Intraday Volatility-Volume Joint Modeling and

Forecasting: A State-space Approach

Shengjie Xiu and Daniel P. Palomar
Department of Electronic and Computer Engineering
The Hong Kong University of Science and Technology
Clear Water Bay, Hong Kong
sxiu@[Link], palomar@[Link]

Abstract—Intraday volatility is a crucial indicator for short- a multiplicative component GARCH (mcsGARCH). However,
term price movements of financial assets, playing a pivotal role in mcsGARCH is impractical in practice due to the potential
risk management and other financial applications in the era of error accumulation in multistep estimation. Therefore, intraday
high-frequency algorithmic trading. However, existing methods
fail to capture complex intraday volatility patterns, leading to volatility modeling and forecasting remain under-explored
inadequate forecast capabilities. To address these shortcomings, research areas.
we propose a novel multiplicative component framework that In recent years, scholars have been exploring the role of
leverages the close relationship between volatility and trading trading volume in advancing the understanding of intraday
volume to analyze intraday volatility. We further introduce a volatility. The positive relationship between volatility and vol-
state-space approach to implement this framework for modeling
and forecasting. Empirical experiments demonstrate that our ume is widely acknowledged, with two established theoretical
proposed method outperforms traditional models by providing explanations. The sequential information arrival hypothesis
more comprehensive insights and achieving higher forecasting (SIAH) assumes that traders react to new information se-
accuracy for intraday volatility. quentially, thus, lagged values of volatility and volume can
Index Terms—Intraday volatility modeling and forecasting, predict current ones [12], [13]. The other explanation for the
volatility-volume relation, state-space model
positive correlation is the mixture of distributions hypothesis
(MDH), which posits that stock price and volume are jointly
I. I NTRODUCTION dependent on common information flow [14], [15]. While
In the last half-century, stock market volatility has been incorporating trading volume information may be beneficial,
extensively discussed. Volatility, usually measured by the how to effectively model and exploit the intraday volatility-
standard deviation of stock returns, is a crucial factor as volume relationship remains an open question.
it characterizes the level of risk associated with an asset. Our study contributes to the existing literature in the fol-
With the emergence of algorithmic trading, financial assets lowing ways:
are now subject to substantial price fluctuations within min- • We propose a new multiplicative component framework
utes. Therefore, intraday volatility has become increasingly for intraday volatility that accounts for the volatility-
important to academia and the financial industry. Accurate volume relationship.
intraday volatility modeling and forecasting assist traders in • We develop a state-space approach for intraday volatility
assessing their risk exposure and identifying profitable trading modeling and forecasting.
opportunities [1]. They are also fundamental in pricing options • We conduct extensive empirical experiments that demon-
and other derivatives [2], [3]. strate our proposed model’s superior financial interpreta-
In the literature, daily volatility modeling has been well tion and forecasting performance.
researched, with three classes of models receiving significant
attention: historical volatility [4], ARCH-GARCH class con- II. P RELIMINARIES AND M OTIVATION
ditional volatility (e.g. [5], [6]), and stochastic volatility (e.g.
A. Intraday Volatility
[7]). However, intraday volatility modeling and forecasting
pose greater challenges. Although ARCH class models can In this study, we use 15-minute intervals to model intraday
be applied directly to intraday volatility [8], [9], they have volatility, where each trading day is indexed by t ∈ {1, 2, . . . }
been criticized for their inability to capture complex structures, and each bin within a day is indexed by n ∈ {1, . . . , N }. We
such as intraday regularities. To address this limitation, [10] denote the asset price at the end of bin n of day t by Pt,n and
proposes filtering out the intraday seasonal pattern before compute the return in every 15 minutes as
applying GARCH models. Alternatively, [11] suggests rep- (
log (Pt,n /Pt,n−1 ) n ≥ 1
resenting intraday volatility as a product of daily, intraday rt,n = (1)
seasonal, and intraday dynamics components and developing log (Pt,1 /Pt−1,N ) n = 0.

This work was supported by the Hong Kong GRF 16207820 research grant. The R package intradayModel will support the proposed algorithm.

ISBN: 978-9-4645-9360-0 1395 EUSIPCO 2023

To measure intraday volatility, we use the sum of higher- Stock Exchange common stocks [17] and nine national mar-
frequency squared returns (realized volatility), expressed as kets [18]. Similarly, the positive correlation can be observed
v
u1/∆ in intraday data [19], [20], which can be identified in Figure
uX
2
1. The widely researched MDH hypothesis builds a bivariate
σt,n = t r̂t,n−1+j∆ , (2)
normal mixture model of daily price change ∆P and trading
j=1
volume V , expressed as
where ∆ ∈ (0, 1] is a possibly shorter horizon standardized √
by 15 minutes, for example, ∆ = 1/3 stands for 5-minute ∆P = σ1 IZ1
√ (5)
intervals. r̂t,n−1+j∆ is the higher-frequency return satisfying V = µ2 I + σ2 IZ2 ,
P1/∆
(1) and rt,n = j=1 r̂t,n−1+j∆ . As shown in the top panel of where Z1 and Z2 are N (0, 1) random variables. This model
Figure 1, the time series of intraday volatility exhibits volatility directly and positively links volatility and volume with the
clustering and persistence across different time scales, which information flow I.
are crucial for accurate forecasting. The MDH hypothesis, while appealing from an information
flow standpoint, oversimplifies the intraday relationship as one
−5 latent variable I and suffers two significant limitations: 1)
Volatility
Intraday

−6
−7
−8
It fails to fully capture the volatility and volume intraday
−9 dynamics, including unexpected movements [20]. 2) It can-
16 not account for the long-term variations in daily volatility
Intraday
Volume

15
14
and volume [21]. Information flow is associated only with
13 medium-term volatility persistence, highlighting the need for
0 25 50 75 100
Day
a more sophisticated modeling approach to incorporate volume
information in intraday volatility modeling.
Fig. 1. Intraday volatility and volume series of AMZN stock in log scale. III. P ROPOSED S TATE - SPACE A PPROACH
A. New Multiplicative Component Framework
B. Multiplicative Component GARCH
We propose a novel framework for intraday volatility mod-
Engle and Sokalska propose the mcsGARCH model for
eling that decomposes it into multiple components: seasonal,
intraday financial returns [11], which has the expression
daily (long-term), information flow (medium-term), unex-
rt,n = σ̂t,n εt,n , and εt,n ∼ N (0, 1) , (3) pected volatility (short-term), and noise components. Infor-
mation flow and unexpected volatility correspond to "good"
where the conditional variance is decomposed into three and "bad" intraday volatilities, respectively. "Good" volatility
components: is directional, persistent, relatively easy to anticipate, and
2
σ̂t,n = dailyt × seasonaln × intraday dynamicst,n . (4) accompanied by sufficiently high volume. "Bad" volatility is
erratic, relatively difficult to predict, and less associated with
The daily component is the systematic volatility variation volume [20]. Intraday volume is decomposed similarly but
across different trading days and adjusts the average level without the unexpected volatility component.
of intraday volatility. It can be obtained from the volatility Therefore, our framework is expressed as
forecasts such as the multifactor risk model or daily GARCH.
The seasonal component is the average volatility within the σt,n = dt × sn × It,n × ut,n × εt,n
(6)
same bin across different days. It exhibits a U-shaped pattern Vt,n = d∗t × s∗n × It,n × ε∗t,n ,
due to the higher trading activity at the opening and closing where
hours. The intraday dynamics component captures the intraday ∗
• dt (dt ) is the daily volatility (volume) component;
variation of volatility, such as due to unexpected news or ∗
• sn (sn ) is the intraday seasonal volatility (volume) com-
sudden changes in market sentiment.
ponent;
While the mcsGARCH has reasonable financial interpreta-
• It,n is the shared information flow;
tions, it also has several practical limitations that deteriorate its
• ut,n is the unexpected volatility component;
forecasting ability: 1) Errors in the one-day-ahead prediction ∗
• εt,n (εt,n ) is the log-normal noise term in volatility
of the daily component can propagate and cause errors in the
(volume).
intraday GARCH model [16]. 2) Intraday GARCH cannot
capture the longer-lasting effects of market shocks [4]. To Specifically, the intraday seasonal components are estimated
address these issues, incorporating new sources of information, separately based on historical averages.
such as trading volume, is necessary. Our proposed model for intraday volatility modeling of-
fers several improvements over the traditional mcsGARCH
C. Mixture of Distribution Hypothesis model. First, incorporating information flow can better extract
The positive relationship between daily volatility and trad- medium-term volatility persistence with the knowledge of
ing volume has been extensively investigated in New York trading volume. Second, we gain greater flexibility in capturing

1396
intraday dynamics by modeling it as a stochastic process where
instead of a deterministic function. Third, our state-space •
⊺
yτ = [ln στ , ln Vτ ] is the observation vector of intraday
approach, introduced in Section III-B, enables us to extract log-volatility and log-volume;
all components simultaneously. ⊺
• xτ = [ln dτ , ln d∗τ , ln Iτ , ln uτ ] is the state vector con-
Compared to the MDH hypothesis, our framework provides
taining daily, information flow, and unexpected volatility
a more precise identification of the information flow by filter-
ing out components that cannot be well captured, such as daily components;
patterns with long-term differences and unexpected volatility. • W is the observation matrix:
These improvements enable us to identify volatility persistence
 
1 0 1 1
with different frequencies, leading to richer interpretations of W= ;
0 1 1 0
market microstructure.
⊺
• sτ = [ln sτ , ln s∗τ ] is the vector of intraday seasonal
B. State-space Approach components corresponding to specific bins;
To facilitate component extraction in the model (6), we • Fτ is the transition matrix
propose a state-space formulation that employs the Kalman (
diag (ad , a∗d , aI , au ) τ = kN
filter and smoother. Notably, a key issue with intraday data is Fτ =
its heavy-tailed distribution. As demonstrated by the quantile- diag (1, 1, aI , au ) otherwise;
quantile plot (Q-Q plot) in Figure 2a, intraday volatility • ετ ∼ N (0, R) is the i.i.d. Gaussian noise with
deviates significantly from the straight line, indicating a non-
R = diag rσ2 , rv2 ;


Gaussian distribution. This finding is consistent with previous
studies of intraday volume [22]. To address this issue, we
apply logarithmic transformations to obtain a more Gaussian- • η τ ∼ N (0, Qτ ) is the i.i.d. Gaussian noise with
like distribution, as shown in Figure 2b [23], [24]. (
diag qd2 , (qd∗ )2 , qI2 , qu2


τ = kN
Qτ = 2 2


diag 0, 0, qI , qu otherwise.
0.012 l l
l
−4
l l l
l l
Sample Quantiles

Sample Quantiles

lll l l
lll
lllllllll

ll
l
ll
ll l
−5 llllllll
lllllll
llll
llllllll
lllllllll

lllllll
0.008 lllllll
ll
llll lllllll
lll lllllll
lll llllllll
lllllllll

lllll
llll
llll
l
lll
lllll
l
ll
ll
lll

−6
llllll
lllllll
llllllll
lllllllll
lllllllll
lllllllll
lllllllll
l
llllllll
lll
llllllllll

The parameters in model (8) include

llll llllllll
llll lllllll
llll llllllll
llll llllllll
0.004 lllll
lllll
lllll −7 llllllll
llllllll
lllllll
lllllllll

θ = (ad , a∗d , aI , au , rσ , rv , qd , qd∗ , qI , qu ) ,

llllll lllllll
lllllll ll
llllll l
llllllll
l
lll
llllllll llllllll
llllllll lllllll
lllllllll
l
l
llllllll
lll
ll lllllll
lllllllllll lllll
lllllllllll lllll

llllllllllllllllllllllll
lllllllllllllllllllllllllll
llllllllllllllllllllllllllllll
llllllllllll
llllllllllllllll
lllllllllllllllll −8 llll
lllll
llll
llll
llllll

0.000 l l l l l lllllllllllllllllllllllllllllllllllllllll
lllll
lllllll
lllll
lllllll
l ll

−9
l
l l l
l
and they are estimated using the expectation–maximization
−2 0 2 −2 0 2
Normal Theoretical Quantiles Normal Theoretical Quantiles (EM) algorithm [26].
(a) Q-Q plot of volatility. (b) Q-Q plot of log-volatility. 1) Modeling Procedure: Once the parameters are estimated
Fig. 2. Q-Q plot of intraday volatility and log-volatility of AAPL stock. and fixed, we can extract the hidden components from the
observed intraday volatility signal. The Kalman smoother is
Therefore, we reformulate the model (6) as utilized to estimate the optimal decomposition, the latent
variable xτ |T , which is the optimal estimate of the state
ln σt,n = ln dt + ln si + ln It,i + ln ut,i + ln εt,i conditioned on all data from τ = 1 to τ = T .
(7)
ln Vt,n = ln d∗t + ln s∗i + ln It,i + ln ε∗t,i . 2) Forecasting Procedure: In addition to component extrac-
We make several assumptions to specify the evolution of each tion, the proposed state-space model also allows for intraday
latent variable in the model (7). Firstly, the daily compo- volatility forecasting. This study focuses on one-bin-ahead
nents change only during the close of markets and remain forecasting, where the volatility at a specific bin is predicted
constant during trading hours, while the seasonal components based on all the information up to the last bin. Specifically,
are assumed to be unchanged throughout. Additionally, the we use the Kalman filter to generate the one-bin-ahead state
information flow and unexpected volatility components are prediction, denoted as xτ +1|τ , and then obtain the volatility
modeled as autoregressive processes with order 1 (AR(1)), as prediction by adding up the components:
both the informed trades and uninformed trades are highly ln στ +1|τ = ln dτ +1|τ + ln sτ +1 + ln Iτ +1|τ + ln uτ +1|τ . (9)
history dependent (e.g., [25]).
To accommodate these assumptions, we propose a state- IV. E XPERIMENT
space model inspired by prior work [22]. For ease of notation, A. Experiment Set-Up
we index the subscript (t, n) as τ = N × (t − 1) + n, τ =
This empirical study examines the intraday price and vol-
1, 2, . . . , T . Subsequently, we can formulate the state-space
ume data of S&P 500 stocks from January 2019 to June 2021.
model as follows,
The stocks are divided into five batches based on alphabetical
yτ = Wxτ + sτ + ετ , order. Each trading day is divided into N 15-minute bins,
(8)
xτ +1 = Fτ xτ + η τ , and the actual volatility is calculated using 5-minute price

1397
data, as shown in Equation (2). To evaluate the modeling volatility persistence, respectively. Conversely, the unexpected
and forecasting performance, we split each stock’s data into a volatility component has a flatter spectrum, indicating its
training set to estimate the model parameters and an out-of- higher randomness and tendency to capture unexpected volatil-
sample testing set. Table II provides further details about the ity changes.
dataset.
0.020
B. Modeling Results
Components
The volatility components of the sample data are presented 0.015 Information
in Figure 3. The daily component captures the average level

Magnitude
Unexpected
of daily volatility and demonstrates long-term persistence, 0.010
Daily
contributing to the volatility clustering across trading days.
The information flow captures most intraday patterns, while 0.005
the unexpected volatility component is less informative with
a small magnitude. 0.000
0e+00 2e−04 4e−04
Frequency
−5
Volatility
Intraday

−6
−7
−8 Fig. 4. Frequency spectrum of volatility components of ABT stock after
−9 smoothing.
0.50
0.25
Daily

0.00
−0.25
−0.50
C. Forecasting Results
−5.5
Seasonal

−6.0
−6.5 1) Competing Methods: This study presents three groups of
−7.0
candidate methods for intraday volatility modeling discussed
0.50
in the literature [4], [27]. The first group includes traditional
Information

0.25
Flow

0.00
−0.25
time-series models such as random walk, moving average
−0.50
(MA), and exponential weighted moving average (EWMA).
0.50
Unexpected

The second group removes seasonal effects first and im-

Volatility

0.25
0.00
−0.25 plements classical volatility models, including ARCH and
−0.50
0 1000 2000 GARCH models [10]. Finally, the third group is the specific-
τ
designed models for intraday volatility, i.e., the multiplicative
component GARCH (mcsGARCH) [11].
Fig. 3. Decomposition of intraday volatility of ABT stock in log scale.
2) Evaluation Measures: We evaluate the one-bin-ahead
Table I summarizes the estimated parameters from S&P 500, forecasting performance in the out-of-sample testing set, which
indicating a long-term daily volatility persistence with a large consists of M bins. Mean absolute percent error (MAPE) and
parameter ad . The information flow has a smaller parameter Theil-U statistic are used as indicators of the forecast accuracy
aI , supporting the presence of autocorrelation in information [4]. They are defined as
flow and medium-term volatility persistence. The unexpected M PM 2
1 X |σ̂i − σi | (σ̂i − σi )
volatility component has the smallest parameter au . However, MAPE = , Theil-U = PMi=1 2
.
M i=1 σi i=1 (σi−1 − σi )
it indicates the existence of some intraday patterns that cannot
be explained by the joint volatility-volume movement, aligning (10)
with the critics in [20]. Specifically, the Theil-U statistic is the prediction error stan-
dardized by random walk error.
TABLE I The results are summarized in Table II. Our proposed
E STIMATED PARAMETERS OF VOLATILITY COMPONENTS IN S&P 500. method demonstrated the highest level of accuracy when
compared to the other competing methods, as indicated by
Parameter Mean Standard error t-ratio both measures. While the ARCH and GARCH models have
ad 0.833 0.00503 165.462 lower accuracy than traditional time-series models concern-
aI 0.686 0.00375 182.791 ing MAPE, they outperform the latter concerning Theil-U.
au 0.449 0.01708 26.299 This can be attributed to the ARCH-GARCH class models’
sensitivity to unexpected movements, compared to traditional
To further confirm the features of each component, we methods based on historical smoothing that typically have a
perform spectral analysis through Fourier Transform. As il- lag in response. The mcsGARCH performs worst likely due to
lustrated in Figure 4, the daily component dominates the its multistep estimation strategy. Overall, our proposed method
low-frequency spectrum, while the information flow has a demonstrates its efficacy for accurate intraday volatility fore-
higher frequency, responsible for long-term and medium-term casting.

1398
 TABLE II
F ORECASTING PERFORMANCE IN S&P 500.

Stocks A - CMCSA CME - GILD GIS - META MGM - ROP ROST - ZTS
Training Period Jan - May, 2019 Jul - Nov, 2019 Jan - May, 2020 Jul - Nov, 2020 Jan - May, 2021
Forecast Period Jun, 2019 Dec, 2019 Jun, 2020 Dec, 2020 Jun, 2021
Measure MAPE Theil-U MAPE Theil-U MAPE Theil-U MAPE Theil-U MAPE Theil-U
Proposed 0.547 0.476 0.591 0.482 0.512 0.535 0.550 0.470 0.568 0.474
Random walk 0.752 1.000 0.839 1.000 0.703 1.000 0.745 1.000 0.768 1.000
MA(5) 0.722 0.796 0.860 0.802 0.644 0.760 0.704 0.794 0.724 0.800
EWMA(5) 0.729 0.734 0.884 0.740 0.654 0.712 0.713 0.738 0.734 0.740
ARCH(5) 0.821 0.578 1.067 0.702 0.786 0.844 0.915 0.695 1.162 0.874
GARCH(3,2) 0.788 0.543 0.968 0.616 0.731 0.670 0.833 0.605 0.950 0.649
mcsGARCH 0.926 2.002 0.922 1.871 0.889 2.200 0.910 2.035 0.922 1.939

V. C ONCLUSION [12] T. E. Copeland, “A model of asset trading under the assumption of

sequential information arrival,” The Journal of Finance, vol. 31, no. 4,
In this paper, we introduced a novel multiplicative frame- pp. 1149–1168, 1976.
[13] R. H. Jennings, L. T. Starks, and J. C. Fellingham, “An equilibrium
work for intraday volatility modeling that decomposes volatil- model of asset trading with sequential information arrival,” The Journal
ity into long-, medium-, and short-term components. Our of Finance, vol. 36, no. 1, pp. 143–161, 1981.
proposed state-space approach enables the extraction of hid- [14] P. K. Clark, “A subordinated stochastic process model with finite vari-
ance for speculative prices,” Econometrica: journal of the Econometric
den components and has demonstrated improved forecast Society, pp. 135–155, 1973.
accuracy compared to existing methods, as shown through [15] G. E. Tauchen and M. Pitts, “The price variability-volume relationship
empirical experiments on S&P 500 stocks. Furthermore, this on speculative markets,” Econometrica: Journal of the Econometric
Society, pp. 485–505, 1983.
work offers valuable tools and insights for future research in [16] H. Zhang and A. Dufour, “Modeling intraday volatility of European bond
microeconomics, such as identifying which latent demands markets: A data filtering application,” International Review of Financial
are translated into intraday prices and volumes, as well as Analysis, vol. 63, pp. 131–146, 2019.
[17] T. W. Epps and M. L. Epps, “The stochastic dependence of security
in macroeconomics, particularly in studying the relationship price changes and transaction volumes: Implications for the mixture-
between systematic risk and global information flow. of-distributions hypothesis,” Econometrica: Journal of the Econometric
Society, pp. 305–321, 1976.
[18] G.-m. Chen, M. Firth, and O. M. Rui, “The dynamic relation between
R EFERENCES stock returns, trading volume, and volatility,” Financial Review, vol. 36,
no. 3, pp. 153–174, 2001.
[1] K. Northington, Volatility-based technical analysis: strategies for trading [19] O. A. Gwilym, D. McMillan, and A. Speight, “The intraday relationship
the invisible. John Wiley & Sons, 2009. between volume and volatility in LIFFE futures markets,” Applied
[2] J. A. Stephan and R. E. Whaley, “Intraday price change and trading Financial Economics, vol. 9, no. 6, pp. 593–604, 1999.
volume relations in the stock and stock option markets,” The Journal of [20] P. Giot, S. Laurent, and M. Petitjean, “Trading activity, realized volatility
Finance, vol. 45, no. 1, pp. 191–220, 1990. and jumps,” Journal of Empirical Finance, vol. 17, no. 1, pp. 168–175,
[3] B. Feunou and C. Okou, “Good volatility, bad volatility, and option 2010.
pricing,” Journal of financial and quantitative analysis, vol. 54, no. 2, [21] S. Darolles, G. Le Fol, and G. Mero, “Mixture of distribution hypothesis:
pp. 695–727, 2019. Analyzing daily liquidity frictions and information flows,” Journal of
[4] S.-H. Poon and C. W. J. Granger, “Forecasting volatility in financial Econometrics, vol. 201, no. 2, pp. 367–383, 2017.
markets: A review,” Journal of economic literature, vol. 41, no. 2, pp. [22] R. Chen, Y. Feng, and D. Palomar, “Forecasting intraday trading volume:
478–539, 2003. a Kalman filter approach,” Available at SSRN 3101695, 2016.
[5] R. F. Engle, “Autoregressive conditional heteroscedasticity with esti- [23] T. G. Andersen, T. Bollerslev, F. X. Diebold, and P. Labys, “Modeling
mates of the variance of United Kingdom inflation,” Econometrica: and forecasting realized volatility,” Econometrica, vol. 71, no. 2, pp.
Journal of the econometric society, pp. 987–1007, 1982. 579–625, 2003.
[6] T. Bollerslev, “Generalized autoregressive conditional heteroskedastic- [24] S. Alizadeh, M. W. Brandt, and F. X. Diebold, “Range-based estimation
ity,” Journal of econometrics, vol. 31, no. 3, pp. 307–327, 1986. of stochastic volatility models,” The Journal of Finance, vol. 57, no. 3,
[7] N. Shephard and T. G. Andersen, “Stochastic volatility: origins and pp. 1047–1091, 2002.
overview,” in Handbook of financial time series. Springer, 2009, pp. [25] D. Easley, N. M. Kiefer, and M. O’Hara, “One day in the life of a very
233–254. common stock,” The Review of Financial Studies, vol. 10, no. 3, pp.
[8] S. J. Taylor and X. Xu, “The incremental volatility information in one 805–835, 1997.
million foreign exchange quotations,” Journal of Empirical Finance, [26] J. Durbin and S. J. Koopman, Time series analysis by state space
vol. 4, no. 4, pp. 317–340, 1997. methods. OUP Oxford, 2012, vol. 38.
[9] T. G. Andersen and T. Bollerslev, “Intraday periodicity and volatility [27] J. Yu, “Forecasting volatility in the New Zealand stock market,” Applied
persistence in financial markets,” Journal of empirical finance, vol. 4, Financial Economics, vol. 12, no. 3, pp. 193–202, 2002.
no. 2-3, pp. 115–158, 1997.
[10] M. Martens, Y.-C. Chang, and S. J. Taylor, “A comparison of seasonal
adjustment methods when forecasting intraday volatility,” Journal of
Financial Research, vol. 25, no. 2, pp. 283–299, 2002.
[11] R. F. Engle and M. E. Sokalska, “Forecasting intraday volatility in
the US equity market. Multiplicative component GARCH,” Journal of
Financial Econometrics, vol. 10, no. 1, pp. 54–83, 2012.

1399