ 2 , XT + 3 σt3 dWt + σt−1 μt dt
0
2 0
T T 1/2
= F σv2 2σv2 fv dv
0 + 6 σt6 dt Z,
0
= X, F (σ 2 )T .
where Z is a standard normal random variable. This is a biased
With all index notation kept the same as in Section 2.3, the and inconsistent estimator, but it is interesting to find that lever-
estimator of Z, F (s 2 )T can be constructed similar to Equa- age effect appears on the right-hand side. When the mean is
tion (5), with one adaptation: removed from blocks of size M, this empirical skewness con-
n −2
verges to the leverage effect plus a mixed normal error:

K
Z,
F (s 2 )T = 2 (X ◦ G((i + 1)Mn t) − X ◦ G(iMn t)) n
i=0 lim (Xtn,i+1 − local mean of X)3
2 2 T
× F ŝ(i+1)Mn t
− F ŝiMn t
tn,i+1 ≤T
T 1/2
L 3 M −1 18 15
and (30) → σ , XT +
2
6+ − 2 6
σt dt Z.
2 M M M 0
1
2
ŝiMn t
= M is chosen differently in Mykland and Zhang (2009) from that
τn,i+1
tn,j +1 ∈ in this article. It is a constant instead (i.e., M does not grow with
(iMn t, (i + 1)Mn t]
n). This relationship tells us that in the case where skewness is
× (X ◦ G(tn,j +1 ) − X ◦ G(tn,j ))2 . hard to estimate directly, the consistent estimation of leverage
effect proposed by this article provides an alternative way to
The CLT, estimation of asymptotic variance, and feasible estimate skewness.
CLT follow for the estimator Z, F (s 2 T as in the equidistant To further emphasize that we are indeed estimating a form of
case without microstructure noise. The results for the case with skewness by the leverage effect, we now consider the predictable
microstructure noise can be derived analogously. Considering instantaneous skewness:
the contaminated process Y ◦ G(t) = X ◦ G(t) + ◦ G(t) =
n
Zt + ◦ G(t), with all index notation kept the same as in sec- p-skew := E Xt3n,i+1 |Fti .
tion 4, the estimator of the leverage effect can be constructed as T t ≤T
n,i+1
We obtain between volatility and return in the regression model. One way
to extrapolate this intraday behavior to between-day volatility
Proposition 2. Subject to regularity conditions, as n → ∞,
prediction is to include the previous day’s return but scaled by a
p 3 time-varying leverage effect. Technically, all regressors are now
p-skew → σ 2 , XT .
2 in the drift term.
Since we are not trying to discover the best model calibration
It should be noted that since E(Xt3n,i+1 |Fti ) is an unobserv- for volatility prediction, but rather to investigate the predictive
able quantity, this proposition does not yield a method of estima- power of return scaled by leverage effect, the prediction model
tion. It does, however, clarify the relationship between skewness is simply a linear regression (or AR(2)). Though this may not
and leverage effect. be a very sophisticated model, the results can still improve
The existence of a connection between skewness and the understanding the role of leverage effects in volatility prediction:
leverage effect has previously been noted in Meddahi and ti+1 ti ti−1
Renault (2004); see the discussion following Proposition 3.4 σt dt = α0 + α1
2
σt dt + α2
2
σt2 dt + α3 Xt2i −
(p. 370). ti
ti
ti−1 ti−2
+ α4 2ft dt × Xti + i .
8. LEVERAGE EFFECTS AND REGRESSION ti−1
COEFFICIENTS

t
The estimation of leverage effects also has an application to • The integrated volatility ti i+1 σt2 dt can be estimated by
estimating the regression coefficient of the volatility on its own various methods. In this empirical study, the preaveraging
log return. On one hand, the existence of leverage effect implies method (Jacod et al. 2009) is adopted.
the relation of volatility and the log return as stated below: • Xti− denotes the overnight log return.

ti
dσt2 = 2ft dXt + 2σt gt dBt + 2σt at − 2ft μt + ft2 + gt2 dt, • ti−1 2ft dt can be estimated by the proposed leverage effect
(31) estimator in this article by setting F (x) = 12 log(x).
• The inclusion of lagged volatilities and overnight returns
and is due to the empirical finding of volatility clustering.

d X, σt2 t
= 2ft . (32) In this study, since we do not consider the case with jumps
d X, X t involved, we first remove the days with jump activities by the
On the other hand, the leverage effect specified as X, log σ jump test from Lee and Mykland (2012).11 Alternatively, one
takes the following form: can apply the jump tests as in Aı̈t-Shalia and Jacod (2009),
Barndorff-Nielsen and Shephard (2004), Barndorff-Nielsen and
2dX, log σ Shephard (2006), Mancini (2001), Podolskij and Ziggel (2010),
= 2ft . (33)
dt and other works by the same authors.
Equations (33) and (34) suggest two ways of applying the esti- We also follow the convention by preaveraging the data over
mation of leverage effects (F (x) = x and F (x) = 12 log(x)) to every 5 min in the first step. After preaveraging, we take c = 1.
estimating the regression coefficient of the volatility process on We explore the volatility prediction over a two-day period, a
its own log-return process. The second method only involves longer period than one day, because of the comparatively slow
lower orders of the volatility process, and is thus comparatively convergence rate of the estimators as discussed in Section 5. To
robust. We will use this method in the next section. check the robustness, we also repeat the regression replacing
return scaled the leverage effect by return itself. The prediction
9. EMPIRICAL STUDY results are shown in Table 3, and the time series plot of the
estimated leverage effect is given in Figure 1.
In the empirical study, we employ Microsoft stock trades “SS explained” is the main indicator to show whether return
data from the New York Stock Exchange (NYSE TAQ). The scaled by leverage effect has a big contribution to the prediction
years under study are 2008 through 2011. Even though the of leverage effect. For each year, the first column of p-valued
stock is traded between 9:30 am and 4:00 pm, the window based on t-test is given only as reference. Of course, since the
9:45 am–3:45 pm is chosen in the empirical analysis. The reason covariates are not independent from the response variable, these
for choosing this window is that a vast body of empirical studies statistics cannot really tell whether the corresponding covariate
documents increased return volatility and trading volume at the should be included in the model. The third column gives the
open and close of the stock market (Wood et al. 1985; Chan collinearity diagnostic by variance inflation factors (vif) (see,
et al. 2000). A 15-min cushion at the open and close may strike e.g., Weisberg 2004).
a good balance between avoiding abnormal trading activities in Since almost all vif values are close to 1, one can consider the
the market and preserving enough data points to perform the covariates not to be collinear with each other. The sum of squares
estimation procedures in a consistent way. On average, there are explained by the return scaled by the leverage effect (RLE)
currently several hundred thousand trades of Microsoft during are substantial even when this term is included into the model
each trading day. There are frequently multiple trades in each
second.
In Section 8, we explored intraday high-frequency regression. 11The total numbers of days removed are 21 for 2008, 27 for 2009, 60 for 2010,
It is clear that two forms of the leverage effect reveal the relation and 43 for 2011.
Table 3. Two-day ahead volatility prediction results with Microsoft 2007–2010 data
2008 2009
P(> |t|) SS explained vif P(> |t|) SS explained vif
α0 0.003909 0.035226
RVt−1 1.75 · 10−7 7.05 · 10−5 * 2.050335 0.000752 1.7633 · 10−5 ∗ ∗∗ 5.603527
RVt−2 0.369974 1.05 · 10−5 * 1.461129 0.658205 1.9960 · 10−6 1.349047
2
Rt− 0.0493 9.198 · 10−6 1.142936 0.570849 7.09 × 10−7 1.062036
RLEt−1 0.000454 2.0697 · 10−5 ∗ 1.551306 0.027969 4.827 · 10−6 ∗ 4.855324
or Rt−1 0.821 9.0 · 10−8 ∗ 1.003209 0.280060 1.1866 · 10−6 1.198528
2010 2011
P(> |t|) SS explained vif P(> |t|) SS explained vif
α0 0.000706 0.000747
RVt−1 0.833737 1.18 · 10−7 1.453209 5.61 · 10−5 2.2341 · 10−6 ∗ ∗∗ 1.622136
RVt−2 0.180635 2.066 · 10−7 1.008225 0.215383 8.6345 × 10−7 ∗ 1.178455
2
Rt− 4.32 · 10−10 1.4117 · 10−5 ∗ ∗∗ 1.011785 0.270385 1.7367 · 10−7 1.034753
RLEt−1 0.5254 1.172 · 10−7 1.455641 0.000486 2.1891 · 10−6 ∗ ∗∗ 1.463627
or Rt−1 0.870061 7.8 · 10−9 1.068818 0.102010 5.0273 · 10−7 1.049329
RVt denotes the estimated integrated volatility at day t; Rt− denotes overnight return for day t; RLE denotes the log return scaled by leverage effect at day t (estimated leverage effect ×
log return). Rt−1 denotes the previous period return itself without scaling. “SS explained” denotes the sum of squares gained by adding each covariate in the order presented in the table.
tsplot for the leverage effect in 2008 tsplot for the leverage effect in 2009
1
1.0
the leverage effect
the leverage effect
0
0.0
-1
-1.0
-2
-2.0
-3
0 20 40 60 80 100 120 0 20 40 60 80 100

time time
tsplot for the leverage effect in 2010 tsplot for the leverage effect in 2011
0.4
1.0
the leverage effect
the leverage effect
0.2
0.5
0.0
0.0
-0.2
0 20 40 60 80 100 0 20 40 60 80 100
time time
Figure 1. TS-plot of the estimated leverage effects: The black curves present the time series plots of the estimated leverage effects. The red
curves give the 95% confidence intervals of the estimated values. The values on the vertical axes are different from one year to another. That is
due to the different magnitudes of the estimated leverage effects. Apparently year 2008 and 2009 display the biggest negative leverage effects.
This observation coincides with the empirical fact during the financial crisis.
last. Most of these sums of squares are comparable to the sum They not only provide the consistency of the estimation, but also
of squares explained by the previous day’s volatility, which is imply that simple covariance estimators tend to underestimate
believed to be the most significant factor for volatility prediction the leverage effect, especially when the values of the leverage
(Engle 1982; Bollerslev 1986). In all cases, RLE has stronger effect are close to zero. The amplifying factors play a vital role
predictive power than the two-period ahead integrated volatility of bias correction in the estimation.
does. This strongly suggests the inclusion of return scaled by The empirical studies demonstrate the importance of the
leverage effect into any model trying to predict next-period’s leverage effect in volatility prediction. Even though the sim-
volatility. The predictive power of a time-varying leverage effect ple regression (or AR(2)) model is adopted in the study, the
estimator12 is consistent with the earlier work by Engle and Ng explanatory power of RLE is surprisingly high. The power is
(1993) and Chen and Ghysels (2011), but here appears in a new almost of the same magnitude as the predictive power of the pre-
form. In addition, the previous period return does not contribute vious period volatility which is widely considered to be the main
to the sum of squares as much as the one scaled by the leverage source of variation in volatility prediction. This high explana-
effect. In some case, the previous period return is not significant tory power suggests that time-varying leverage effects should
while the return scaled by the leverage effect explains significant be included additionally to explain the variation and clustering
amount of sum of squares. in volatility prediction models.
Even though we have provided a way to deal with irregularly
10. CONCLUSION spaced data, it is important to study the estimation of leverage
effects and the asymptotic properties of estimators when time is
This article provides nonparametric estimators of the leverage
endogenous (as in Li et al. 2013). Different methods of dealing
effect, and analyzes them both theoretically and in simulation.
with microstructure noise should also be studied and compared
The definition of the stochastic parameter of the leverage effect

with the ones in this article. Our findings create the important
involves a twice differentiable monotone function. Even though
necessary foundation for further analysis both theoretically and
the reliance of the estimation on higher moments of volatility
empirically, as well as an investigation of how to carry out risk
is of concern in practice, the carefully chosen function F can
management in the presence of leverage effects.
help to reduce the order of moments required and provide ro-
Finally, as discussed in Section 2.2, many open questions
bust results. Other benefits of this function can be easily seen
remain in terms of model specification (continuous vs. jumps,
in the discussion of the connection between the leverage ef-
semimartingale vs. long range dependence), with reference to
fect and skewness. While the sum of intraday cubic returns is
the papers cited in that section. In particular, the connection to
not a consistent estimator of skewness, the p-skewness (Sec-
semivariance remains to be explored.
tion 7) can instead be estimated consistently by the leverage
effect estimators proposed by this article. The related properties
of p-skewness can also be studied by the properties of the lever- APPENDIX A: PROOF OF THEOREM 1
age effect. Clever choices of the function F can also reduce the
work of estimation, such as the estimation of high-frequency re- A.1 Preliminaries
gression coefficients. Instead of estimating both leverage effects In the following, by Fj we mean Ftn,j , and p is a positive integer.
and realized volatility, a different form of the leverage effect can Without loss of generality, we will set μt = 0, see Mykland and Zhang
serve as the estimated coefficient (Section 8). If the properties (2009), sec. 2.2, as well as our current Assumption 1 and associated
of the estimated coefficient are of interest, it is more attractive Footnote 1. Recall that
to apply the method in Equation (34), whose statistical proper-
ties have already been studied in this article, than to apply the X,F (σ 2 )T = 2 Xτn,i+1 − Xτn,i F σ̂τ2n,i+1 − F σ̂τ2n,i (A.1)
i
first ratio statistic in Equation (33) whose statistical properties
require further efforts to investigate. and
The bias correction factors in the estimators contribute to 1
an important finding in this article and are previously unknown. σ̂τ2n,i = Xt2n,j +1 . (A.2)
Mn × t tn,j +1 ∈(τn,i ,τn,i+1 ]
12The main motivation for us to include “leverage effect scaled returns” is the Lemma 1. Under Assumption 1,
earlier empirical findings on asymmetric impact of positive or negative returns
2 p
on the volatility process. To capture this asymmetric impact in the prediction E Xtn,k − Xtn,j σtn,m Fj
model, we include the extra term. We realize that Jacod (1994, 1996), Barndorff-
p+2 p
Nielsen, and Shephard (2005), and others, showed that the correlation (leverage = σtn,j (tn,k − tn,j ) + p 2 ft2n,j σtn,j (tn,k − tn,j )2
effect) has no impact on the asymptotic distribution. These observations seem p
to suggest that asymmetries do not matter for forecasting, but that is not so. + pσtn,j X, f tn,j (tn,k − tn,j )2
The concept of the news impact curve (Engle and Ng 1993) was originally 1 p
formulated within the context of daily ARCH-type models. In the models of + p(p − 1) ft2n,j + gt2n,j σtn,j (tn,k − tn,j )(tn,m − tn,j )
Engle and Ng (1993), the returns are included in the volatility prediction models 2
by differently scaling the positive or negative returns. Here we applied the intra- 1 2 p
+ p+ ftn,j + gt2n,j σtn,j (tn,k − tn,j )2
day data over small time interval to estimate integrated volatility, but predict 2
the next volatility over a much longer time interval (two-day ahead). Therefore, p+1
the prediction falls into a comparative low frequency setting. The leverage + atn,j σtn,j (tn,k − tn,j )2
effect turns out to have a significant impact on the volatility prediction which p+1
+ patn,j σtn,j (tn,k − tn,j )(tn,m − tn,j ) + Op (t 5/2 ), (A.3)
is supported by our empirical finding. This inclusion of leverage effect scaled
p
− Xtn,j σtn,m Fj = pσtn,j ftn,j (tn,k − tn,j ) + Op (t 3/2 ),
returns in the volatility prediction model is further supported by the findings in p
Chen and Ghysels (2011), where they are dealing with two different frequencies
E Xtn,k
and reaching a similar conclusion. (A.4)
⎛
and
⎜ p+2 p+1
E 2 Xtn,j − Xtn,i σt2n,j +1 Xtn,k − Xtn,j σt2n,k+1 Fi = E ⎝σtn,j (tn,k − tn,j ) + σtn,j atn,j (tn,k − tn,j )2
= 16σt4n,i ft2n,i (tn,k − tn,j +1 )(tn,j − tn,i )

1 p
+ 4σt4n,i X, f tn,i (tn,k − tn,j +1 )(tn,j − tn,i ) + Op (t 5/2
), (A.5) + + p ft2n,j + gt2n,j σtn,j (tn,k − tn,j )2
2
1 p(p − 1) 2 p
where tn,i < tn,j < tn,k ≤ tn,m , tn,m − tn,j = Op (t 2 ) and t = T
n
. + ftn,j + gt2n,j σtn,j (tn,k − tn,j )(tn,m − tn,j )
2
tn,k
Proof of Lemma 1: p
+ p 2 ftn,j (tn,k − t)2σt ft dt
tn,j
1. For (A.3), note first that p p+1
+ pσtn,j X, f tn,j (tn,k − tn,j )2 + pσtn,j atn,j (tn,k − tn,j )
p
σtn,m −
p
σtn,j ⎞
tn,m
p(p − 1) p−2 ⎟
tn,m
= p σt
p−1
dσt + σt ft2 + gt2 dt, ×(tn,m − tn,j )|Fj ⎠ + Op (t 5/2 )
tn,j 2 tn,j
and p+2 p+1

= σtn,j (tn,k − tn,j ) + atn,j σtn,j (tn,k − tn,j )2
2
Xtn,k − Xtn,j 1 2 p
tn,k + p+ ftn,j + gt2n,j σtn,j (tn,k − tn,j )2
tn,k 2
= 2 Xt − Xtn,j dXt + σt2 dt p(p − 1) 2 p
tn,j tn,j + ftn,j + gt2n,j σtn,j (tn,k − tn,j )(tn,m − tn,j )
2
tn,k tn,k p
= 2 Xt − Xtn,j dXt + (tn,k − t)d σt2 − σt2n,j + p 2 ft2n,j σtn,j (tn,k − tn,j )2
tn,j tn,j p
+ pσtn,j X, f tn,j (tn,k − tn,j )2
+ σt2n,j (tn,k − tn,j )
tn,k p+1
+ patn,j σtn,j (tn,k − tn,j )(tn,m − tn,j ) + Op (t 5/2 ).
tn,k
= 2 Xt − Xtn,j dXt + (tn,k − t)2σt dσt
tn,j tn,j
2. For (A.4):
tn,k
+ (tn,k − t) ft2 + gt2 dt

E (Xtn,k − Xtn,j )σtn,m Fj
tn,j p
+ σt2n,j (tn,k − tn,j ).
tn,k
p−1
= E p σt dX, σ t |Fj + Op (t 3/2 )
Hence, tn,j
2 p
E Xtn,k − Xtn,j σtn,m Fj tn,k
p
= E p σ t ft dt|Fj + Op (t 3/2 )
p p(p − 1) tn,m p−2 2 tn,j
= E σtn,j + σt ft + gt dt
2
p
2 tn,j = pσtn,j ftn,j (tn,k − tn,j ) + Op (t 3/2 ).

tn,k
× 2 Xt − Xtn,j dXt + σt2n,j (tn,k − tn,j ) 3. (A.5) is the direct consequence of (A.3) and (A.4). This com-
tn,j
pletes the proof.
tn,k tn,k
+ (tn,k − t)2σt at dt + (tn,k − t) ft2 + gt2 dt
tn,j tn,j Later in the derivation of limit theorem, it is easy to see that the proof
tn,k p−1 and calculations strongly depend on Lemma 1, which will be applied
+p 2 Xt − Xtn,j σt dX, σ t recursively.
tn,j
tn,k
p
+p 2(tn,k − t)σt dσ, σ t + σt2n,j (tn,k − tn,j )
tn,j
A.2 Main Martingale Representation and Argument
tn,m
×p
p−1
σt at dt|Fj + Op (t 5/2
) to Prove Theorem
tn,j
The proof of Theorem 1 is provided in this section, with supporting
p+2 p+1
= E σtn,j (tn,k − tn,j ) + σtn,j atn,j (tn,k − tn,j )2 details in the following sections. Construct an approximate martingale
(MG) on the grid of the τn,i ’s as follows:
1 2 p
+ f + gt2n,j σtn,j (tn,k − tn,j )2
2 tn,j 1
p(p − 1) 2 p √ X,
F (σ 2 ))t
+ ftn,j + gt2n,j σtn,j (tn,k − tn,j )(tn,m − tn,j ) Mn t
⎧ ⎫
2
tn,k p 2 ⎨ K n −2
2 2 ⎬
+ pftn,j 2 Xt − Xtn,j σt dt = √ Xτn,i+1 − Xτn,i F σ̂τn,i+1 − F σ̂τn,i
tn,j Mn t ⎩i=0,τ ≤t ⎭
⎧
n,i+1
tn,m
+
p
pσtn,j 2 Xt − Xtn,j ft − ftn,j dt 2 ⎨ K n −2

tn,j = √ Xτn,i+1 − Xτn,i F σ̂τ2n,i+1 − F στ2n,i+1
Mn t ⎩i=0,τ ≤t
p
tn,k n,i+1
+p 2(tn,k − t)σt ft2 + gt2 dt ⎫
⎪
tn,j
2 2 2 2 ⎬
+F στn,i+1 − F στn,i + F στn,i − F σ̂τn,i
p+1
+ pσtn,j atn,j (tn,k − tn,j )(tn,m − tn,j )|Fj + Op (t 5/2 ) ⎪
⎭
⎧
⎨ K n −2
particular, Heath (1977) as well as Mykland (1995) and the references
2 therein. This interpolation is closely related to Skorokhod embedding
= √ Xτn,i+1 − Xτn,i F στ2n,i+1 − F στ2n,i
⎩
Mn t i=0,τ ≤t in Brownian motion; see, for example, Appendix I of Hall and Heyde
n,i+1
n −2
(1980). Then we only need to prove the CLT for the interpolated con-

K

+ Xτn,i+1 − Xτn,i F σ̂τ2n,i+1 − F στ2n,i+1 tinuous martingale. Therefore, we can apply Theorem 2.28 (p. 152) in
i=0,τn,i+1 ≤t Mykland and Zhang (2012) to prove the CLT. The conditions of cited
⎫ theorem will follow from the development in the rest of Appendix 10,

Kn −2
2 2 ⎬ in particular (A.16) and (A.19), and the approximation in Lemma 2.
− Xτn,i+1 − Xτn,i F σ̂τn,i − F στn,i
⎭ Alternatively, one can develop a functional argument for the CLT as
i=0,τn,i+1 ≤t
⎧ in Jacad (2009), Jacod and Shiryaev (2003), Jacod and Protter (2011),
2 ⎨ K n −2
and Podolskij and Vetter (2009b).
= √ Xτn,i+1 − Xτn,i F στ2n,i+1 − F στ2n,i
⎩
Mn t i=0,τ ≤t
n,i+1 A.3 The Aggregate Conditional Variance
n −1
K
To calculate the quadratic variation, we will calculate the aggregate
+ Xτn,i − Xτn,i−1 F σ̂τ2n,i − F στ2n,i
i=1,τn,i+1 ≤t
conditional variance of Mtn
⎫

Kn −2
2 2 ⎬ var Mτni+1 Fi
− Xτn,i+1 − Xτn,i F σ̂τn,i − F στn,i τi+1 ≤t
⎭ 2
i=0,τn,i+1 ≤t
⎧ = E Mτni+1 Fi
⎪
⎪ τi+1 ≤t
⎪
⎪ )
⎪
⎨ 2 2 *
n −2
E Vτnn,i+1 Fi ) − E Vτnn,i+1 Fi
K
2 = . (A.8)
− Xτn,i στ2n,i+1 − στ2n,i F στ2n,i
= √ Xτn,i+1
Mn t ⎪
⎪
⎪i=0,τn,i+1 ≤t
τi+1 ≤t
⎪
⎪
⎩% &' ( We will first prove that τi+1 ≤t (E(Vτnn,i+1 |Fi ))2 is negligible, of order
(1)
n −1
Op (Mn2 t 2 ). Except a few terms at the edge, the following could be

K

+ Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i F στ2n,i established:

i=1,τn,i+1 ≤t
% &' ( E Vτnn,i+1 Fi
(2)
⎫ 2
= √ E Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i F στ2i
⎪
⎪ Mn t
⎪
⎪
⎪ + Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i F στ2i

Kn −2
⎬
− Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i F στ2n,i + op (1).
⎪
⎪ − Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i F στ2i Fi
i=0,τn,i+1 ≤t ⎪
% &' (⎪
⎪
⎭ 2F στ2n,i τi+1
(3) = √ E σt2 ft dt − Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i
Mn t τi
It follows that

1 + (Xτn,i − Xτn,i−1 ) σ̂τ2n,i − στ2n,i |Fi + Op (Mn t)
√ (X,
F (σ 2 ))t − X, F (σ 2 ))t ) = Vτnn,i+1 + op (1),
Mn t i=0,τ n,i+1 ≤t 2F στ2n,i τi+1
= √ E σt2 ft − στ2n,i fτn,i dt
where, except for a few terms at the edge, Mn t τi
⎧ ⎞
M tn,j
⎪
⎪ 2
2 ⎨ − 2 σt ft − στ2 fτn,i dt|Fi ⎠
n,i
Vτn,i+1 = √
n
Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i F στ2n,i j =1 τn,i
⎪
Mn t ⎪⎩% &' (
2F στ2n,i τn,i+1
(1)
+ √ Xτn,i − Xτn,i−1 E σt − στn,i dt|Fi
2 2
+ Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i F στ2n,i Mn t τn,i
% &' (
(2)
+ Op (Mn t). (A.9)

− Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i F στ2n,i We derive that
% &' ( ⎛ ⎞

(3)
E⎝ E Vτnn,i+1 Fi ⎠
2
τi+1
− F σt2 σt2 ft dt . τn,i+1 ≤t
%
τi
&' ( 2 +
12F στ2i τi+1 2 2
(4)
≤ E E σt ft − στ2n,i fτn,i dt Fi
Mn t τ ≤t τi
i+1
And so the martingale increment becomes ⎛ ⎞2
tn,j
M
2
Mτni+1 = Vτnn,i+1 − E Vτnn,i+1 |Fi . (A.6) +⎝ E 2 σt ft − στ2n,i fτn,i dt Fi ⎠
j =1 τn,i
The martingale up to time t is 2 ⎫
) * 2 τn,i+1 ⎬
Mtn = Vτnn,i+1 − E Vτnn,i+1 Fi . (A.7) + Xτn,i − Xτn,i−1 E σt2 − στ2n,i dt Fi
τn,i ⎭
τn,i+1 ≤t
Although M n above is observed in discrete time, we can interpolate + Op ((Mn t)2 )

the martingale into a continuous martingale up to any time t; see, in ≤ K(Mn t)2 . (A.10)
With this result, we can calculate the conditional variance by only The above three equations lead to the following convergence:
considering conditional second moments instead.
p 16 t 2 2 6
var Mτni+1 Fi → 2 F σs σs ds
A.4 Aggregate Conditional Second Moment τi+1 ≤t
c t 0
t
2 2 4 44 2 32 2
+ F σs σs fs + gs ds.
2 0 3 3
E Vτni+1 Fi (A.16)
τn,i+1 ≤t
⎛⎡ To finalize the proof of Theorem 1, we need the quadratic varia-
4 ⎜⎢ tion of the interpolated martingale. The following lemma shows that
= E ⎝⎣ Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i F στ2i
Mn t τ ≤t the asymptotics of the quadratic variation is the same as that of the
i+1
aggregate conditional variance as in Equation (A.16)
+ Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i F στ2i
Lemma 2.
− Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i F στ2i
⎤2 ⎞ [M n , M n ]t − var Mτnn,i Fi = op (1) (A.17)
τi+1
⎥ ⎟ τn,i ≤t
− F σt2 σt2 ft dt ⎦ Fi ⎠
τi
Proof of Lemma 2:
2
4F στ2i 2 ⎛ ⎞2
= E((Xτn,i+1 − Xτn,i ) στ2n,i+1 − στ2n,i
2

Mn t τ ≤t
i+1 E ⎝[M n , M n ]t − var Mτnn,i Fi ⎠
2 2
τn,i ≤t
+ Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i ⎛ ⎞2
2 2
+ Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i 2
= E⎝ Mτnn,i − var Mτnn,i Fi ⎠
2
− 2 Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i σ̂τ2n,i − στ2n,i τn,i ≤t τn,i ≤t
τi+1 2 2 2 2 32
= E Mτnn,i − E Mτnn,i Fi
+ σt2 ft dt − 2(Xτn,i+1 − Xτn,i ) στ2n,i+1 − στ2n,i τn,i ≤t
τi
τi+1 (by MG property, cross product term has expectation 0)
× σt2 ft dt|Fi + Op (Mn t). 4 2 32
τi
(A.11) ≤ E Mτnn,i + E E (Mτnn,i )2 Fi
τn,i ≤t
Applying Lemma (1), we can prove the following results 4
≤ E Mτnn,i (by Jensen’s inequality for
1 2 2 τn,i ≤t
E Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i Fi conditional expectation)
Mn t τ ≤t 2 2 3
n,i+1
≤ E sup Mτnn,i × M n , M n t
2 4 4 2
= 2 σ 6 Mn t + σ f + gτ2n,i Mn t → 0, (A.18)
Mn t τ ≤t τn,i 3 τ ≤t τn,i τn,i
i+1 i+1
+ Op (Mn t), (A.12) since στi , fτi , gτi , [M n , M n ]t are assumed to be bounded, Mn t =
Op (n−1/2 ).
2 A.5 Elimination of the Bias Term

1 τi+1
E σt2 ft dt − 2(Xτn,i+1 − Xτn,i ) The bias in the limit of X, F (σ 2 )T also depends on the limit of
Mn t τ ≤t (i)
n,i+1
τi
[M , W ]t , where either Wt = Wt or W (i) is orthogonal to Wt , for
n (i)
τi+1
2 any t ∈ (0, T ]. For the second case, it is obvious that [M n , W (i) ]t = 0.
× στn,i+1 − στ2n,i σt2 ft dt|Fi We only need to study the first case Wt(i) = Wt . In this case, since the
τi
covariance between Op (Mn t) terms and Wt will be of even higher
= − στ4n,i fτ2n,i Mn t + Op (Mn t), (A.13)
order (at least Op ((Mn t)3/2 )), those are negligible in the limit. Thus we
τi+1 ≤t
only need to consider the following aggregate conditional expectation:
1
1 2 2 √ cov Mτnn,i+1 , Wτn,i+1 Fi
E Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i Mn t τ ≤t
Mn t τ ≤t i+1

n,i+1
2 1
− 2 Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i σ̂τ2n,i − στ2n,i Fi = √ F στ2n,i
Mn t

= Op (Mn t), (A.14)
× E Wτn,i+1 Xτn,i+1 − Xτn,i στ2n,i+1 − στ2n,i
and τi+1 ≤t i

1 2 2 + Wτn,i+1 Xτn,i − Xτn,i−1 σ̂τ2n,i − στ2n,i
E Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i Fi
Mn t τ ≤t − Wτn,i+1 Xτn,i+1 − Xτn,i σ̂τ2n,i − στ2n,i
n,i+1
2 4 + Op ((Mn t)3/2 )
= σ 6 Mn t + στ4n,i fτ2n,i + gτ2n,i Mn t
τi+1 ≤t
Mn2 t τn,i 3 = Op ((Mn t)3/2 ) (By Itô’s formula and Lemma 1). (A.19)
+ 2στ4n,i fτ2n,i Mn t + Op (Mn t), (A.15) Thus, [M n , W ]t = Op ((Mn t)3/2 ) for any t ∈ (0, T ].

All the proofs above can be easily extended to the case for any of later calculations, we always assume τn,j +1 ∈ (λn,i , λn,i+1 ], τn,j +1 ∈

t ∈ (0, T ]. Thus, as outlined at the end of Section A.2, with (A.3), (λn,i+1 , λn,i+2 ] and τn,j +1 and τn,j +1 are corresponding (j + 1)th obser-
(A.16), (A.18), and (A.19), the proof of Theorem 1 is completed by vation time in the consecutive two big λ-blocks.
applying the Central Limit Theorem for semimartingales (refer to the
version in Mykland and Zhang 2012, thm 2.28).
D.1 Aggregate Conditional Expectation of the Estimator

APPENDIX B: PROOF OF THEOREM 2 According to the contiguity to Gaussian noise (Mykland and Zhang
With Taylor expansion, the first-order difference of the 2011a), the process can be simplified as follows:
2 3
two estimators is: τi+1 ≤t Mn (Mn −1)t F (στn,i )(Xτn,i+1 ) and
2
F
(σ )X (X ) 2
. In order for any term 1 1
τi+1 ≤t Mn (Mn −1)t
Xtn,j + M −1/2 Zτn,i
τ τ τ
n,i n,i+1 n,i+2
to make a difference in the limit, the term must be of order lower than X̂τn,i = Ytn,j =
M tn,j ∈[τi ,τi+1 )
M tn,j ∈[τi ,τi+1 )
Op (Mn t). However, the difference terms are of order higher than
Op (Mn t) by BDG inequality as shown in the proof of Proposition 2 = X̄τn,i + M −1/2 Zτn,i ,
at the end of the article: 1 2
σ̂λ2n,i+1 = X̂τn,j +1 − X̂τn,j
2 3 1 LMt τ
F στn,i E Xτn,i+1 Fi = Op n− 2 , ∈(λ ,λ
n,j +1 n,i n,i+1 ]
Mn (Mn − 1)t 1 2
τi+1 ≤t
= X̄τn,j +1 + Zτn,j +1 M −1/2 ,
2 2 1 LMt
F στn,i Xτn,i+1 Xτn,i+2 = Op n− 2 . τn,j +1 ∈(λn,i ,λn,i+1 ]
τ ≤t
Mn (Mn − 1)t (D.21)
i+1
Theorem 2 is easily proved with this result and a slight variation of the
proof of Theorem 1. and
n −2
K

APPENDIX C: PROOF OF THEOREM 3
X, σ 2 T = 3 X̂λn,i+1 − X̂λn,i σ̂λ2n,i+1 − σ̂λ2n,i
i=0
As we saw in the proof of Theorem 1 that
K −2
3 n

1 2 2 = X̄λn,i+1 + Zλn,i+1 M −1/2
E Xτn,i+1 − Xτn,i σ̂τ2n,i+1 − σ̂τ2n,i Fi LMt i=0
i
M n t ⎛
T T
4 2 2 6 2 2 4 14 2 8 2
×⎝ M −1/2 )2
p
→ 2 F σt σt dt + F σt σt ft + gt dt. (X̄τn,j

+1
+ Zτn,j

+1
c T 0 0 3 3
τn,j +1
⎞
To prove the consistency of Equation (11), we can again apply the

X̄τn,j +1 + Zτn,j +1 M −1/2 ⎠ ,
martingale convergence argument. According to Lemma 2, it suffices 2
−
to check wether the conditional variance of the martingale converges τn,j +1
to 0. By applying Lemma 1, we can obtain:
1 2 2
E Xτn,i+1 − Xτn,i σ̂τ2n,i+1 − σ̂τ2n,i where the Zτn,i ’s are iid N (0, a 2 ). This iid normality leads to the fol-
i
Mn t lowing aggregate conditional expectation of the estimator:

1 2 2 2
− E Xτn,i+1 − Xτn,i σ̂τ2n,i+1 − σ̂τ2n,i Fi Fi ⎛ ⎛
Mn t n −2

K
1 4 2 4 E⎝
3
X̄λn,i+1 ⎝ X̄τ2
= E Xτn,i+1 − Xτn,i σ̂τn,i+1 − σ̂τn,i Fi
2
LMt i=0
n,j +1
(Mn t)2 τn,j +1 ∈(λn,i+1 ,λn,i+2 ]
⎞ ⎞
i
2
1 2 2
−E Xτn,i+1 − Xτn,i σ̂τ2n,i+1 − σ̂τ2n,i Fi −
⎟
X̄τ2n,j +1 ⎠ Fi ⎠ + Op (LMt)
Mn t
τn,j +1 ∈(λn,i ,λn,i+1 ]
= Op Mn2 t 2 . (C.20) ⎛
p
T
T
This proves G1n → 8c 0 (F (σt2 ))2 σt6 dt + cT 0 (F (σt2 ))2 σt4 ⎜
⎜
( 3 ft + 3 gt ) dt.
28 2 16 2 ⎜

T ⎜
p ⎜
By similar argument, one can prove G2n → 8c 0 (F (σt2 ))2 σt6 dt +
Kn −2 ⎜ 1

T 2 2 4 16 2 3 ⎜
cT 0 (F (σt )) σt 3 (ft + gt2 ) dt. = E⎜ 3
LMt i=0 ⎜ M
These two convergences, together with Theorem 1, prove the stable ⎜
τn,j +1 ∈ (λn,i+1 , λn,i+2 ]
⎜
convergence in Theorem 3 . ⎜ tn,l ∈ [λn,i , τn,1 )
⎜
tn,l
∈ [λn,i+1 , τn,1 )
⎝
tn,k ∈ [τn,j
, τn,j +1 )

tn,k ∈ [τn,j +1 , τn,j +2 )
APPENDIX D: PROOF OF THEOREM 4 2
× Xtn,l
− Xt
n,l Xtn,k
− Xtn,k
The proof of Theorem 4 will be done similarly to that of Theorem 1
1 in a manner to compare the difference between (Xtn,j +1 − Xtn,j ) and − 3 − Xt
Xtn,l n,l
M
(X̄tn,j +1 − X̄tn,j ) in each step. As seen in the case without microstructure tn,l ∈ [λn,i , τn,1 )

tn,l ∈ [λn,i+1 , τn,1 )
noise, it is enough to prove the case F (x) = x. For the convenience
⎞
where, except a few terms at the edge,

⎛ ⎞2 ⎟ ⎧
⎟
⎟ ⎪
⎨
⎜ ⎟ ⎟ 3
⎜ ⎟ ⎟ λn = √ λn,i+1 − Xλn,i σλ2 − σλ2n,i
⎜ ⎟ ⎟ V
⎪
X
× ⎜ − Xtn,k ⎟ ⎟ LMt ⎩
n,i+1 n,i+1
⎜ Xtn,k

⎟ Fi ⎟ + Op (LMt)
⎜τ ⎟ ⎟ ⎛ ⎞
⎝ n,j +1 ∈ (λ n,i , λ n,i+1 ] ⎠ ⎟
tn,k ∈ [τn,j , τn,j +1 ) ⎟ 1

∈ [τn,j +1 , τn,j +2 ) ⎟ + X λn,i − X
λn,i−1 ⎝ X̄τ2i − σλ2n,i ⎠
tn,k ⎠
LMt τ ∈(λ ,λ ]
i n,i n,i+1
⎛ ⎞

Kn −2 1
− X λn,i+1 − Xλn,i ⎝ X̄τi − σλn,i ⎠
2 2
= 2σλ2n,i fλn,i LMt + Op (LMt). (D.22) LMt τ ∈(λ ,λ ]
i=0 i n,i n,i+1
λi+1
2 2
The last step applies Lemma 1 (A.3) and (A.4), and we conclude that: − σ ft dt
λi 3 t
⎛
T 1

X,
p
σ 2 T → 2σt2 ft dt = X, σ 2 T . (D.23) + X λn,i+1 − Xλn,i ⎝ Zτ2i M −1
0
LMt τ ∈(λ ,λ ]
i n,i+1 n,i+2
⎞
There is another way to look at X̄τn,j +1 , which leads to a continuous
− Zτ2 M −1 ⎠ i
approximation and provides a simpler way to prove the results. τi ∈(λn,i ,λn,i+1 ]
⎛ ⎞

X̄τn,j +1 − X̄τn,j λn,i−1 )⎝ 1

λn,i − X −1/2 ⎠
+ (X X̄τi Zτi M
LMt
M
M −k τi ∈(λn,i ,λn,i+1 ]
= Xτn,j +1 − Xτn,j + Xtn,k+1
− Xtn,k

M λn,i+1 − X
− (X λn,i )

k=1
⎛ ⎞⎫
tn,k ∈ [τn,j +1 , τn,j +2 )
⎪
⎬
1

M
M −k × ⎝ X̄τi Zτi M −1/2 ⎠ .
− Xtn,k+1 − Xtn,k LMt τ ∈(λ ,λ ⎪
⎭
i n,i n,i+1 ]
M
k=1
tn,k ∈ [τn,j , τn,j +1)
τn,j +2
Again, we will interpolate this martingale into a continuous martin-
1 gale up to any time t ∈ (0, T ] as in the proof of Theorem 1.
Xτn,j +1 + (τn,j +2 − t) dXt
Mt τn,j +1 From this point on, the continuous approximation of X̄ with inte-
τn,j +1
1 gral will be adopted for simplicity and transparency. However, all the
− (τn,j +1 − t) dXt , (D.24) calculation can be checked by applying Lemma 1 to the direct proof
Mt τn,j
without continuous approximation.
and
Lemma 3.
⎛ ⎞
X̄λn,i+1 − X̄λn,i τn,j +1 2

⎜ 1 ⎟
M
M −k E⎝ (τn,j +1 − t) dXt στpn,m Fj ⎠
= Xλn,i+1 − Xλn,i + Xtn,k+1
− Xtn,k
Mt τn,j
M
k=1 4

tn,k
∈ [λn,i+1 , τn,1 ) 1 p(p − 1)
= στp+2 (τn,j +1 − τn,j ) + (τn,m − τn,j )(τn,j +1 − τn,j )
3 n,j 6
M
M −k
− Xtn,k+1 − Xtn,k 2
p p 2
M + σ f (τn,j +1 − τn,j )2
k=1 4 τn,j τn,j 5
tn,k ∈ [λn,i , τn,1 ) 2p + 1

τn,1 + (τn,j +1 − τn,j )2 στpn,j fτn,j + gτ2n,j
1 12
Xλn,i+1 − Xλn,i + (τn,1 − t) dXt p 3
Mt
λn,i+1 + στpn,j X, f τn,j (τn,j +1 − τn,j )2 + Op t 2 , (D.27)
τn,1 4
1
− (τn,1 − t) dXt . (D.25)
Mt λn,i
⎛ ⎞

⎜ 1
One can verify that, to the relevant order, the aggregate conditional τn,k+1
p ⎟
expectation by this continuous approximation gives the same result but E ⎝ Xτn,j − Xτn,k (τn,k+1 − t) dXt στn,m Fk ⎠
Mt
less complicated calculations.
τn,k

2
1 p 2 p
D.2 Conditional Variance of the Approximate Martingale = στp+2 (τn,k+1 − τ n,k ) + σ p
τn,k f + X, f τn,k
2 n,k 2 τn,k 2
2 3
Similarly as in the proof of Theorem 1,the martingale is constructed × (τn,k+1 − τn,k ) + (τn,k+1 − τn,k )(τn,j − τn,k+1 )
2
as follows: p(p − 1) p 2
Up to time t, + στn,k fτn,k + gτ2n,k (τn,k+1 − τn,k )(τn,m − τn,k )
4
) 2p + 1 p 2 3
* + στn,k fτn,k + gτ2n,k (τn,k+1 − τn,k )2 + Op t 2 ,
Mtn = λn
V λn Fi ,
− E V (D.26) 6
n,i+1 n,i+1
λn,i+1 ≤t (D.28)
n,process
⎛ ⎞ var(Mλn,i+1 |Fi ). Applying Lemma 1 and Lemma 3, we can derive:
τn,j +1

⎜ 1 p ⎟
E ⎝ Xτn,j − Xτn,k (τn,j +1 − t) dXt στn,m Fk ⎠ noise
var Mλn,n,i+1 Fi
Mt τn,j
6
2 9
= 3 3 3 E X λn,i+1 − X
λn,i 2
p 2 p L M t λ ≤t
= στpn,k f + X, f τn,k (τn,j +1 − τn,j )(τn,j − τn,k )
2 τn,k 2 ⎛
n,i+1
⎞2
3
+ Op t 2 , (D.29)
× ⎝ Zτ2i M −1 − Zτ2i M −1 ⎠
τi ∈(λn,i+1 ,λn,i+2 ] τi ∈(λn,i ,λn,i+1 ]
and ⎛ ⎞2
2
λn,i − X
+ X λn,i−1 ⎝ X̄τi Zτi M −1/2 ⎠
τn,k+1
1
E (τn,k+1 − t) dXt τi ∈(λn,i ,λn,i+1 ]
Mt τn,k ⎛ ⎞2 ⎞
⎞ 2
τn,j +1 λn,i+1 − X
λn,i ⎝ ⎟
+ X X̄τi Zτi M −1/2 ⎠ Fi ⎠
1 p ⎟
× (τn,j +1 − t) dXt στn,m Fk ⎠ τi ∈(λn,i ,λn,i+1 ]
Mt τn,j

2 + Op (LMt)
p 2 p 96a 2
= στpn,k fτn,k + X, f τn,k (τn,j +1 − τn,j )(τn,k+1 − τn,k ) = σλ4n,i (LMt)2
4 4 L3 M 4 t 3
3 λn,i ≤t
+ Op t 2 , (D.30)
324a 4
E X̄λ2n,i+1 Fi + Op (LMt)
+
L2 M 5 t 3 λn,i+1 ≤t
where τn,k < τn,j ≤ τn,m , Fk = Fn,k .
T T
p 96a 2 216a 4
→ 2 2 2 σt4 dt + σt2 dt, (D.31)
The proof of this Lemma is similar to that of Lemma 1. c c1 T 0 c2 c14 T 3 0
It is easy to see that the aggregate conditional expectation term in
Mtn is of order Op (LMt) from (D.22). So the aggregate conditional and
variance will be the same as the second moment of the term before the
n, process
aggregate conditional expectation. var Mλn,i+1 |Fi
⎛

var Mλnn,i+1 Fi 9 ⎜
2 2 2
= E⎝ X λn,i+1 − Xλn,i σλn,i+1 − σλ2n,i
λn,i+1 ≤t LMt λ ≤t
⎛ n,i+1
2 2
9 ⎜
2 2 2 + X̂λn,i − X̂λn,i−1 σ̂λ2n,i − σλ2n,i
= E⎝ X λn,i+1 − Xλn,i σλn,i+1 − σλ2n,i 2 2
LMt λn,i+1 ≤t + X̂λn,i+1 − X̂λn,i σ̂λ2n,i − σλ2n,i
2 2 2
+ X̂λn,i − X̂λn,i−1 σ̂λ2n,i − σλ2n,i − 2 X̂λn,i+1 − X̂λn,i σ̂λ2n,i − σλ2n,i σλ2n,i − σλ2n,i
2 2 λi+1 2 λi+1
+ X̂λn,i+1 − X̂λn,i σ̂λ2n,i − σλ2n,i 2 2 4
+ σt ft dt − σt2 ft dt X λn,i+1 − X
λn,i
2 3 3
− 2 X̂λn,i+1 − X̂λn,i σ̂λ2n,i − σλ2n,i σλ2n,i − σλ2n,i λi
⎞
λi
λi+1 2 λi+1 ⎟
2 2 4
+ σ ft dt − 2
σt ft dt × σλ2n,i+1 − σλ2n,i |Fi ⎠ + Op (LMt)
λ 3 t 3 λi
i 2
λn,i σλ
λn,i+1 − X
× X − σλ2n,i 16 44
n,i+1
⎛ = σ 6 LMt + σλ4n,i fλ2n,i LMt
2 1 L 2 Mt λn,i 3
+ X λn,i
λn,i+1 − X ⎝ Zτ2i M −1 λn,i ≤t

2
L M t 2 2
τi ∈(λn,i+1 ,λn,i+2 ] 32 4 2
⎞2 + σ g LMt + Op (LMt). (D.32)
3 λn,i λn,i

− Zτ2 M −1 ⎠
τi ∈(λn,i ,λn,i+1 ]
i
These results prove that, for any t ∈ (0, T ]:
⎛ ⎞2 t
1 p 16 t
44 32
+ Xλn,i − X
λn,i−1 2 ⎝ X̄τi Zτi M −1/2 ⎠ [M, M]t → σs6 ds + σs4 ( fs2 + gs2 ) ds
LMt c2 t0 0 3 3
τi ∈(λn,i ,λn,i+1 ] t
⎛ ⎞2 ⎞ 96a 2
+ 2 2 2
216a 4 t 2
σ ds + 2 4 3
4
σ ds. (D.33)
c c1 t 0 s c c1 t 0 s
λn,i )2 ⎝ 1
λn,i+1 − X ⎟
+ (X X̄τi Zτi M −1/2 ⎠ Fi⎠
LMt
τi ∈(λn,i ,λn,i+1 ] By similar methods, we can prove that the bias term also converges
+ Op (LMt). to zero (here we omit the lengthy proof):
1
We will separately calculate the conditional variance involving the √ Cov Mλnn,i+1 , Wλn,i+1 Fi = Op ((LMt)3/2 ).
microstructure noise ZM −1/2 and the variance involving only the LMt i
semimartingale process, denoted accordingly as var(Mλn,noise
n,i+1
|Fi ) and (D.34)
Thus, [M n , W ]t = Op ((LMt)3/2 ) for any t ∈ (0, T ]. This completes REFERENCES

the proof of Theorem 4. Aı̈t-Sahalia, Y., Fan, J., and Li, Y. (2013), “The Leverage Effect Puzzle: Dis-
entangling Sources of Bias at High Frequency,” Journal of Financial Eco-
nomics, 109, 224–249. [198]
APPENDIX E: PROOF OF THEOREM 5 Aı̈t-Shalia, Y., and Jacod, J. (2009), “Testing for Jumps in a Discretely Observed
Process,” The Annals of Statistics, 37, 184–222. [199,205]
Theorem 5 can be easily proved by the consistency of asymptotic Andersen, T. G., and Bollerslev, T. (1998), “Answering the Skeptics: Yes, Stan-
variance (27) and stable convergence in Theorem 4. We only need dard Volatility Models do Provide Accurate Forecasts,” International Eco-
nomic Review, 39, 885–905. [197]
to establish the convergence in probability of Equation (27). This Andersen, T. G., Bollerslev, T., Diebold, F. X., and Ebens, H. (2001), “The
convergence can be proved by adopting the similar martingale Distribution of Realized Stock Return Volatility,” Journal of Financial Eco-
technique as in the proof of Theorme 3, and by applying Lemma 1, nomics, 61, 43–76. [197]
Lemma 2, and Lemma 3. We can prove: E( 4L2 M812 t 2 i (X̂λn,i+1 − Andersen, T. G., Bollerslev, T., Diebold, F. X., and Labys, P. (2000), “Great
Realizations,” Risk, 13, 105–108. [201]
X̂λn,i )4 (F (σ̂λ2n,i+1 ) − F (σ̂λ2n,i ))4 |Fi )−(E( 2LMt
9
i (X̂λn,i+1 − X̂λn,i ) (F
2
Ball, C. A., and Roma, A. (1994), “Stochastic Volatility Option Pricing,” The
(σ̂λn,i+1 ) − F (σ̂λn,i )) |Fi )) = Op (LMt). This equation gives the
2 2 2 2
Journal of Financial and Quantitative Analysis, 29, 589–607. [197,199]
consistency of Gn,1 in Equation (27). By similar argument, the Barndorff-Nielsen, O., Graversen, S., Jacod, J., Podolskij, M., and Shephard,
N. (2006), “A Central Limit Theorem for Realised Power and Bipower
consistency of Gn,2 can be proved and consequently completing the
Variations of Continuous Semimartingales,” in From Stochastic Calculus
proof of the convergence of Equation (27) and Theorem 5. to Mathematical Finance, The Shiryaev Festschrift, eds. Y. Kabanov, R.
Liptser, and J. Stoyanov, Berlin: Springer Verlag, pp. 33–69. [199]
Barndorff-Nielsen, O., and Veraart, A. (2009), “Stochastic Volatility of Volatility
APPENDIX F: PROOF OF PROPOSITION 2 in Continuous Time,” CREATES Research Paper No. 2009-25. [199]
Barndorff-Nielsen, O. E., Hansen, P. R., Lunde, A., and Shephard, N. (2008a),
We will assume the equivalent measure where both Xt and σt are “Designing Realized Kernels to Measure Ex-Post Variation of Equity Prices
in the Presence of Noise,” Econometrica, 76, 1481–1536. [198,203]

martingales. This can be done in analogy with the development in
Barndorff-Nielsen, O. E., Kinnebrock, S., and Shephard, N. (2008b), “Measur-
Mykland and Zhang (2009), sec. 2.2, by a stopping argument and so ing Downside Risk—Realised Semivariance,” Department of Economics
long as the instantaneous correlation between Xt and σt is not ±1. (In Working Paper 382, University of Oxford. [198,199]
the latter case, the proof is different but straightforward.) Barndorff-Nielsen, O. E., and Shephard, N. (2002), “Econometric Analysis of
By the third Bartlett identity (Mykland 1994) Realized Volatility and its Use in Estimating Stochastic Volatility Models,”
Journal of The Royal Statistical Society, Series B, 64, 253–280. [199]
——— (2004), “Power and Bipower Variation With Stochastic Volatility and
E Xt3n,i+1 Fti = 3E Xtn,i+1 X, X|Fti Jumps,” (with discussion), Journal of Financial Econometrics, 2, 1–48.
tn,i+1 [205]
——— (2005), “Variation, Jumps, Market Frictions and High Frequency Data
= 3E Xtn,i+1 2
σu du|Fti
tn,i
in Financial Econometrics,” Technical Report, University of Aarhus. [11]
——— (2006), “Econometrics of Testing for Jumps in Financial Economics
tn,i+1 Using Bipower Variation,” Journal of Financial Econometrics, 4, 1–30.
= 3E Xtn,i+1 σu2 − σt2n,i du|Fti [205]
tn,i Bekaert, G., and Wu, G. (1997), “Asymmetric Volatility and Risk in Equity
(by MG property of Xt ). (F.35) Markets,” NBER Working Paper No. 6022, National Bureau of Economic
Research, Inc. [197]
Black, B. (1976), “Studies of Stock Price Volatility Changes,” in Proceedings
By the Itô formula, d(tn,i+1 − t)(σu2 − σt2n,i ) = −(σu2 − σt2n,i ) dt + of the 176 Meetings of the American Statistical Association, Business and
(tn,i+1 − t) dσu2 , we obtain: Economic Statistics, pp. 177–181. [197]
Bollerslev, T. (1986), “Generalized Autorgeressive Conditional Heteroskedas-
ticity,” Journal of Econometrics, 31, 307–327. [197,207]
tn,i+1
(F.35) = 3E Xtn,i+1 (tn,i+1 − t) dσu2 |Fti Bollerslev, T., Litvinova, J., and Tauchen, G. (2006), “Leverage and Volatility
tn,i Feedback Effects in High-Frequency Data,” Journal of Financial Econo-
metrics, 4, 353–384. [198,199]
tn,i+1 Bouchaud, J.-P., Matacz, A., and Potters, M. (2001), “The Leverage Effect in
= 3E (tn,i+1 − t) d X, σu2 |Fti . Financial Markets: Retarded Volatility and Market Panic,” Science & Fi-
tn,i
nance (CFM) Working Paper Archive 0101120, Science & Finance, Capital
Fund Management. [198]
Brockwell, P. J., and Marquardt, T. (2005), “Lévy-Driven and Fractionally
Hence, Integrated ARMA Processes With Continuous Time Parameter,” Statistica
Sinica, 477–494. [199]
n n3 Campbell, J. Y., and Hentschel, L. (1991), “No News is Good News: An Asym-
E Xt3n,i+1 Fti = 2
tn,i+1 X, σ 2 tn,i metric Model of Changing Volatility in Stock Returns,” NBER Working
T tn,i+1 ≤T
T 2 t ≤T Paper No. 3742, National Bureau of Economic Research, Inc. [197]
n,i+1
+ higher order terms Chan, K., Chockalingam, M., and Lai, K. W. L. (2000), “Overnight Information
T and Intraday Trading Behavior: Evidence From NYSE Cross-Listed Stocks
p 3
→ X, σ 2 u du and Their Local Market Information,” Journal of Multinational Financial
2 0 Management, 10, 495–509. [205]
3 Chen, X., and Ghysels, E. (2011), “News—Good or Bad—and the Impact on
= X, σ 2 T . Volatility Predictions Over Multiple Horizons,” Review of Financial Studies,
2 24, 46–81. [198,199]
Christie, A. A. (1982), “The Stochastic Behavior of Common Stock Variances:
Value, Leverage and Interest Rate Effects,” Journal of Financial Economics,
We have here used that ti = Tn (equidistant spacing); for a more 10, 407–432. [197]
general case, the expression will involve the quadratic variation of Comte, F., Coutin, L., and Renault, E. (2010), “Affine Fractional Stochastic
time. Volatility Models,” Recherche, 13, 1–35. [199]
Comte, F., and Renault, E. (1998), “Long Memory in Continuous-Time Stochas-
tic Volatility Models,” Mathematical Finance, 8, 291–323. [199]
Delbaen, F., and Schachermayer, W. (1994), “A General Version of the Funda-
mental Theorem of Asset Pricing,” Mathematische Annalen, 300, 463–520.
[Received September 2011. Revised August 2013.] [198]
——— (1995), “The Existence of Absolutely Continuous Local Martingale Mandelbrot, B. (1963), “The Variation of Certain Speculative Prices,” Journal
Measures,” Annals of Applied Probability, 5, 926–945. [198] of Business, 36, 394. [197]
——— (1998), “The Fundamental Theorem of Asset Pricing for Unbounded Meddahi, N., and Renault, E. (2004), “Temporal Aggregation and volatility
Stochastic Processes,” Mathematische Annalen, 312, 215–250. [198] Models,” Journal of Econometrics, 119, 355–379. [205]
Engle, R. F. (1982), “Autogregressive Conditional Heteroskedasticity With Es- Mykland, P. A. (1994), “Bartlett Type Identities for Martingales,” The Annals
timates of the Variance of U.K. Inflation,” Econometrica, 50, 987–1008. of Statistics, 22, 21–38. [214]
[197,207] ——— (1995), “Embedding and Asymptotic Expansions for Martingales,”
——— (2000), “The Econometrics of Ultra-High Frequency Data,” Economet- Probability Theory and Related Fields, 103, 475–492. [209]
rica, 68, 1–22. [197] Mykland, P. A., and Zhang, L. (2006), “ANOVA for Diffusions and Itô Pro-
Engle, R. F., and Ng, V. K. (1993), “Measuring and Testing the Impact of News cesses,” The Annals of Statistics, 34, 1931–1963. [198]
on Volatility,” Journal of Finance, 48, 1749–1778. [197] ——— (2009), “Inference for Continuous Semimartingales Observed at High
Fama, E. F. (1965), “The Behavior of Stock-Market Prices,” The Journal of Frequency,” Econometrica, 77, 1403–1455. [201,204,207,214]
Business, 38, 34–105. [197] ——— (2011a), “Between Data Cleaning and Inference: Pre-Averaging and
Figlewski, S., and Wang, X. (2001), “Is the ’Leverage Effect’ a Leverage Ef- other Robust Estimators of the Efficient Price,” Working Paper, University
fect?,” Review of Financial Studies, 24, 46–81. [197] of Illinois at Chicago and University of Chicago. [201,202,211]
French, K. R., Schwert, G. W., and Stambaugh, R. F. (1987), “Expected Stock ——— (2011b), “The Double Gaussian Approximation for High Frequency
Returns and Volatility,” Journal of Financial Economics, 19, 3–29. [197] Data,” Scandinavian Journal of Statistics, 38, 215–236. [199]
Glosten, L. R., Jagannathan, R., and Runkle, D. E. (1993), “On the Relation ——— (2012), “The Econometrics of High Frequency Data,” in Statistical
Between the Expected Value and the Volatility of the Nominal Excess Return Methods for Stochastic Differential Equations, eds. M. Kessler, A. Lindner,
on Stocks,” Journal of Finance, 48, 1779–1801. [197] and M. Sørensen, Boca Raton, FL: Chapman and Hall/CRC Press. [209,211]
Gloter, A., and Hoffmann, M. (2004), “Stochastic Volatility and Fractional Nelson, D. B. (1991), “Conditional Heteroskedasticity in Asset Returns: A New
Brownian Motion,” Stochastic Processes and Their Applications, 113, Approach,” Econometrica, 59, 347–370. [197]
143–172. [199] Neuberger, A. (2012), “Realized Skewness,” The Review of Financial Studies,
Hall, P., and Heyde, C. C. (1980), Martingale Limit Theory and Its Application, 25, 3423–3455. [198]
Boston: Academic Press. [209] Nualart, D. (2006), Fractional Brownian Motion: Stochastic Calculus and Ap-
Harris, L. (1990), “Statistical Properties of the Roll Serial Covariance Bid/Ask plications, Zürich: The European Mathematical Society. [199]
Spread Estimator,” Journal of Finance, 45, 579–590. [201] Officer, R. R. (1973), “The Variability of the Market Factor of the New York
Hasanhodzic, J., and Lo, A. W. (2011), “Black’s Leverage Effect is Not Due to Stock Exchange,” The Journal of Business, 46, 434–453. [197]
Leverage,” Working Paper. [197] O’Hara, M. (1995), Market Microstructure Theory, Cambridge, MA: Blackwell
Hasbrouck, J. (1996), “Modeling Market Microstructure Time Series,” in Hand- Publishers. [201]
book of Statistics, volume 14, eds. C. R. Rao and G. S. Maddala, Amsterdam: Pindyck, R. S. (1984), “Risk, Inflation, and the Stock Market,” American Eco-
North-Holland, pp. 647–692. [201] nomic Review, 74, 335–351. [197]
Heath, D. (1977), “Interpolation of Martingales,” Annals of Probability, 5, Podolskij, M. and Vetter, M. (2009a), “Bipower-Type Estimation in a Noisy
804–806. [209] Diffusion Setting,” Stochastic Processes and Their Applications, 119,
Heston, S. (1993), “A Closed-Form Solution for Options With Stochastic 2803–2831. [201]
Volatility With Applications to Bonds and Currency Options,” Review of ——— (2009b), “Understanding Limit Theorems for Semimartingales: A Short
Financial Studies, 6, 327–343. [197,199,201] Survey,” Statistica Neerlandica, 64, 329–351. [209]
Hull, J., and White, A. (1987), “The Pricing of Options on Assets With Stochastic Podolskij, M., and Ziggel, D. (2010), “New Tests for Jumps in Semimartingale
Volatilities,” Journal of Finance, 42, 281–300. [197,199] Models,” Statistical Inference for Stochastic Processes, 13, 15–41. [205]
Jacad, J. (2009), “On Continuous Conditional Gaussian Martingales and Stable Reiss, M. (2010), “Asymptotic Equivalence and Sufficiency for Volatility Esti-
Convergence in Law,” Seminaire de Probabilites, 232–246. [209] mation Under Microstructure Noise,” ArXiv:1001.3006. [5]
Jacod, J. (1994), “Limit of Random Measures Associated With the Increments Renault, E., and Werker, B. J. (2011), “Causality Effects in Return Volatility
of a Brownian Semimartingale,” Technical Report, Université de Paris VI. Measures With Random Times,” Journal of Econometrics, 160, 272–279.
[11] [198,203]
——— (1996), “La Variation Quadratique du Brownien en Présence d’Erreurs Roll, R. (1984), “A Simple Model of the Implicit Bid-Ask Spread in an Efficient
d’Arrondi,” Astérisque, 236, 155–162. [11] Market,” Journal of Finance, 39, 1127–1139. [201]
——— (2008), “Asymptotic Properties of Realized Power Variations and Re- Stein, E. M., and Stein, J. C. (1991), “Stock Price Distributions With Stochastic
lated Functionals of Semimartingales,” Stochastic Processes and Their Ap- Volatility: An Analytic Approach,” Review of Financial Studies, 4, 727–752.
plications, 118, 517–559. [199] [197,199]
Jacod, J., Li, Y., Mykland, P. A., Podolskij, M., and Vetter, M. (2009), Tauchen, G., and Zhang, L. M. (1996), “Volume, Volatility, and Leverage: A
“Microstructure Noise in the Continuous Case: The Pre-Averaging Ap- Dynamic Analysis,” Journal of Econometrics, 74, 177–208. [197]
proach,” Stochastic Processes and Their Applications, 119, 2249–2276. Weisberg, S. (2004), Applied Linear Regression (3rd ed.), New York:
[201,202,205] Wiley/Interscience. [205]
Jacod, J., and Protter, P. (2011), Discretization of Processes, New York: Springer. Wood, R. A., McInish, T. H., and Ord, J. K. (1985), “An Investigation of
[209] Transactions Data for NYSE Stocks,” Journal of Finance, 40, 723–739.
Jacod, J., and Shiryaev, A. N. (2003), Limit Theorems for Stochastic Processes [205]
(2nd ed.), New York: Springer-Verlag. [209] Wu, G. (2001), “The Determinants of Asymmetric Volatility,” Review of Finan-
Kim, D., and Kon, S. J. (1994), “Alternative Models for the Conditional cial Studies, 14, 837–859. [197]
Heteroscedasticity of Stock Returns,” Journal of Business, 67, 563– Wu, G., and Xiao, Z. (2002), “A Generalized Partially Linear Model of Asym-
598. [197] metric Volatility,” Journal of Empirical Finance, 9, 287–319. [197]
Kristensen, D. (2010), “Nonparametric Filtering of the Realized Spot Volatility: Xiu, D. (2010), “Quasi-Maximum Likelihood Estimation of Volatility With
A Kernel-Based Approach,” Econometric Theory, 26. [3] High Frequency Data,” Journal of Econometrics, 159, 235–250. [5]
Lee, S., and Mykland, P. A. (2012), “Jumps in Equilibrium Prices and Market Zakoian, J.-M. (1994), “Threshold Heteroskedastic Models,” Journal of Eco-
Microstructure Noise,” Journal of Econometrics, 168, 396–406. [205] nomic Dynamics and Control, 18, 931–955. [197]
Li, Y., Mykland, P., Renault, E., Zhang, L., and Zheng, X. (2013), “Realized Zhang, L. (2006), “Efficient Estimation of Stochastic Volatility Using Noisy
Volatility When Endogeniety of Time Matters,” Econometric Theory, to Observations: A Multi-Scale Approach,” Bernoulli, 12, 1019–1043. [5]
appear. [207] Zhang, L., Mykland, P. A., and Aı̈t-Sahalia, Y. (2005), “A Tale of Two
Mancini, C. (2001), “Disentangling the Jumps of the Diffusion in a Geometric Time Scales: Determining Integrated Volatility With Noisy High-Frequency
Jumping Brownian Motion,” Giornale dell’Istituto Italiano degli Attuari, Data,” Journal of the American Statistical Association, 100, 1394–
LXIV, 19–47. [205] 1411. [203]

23.03.29 SW the-Estimation-Of-Leverage-Effect-With-High-Frequency-Data (2014, Christina D. Wang, Leverage Influences Skewness and Volatility)

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

23.03.29 SW the-Estimation-Of-Leverage-Effect-With-High-Frequency-Data (2014, Christina D. Wang, Leverage Influences Skewness and Volatility)

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [134.117.10.

Journal of the American Statistical Association

The Estimation of Leverage Effect With High-Frequency

To link to this article: http://dx.doi.org/10.1080/01621459.2013.864189

PLEASE SCROLL DOWN FOR ARTICLE

1. INTRODUCTION based on the “leverage effect” hypothesis: A drop in the value

Another natural estimator analogous to X,

By the same methods as in the proof of Theorem 1, we have

Table 1. The summary statistics do exhibit the target normality

MSE Mean Median Q1 Q3

n = 390, T = 1/250 infeasible 1.097112 −0.02108 −0.005707 −0.676 0.6496

c = arg min{G1n + G2n } over a suitable grid of c’s. If the grid

MSE Mean Median Q1 Q3

n = 5 days, T = 1/50 infeasible 1.315581 −0.05502 −0.01236 −0.71910 0.66330

Assumption 4. The observation times (tn,i ) satisfy the follows:

and (29) We emphasize that this estimator is feasible (observable), since

P(> |t|) SS explained vif P(> |t|) SS explained vif

P(> |t|) SS explained vif P(> |t|) SS explained vif

the leverage effect

0 20 40 60 80 100 120 0 20 40 60 80 100

the leverage effect

The definition of the stochastic parameter of the leverage effect

= 16σt4n,i ft2n,i (tn,k − tn,j +1 )(tn,j − tn,i )  

and p+2 p+1

Although M n above is observed in discrete time, we can interpolate + Op ((Mn t)2 )

 2 A.5 Elimination of the Bias Term

D.1 Aggregate Conditional Expectation of the Estimator

X̄τn,j +1 − X̄τn,j λn,i−1 )⎝ 1

Thus, [M n , W ]t = Op ((LMt)3/2 ) for any t ∈ (0, T ]. This completes REFERENCES

in the Presence of Noise,” Econometrica, 76, 1481–1536. [198,203]

You might also like

= 16σt4n,i ft2n,i (tn,k − tn,j +1 )(tn,j − tn,i )

2 A.5 Elimination of the Bias Term