Accepted Manuscript

On high frequency estimation of the frictionless price: The use of

observed liquidity variables

Selma Chaker

PII: S0304-4076(17)30104-5
DOI: http://dx.doi.org/10.1016/j.jeconom.2017.06.018
Reference: ECONOM 4397

To appear in: Journal of Econometrics

Received date : 27 March 2015

Revised date : 25 June 2017
Accepted date : 25 June 2017

Please cite this article as: Chaker, S., On high frequency estimation of the frictionless price: The
use of observed liquidity variables. Journal of Econometrics (2017),
http://dx.doi.org/10.1016/j.jeconom.2017.06.018

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to
our customers we are providing this early version of the manuscript. The manuscript will undergo
copyediting, typesetting, and review of the resulting proof before it is published in its final form.
Please note that during the production process errors may be discovered which could affect the
content, and all legal disclaimers that apply to the journal pertain.
On High Frequency Estimation of the Frictionless Price:
The Use of Observed Liquidity Variables

Selma Chaker∗†

Observed high-frequency prices are always contaminated with liquidity costs or market

microstructure noise. Inspired by the market microstructure literature, I explicitly model

this noise and remove it from observed prices to obtain an estimate of the frictionless price.

I then formally test whether the prices adjusted for the estimated liquidity costs are either

totally or partially free from noise. If the liquidity costs are only partially removed, the

residual noise is smaller and closer to an exogenous white noise than the original noise is.

To illustrate my approach, I use the adjusted prices to improve volatility estimation.

JEL Codes: C, C1, C14, C5, C51, C58, G, G2, G20.

KEY WORDS : Stochastic volatility; Hidden semimartingale model; Infill regression ; En-

dogenous noise; Semiparametric volatility estimation.

∗
Université de Tunis, Tunis Business School, BP 65, Bir El Kassaa 2059, Tunisia, Phone: +216
79 409 400 Extension 178, Fax +216 79 409 119, E-mail: selma.chaker@tbs.rnu.tn.
†
Université de Carthage, École Supérieure de la Statistique et de l’Analyse de l’Information,
MASE (Unité Modélisation et Analyse Statistique et Économique), 6 Rue des Métiers, Charguia II,
BP 675, Tunis 1080, Tunisia, E-mail: selma.chaker@gmail.com.
 1. INTRODUCTION

The frictionless price, also referred to as the true price, the efficient price or the
equilibrium price, is the expectation of the asset’s true value conditional on all pub-
licly available information. However, the frictionless price is latent as observed prices
are contaminated with market microstructure noise.
Why then is it important to estimate the frictionless price? From a high frequency
financial econometrics perspective, specifically nonparametric methods, the friction-
less price is either treated as observable or suffering from a measurement error. In
this literature, the object of interest varies from integrated volatility, pioneered by An-
dersen et al. (2003a), to spot volatility, leverage effects, integrated betas and jumps.
The financial applications range from risk management to options hedging, execution
of transactions, portfolio optimization, and forecasting. Therefore, the frictions in
the observed price will impact the estimation of the object of interest as well as the
application. In this paper, I show that estimating the frictionless price before using it
to measure the integrated volatility not only improves the accuracy but also relaxes
the assumptions underlying the traditional robust-to-noise volatility estimators. In
the empirical illustration, I find that for more than half of the business days in 2010,
I can recover the frictionless price for Alcoa Aluminum.
From a market microstructure perspective, several papers provide estimators of
the liquidity or transaction costs which may induce a measure of the frictionless
price. However, the underlying assumptions for the frictionless price usually restrict
its volatility to be constant; key models are surveyed in Hasbrouck (2007). This paper
allows the volatility of the frictionless price to be time-varying as a result of adopting
the standard model in financial econometrics, the hidden semimartingale model. The
latter model assumes that the observed price is the sum of a semimartingale, which
usually has stochastic volatility, and a noise term.

2
 I introduce the liquidity costs in the context of a model that is consistent with both
the standard additive price model of high-frequency financial econometrics and sev-
eral transaction-cost models from the market microstructure literature. The standard
model is given by
pt = p∗t + εt , t ∈ [0, 1], (1)

where pt is the observed log price, p∗t is the log of the frictionless price and εt is a
measurement error term summarizing the market microstructure noise generated by
the trading process1 . The fixed interval [0, 1] is a day, for example. In this context,
the observed price is the sum of two unobservable components, the frictionless price
and the noise.
Regarding the noise, within the market microstructure literature, Stoll (2000)
studies various sources of trading frictions. The presence of a bid-ask spread and
the corresponding bounces is one such source modeled by Roll (1984). Glosten and
Harris (1988) extend Roll’s model by adding a trading volume component to capture
size-varying costs of providing liquidity service. This model is nested in (1) and is
given by
pt = p∗t + β1 qt + β2 qt vt , (2)

where qt is the trade-direction indicator, which takes the value +1 if the trade is
buyer-initiated and -1 if the trade is seller-initiated, and vt is the trading volume. β1
and β2 are two parameters to be estimated.
If an estimator of the noise is denoted by εbt , then

pb∗t = pt − εbt (3)

1
In a similar framework, Aı̈t-Sahalia and Yu (2009) relates statistical measures of the noise ε to
financial measures of the stock liquidity.

3
is an estimator of the frictionless price. For example, in the context of model (2), εbt
would be written as βb1 qt + βb2 qt vt where βb1 and βb2 are consistent estimators of β1 and
β2 , respectively.
The rest of this paper is organized as follows. Section 2 describes the model for
market microstructure noise based on liquidity costs. In Section 3, I discuss the
estimation of this model and describe a test for the performance of the frictionless-
price measure. In Section 4, I study volatility estimation based on adjusting prices for
the liquidity measure introduced in Section 2. To study the finite-sample properties
of the estimators provided in Section 3 and 4, I conduct a Monte Carlo experiment in
Section 5. Section 6 is an empirical application to assess with data the performance
of the price model of Section 2. In Section 7, I offer several conclusions.

2. PRICE MODEL AND SETUP

A generalized model of Glosten and Harris (1988) introduced in (2) is given by

0
pt = p∗t + Ft β, (4)

where F is an M-vector of liquidity-cost variables and β is a parameter to be estimated

from the data. In addition to qt and qt vt , F could contain the bid-ask spread, a natural
measure of frictions, and the quoted depths2 . Indeed, in Kavajecz (1999), the depths
are used to capture inventory-control costs as well as asymmetric-information costs.
It is perhaps too strong an assumption that a linear function of few explanatory
0
variables can absorb all the noise. Therefore, I allow for the possibility that Ft β only
partially absorb the noise εt . In this more realistic scenario, the model of this paper
2
The ask (bid) depth specifies the maximum quantity for which the ask (bid) price applies.

4
accounts for other noise sources in the following way:

0
pt = p∗t + Ft β + ξt . (5)

The residual noise ξt captures all the trading frictions that are not captured by the
0
Ft β form. The magnitude of ξt could also be seen as a measure of the performance of
0 0
the liquidity costs Ft β. If ξt is small, then Ft β is a good measure of liquidity costs.
To present the model in discrete time, I introduce the following notation. I consider
n+1 equidistant price observations at i = 0, 1, .., n over [0,1]. Under the infill sampling
design, an intraday observation of a variable Y is denoted Yiδ where the time span δ
is equal to 1/n. The variation of the variable Y is denoted by ∆Yiδ = Yiδ − Y(i−1)δ .
∗
Then, I denote riδ and riδ the intraday observed and latent returns piδ − p(i−1)δ and
p∗iδ − p∗(i−1)δ , respectively. Finally, the first differences or variations of the regressors
are denoted by Xiδ = Fiδ − F(i−1)δ . Using the model (5), the high-frequency returns
are written as
∗ 0
riδ = riδ + Xiδ β + ∆ξiδ , i = 1..n. (6)

Next, I turn to the assumptions underlying the frictionless price and liquidity
costs. I make the arbitrage-free semimartingale assumption for the frictionless price
as in the standard hidden semimartingale model. This one-dimensional price process,
which evolves in continuous time over the fixed interval [0,1], is defined on a complete
probability space (f, F , P). I consider an information filtration, the increasing family
of σ-fields (Ft )t∈[0,1] ⊆ F , which satisfies the usual conditions of P-completeness
and right continuity. The prices and noise explanatory variables are included in the
information set Ft .
Assumption 1

5
The frictionless price p∗ follows the dynamics

dp∗t = µt dt + σt dWt , (7)

where Wt is a standard Brownian motion, σt is a càdlàg volatility function, and µt is

the drift coefficient.
I make the following set of assumptions for the different components of the noise.
Assumption 2
(i) The increments of Ft are OP (1) and E[Ft ] = 0.
(ii) The fourth moment of Ft is bounded.
Essentially, Assumption 2 allows the first component of the noise to be endogenous
with the frictionless return, autocorrelated and heteroskedastic3 . In fact, the return-
noise endogeneity is empirically evidenced (see, e.g., Hansen and Lunde 2005) and
theoretically modeled (see, e.g., Diebold and Strasser 2013).
Assumption 3
(i) ξt is independent from p∗s and Fs , for all t, s in [0, 1].
(ii) ξt is normally identically distributed with E[ξt ] = 0.
In Assumption 3, the residual noise is an exogenous white noise.

3. FRICTIONLESS-PRICE ESTIMATION

In this section, I estimate the liquidity costs which yield an estimate of the friction-
less price, as in (3). To check whether the proposed liquidity-cost model is reliable, I
derive a formal econometric test to distinguish between models (4) and (5).
The idea of the estimation is to write the price-impact regression in (6) such that
3
For example, the trading volume is highly persistent because of the clustering of small-size trades,
heteroskedastic as a result of its U-shaped intraday pattern, and endogenous with the frictionless
price as modeled in Glosten and Harris (1988).

6
all latent variables, including the frictionless return, are in the regression’s residual:

0 ∗
riδ = Xiδ β + (riδ + ∆ξiδ ), i = 1, ..., n. (8)

In matrix notation, the regression (8) is written as r = Xβ + r∗ + ∆ξ where r =

0 0 0 (m) (m)
(rδ , .., rnδ ) , X = (X(1) , .., X(M ) ), X(m) = (Xδ , .., Xnδ ) for m = 1..M , and ∆ξ =
0
(∆ξδ , .., ∆ξnδ ) . Using a similar notation, I also write the price model in matrix form:
p = p∗ + ε = p∗ + Fβ + ξ.

3.1. Asymptotic theory

In this subsection, I show the consistency and the asymptotic normality of the
estimator of β. Let βb be the ordinary least squares (OLS) estimator of β, defined by

0 0
βb = (X X)−1 X r. (9)

I make the following assumptions.

0
XX P
Assumption A n
−
→ Ω, a matrix of rank M.
0 P
Assumption B X V ar[r∗ |X]X −
→ Ω∗ , a positive definite matrix.
0
X V ar[∆ξ]X P
Assumption C n
−
→ S, a positive definite matrix.
Assumption A concerns the regressors in (8), whereas Assumptions B and C are re-
lated to the residual of the price-impact regression.
I next derive the asymptotic theory for the estimator of the liquidity-cost param-
eters. All proofs are given in Appendix A. Convergence in probability is denoted by
P L
−
→, whereas convergence in law is denoted by −
→.

Theorem 1 Suppose Assumptions 1 and 2 hold.

(i) If V ar[ξt ] = 0:
L
Under Assumptions A and B, n(βb − β) −
→ N (0, Ω−1 Ω∗ Ω−1 ).

7
(ii) If V ar[ξt ] 6= 0:
√ L
Under Assumptions 3, A and C, n(βb − β) −
→ N (0, Ω−1 SΩ−1 ) .

In Theorem 1 (i), consistency is obtained with a faster rate of convergence than the
√
usual n. In that case, the residual is the frictionless return, which is very small
at high frequencies. On the other hand, the noise Xβ is relatively big. Therefore,
the regression performs well and βb is super consistent. In Stock (1987), the super-
convergence rate is obtained in a similar setting.
√
For the case where V ar[ξt ] 6= 0 in Theorem 1 (ii), I obtain the usual n rate of
convergence because the regression residual r∗ + ∆ξ is OP (1). The frictionless-return
b Indeed, the stochastic
moments do not appear in the asymptotic variance of β.
magnitude of the frictionless return is negligible compared to ∆ξ.
0
Once β is consistently estimated, εb = F βb is the liquidity-costs measure proposed
in this paper. By subtracting the liquidity-costs measure from the observed prices, I
decontaminate the latter from noise and obtain a proxy for the frictionless price. Let
the adjusted price pb∗ and the adjusted return rb∗ be defined, respectively, as

0
b
pb∗iδ = piδ − Fiδ β,
(10)
0
∗
rbiδ b
= riδ − Xiδ β.

Finally, the consistency results for βb in Theorem 1 would not be achievable in

a standard setting because of endogenous regression residual; instrumental variables
would be needed to achieve consistency. However, in the setting of this paper, the
endogeneity of the residual does not cause inconsistency because its source, r∗ , is
asymptotically negligible. In finite sample, to consistently estimate β, it might be
important to use instruments. The lag of the regressor X = ∆F would be a valid
instrument.

8
3.2. Testing for the absence of the residual noise ξ

In this subsection, I formally test whether the adjusted returns still have a noise
component. The null hypothesis H0 and the alternative hypothesis H1 are, respec-
tively,

H0 : V ar[ξt ] = 0,
(11)
H1 : V ar[ξt ] 6= 0.

Pn
p∗ , δ) =
I denote by RV (b i=1 r
∗2
biδ the realized variance of the adjusted returns us-
Pn Pn
∗ p∗2iδ −b
i=1 (b p∗2(i−1)δ )2 + p∗(2i+1)δ −b
i=1 (b p∗(2i−1)δ )2
ing δ-frequency returns and RV (b
p , 2δ) = 2
a sim-
ilar object using the sparse 2δ-frequency returns. The objective of taking the average
of two sparse sampling realized variances is to use all the prices and to be able to derive
P
an asymptotic distribution. In fact, the first realized variance ni=1 (b p∗2iδ − pb∗2(i−1)δ )2
P
p∗0 , pb∗2 , pb∗4 , ...} and the second realized variance ni=1 (b
uses the grid {b p∗(2i+1)δ − pb∗(2i−1)δ )2
p∗1 , pb∗3 , pb∗5 , ...}.
uses the grid {b
p∗ , δ) and RV (b
Under H0 , RV (b p∗ , 2δ) are consistent estimators of the integrated vari-
R1 p∗ ,δ) ∗
ance IV = 0 σt2 dt. However, under H1 , RV 2n(b
and RV (bpn ,2δ) are consistent estimators
p∗ ,δ)
RV (b
of V ar[ξt ]. Therefore, the ratio Sn = p∗ ,2δ)
RV (b
converges to 1 under H0 and to 2 under
H1 and a test could be constructed based on Sn .
In the next theorem, I formally state the asymptotic distribution of an Sn -based test
statistics.

Theorem 2
(i) Suppose Assumptions 1, 2, A and B hold.
√
n(Sn −1) L
Under H0 : V ar[ξt ] = 0, r
c
−
→ N (0, 1).
2IQ
c2
IV
(ii) Suppose Assumptions 1-3, A and C hold.
√
n(Sn −1) P
Under H1 : V ar[ξt ] 6= 0, r
c
−
→ +∞.
2IQ
c2
IV

9
 c V\ar[ξt ], V\
c , IQ, P
(iii) P ower(n, c1−α , IV ar[ξt2 ]) −
→ 1,
R1 R1
where 0 σt2 dt and 0 σt4 dt are the integrated variance IV and the integrated quarticity
c and IQ
IQ, respectively. The quantities IV c are consistent estimators of IV and IQ,

ar[ξt ] is a consistent estimator of V ar[ξt ] and V\

respectively. V\ ar[ξt2 ] is a consistent
estimator of V ar[ξt2 ]. The function Power(.) is the power of the test defined as the
probability to reject H0 conditional on H1 being true. The 1 − α– quantile for the
N (0, 1) distribution is denoted c1−α .

According to Theorem 2 (i) and (ii), I reject H0 at the confidence level α when
√
n(Sn −1)
r
c
> c1−α since the test is one sided. In (iii), the asymptotic power of the test
2IQ
c2
IV
converges to the maximum of 1.
Finally, from a market microstructure perspective, the proposed test might be
0
b If this is a
interpreted as a test for the quality of the trading-costs measure F β.
good measure of the noise ε, then the residual noise ξ should go to zero. Otherwise,
the trading-costs measure does not capture all the frictions and the term ξ does not
vanish.
In the context of this paper, rejecting the hypothesis that the adjusted prices still
contain noise is a more desirable outcome. Therefore let’s examine reversing the null
and alternative hypotheses.

H0 : V ar[ξt ] 6= 0,
(12)
H1 : V ar[ξt ] = 0.

Similarly to Theorem 2 and using the same notation, I formally state the asymptotic
distribution of an Sn -based test statistics in Theorem 3.

Theorem 3
(i) Suppose Assumptions 1-3, A and C hold.
√ L 3V ar[ξt2 ]+7V ar[ξt ]2
Under H0 : V ar[ξt ] 6= 0, n(Sn − 2) −
→ N (0, 4 V ar[ξt ]2
).

10
(ii) Suppose Assumptions 1-3, A and C hold.
√ P
Under H1 : V ar[ξt ] = 0, n(Sn − 2) −
→ −∞.
c V\ar[ξt ], V\
c , IQ, P
(iii) P ower(n, c1−α , IV ar[ξt2 ]) −
→ 1.
r
√ 3V\
ar[ξt2 ]+7V\
ar[ξt ]
2

I reject H0 at the confidence level α when n(Sn − 2) < −c1−α 4 \

2
V ar[ξt ]
since the test is one sided, as shown in Theorem 3 (i) and (ii). It is important to note
that testing for the presence of the noise instead of the absence of the noise needs
consistent estimators of V ar[ξt ] and V ar[ξt2 ] to conduct the test. Unfortunately, the
estimators of the moments of the noise are instable as showed in Table 2 of Hansen and
Lunde (2005) and therefore can generate unreliable results. Nevertheless, in Theorem
3 (iii) the asymptotic power of the test converges to the maximum of 1.

4. APPLICATION TO VOLATILITY ESTIMATION

R1
The object of interest in this section is the integrated variance IV = 0
σu2 du. To
simplify the notation I denote the realized variance by RV (.) of a given process if the
return frequency is δ, the highest possible. If the frictionless return were observed,
P
then the realized variance RV (p∗ ) = ni=1 riδ
∗2
would be a consistent estimator of IV,
as first shown in Meyer (1967). However, the realized variance based on observed
P
prices, RV (p) = ni=1 riδ
2
, is inconsistent for IV because of the market microstructure
noise.
The first consistent IV estimator which is robust to noise is the two time scales estima-
tor of Zhang et al. (2005). This estimator relies on standard noise assumptions rather
than endogenous, autocorrelated and heteroskedastic noise, as in this paper. In the
same line of nonparametric volatility estimators, Barndorff-Nielsen et al. (2008a) and
Jacod et al. (2009) derive the kernel and the pre-averaging estimators, respectively.
All the mentioned three estimators are inconsistent for IV if applied to the price model
(5) because of the endogeneity in the noise4 . However, the noise heteroskedasticy is
4
See Li and Mykland (2007) for a sensitivity analysis of a specific endogeneity form.

11
allowed only for the pre-averaging estimator. As for the autocorrelation in the noise,
it could be accommodated for in the two time scales estimator and the pre-averaging
estimator; see the extensions in Aı̈t-Sahalia et al. (2011) and Hautsch and Podolskij
(2013), respectively.
In this section, I propose a novel volatility estimator, based on adjusted prices defined
in (10), which relaxes the underlying noise assumptions of the aforementioned non-
parametric volatility estimators. Improved volatility estimation is due to the fact that
the adjusted price pb∗ is closer to the frictionless price p∗ . And more importantly, pb∗
fits the assumptions justifying the use of model-free volatility estimators better than
p if and only if the DGP is indeed given by the model of this paper.
In Theorem 3, I show that if the liquidity costs are removed or H0 is not rejected, the
P
realized variance based on adjusted returns, RV (b p∗ ) = ni=1 rbiδ
∗2
, is consistent for IV.
For the case where the liquidity costs are only partially absorbed or H0 is rejected, a
robust-to-noise volatility estimator is needed. I show that, based on adjusted rather
p∗ ), becomes con-
than observed prices, the two time scale estimator, denoted by T S(b
sistent for IV.
p∗ ) is defined in Appendix B and in the proof of Theorem 3. For mixed normal-
T S(b
st
limit distributions, I denote the stable convergence5 as −
→.

Theorem 4 Suppose Assumptions 1 and 2 hold.

(i) Suppose V ar[ξt ] = 0:
√ st
Under Assumptions A and B, p∗ ) − IV ) −
n(RV (b → N (0, 2 IQ),
R1
where IQ = 0 σu4 du.
(ii) Suppose V ar[ξt ] 6= 0:
Under Assumptions 3, A, C and if the increments of Ft have bounded fourth moments,
st
p∗ ) − IV ) −
n1/6 (T S(b → N (0, Γξ ),
 −1/3
8
where Γξ = c2
E[ξ 2 ]2 + 34 cIQ, c = 1
12E[ξ 2 ]2
IQ .
5
The stable convergence concept is discussed in Aldous and Eagleson (1978).

12
According to Theorem 3, the estimation error in β̂ impacts neither the consistency
nor the asymptotic distribution of the estimator based on the adjusted returns. In
case (i), the result is similar to that where p∗ is observed. And for case (ii), the result
is similar to that where p∗ + ξ is observed.
p∗ ) of Theorem 3 (i) has no c-like tuning parameter
The new volatility estimator RV (b
as it would be the case if T S(p) were used. The result is obtained because the adjusted
return could be written as rb∗ = r∗ +oP (1/n), where the oP (1/n) term is a consequence
of the rate of convergence of β̂ in Theorem 1 (i). Finally, the rate of convergence of
√
RV (bp∗ ) is n, which is not achievable using any robust-to-noise volatility estimator if
the comparison abstracts from the assumption of observing the noise. Indeed, Gloter
and Jacod (2001) show that the rate of convergence of any robust-to-noise integrated
volatility estimator is bounded by n−1/4 , where n is the sample size.
p∗ ) has a limit distribution as if
In Theorem 3 (ii), the new volatility estimator T S(b
the noise were ξ (exogenous, independent and identically distributed) instead of ε
(endogenous6 , autocorrelated and heteroskedastic). The idea of the two time scales
estimator is to combine a slow-scale sampling return frequency (to mitigate the noise
effect) with a high-scale one (to correct the noise-induced bias). The same idea still
holds if the adjusted returns are used. First, for the slow-scale sampling return, low-
√
frequency adjusted returns contain an oP (1/ n) term resulting from the estimation
of β̂ in Theorem 1 (ii). This term is negligible compared to low-frequency frictionless
√
returns, which are oP (1/ n), where n is the slow frequency verifying n < n. Second,
√
for the high-scale frequency, high-frequency adjusted returns also contain an oP (1/ n)
term resulting from β̂ estimation in Theorem 1 (ii). This term is negligible compared
to the residual noise component ∆ξ, which is OP (1). Therefore, the bias correction is
not impacted.
To summarize, the new liquidity-based estimator of integrated variance denoted by
6
Within the statistical approach of the nonparametric estimation of volatility, Kalnina and Linton
(2008) propose an alternative specification of the return-noise endogeneity.

13
LBE is written as


 RV (b
p∗ ), if V ar[ξt ]=0;
LBEts (p) = (13)

 T S(b
p∗ ), if V ar[ξt ] 6= 0.

Confidence intervals for LBEts (p) are computed using Theorem 3. And, as in Gonçalves
and Meddahi (2009), more accurate confidence intervals could be obtained using the
bootstrap method.

5. MONTE CARLO EVIDENCE

In this section, I show that the finite-sample simulation results are consistent with
those predicted by the aforementioned asymptotic theory. I find that the test has
a good performance. The new IV estimator is more accurate than most estimators
proposed in the literature.
I first describe the data-generating process for the frictionless price, the spot volatility
and liquidity-cost variables. Second, I report the simulation results for the liquidity-
cost estimation as well as the return variance estimation.

5.1. The artificial data

For the frictionless price, I use the Heston (1993) model given in discrete time

∗ (1)
riδ = (ϕ − νiδ /2)δ + σiδ ∆Biδ
(14)
1/2 (2)
dνiδ = κ(α − νiδ )δ + γνiδ ∆Biδ ,

where ν = σ 2 and the parameters (ϕ, κ, α, γ, ρ) are set to (0.05, 5, 0.04, 0.5, −0.5) as
in Zhang et al. (2005). The parameter ρ is the correlation coefficient between the two
Brownian motions dB (1) and dB (2) .

14
 I set the noise εiδ = βFiδ + ξiδ . For the liquidity cost, I use the following model
to simulate a one factor variable F that is correlated with the frictionless return, for
i = 1..n,

∗
βFiδ = βqiδ + %riδ (15)
|{z} |{z}
exogenous endogenous

This type of dependence has been introduced in Kalnina and Linton (2008) and
Barndorff-Nielsen et al. (2008a). In Hansen and Lunde (2005), empirically the sign
of % is negative. In the simulations, the main scenario refers to (β, %) = (0.01, −1)
but I run eight other scenarios corresponding to different values of (β, %) (see Table 3
commented in Section 5.2).
The direction of the trade q is triggered by a Bernoulli process with clustering.
Trades cluster since buys are likely followed by buys, and sells are likely followed by
sells. Moreover, some big-volume trades are divided into small-volume trades and
executed consecutively as a series of sells or buys. The Bernoulli process is originally
a sequence of random binary variables, which are independent. A generalization of a
Bernoulli process that incorporates a dependence structure is given by Klotz (1972),
in which he considers q, as a stationary two-state Markov chain with state space
{−1, 1}.The parameter of the process is λ which measures the degree of persistence
in the chain. The transition matrix is given by
 
 λ 1−λ 
T (λ) =  ,
1−λ λ

where λ = 0.9 to simulate the clustering of trades.

15
5.2. Simulation results

I run 2, 500 replications or days. For each day, a trade occurs every second leading
to 23, 400 observations. A business day has 6.5 working hours. I compare the esti-
mations of OLS and the instrumental variable denoted INST. The adjusted return
0 0 0
∗
expression becomes rbiδ = riδ − Xiδ βbinst where βbinst = (Z X)−1 Z r and Z denotes
0
the lag of X. Z is a valid instrument since Z X is nonsingular, as the liquidity-cost
variables are persistent, and it is uncorrelated with the regression residual at high
frequencies, as a consequence of the p∗ semimartingale property.
For the main scenario, Table 1 shows that the price-impact regression parameters, β,
are estimated very accurately. There is evidence of INST estimation better than OLS
estimation in terms of bias and RMSE reduction if the test of Theorem 2 is used. For
the reverse test of Theorem 3, the variance and RMSE are higher than those obtained
using the test of Theorem 2. However, both tests achieve consistency even in finite
sample.
Regarding the omitted factor scenario, which refers to the noise model ε = 0.01N (0, 1)+
0.01F + ξ where the factor N (0, 1) is omitted, the variance and rmse in Table 1 are
smaller than those of the main scenario. However, as the asymptotic theory derived in
Section 3.1 is not valid under the omitted factor scenario, then the asymptotic values
of bias, variance, RMSE, test level and power are not known any more. Moreover, the
small sample level using the test of Theorem 2 is close to 1 strongly rejecting H0 due
to the bias induced by the omitted factor. For the same reason7 , the small sample
power using the test of Theorem 3 is close to 0.
The extra factor scenario differs from the main scenario by including a factor of iid
N (0, 1) series which is unrelated to the noise, the aforementioned asymptotic theory
does not apply though. The results reported in Table 1 show that the bias and the
7
If the coefficient of the omitted factor is decreased as compared to the coefficient of the included
factor, the results are closer to the main scenario and are available upon request.

16
RMSE are close to those of the main scenario for the OLS estimation whether the
test or its reverse is used. For INST, the bias is stable but the variance and RMSE
are higher than those of the main scenario whether the test or its reverse is used. The
small sample test level and power are similar to the main scenario.
Concerning the volatility application, I compare the new family of integrated vari-
ance estimators of this paper which are liquidity based to the following estimators
from the literature. For the nonparametric family, I use the two time scales estima-
tor (T S) of Zhang et al. (2005), the multi scale estimator8 (M S) of Zhang (2006)
achieving the optimal convergence rate, the pre-averaging estimator9 (P RE) of Jacod
et al. (2009), the kernel estimator10 (KER) of Barndorff-Nielsen et al. (2008b). For
the parametric family, I use the maximum likelihood estimator (M LE) of Aı̈t-Sahalia
et al. (2005) and the quasi-maximum likelihood estimator (QM LE) of Xiu (2010).
The details of the estimators are provided in the Appendix B.
Table 2 shows the IV estimation results. The common IV estimator are dominated
by the LBE estimators of the main scenario according to the bias, variance, RMSE
and coverage probability criteria. This improvement is evidenced also for the omitted
factor and the extra factor scenarios according to the variance, RMSE and coverage
probability criteria. Moreover, using OLS or INST, the test or its reverse, does not
matter to improve IV estimation (for the main scenario, refer to the first column as
compared to each one of the following four columns). Across scenarios even though
the variance is bigger for the omitted factor and the extra factor scenarios than the
main scenario, the rmse is comparable.
Within the class of LBE volatility estimators, adjusting the returns using INST pro-
8
In Aı̈t-Sahalia et al. (2011), the multi scale estimator is extended to account for dependent noise.
The authors show that the results are very similar to the original multi scale estimator.
9
In Hautsch and Podolskij (2013), the pre-averaging estimator is extended to account for de-
pendent noise. Unfortunately, I find that the resulting integrated quarticity estimator performs
badly compared to the original pre-averaging estimator of quarticity. Therefore, I use the original
pre-averaging estimator throughout the paper.
10
In Barndorff-Nielsen et al. (2011), the authors show that subsampling realized kernels gives
similar results to the original kernel estimator.

17
vides better results than using OLS in terms of higher coverage probability for all cases
(for the main and the extra factor scenarios, refer to the second column as compared
to fourth column and the third column as compared to the fifth column). According
to the rmse, and for the omitted factor and the extra factor scenarios, INST gives
better results than OLS.
To compare the two tests, the main scenario shows that the test of Theorem 2 per-
forms better than the test of Theorem 3 according to the bias, variance and RMSE
criteria. Globally, the results of the reverse test are stable since the simulated residual
noise ξ follows a Normal distribution and for that specific case the asymptotic variance
√ 3V\
ar[ξt2 ]+7V\
ar[ξt ]
2

in Theorem 3 (i) should be 4 13 instead of 4 \

2 . In the simulations, I
V ar[ξt ]

use the estimators V\ ar[ξt ] = RV (b p∗ )/2n and V\ p∗ )/2n − 4V\

ar[ξt2 ] = RQ(b ar[ξt ] where
P P
RV (bp∗ ) = ni=1 rbiδ
∗2
and RQ(b p∗ ) = ni=1 rbiδ
∗4
. However, if ξ does not follow a Normal
distribution, the error in estimating its moments could make the test unreliable.
Finally, Table 3 reports the variation of the coverage probability of the TS-related
estimators for different values of (β, %). The confidence interval of the traditional TS
estimator fails to contain the true IV. However, LBEts achieves higher coverage prob-
ability reaching 78% in many cases. If OLS and INST are compared, mostly INST
gives better coverage. However the test of Theorem 2 and its reverse of Theorem 3
give similar results.

18
 OLS INST
as ss as ss
Main scenario
Bias ×104 0 -3.60 0 -0.03
Test Reverse Test Reverse
Variance ×1010 1.24 1.52 2.40 68.86 65.45 116.35
RMSE ×105 1.11 39.03 39.04 8.29 8.11 40.47
Level 0.050 0.000 0.031 0.050 0.048 0.037
Power 1 1 1 1 1 1
Omitted factor
Bias ×104 - -3.88 - 0.32
Test Reverse Test Reverse
Variance ×1010 - 0.02 0.09 - 0.25 0.30
RMSE ×105 - 0.38 0.38 - 3.31 39.27
Level - 0.998 0.000 - 0.998 0.000
Power - 1 0.001 - 1 0.004
Extra factor
Bias ×104 - -3.88 - 0.19
Test Reverse Test Reverse
Variance ×1010 - 35.51 35.39 - 2094.61 2084.02
RMSE ×105 - 39.29 39.29 - 45.80 59.93
Level - 0.000 0.029 - 0.058 0.050
Power - 1 1 - 1 1
Table 1: β estimation and the test rejection rate. For the main scenario β = 0.01
and % = −1. INST refers to the instrumental variable estimation. The omitted factor
model for the noise is ε = 0.01N (0, 1) + 0.01F + ξ where the factor N (0, 1) is omitted.
The extra factor scenario adds an unrelated iid N (0, 1) series to the main scenario.
Asymptotic and small sample are denoted as and ss, respectively. Test and Reverse
refer to the tests of Theorem 2 and 3, respectively.

19
 Main scenario Omitted factor Extra factor
Estimators IV LBEIV IV LBEIV IV LBEIV
OLS INST OLS INST OLS INST
c c c
c c c

The null hypothesis H0 H1 H0 H1 H0 H1 H0 H1 H0 H1 H0 H1

TS: Bias ×102 30.03 -0.27 1.87 -0.22 2.42 -13.16 -10.12 -9.94 0.12 1.26 3.61 -10.40 -8.36 -0.25 3.45
Variance ×107 55.71 12.44 56.58 12.75 65.28 2 105 4 104 4 104 1 105 1 105 2 105 4 104 4 104 1 105 1 105
RMSE ×102 30.03 0.29 2.41 0.31 2.72 38.77 10.25 10.43 0.96 2.10 55.92 10.43 12.46 0.70 4.13
Coverage Prob. 0 20.16 20.04 59.64 60.52 0 68.40 68.36 73.32 73.28 0 22.68 22.32 59.16 59.36
MS: Bias ×102 31.65 -0.13 2.00 0.01 2.55 8.63 -9.75 -9.57 0.48 1.61 26.61 -10.25 -8.21 0.01 3.59
Variance ×107 149.40 13.85 57.81 14.32 66.54 5 105 4 10 4 4 10 4 1 10 5 1 105 5 105 4 10 4 4 10 4 1 10 5 1 105
RMSE ×102 31.65 0.22 2.34 0.17 2.67 18.11 10.10 10.28 0.88 2.02 36.10 10.35 12.38 0.54 4.07
Coverage Prob. 0 36.16 35.32 72.08 72.52 0 93.96 93.92 86.68 86.64 0 39.20 38.40 70.80 70.80
KER: Bias ×102 24.73 -0.12 2.01 0.007 2.56 11.37 -9.69 -9.51 0.55 1.69 24.98 -10.24 -8.20 0.04 3.60
Variance ×107 68.34 13.28 57.28 13.69 65.98 5 105 4 104 4 104 1 105 1 105 5 105 4 104 4 104 1 105 1 105
RMSE ×102 24.73 0.22 2.34 0.15 2.67 11.66 10.10 10.29 0.91 2.04 25.26 10.35 12.38 0.52 4.07
Coverage Prob. 0 36.32 35.52 74.08 71.96 0.08 92.04 92.00 80.12 80.08 0.08 38.76 37.96 73.80 70.44

20
PRE: Bias ×102 3.45 -0.13 2.00 0.002 2.53 10.18 -9.75 -9.57 0.41 1.54 10.19 -10.25 -8.21 0.01 3.58
Variance ×107 231.50 27.69 70.15 31.01 78.81 1 107 4 104 4 104 1 105 1 105 1 107 4 104 4 104 1 105 1 105
RMSE ×102 3.48 0.25 2.37 0.18 2.69 10.91 10.16 10.34 0.90 2.04 10.92 10.38 12.41 0.55 4.09
Coverage Prob. 0.04 36.92 36.04 77.92 74.96 0.08 94.20 94.16 93.52 93.40 0.08 40.00 39.12 77.60 73.36
MLE: Bias ×102 64.19 -0.12 2.01 0.03 2.56 34.43 -9.73 -9.55 0.53 1.67 64.25 -10.24 -8.20 0.04 3.60
Variance ×107 983.76 10.74 54.89 11.18 63.78 5 103 4 10 4 4 10 4 1 10 5 1 105 1 103 4 10 4 4 10 4 1 10 5 1 105
RMSE ×102 64.20 0.21 2.33 0.15 2.66 34.45 10.06 10.25 0.88 2.02 64.26 10.34 12.38 0.52 4.07
Coverage Prob. 0 35.76 34.96 72.00 71.48 0 91.28 91.24 79.08 79.04 0 38.12 37.40 71.04 69.20
QMLE:Bias ×102 64.19 -0.12 2.01 0.03 2.56 34.43 -9.73 -9.55 0.53 1.67 64.25 -10.24 -8.20 0.04 3.60
Variance ×107 379.24 12.51 56.57 12.79 65.30 2 104 4 10 4 4 10 4 1 10 5 1 105 4 102 4 10 4 4 10 4 1 10 5 1 105
RMSE ×102 64.20 0.21 2.34 0.15 2.67 34.45 10.08 10.26 0.90 2.03 64.25 10.34 12.38 0.52 4.07
Coverage Prob. 0 36.52 35.68 72.92 72.20 0 93.84 93.80 82.24 82.20 0 39.12 38.36 72.08 70.16

Table 2: Simulation results for the volatility application. Coverage Prob. is the number of times in percentage the estimated
confidence interval includes IV. For the main scenario β = 0.01 and % = −1. INST refers to the instrumental variable
estimation. The omitted factor model for the noise is ε = 0.01N (0, 1) + 0.01F + ξ where the factor N (0, 1) is omitted. The
extra factor scenario adds an unrelated iid N (0, 1) series to the main scenario. When the null hypothesis is H0 : V ar[ξt ] = 0
(H1 : V ar[ξt ] 6= 0), the test (the reverse test) of Theorem 2 (Theorem 3) is used. The expressions of the volatility estimators
are provided in Appendix B.
 %
β -0.1 -0.5 -1 -5
0.005 TS 0 0 0 0
H0 H1 H0 H1 H0 H1 H0 H1
LBEts − OLS 76.00 76.56 20.16 20.04 12.76 12.52 11.20 11.12
LBEts − IN ST 76.88 77.64 59.64 60.52 52.32 53.24 53.28 53.52
0.01 TS 0 0 0 0
H0 H1 H0 H1 H0 H1 H0 H1
LBEts − OLS 76.20 77.08 67.52 67.48 20.16 20.04 12.72 12.56
LBEts − IN ST 77.76 78.52 70.64 71.64 59.64 60.52 52.52 52.88
0.05 TS 0 0 0 0
H0 H1 H0 H1 H0 H1 H0 H1
LBEts − OLS 76.36 77.80 76.20 77.08 76.00 76.56 20.16 20.04
LBEts − IN ST 77.40 78.52 77.76 78.52 76.88 77.64 59.64 60.52

Table 3: The coverage Probability for the TS-related estimators in %. In the first line,
the coverage probability of the traditional TS estimator is reported. The coverage
probability of LBEts is reported for OLS and INST. When the null hypothesis is
H0 : V ar[ξt ] = 0 (H1 : V ar[ξt ] 6= 0), the test (the reverse test) of Theorem 2 (Theorem
3) is used.

6. EMPIRICAL ILLUSTRATION

In this section, I assess with data the performance of the model presented in Section
2. I use Alcoa Aluminum stock, listed on the NYSE, and the high frequency dataset
covers the 2010 period. First, I describe the liquidity variables. Then, I provide the
results for the frictionless price estimation. Finally, I compare the liquidity based
variance estimator of this paper to the common estimators of the literature.

6.1. Liquidity variables’ description

To clean this data, I apply the same procedure as in Barndorff-Nielsen et al.

(2008b). I use five explanatory variables to capture the liquidity costs: the inferred
trade-direction indicator11 (q), trading volume (v), bid-ask spread (s), ask depth (da )
11
I infer the binary series qt from observed trade and quote prices using the Lee and Ready (1991)
trade classification algorithm. Trade classification requires that the trade series be matched with the
quote series because in the Trade and Quote (TAQ) database the two series are offered separately. I

21
 Correlation q v s da db
q 1.0000
v 0.0097 1.0000
s 0.1539 -0.0274 1.0000
da -0.1440 0.4574 -0.0305 1.0000
db 0.1474 0.4571 -0.0324 0.7200 1.0000
Table 4: The correlation matrix for the liquidity variables.

and bid depth (db ). The last four variables are used in logarithmic terms and I demean
the five variables.
In order to check that the liquidity-cost variables are valid candidates as noise
explanatory variables, I use the signature plot tool. The volatility signature plot of
Andersen et al. (2000) draws the average of daily realized variances across the sam-
pling frequency of the underlying returns. An explanatory variable is valid (i.e., O(1))
if its quadratic variation explodes at high frequencies, as in Assumption 2(i). The sig-
nature plots of Figure 1 illustrate the validity of the liquidity variables considered.
The autocorrelation functions (ACF) of the five noise explanatory variables are
plotted in Figure 1. Each plot displays the average autocorrelation across days. The
first-order autocorrelation is the highest for all the variables, and the autocorrelation
decays when the lag increases except for the bid-ask spread. Finally, Table 4 shows
the correlation between the liquidity variables.

6.2. The frictionless price estimation

I use OLS and instrumental variable to estimate the model (6). Indeed, as
mentioned in Section 3.1, using instrumental variables might be important in fi-
nite sample. When instruments are used, the adjusted return expression becomes
0 0 0
∗
rbiδ = riδ − Xiδ βbinst where βbinst = (Z X)−1 Z r and Z denotes the lag of X. Z is a valid
0
instrument since Z X is nonsingular, as the liquidity-cost variables are persistent, and
match trades and quotes by assuming a zero time lag.

22
it is uncorrelated with the regression residual at high frequencies, as a consequence of
the p∗ semimartingale property.
To highlight the importance of the use of instrumental variables, I plot in Figure
2 the first-order realized autocovariance of the observed returns and adjusted returns.
For Alcoa in 2010, the first-order realized autocovariance using adjusted returns with
instrumental variables is mostly negative and close to zero. This is an evidence that
any residual noise would be an exogenous white noise. However, OLS-based realized
autocovariance could be positive in contradiction with a residual noise that is an ex-
ogenous white noise. This Figure also shows that the stylized fact of the negativity
of the first-order autocovariance of the high-frequency returns disappears, or at least
becomes much less pronounced, by adjusting the returns for liquidity costs.
For the test of Theorem 2, I find that the liquidity-cost measure absorbs all the
noise in more than 70% of the sample days compared to less than 10% if the prices
are not adjusted. In Table 5, I report the rejection rates for two different estimators
of IV: the PRE and the MS. To estimate the quarticity IQ, I use the pre-averaging
method. These classical estimators are described in Appendix B. The results of the
test are consistent across the two measures of IV and instrumental variable estimation
allows to absorb more noise than the OLS estimation.
By adding an extra unrelated factor such as a simulated iid N (0, 1) process, the re-
jection rate remains the same as the five factor model as shown in Table 5. The five
factor model achieves the lowest rejection rate, which is less than 4%, if OLS and test
of Theorem 2 are used. For INST, the Roll and five factor models deliver similar best
results with a rejection rate less than 26%.
However, for the reverse test of Theorem 3, the rejection rates reported in Table 5
p∗ )/2n and
are highly instable across two different measures of V ar[ξt ] which are RV (b
p∗ )/n. Moreover, using original or adjusted prices does not make any difference
−RC(b
in terms of rejection rate if the RV based estimator of the noise variance is used.

23
 Rejection rate in % Original Roll Glosten and Harris Five factors Extra factor
OLS INST OLS INST OLS INST OLS INST
c
IV Test: H0 : V ar[ξt ] = 0
PRE 90.47 13.88 24.60 7.93 32.53 3.57 25.39 3.57 25.39
MS 90.07 11.11 21.42 6.34 28.17 1.98 22.61 1.98 22.22
Noise variance Reverse test: H0 : V ar[ξt ] 6= 0
RV/2n 98.80 98.80 98.80 98.80 98.80 98.80 98.80 98.80 98.80
-RC/n 24.60 1.19 0 2.77 0 5.95 0 5.95 0

Table 5: Rejection rates for the test and the reverse test. The Roll noise model
contains one factor which is q. The Glosten and Harris model adds the factor qv to
the Roll model. For the five factor model, the factors (q, qv, qs, da , db ) are used. And
the extra factor model adds an unrelated iid series to the five factor model.

Finally, if the RC based estimator is used, the reverse test rejects more V ar[ξt ] 6= 0
for original prices than for adjusted prices which is problematic. Based on the afore-
mentioned reasons, I will not use the reverse test for the rest of the paper.
I find that the noise explanatory-variable coefficients are mostly significant at the
95% confidence level for almost all of the business days (see Figure 2). However, the
trade direction indicator is the most relevant variable which makes the Roll model im-
portant. The confidence intervals are computed using Theorem 1 and adapted to the
use of instrumental variable for the semimartingale-based model. The trade indicator
q coefficient is positive for all days. The signed-volume qv coefficient is negative for all
days. A transaction with a higher number of shares generates a lower cost per share.
For the signed spread qs, the coefficient is mostly negative in 2010. A wider spread is
associated with a smaller buy price and a bigger sell price. The quoted depths coef-
ficients are negative. This is consistent with the presence of inventory-control costs.
If the ask volume increases, the price rises in an attempt to elicit sales. The same is
true for the bid volume.
I randomly choose one day for which H0 is not rejected (Monday September 27,
2010, n=7069) and one day for which H0 is rejected (Tuesday August 24, 2010,
n=1728). For both days, I plot in Figure 3 the histograms of the original and adjusted

24
returns as well as the histograms of fitted noise.
For the whole sample, and to visualize the improvement of adjusted returns over
original returns, I plot in Figure 3 the ACF function of the returns as well as the fitted
noise. And I plot in Figure 2, the return-noise correlation and the noise-to-signal ra-
tio. Finally, the signature plot in Figure 2 shows a flatter curve for adjusted than for
original returns except at the highest frequencies. However, as shown the signature
plot is almost flat if the price impact regression is augmented by the factors F in ad-
dition to the factors ∆F so ten factors in total. The detailed study of this augmented
model is left for future research. In the Figure, the five factor model is refered as the
semimartingale model and the ten factor model is refered as the information based
model.

6.3. Application to volatility estimation

Using the same common IV estimator as in the Monte Carlo exercise and described
in Appendix B, I quantify in Table 6 the improvement of the LBE in terms of precision
measured by the confidence interval width. In order to highlight the difference between
c
c and IV , I also report the overlapping of their confidence intervals. The
LBEIV
precision of the LBE estimators is better than the usual estimators expect for the
MLE estimator. Moreover, the confidence interval overlapping is about 50% using
MLE and QMLE but lower for TS which has the least overlap, but also MS, KER
and PRE. Therefore LBEIV c
c and IV are statistically different objects.

To compare OLS and INST, the Roll model using INST achieves higher confidence
interval precision than if OLS is used. But this is not the case for the five factor model
where OLS is more precise than INST. Except for the TS estimator, the confidence
intervals overlap is higher for INST than for OLS.
Finally, across models, the five factor model achieves the smallest confidence interval
width as compared to the Roll and Glosten and Harris models if OLS is used. This is

25
not the case if INST is used instead of OLS. Moreover adding an extra unrelated factor
to the five factor model does not deteriorate the estimation quality of the volatility
compared to the performance of the five factor model.

26
 4000 0.5

0.45
3500

Average ACF − Trade direction indicator

Average RV − Trade direction indicator

0.4
3000
0.35
2500
0.3

2000 0.25

0.2
1500
0.15
1000
0.1
500
0.05

0 0
0 50 100 150 200 0 5 10 15 20 25 30 35 40
Inverse sampling frequency Autocorrelation lag

10000 0.5

9000 0.45

8000 0.4

Average ACF − Trade volume

Average RV − Trade volume

7000 0.35

6000 0.3

5000 0.25

4000 0.2

3000 0.15

2000 0.1

1000 0.05

0 0
0 50 100 150 200 0 5 10 15 20 25 30 35 40
Inverse sampling frequency Autocorrelation lag

−7 −4
x 10 x 10
1.4

1.2 5
Average ACF − Bid−ask spread
Average RV − Bid−ask spread

4
1

3
0.8
2
0.6
1

0.4
0

0.2 −1

0 −2
0 50 100 150 200 0 5 10 15 20 25 30 35 40
Inverse sampling frequency Autocorrelation lag

14000 0.5

0.45
12000
0.4

10000
Average ACF − Ask depth
Average RV − Ask depth

0.35

0.3
8000
0.25
6000
0.2

4000 0.15

0.1
2000
0.05

0 0
0 50 100 150 200 0 5 10 15 20 25 30 35 40
Inverse sampling frequency Autocorrelation lag

14000 0.5

0.45
12000
0.4

10000
Average ACF − Bid depth

0.35
Average RV − Bid depth

0.3
8000
0.25
6000
0.2

4000 0.15

0.1
2000
0.05

0 0
0 50 100 150 200 0 5 10 15 20 25 30 35 40
Inverse sampling frequency Autocorrelation lag

Figure 1: The signature plots (left) and the autocorrelation functions (right) for the
liquidity variables.
27
 −3 −4
x 10 x 10
2.5 4
95% confidence interval lower bound RC(1) using original returns
Trade direction indicator coefficient 3 RC(1) using INST adjusted returns
95% confidence interval upper bound RC(1) using OLS adjusted returns
2
2
Estimated coefficient

1
1.5

Daily RC(1)
0

1
−1

−2
0.5
−3

0 −4
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

−5
x 10
4 1
95% confidence interval lower bound
2 Signed trade volume coefficient 0.8
95% confidence interval upper bound

0 0.6

Correlation with frictions

Estimated coefficient

0.4
−2

0.2
−4

0
−6
−0.2
−8
−0.4
−10 Original returns
−0.6 Adjusted returns
−12
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec
2010 business days

0.5 0.5 Original returns

Adjusted returns
Noise−to−signal ratio
Estimated coefficient

0 0.4

−0.5 0.3

−1 0.2
95% confidence interval lower bound
Signed spread coefficient
−1.5 0.1
95% confidence interval upper bound

−2 0
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec

−4 −4
x 10 x 10
1.5
95% confidence interval lower bound Original returns
Signed ask depth coefficient Returns adjusted for real frictions
Average RV − Semimartingale−based model

1 95% confidence interval upper bound

Estimated coefficient

0.5 5

−0.5 4

−1

−1.5 3
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 0 50 100 150 200 250 300
Inverse sampling frequency

−4 −4
x 10 x 10
1.5
95% confidence interval lower bound Original returns
Signed bid depth coefficient Returns adjusted for real frictions
95% confidence interval upper bound
Average RV − Information−based model

1
Estimated coefficient

0.5 5

−0.5 4

−1

−1.5 3
Jan Feb Mar Apr May Jun Jul Aug Sep Oct Nov Dec 0 50 100 150 200 250 300
Inverse sampling frequency

Figure 2: The estimated coefficients with confidence intervals for the five factor model
(left). On the right: the first-order realized autocovariance, the return-noise correla-
tion, the noise-to-signal ratio, the signature plot for the five factor and the ten factor
model. 28
 6000 1500 450
Original return Adjusted for real frictions return Real frictions
Histogram − Semimartingale−based model 400

Histogram − Semimartingale−based model

5000
350

4000 1000 300

Histogram

250
3000
200

2000 500 150

100
1000
50

0 0 0
−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 −1 −0.5 0 0.5 1
−3 −3 −3
x 10 x 10 x 10

1200 500 100

Original return Adjusted for real frictions return Real frictions
450 90
Histogram − Semimartingale−based model

Histogram − Semimartingale−based model

1000
400 80

350 70
800
300 60

600 250 50

200 40
400
150 30

100 20
200
50 10

0 0 0
−6 −4 −2 0 2 4 6 8 10 −6 −4 −2 0 2 4 6 8 10 −1 −0.5 0 0.5 1
−3 −3 −3
x 10 x 10 x 10

−4 −4 −3
x 10 x 10 x 10
Average ACF adjusted return − Semimartingale−based model

5 5 3
Average ACF real frictions − Semimartingale−based model

2.5
0 0
Average ACF − Original return

2
−5 −5

1.5

−10 −10
1

−15 −15
0.5

−20 −20 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
Autocorrelation lag Autocorrelation lag Autocorrelation lag

Figure 3: Histograms for the returns and fitted noise on Monday September 27, 2010
(top), on Tuesday August 24, 2010 (middle) and the autocorrelation function for
returns and for fitted noise (bottom).

29
 Original Roll Glosten and Harris Five factors Extra factor
Estimators IV LBEIV LBEIV LBEIV LBEIV
OLS INST OLS INST OLS INST OLS INST
c c c c
c

TS: Estimate ×104 1.3101 3.2828 3.2574 3.2727 3.0958 3.2591 3.2982 3.2590 3.2992
Width CI ×105 5.8663 5.5206 5.4538 5.5168 5.4410 5.5111 5.5548 5.5108 5.5566
Overlap CI in % 100 7.64 5.67 7.64 6.93 7.64 7.55 7.64 7.71
MS: Estimate ×104 3.4796 3.6636 3.6361 3.6536 3.5969 3.6400 3.7049 3.6399 3.7074
Width CI ×105 8.2997 5.9408 5.8554 5.9370 5.9815 5.9313 5.9963 5.9310 5.9983
Overlap CI in % 100 42.45 46.68 41.88 45.36 40.89 41.30 40.88 41.14
KER: Estimate ×104 3.8489 3.7364 3.7111 3.7263 3.6937 3.7127 3.7815 3.7126 3.7839
Width CI ×105 8.2015 5.9207 5.8480 5.9169 5.9670 5.9112 5.9776 5.9109 5.9795
Overlap CI in % 100 49.23 54.03 48.76 53.45 48.28 51.96 48.28 52.02
3.7918 3.7105 3.7123 3.7005 3.6841 3.6869 3.7611 3.6868 3.7631

30
PRE: Estimate ×104
Width CI ×105 17.344 7.5594 7.4890 7.5556 8.1282 7.5499 7.7355 7.5497 7.7370
Overlap CI in % 100 40.30 41.57 40.02 47.02 39.60 41.33 39.60 41.33
MLE: Estimate ×104 3.8298 3.7606 3.7174 3.7506 3.7048 3.7369 3.8059 3.7368 3.8086
Width CI ×105 4.8538 5.3564 5.2519 5.3526 5.1867 5.3469 5.3699 5.3467 5.3727
Overlap CI in % 100 53.36 59.14 52.85 56.41 52.45 56.87 52.45 56.79
QMLE: Estimate ×104 3.8298 3.7606 3.7174 3.7506 3.7048 3.7369 3.8059 3.7368 3.8086
Width CI ×105 5.4584 5.4037 5.3395 5.3999 5.2934 5.3942 5.4272 5.3939 5.4293
Overlap CI in % 100 52.02 58.80 51.50 55.53 51.13 55.38 51.13 55.34

Table 6: IV estimation results. CI denotes the confidence interval. Overlap CI is the overlapping between the CI of IV
the CI of LBEIV
one factor which is q. The Glosten and Harris model adds the factor qv to the Roll model. For the five factor model, the
c and
c . The expressions of the volatility estimators are provided in Appendix B. The Roll noise model contains

factors (q, qv, qs, da , db ) are used. And the extra factor model adds an unrelated iid series to the five factor model.
 7. CONCLUSION

In light of the market microstructure literature that provides economic drivers for
trading frictions or noise, I propose a semiparametric price model. Thus, by exploiting
a much bigger set of available trade and quote data, I estimate the frictionless price.
I derive a new volatility measure using the estimated frictionless price. Compared to
common robust-to-noise volatility estimators, this new volatility estimator does not
rely on the absence of endogeneity for the noise, and allows by construction for het-
eroskedastic and autocorrelated noise. Moreover, if the noise is completely removed
by the liquidity-cost variables considered, then the new volatility estimator is as ac-
curate as if the frictionless price were observed.
There are many possible extensions to this work. First, by decomposing the per-
manent component of the return into a public information-based part and a pri-
vate information-based part, one could estimate each part. This decomposition is
promising for future research. In line with Evans and Lyons (2002) and Andersen
et al. (2003b), analysing the high frequency micro decomposition into a macro con-
text would help us understand the role of public versus private information. Second,
adding jumps to the frictionless return dynamics would make the analysis more real-
istic. Finally, generalizing the parametric price impact regression model of this paper
to a nonparametric model is theoretically challenging.

31
 APPENDIX A: PROOFS OF RESULTS
Proof of Theorem 1

By substituting the return expression into the definition of βb given in (9), I obtain

0 0
βb − β = (X X)−1 X r − β
0 0
= (X X)−1 X (r∗ + Xβ + ∆ξ) − β
0 0 0 0 (A.1)
= (X X)−1 X r∗ + (X X)−1 X ∆ξ
0
!−1 0 0
!−1 0
XX X r∗ XX X ∆ξ
= + .
n n n n

• Consistency:

(i) If V ar[ξt ] = 0, then equation (A.1) becomes

0
!−1 0
XX X r∗
βb − β = + oP (1). (A.2)
n n

0
XX
The non negligible term of (A.2) is the product of n - which has a finite limit, Ω,
0
according to Assumption A - and the vector X r∗ , given by

 
Pn (1)
∗
n  i=1 Xiδ riδ 
0
X  .. 
Xr = ∗ ∗
Xiδ riδ =
 . .
 (A.3)
i=1  P 
n (M ) ∗
i=1 Xiδ riδ

0
I apply the Cauchy-Schwartz inequality for each element of the vector X r∗ , for

m=1..M,
Pn (m) ∗
!2 Pn (m)2
! n
!
X
i=1 Xiδ riδ 1 i=1 Xiδ ∗2
≤ riδ . (A.4)
n n n
i=1
Pn (m)2
Xiδ
Using Assumption A, → Ω(m, m), where Ω(m, m) is the mth diagonal
i=1
N
P P
element of the matrix Ω. On the other hand, the realized variance ni=1 riδ
∗2 −
→ IV .
0
X r∗ P P
Therefore n − → 0, which implies along with (A.2) and Assumption A that βb −
→ β.

32
 (ii) If V ar[ξt ] 6= 0, I show that both terms in (A.1) converge to zero. For the first
 0 −1 0 ∗
term, I use the consistency result (i) demonstrated above; XnX Xr
n = op (1).
 0 −1 0 0
X ∆ξ P
For the second term - XnX n - I need to show that XN∆ξ − → 0 to obtain
0
XX P
the consistency result since n −
→ Ω according to Assumption A. By applying the
Pn m ∆ξ
Xiδ iδ
Law of Large numbers to each element i=1
n of the sample mean, I obtain the
0
X ∆ξ P
outcome n −
→ E[Xt ∆ξt ]. The limit is zero since E[Xt ∆ξt ] = E[Xt ]E[∆ξt ] = 0

using Assumption 3 (i) and (ii).

• Asymptotic normality:

In both cases (i) and (ii), the regression residual has a normal distribution because

of the normality of r∗ in Assumption 1 and the normality of ξ in Assumption 3 (ii).

Therefore,

βb ∼ N (E[β],
b V ar[β]).
b (A.5)

Now, I turn to the derivation of the asymptotic variance for each case.

(i) If V ar[ξt ] = 0, (A.1) yields

0 0
V ar[n(βb − β)] = V ar[n(X X)−1 X r∗ ]
0 0
X X −1 0 X X −1 0
=( ) V ar[X r∗ ](( ) )
n n
(A.6)
X X −1  0  X0 X
0
∗
=( ) X V ar[r |X]X ( )−1
| n{z } | {z } | n{z }
P
→ Ω−1 Ω∗ Ω−1 ,
−

which follows from Assumptions A and B.

(ii) If V ar[ξt ] 6= 0, (A.1) gives

√ 1 0 0 √ 0 0
V ar[ n(βb − β)] = V ar[ √ n(X X)−1 X r∗ + n(X X)−1 X ∆ξ]
n
(A.7)
1 0 0 √ 0 0
= V ar[n(X X)−1 X r∗ ] + V ar[ n(X X)−1 X ∆ξ],
n

using Assumption 3 (i). The first term in (A.7) vanishes as a result of (A.6). For the

33
 second term,

0 0 0
√ 0 0 X X −1 V ar[X ∆ξ] X X −1 0
V ar[ n(X X)−1 X ∆ξ] = ( ) (( ) )
n n n
0 0 0
X X −1 X V ar[∆ξ]X X X −1
=( ) ( ) (A.8)
| n{z } | n
{z } | n{z }
P
→ Ω−1 SΩ−1 ,
−

following from Assumptions A and C.

Proofs of Theorem 2 and 3

Parts (i) and (ii) of Theorem 2 and 3 follow directly from Theorem 1 page
 1488 of Barndorff-

p∗ , δ) 
 RV (b
Nielsen et al. (2008a) to derive the asymptotic bivariate distribution of   and
p∗ , 2δ)
RV (b
the Delta method to derive the asymptotic distribution of the ratio Sn :

(i) Suppose Assumptions 1, 2, A and B hold.

√ L 2IQ
Under H0 , n(Sn − 1) −→ MN (0, IV 2 ).

(ii) Suppose Assumptions 1-3, A and C hold.

√ L 3V ar[ξt2 ]+7V ar[ξt ]2
Under H1 , n(Sn − 2) −→ N (0, 4 V ar[ξt ]2
).

34
Concerning part (iii) of Theorem 2, the power of the test denoted P is given by:

√
n(Sn − 1)
P = P rob(reject H0 |H1 ) = P rob( r > c1−α |H1 )
c
2IQ
c2
IV
s
c
2IQ
= P rob((Sn − 1) > c1−α |H1 )
c2
nIV
s
c
2IQ
= P rob((Sn − 2) > c1−α − 1|H1 )
nsc2
IV
√ c
2IQ √
= P rob( n(Sn − 2) > 2 c1−α − n|H1 )
c
IV r
c
2IQ √
√ c − n
n(Sn − 2) c 2 1−α
IV
= P rob( s >s |H1 )
2 2
3V\
ar[ξt2 ]+7V\ ar[ξt ] 3V\
ar[ξt2 ]+7V\ ar[ξt ]
4 2 4 2
V\ ar[ξt ] V\ ar[ξt ]
 r 
2 c
IQ √
 c − n 
 c 2 1−α
IV 

= 1 − Φ s 

 3V\ ar[ξt ] 
ar[ξt2 ]+7V\
2

4 2
V\
ar[ξt ]

c V\
c , IQ,
= P ower(n, c1−α , IV ar[ξt ], V\
ar[ξt2 ]) → 1,

where Φ(.) is the cumulative distribution function for the standard normal distribution.

For the Power in Theorem 3 (iii), it is written as:

35
 v
u 2
√ u 3V\ ar[ξt2 ] + 7V\ ar[ξt ]
P = P rob(reject H0 |H1 ) = P rob( n(Sn − 2) < − 4 t c1−α |H1 )
2
V\ ar[ξt ]
v
u 2
u 3V\ ar[ξt2 ] + 7V\ ar[ξt ]
t
= P rob((Sn − 1) < − 4 c1−α + 1|H1 )
2
nV\ ar[ξt ]
v
u 2
√ u 3V\ ar[ξt2 ] + 7V\ ar[ξt ] √
= P rob( n(Sn − 1) < − 4 t c1−α + n|H1 )
2
V\ ar[ξt ]
s
2
3V\
ar[ξt2 ]+7V\ ar[ξt ] √
√ − 4 2 c1−α + n
n(Sn − 1) \
V ar[ξt ]
= P rob( r < r |H1 )
c
2IQ c
2IQ
2 2
c
IV c
IV
 s 
2
3V\
ar[ξt2 ]+7V\ ar[ξt ] √
− 4 2 c1−α + n
 V\ ar[ξt ] 
= Φ
 r 

 c
2IQ 
2
c
IV

c V\
c , IQ,
= P ower(n, c1−α , IV ar[ξt ], V\
ar[ξt2 ]) → 1.

Proof of Theorem 4

(i) Assume that V ar[ξt ] = 0.

√
p∗ ) − IV )
n(RV (b
n n
√ √ X 0 √ X 0 (A.9)
= n (RV (p∗ ) − IV ) + n (Xi (β − β̂))2 + 2 n ri∗ Xi (β − β̂) .
| {z }
| i=1 {z } | i=1
{z }

√ st
The first term of (A.9) is the usual term: p∗ ) − IV ) −
n(RV (b → MN (0, 2 IQ) as shown in

Barndorff-Nielsen and Shephard (2002). In the following, I show that the second and third

terms are negligible.

36
For the second term of (A.9),

n n
√ X 0 √ X 0 0
n (Xi (β − β̂))2 = n (β − β̂) Xi Xi (β − β̂)
i=1 i=1
Pn 0
√ 0
i=1 Xi Xi
= n n (β − β̂) ( ) (β − β̂) (A.10)
| {z } | n
{z } | {z }
oP (1/n) oP (1/n)
OP (1)

= oP (1),

using Assumption A and Theorem 1 (i).

For the last term of (A.9),

n n
√ X 0 1 X ∗ 0
2 n ri∗ Xi (β − β̂) = 2 √ ri Xi N (β − β̂) = oP (1), (A.11)
n | {z }
i=1 i=1
| {z } oP (1)
OP (1)

using the Cauchy-Schwartz inequality as in (A.4) and Theorem 1 (i).

(ii) In this proof, I use the original notation for indices Yiδ instead of the simplified notation

Yiδ for a given process Y.

The two time scales estimator of Zhang et al. (2005) applied to the adjusted price pb∗ is given

by
n
p∗ ) = RV avg (b
T S(b p∗ ) − p∗ ),
RV (b (A.12)
n

where
K
1 X
p∗ ) =
RV avg (b p∗ ),
RV (k) (b (A.13)
K
k=1

and

nk
X  
(k) ∗ ∗ ∗ 2 n−k+2
RV (b
p )= pk−1+K i − pbk−1+K i−1 ) , nk = F loor
(b , (A.14)
n n K
i=1

K
1 X
n= nk . (A.15)
K
k=1

37
 p∗ ) − IV could be decomposed as
The bias T S(b

p∗ ) − IV = (T S(b
T S(b p∗ ) − RV avg (p∗ )) + (RV avg (p∗ ) − IV ) (A.16)
| {z } | {z }
error due to the noise error due to discretization

My objective is to show

1. for the bias due to noise:

r
K L
p∗ ) − RV avg (p∗ )) −
(T S(b → N (0, 8(E[ξ 2 ])2 ), (A.17)
n

2. and for the bias due to discretization:

r
n st 4
(RV avg (p∗ ) − IV ) −
→ MN (0, IQ). (A.18)
K 3

Combining the two sources of errors (A.17) and (A.18) so that each will be present at the
K n
limit leads to n proportional to K, as in Zhang et al. (2005). Taking

K = cn2/3 ,
(A.19)
−1 1/3
n=c n ,

implies that neither source of error will dominate at the limit. The constant c is a tuning
8
parameter which could be optimally determined. Let Γξ = c2
E[ξ 2 ]2 + c 43 IQ, the limiting
st
p∗ ) − IV ) −
distribution is given by n1/6 (T S(b → N (0, Γξ ). For the optimal choice of c, mini-
 −1/3
1
mizing the asymptotic variance Γξ leads to c = 12E[ξ 2 ]2 IQ .

Let’s start by showing (A.17). The error due to the noise of (A.16) is written as

r r
K ∗ avg ∗ K n
p ) − RV (p )) =
(T S(b p∗ ) − RV (b
(RV avg (b p∗ ) − RV avg (p∗ ))
n n n
r
K n
= p∗ ) − RV avg (p∗ ) − 2nE[ξt2 ] + 2nE[ξt2 ] − RV (b
(RV avg (b p∗ )) (A.20)
n n
r  
K √ RV (bp∗ )
= p∗ ) − RV avg (p∗ ) − 2nE[ξt2 ]) − 2 Kn
(RV avg (b − E[ξt2 ] .
| n {z } | 2n
{z }

38
The following Lemma is needed, which is in the same spirit as Lemma A.2 in Zhang et al.

(2005).

Lemma 1 Under assumptions 1, 2 and 3,

p∗ ) = RV (ξ) + Op (1),
(i) RV (b
 
1
p∗ ) − RV avg (p∗ ) = RV avg (ξ) + Op
(ii) RV avg (b √ .
K

39
 Proof of Lemma 1:

(i) The realized variance of the adjusted price is given by

n
X
p∗ ) =
RV (b rbi∗2
i=1
n 
X 2
0
= b + ∆ξiδ
ri∗ + Xi (β − β)
i=1
n n  2 n (A.21)
X X 0
X
= ri∗2 + b
Xi (β − β) + (∆ξiδ ) 2

i=1 i=1 i=1

Xn n
X n
X
0 0
+2 ri∗ ∆ξiδ + 2 b +2
ri∗ Xi (β − β) b
Xi (β − β)∆ξiδ .
i=1 i=1 i=1

Pn ∗2
Pn 2
Pn ∗
Since i=1 ri = OP (1), i=1 (∆ξiδ ) = RV (ξ) and 2 i=1 ri ∆ξiδ = oP (1) using Assump-

tion 3 then, (A.21) becomes

p∗ ) = RV (ξ) + OP (1)
RV (b
n 
X 0
2 n
X 0
n
X 0
(A.22)
+ b
Xi (β − β) +2 b +2
ri∗ Xi (β − β) b
Xi (β − β)∆ξiδ .
i=1 i=1 i=1

In the following, I show that each of the last three terms of (A.22) is oP (1) which implies

the order OP (1). First:

n 
!
X 0
2 n
X 0 0
b
Xi (β − β) = trace b Xi X (β − β)
(β − β) b
i
i=1 i=1
  0 0
b X X(β − β)
= trace (β − β) b
 
 0
!  (A.23)
√ XX √ 
= trace  b0 b
| n(β{z− β)} n |
n(β − β)
{z }
 
oP (1) | {z } oP (1)
OP (1)

= oP (1),

using Assumption A and Theorem 1 (ii).

40
Second: !
n
X n
0
b = 1 X ∗ 0 √ b = oP (1),
ri∗ Xi (β − β) √ ri Xi n(β − β) (A.24)
n | {z }
i=1 i=1
| {z } oP (1)
OP (1)

using (A.11) and Theorem 1 (ii).

Third: !
n
X n
X
0 0
b
Xi (β − β)∆ξiδ = Xi ∆ξiδ b = oP (1).
(β − β) (A.25)
| {z }
i=1 i=1 √
| {z } oP (1/ n)
√
OP ( n)

using Assumption C and Theorem 1 (ii).

 
p∗ ) − RV avg (p∗ ) = RV avg (ξ) + Op
(ii) I need to show that RV avg (b √1 .
K

I define the average realized covariance of two given processes Ut and Vt by RC avg (U, V ) as

an extension of the average realized variance - RV avg (.) - the notion introduced earlier:

K
1 X
RC avg (U, V ) = RC (k) (U, V ), (A.26)
K
k=1

where

nk
X
(k)
RV (U, V ) = (Uk−1+K i − Uk−1+K i−1 )(Vk−1+K i − Vk−1+K i−1 ). (A.27)
n n n n
i=1

The average realized variance of the adjusted price could be written as

b + ξ)
p∗ ) = RV avg (p∗ + F(β − β)
RV avg (b

b +RV avg (ξ)

= RV avg (p∗ ) + RV avg (F(β − β))
| {z }
(a)
(A.28)

b ξ) + 2RC avg (p∗ , F(β − β))

+ 2RC avg (F(β − β), b + 2RC avg (p∗ , ξ),
| {z } | {z } | {z }
(b) (c) (d)

as a result of the definitions in (B.2) and (B.3) as well as the average realized covariance in

(A.26) and (A.27).

 
Using (A.28), I need to show that each of the terms (a), (b), (c) and (d) is Op √1 to be
K

41
able to prove Lemma 1 (ii).

K
X
b = 1
(a) = RV avg (F(β − β)) b
RV (k) (F(β − β))
K
k=1
nk h
K X
X i2
1 0 0
b
= (Fk−1+K i − Fk−1+K i−1 )(β − β)
K n n
k=1 i=1
XK X nk
1 b 0 (F 0 0
b
= (β − β) k−1+K i − Fk−1+K i−1 )(Fk−1+K i − Fk−1+K i−1 )(β − β)
K n n n n
k=1 i=1

1 XK X nk  (A.29)
= b 0 (F
trace (β − β) i − F i−1 )(F
0
i − F
0
i−1 )(β − b
β)
k−1+K k−1+K k−1+K k−1+K
K n n n n
k=1 i=1
 
nk
K X
X
1  
= trace  (β − b
β)(β − b 0 (F
β) i − F i−1 )(F
0
i − F
0
i−1 )

K | {z }| k−1+K n
k−1+K n k−1+K n k−1+K n 
k=1 i=1 {z }
oP (1/n)
OP (1)
nk
K X K
! √
1 X 1 X n
= oP (1/n) = nk oP (1/n) = n oP (1/n) = oP ( ),
K K n
k=1 i=1 k=1

using the Normality limit distribution of β − βb as well as the rate of convergence of Theorem

1 (ii). I also use the assumption that the fourth moment of Ft is bounded and the definition
√  
of n in (B.4). Finally, as a result of condition (A.19), (a) = oP ( nn ) < Op √1K .

K
avg b 1 X b ξ)
(b) = RC (F(β − β), ξ) = RC (k) (F(β − β),
K
k=1
nk
K X
X
1 0 0
b
= (ξk−1+K i − ξk−1+K i−1 )(Fk−1+K i − Fk−1+K i−1 ) (β − β)
K n n n n | {z } (A.30)
k=1 i=1 | {z } √
oP (1/ n)
OP (1)
K nk K
! r !
1 XX √ 1 X √ √ n
= oP (1/ n) = nk oP (1/ n) = n oP (1/ n) = oP ,
K K n
k=1 i=1 k=1

using Assumption 2 (ii), Theorem 1 (ii), Assumption 3 and the definition of n in (B.4).

42
 q     
n √1 √1
Finally, as a result of condition (A.19), (b) = oP n = op K
< Op K
.

K
b = 1 X b
(c) = RC avg (p∗ , F(β − β)) RC (k) (p∗ , F(β − β))
K
k=1
K nk
1 XX 0 0
b
= (p∗k−1+K i − p∗k−1+K i−1 )(Fk−1+K i − Fk−1+K i−1 ) (β − β)
K | n n
{z n n
} | {z }
k=1 i=1 √
oP (1/ n) (A.31)
<OP (1)
K
!  
1 X √ √ 1
< nk oP (1/ n) = n oP (1/ n) = op √
K K
k=1
 
1
< Op √ ,
K

using Assumptions 1, 2 and Theorem 1 (ii). The derivation of the asymptotic order in (A.31)

results from the definition of n in (B.4) and the condition (A.19).

 
(d) = RC avg (p∗ , ξ) = Op √1K is a direct result from Lemma A.2 part (b) relative to the

multiple grid case, page 1408 from Zhang et al. (2005); which completes the proof of Lemma

1.

Now, I turn to the proof of (A.17) using Lemma 1. Being back to the two components

of (A.20), I derive the joint limit distribution of these two components in order to show

(A.17). In expression (A.15) of their Appendix, Zhang et al. (2005) show that for an i.i.d.

and exogenous noise ξ, the joint distribution of RV (ξ) and RV avg (ξ) is given by

     
1  RV (ξ) − 2N E[ξ 2 ]
 L  0   4V 4E[ξ 4 ]
 ar[ξ 2 ]
√  −→ N   ,   . (A.32)
n RV avg (ξ)K − 2N KE[ξ 2 ] 0 4V ar[ξ 2 ] 4E[ξ 4 ]

Using Lemma 1, the limiting result (A.32) becomes

 
√  RV (bp∗ ) 2]

 n 2n − E[ξ t 
 q

K avg (b ∗ ) − RV avg (p∗ ) − 2nE[ξ 2 ]
n RV p
    (A.33)
4 2
L  0   E[ξ ] 2V ar[ξ ] 
−
→ N   ,   ,
0 2V ar[ξ 2 ] 4E[ξ 4 ]

43
which yields the limit distribution of (A.17).

To complete the proof, it remains to show (A.18). This result is derived in Section 3.4 of

Zhang et al. (2005).

APPENDIX B: VOLATILITY ESTIMATORS

For an observed price series L whether it is p or pb∗ , i.e. L = p∗ +ε(L) where ε(L) could be the
c and IQ
original noise or the decontaminated noise. I denote IV c estimators of the integrated

volatility and quarticity, respectively. V ar[ε(L) ], V ar[ε(L)2 ] and E[ε(L)4 ] are respectively
\
estimated by V ar[ε \
(L) ], V ar[ε \
(L)2 ] and E[ε (L)4 ].

The IV estimators considered as well their respective asymptotic variances are the following.

1.

n
T S(L) = RV avg (L) − RV (L),
n 2
 (B.1)
−1/3 8 \ (L)
4c
V ar[T S(L)] = n V ar[ε ] + c IQ .
c2 3

where the average realized variance estimator is given by

K
1 X
RV avg (L) = RV (k) (L), (B.2)
K
k=1

and

nk
X  
(k) 2 n−k+2
RV (L) = (Lk−1+Kiδ − Lk−1+K(i−1)δ ) , nk = F loor , (B.3)
K
i=1

K
1 X
n= nk . (B.4)
K
k=1

 −1/3
c
IQ
where K = f loor(cn2/3 ) and c = 2 .
\
12V ar[ε(L) ]

44
2.

M
X n−K
αi Xi
M S(L) = (Lkδ+Ki − Lkδ )
Ki
i=1 k=0 (B.5)
48 \ 2 52 c 12 \
V ar[M S(L)] = n−1/2 ( 3 V ar[ε (L) ] + θIQ + V ar[ε(L)2 ]M S(L)),
θ 35 5θ

√ −0.5−1/2M
where M = F loor(θ n) and αi = 12 Mi 2 i/M1−1/M 2 .

3.

n
X H
X   (X
n n
X
)
h−1
KER(L) = ∆2 Yiδ + K ∆Ykδ ∆Y(k−h)δ + ∆Ykδ ∆Y(k+h)δ ,
H
i=1 h=1 k=1 k=1 (B.6)
q
V ar[KER(L)] = n −1/2 \
8.29 V ar[ε c 3/4 ,
(L) ]IQ

 q 
where K(x) = sin( π2 (1 − x)2 )2 \ (L)
and H = F loor 5.74 nV ar[ε ]/KER(L) .

4.
 2
√ N −k X
k  
δ X j ψ1 δ
P RE(L) = ∗ φ ∆Li+j − ∗2 RV (L),
θ ψ2  k  2θ ψ2
i=0 j=1
 4
NX−k X k   
4Φ22 j
V ar[P RE(L)] = n−1/2 ∗ 4 φ ∆Li+j
3θ ψ2 i=0  j=1 k 
 2 (B.7)
  NX−2k X k    i+2k−1
X
4δ Φ12 Φ22 ψ1 j
+ n−1/2 ∗3 3 − 4 φ ∆L i+j ∆Ll
θ ψ2 ψ2  k 
i=0 j=1 l=i+k
 X
2 N −2
δ Φ11 Φ12 ψ1 Φ22 ψ1
+ n−1/2 −2 + ∆L2i ∆L2i+2 ,
θ∗3 ψ22 ψ23 ψ24 i=0

where ∆Lj = Ljδ − L(j−1)δ , √k = θ∗ + O(N −1/4 ) for some θ > 0, and φ(x) =
N

min(x, 1 − x), θ∗ = 0.5, ψ1 = 1, ψ2 = 1/12, Φ11 = 1/6, Φ12 = 1/96 and Φ22 =

151/80640.

45
5.

a2M LE ) = ArgM ax(σ2 ,a2 ) l(σ 2 , a2 ),

(M LE(L), b
(B.8)
V ar[M LE(L)] = 8n−1/2 b
aM LE M LE(L)3/2

 
1 n 1 0
l(σ 2 , a2 ) = − log(det(M)) − log(2π) − ∆L M−1 ∆L
2 2 2

where  
 σ2δ + 2a2 −a2 0 ... 0 
 .. .. 
 −a2 σ2δ + 2a2 −a2 . . 
 
 
 .. 
M= 0 −a2 σ 2 δ + 2a2 . 0 .
 
 .. .. .. .. 
 . . . . −a2 
 
 
0 ... 0 −a2 σ 2 δ + 2a2

6.

QM LE(L) = M LE(L),
! (B.9)
5b
a c
IQ
V ar[QM LE(L)] = n −1/2
p M LE aM LE M LE(L)3/2
+ 3b
M LE(L)

7.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEts (p) = (B.10)

 T S(b
p∗ ), if V ar[ξt ] 6= 0.

8.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEms (p) = (B.11)

 M S(b
p∗ ), if V ar[ξt ] 6= 0.

46
 9.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEker (p) = (B.12)

 KER(b
p∗ ), if V ar[ξt ] 6= 0.

10.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEpre (p) = (B.13)

 P RE(b
p∗ ), if V ar[ξt ] 6= 0.

11.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEmle (p) = (B.14)

 M LE(b
p∗ ), if V ar[ξt ] 6= 0.

12.


 RV (b
p∗ ), if V ar[ξt ] = 0;
LBEqmle (p) = (B.15)

 QM LE(b
p∗ ), if V ar[ξt ] 6= 0.

Acknowledgements: The author is grateful to Nour Meddahi, Sı́lvia Gonçalves and

Ilze Kalnina for helpful discussions and feedback. She thanks Bruno Feunou, Tim
Bollerslev, Andrew Patton, and all participants at the Duke Financial Econometrics
Lunch Group; Torben Andersen and all participants at the NBER-NSF time-series
conference 2011; Rim Lahmandi Ayed, Habib Zitouna, Leila Baghdadi and all par-
ticipants at the biweekly MASE Seminar. She appreciates the comments of Anna
Mikusheva, Mark Podolskij, Asger Lunde, Kim Christensen and Tatevik Sekhposyan.
She also thanks Jean-Sébastien Fontaine, Antonio Diez de los Rios, Gregory H. Bauer
and Federico M. Bandi. This paper grew out of her Ph.D. thesis at University of Mon-
treal and was previously circulated under the title “Volatility and Liquidity Costs”,

47
Bank of Canada Working Paper Series.

48
 References
Y. Aı̈t-Sahalia and J. Yu. High-frequency market microstructure noise estimates and liquidity mea-
sures. Annals of Applied Statistics, 3:422–457, 2009.
Y. Aı̈t-Sahalia, P. A. Mykland, and L. Zhang. How often to sample a continuous-time process in the
presence of market microstructure noise. The Review of Financial Studies, 18(2):351–416, 2005.
Y. Aı̈t-Sahalia, P. A. Mykland, and L. Zhang. Ultra high-frequency volatility estimation with de-
pendent microstructure noise. Journal of Econometrics, 160:160–175, 2011.
D. G. Aldous and G. K. Eagleson. On mixing and stability of limit theorems. The Annals of
Probability, 6:325–331, 1978.
T. G. Andersen, T. Bollerslev, F. X. Diebold, and P. Labys. Modeling and forecasting realized
volatility. Econometrica, 71:579–625, 2003a.
T. G. Andersen, T. Bollerslev, F. X. Diebold, and C. Vega. Micro effects of macro announcements:
Real-time price discovery in foreign exchange. American Economic Review, 93(1):38–62, 2003b.
O. E. Barndorff-Nielsen and N. Shephard. Econometric analysis of realized volatility and its use in
estimating stochastic volatility models. Journal of the Royal Statistical Society, Series B, 64(Part
2):253–280, 2002.
O. E. Barndorff-Nielsen, P. R. Hansen, A. Lunde, and N. Shephard. Designing realized kernels
to measure the ex-post variation of equity prices in the presence of noise. Econometrica, 76:
1481–1536, 2008a.
O. E. Barndorff-Nielsen, P. R. Hansen, A. Lunde, and N. Shephard. Realised kernels in practice:
Trades and quotes. Econometrics Journal, 4:1–32, 2008b.
O. E. Barndorff-Nielsen, P. R. Hansen, A. Lunde, and N. Shephard. Subsampling realised kernels.
The Journal of Econometrics, 160(1):204–219, 2011.
F. X. Diebold and G. H. Strasser. On the correlation structure of microstructure noise: A financial
economic approach. The Review of Economic Studies, 80(4):1304–1337, 2013.
M.D. Evans and R.K. Lyons. Order flow and exchange rate dynamics. The Journal of Political
Economy, 110(1):170–181, 2002.
L. R. Glosten and L. E. Harris. Estimating the components of the bid/ask spread. Journal of
Financial Economics, 21:123–142, 1988.
A. Gloter and J. Jacod. Diffusions with measurement errors. i. local asymptotic normality. ESAIM:
Probability and Statistics, 5:225–242, 2001.
S. Gonçalves and N. Meddahi. Bootstrapping realized volatility. Econometrica, 77(1):283–306, 2009.
P. R. Hansen and A. Lunde. Realized variance and market microstructure noise. Journal of Business
& Economic Statistics, 24:127–161, 2005.
J. Hasbrouck. Measuring the information content of stock trades. The Journal of Finance, 46:
179–207, 1991.
J. Hasbrouck. The dynamics of discrete bid and ask quotes. The Journal of Finance, 54:2109–2142,
1999.
J. Hasbrouck. Empirical market microstructure. New York: Oxford University Press, 2007.
N. Hautsch and M. Podolskij. Pre-averaging based estimation of quadratic variation in the presence
of noise and jumps: Theory, implementation, and empirical evidence. Journal of Business &
Economic Statistics, 31(2):165–183, 2013.
S. Heston. A closed-form solution for options with stochastic volatility with applications to bonds
and currency options. Review of Financial Studies, 6:327–343, 1993.
R.D. Huang and H.R. Stoll. The components of the bid-ask spread: A general approach. The Review
of Financial Studies, 10:995–1034, 1997.
J. Jacod, Y. Li, P. Mykland, M. Podolskij, and M. Vetter. Microstructure noise in the continuous
case: The pre-averaging approach. Stochastic Processes and their Applications, 119:2249–2276,
2009.
I. Kalnina and O. Linton. Estimating quadratic variation consistently in the presence of endogenous
and diurnal measurement error. The Journal of Econometrics, 147:47–59, 2008.

49
K. A. Kavajecz. A specialist’s quoted depth and the limit order book. The Journal of Finance, 54:
747–771, 1999.
S. Kinnebrock and M. Podolskij. A note on the central limit theorem for bipower variation of general
functions. Stochastic Processes and their Applications, 118:1056–1070, 2008.
J. Klotz. Markov chain clustering of births by sex. Proc. Sixth Berkeley Symp. Math. Statist. Prob.,
Univ. of California Press, 4:173–185, 1972.
S. N. Lahiri. On inconsistency of estimators based on spatial data under infill asymptotics. The
Indian Journal of Statistics, 58, Series A, Pt.3:403–417, 1996.
C. M. C. Lee and M. J. Ready. Inferring trade direction from intraday data. The Journal of Finance,
26:733–746, 1991.
Y. Li and P. A. Mykland. Inferring trade direction from intraday data. Bernoulli, 13(3):601–622,
2007.
P. A. Meyer. Intégrales stochastiques–1. Séminaire de probabilités de Strasbourg I, Lecture Notes in
Math., 39:72–94, 1967.
R. Roll. A simple implicit measure of the effective bid-ask spread in an efficient market. The Journal
of Finance, 39:1127–1139, 1984.
J. H. Stock. Asymptotic properties of least squares estimators of cointegrating vectors. Econometrica,
55:1035–1056, 1987.
H. R. Stoll. Friction. The Journal of Finance, 55:1479–1514, 2000.
D. Xiu. Quasi-maximum likelihood estimation of volatility with high frequency data. Journal of
Econometrics, 159:235–250, 2010.
H. Zhang and Zimmerman D. L. Towards reconciling two asymptotic frameworks in spacial statistics.
Biometrika, 92(4):921–936, 2005.
L. Zhang. Efficient estimation of stochastic volatility using noisy observations: a multi-scale ap-
proach. Bernoulli, 12(6):1019–1043, 2006.
L. Zhang, P. A. Mykland, and Y. Aı̈t-Sahalia. A tale of two time scales: Determining integrated
volatility with noisy high-frequency data. Journal of the American Statistical Association, Theory
and Methods, 100(472):1394–1411, 2005.

50