Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453

Contents lists available at ScienceDirect

Commun Nonlinear Sci Numer Simulat

journal homepage: www.elsevier.com/locate/cnsns

Research paper

Detecting the sampling rate through observations

Isao Shoji
Tokyo University of Science, Fujimi, Chiyoda, Tokyo 102-0071, Japan

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a method to detect the sampling rate of discrete time series of diffu-
Received 1 September 2017 sion processes. Using the maximum likelihood estimates of the parameters of a diffusion
Revised 19 January 2018
process, we establish a criterion based on the Kullback–Leibler divergence and thereby es-
Accepted 23 February 2018
timate the sampling rate. Simulation studies are conducted to check whether the method
Available online 26 February 2018
can detect the sampling rates from data and their results show a good performance in the
Keywords: detection. In addition, the method is applied to a financial time series sampled on daily
Sampling rate basis and shows the detected sampling rate is different from the conventional rates.
Stochastic differential equation © 2018 Elsevier B.V. All rights reserved.
Maximum likelihood estimation
Kullback–Leibler divergence

1. Introduction

The time evolution of stochastic dynamic phenomena is formulated by stochastic differential equations (SDE). For prac-
tical use of these equations, we have to estimate their parameters from data. Here, we should pay our attention to the
qualitative difference between the model formulated by an SDE and data used for estimation because the model is ex-
pressed in the continuous time framework while the data are observed at discrete times. To bridge the difference, we have
to develop schemes for discretization to make a model fit to data rather than obtain continuous time data because the latter
is basically difficult due to accuracy of observation by experimental devices. Actually, there are various studies to pursue this
idea; see, for example, [4], [5,6,8,9] and [11].
It is particularly important for the discretization to determine the sampling interval of data t, or the sampling rate 1/t.
When using small t, the discretized model of t would show the short term behavior, whereas the model would show the
longer time behavior when using larger t. Conventionally, we often determine how long t is, depending on the sampling
frequency such as daily, monthly, or annually. This convention, however, is useless to determine the sampling interval. For
example, suppose time series data observed every day. When using 1 day as the unit time scale, t is equal to 1, where as
it is equal to 1/365 when using 1 year as the unit time scale. The discretized model would show the somewhat longer time
behavior in the former case but the short term behavior in the latter case although the data themselves are just identical.
Moreover, from a relativistic point of view, observers in different inertial frames would observe different t because of their
different elapsed time even if we take no account of the unit time scale; one observer at a rest frame O observes
 data every
t, while another observer moving at velocity v relative to O observes the same data every t  = t 1 − v2 /c2 for the
speed of light c.
Additionally, t has a significant influence on the performance of discretization. Deriving a discretized model from an
SDE is basically equivalent to solving the SDE. Since it is generally difficult to obtain the exact solution to such an SDE,
especially as nonlinear stochastic differential equations, approximation methods are often used for the derivation. The per-

E-mail address: shoji@rs.tus.ac.jp

https://doi.org/10.1016/j.cnsns.2018.02.031
1007-5704/© 2018 Elsevier B.V. All rights reserved.
446 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453

formance of the approximation is measured by the rate of convergence in t. For example, when applying the well-known
Euler method to SDEs, O(t) is known as its rate of convergence; see [7] for example. Thus the better the performance of
approximation the shorter t.
Unless running simulations, however, we are unable to take t as short as we want. Rather, in applications we have to
identify how long it is. Given time series, we first assume a statistical model, or a data generating process, which is thought
to generate the time series with specific values of its parameters and t. In parameter estimation we have main interest
in what values those parameters have, but often do not care about what value t is. Since the observed time series are
assumed to follow the data generating process, the value of t as well as the values of the parameters must be estimated
from the time series.
At a first glance, the estimation looks easy. However, we cannot estimate the parameters and t simultaneously. Take
a Brownian motion with drift for example, which are formulated by dXt = μdt + σ dBt , where Bt is the standard Brownian
motion. Here, we want to estimate μ, σ , and t from discrete time series {Xtk }1≤k≤n of its time interval, tk − tk−1 = t. To
this end, the maximum likelihood estimation can be applied because Xtk − Xtk−1 follows the normal distribution with mean
μt and variance σ 2 t. It can be easily seen that we are unable to estimate the parameters and the sampling interval
simultaneously. Alternatively, we could estimate the parameters and the sampling interval separately as follows. Let t be
fixed at some value and then maximize the log-likelihood function with respect to parameters. Repeat this procedure for
one possible sampling interval after another. Among those obtained maximized log-likelihoods, it seems that we have only
to choose the sampling interval at which the maximized log-likelihood attains the maximum. This straightforward approach,
however, is shown to fail to work.
On the basis of the Kullback–Leibler divergence, this paper proposes the mean log-likelihood as criterion for the detec-
tion. In this approach, we first estimate the parameters of an SDE via the maximum likelihood method from one discrete
sample path given a possible sampling rate. And then, we construct a mean log-likelihood from another sample paths in-
dependent of each other. Repeating this procedure by changing one possible sampling rate after another, we choose the
sampling rate at which the mean log-likelihood attains the maximum.
To check the performance of the proposed method numerically, simulation studies are conducted by using well-known
SDEs with several combinations of sampling rates and the number of observations. In addition, we apply the method to
financial time series sampled on daily basis and detect its sampling rate implied by the data. The analysis shows that the
detected sampling rate is different from such rates as implied by conventionally taking 1 week, month, or year as the unit
time scale.
The paper is organized as follows. In Section 2 we show the maximized log-likelihood fails to work as criterion for com-
paring sampling rates, but instead the mean log-likelihood is proposed. In Section 3 the numerical experiments are carried
out to check the numerical performance of the proposed method. In Section 4 an empirical application to the Japanese stock
price index is provided and the concluding remarks are given in the final section.

2. Methods of detecting the sampling rate

Suppose a diffusion process Xt starting at x0 satisfies the following SDE,

dXt = μ(Xt ; θ )dt + σ (Xt ; θ )dBt

X0 = x0 , (1)

where Bt is the standard Brownian motion and μ(Xt ; θ ) and σ (Xt ; θ ) represent functions of Xt with an unknown parameter
vector θ . Here, we have time series with an equidistant sampling interval, or a discrete sample path of the process {Xtk }1≤k≤n
(tk − tk−1 = τ0 ), thereby we estimate the unknown θ . Unlike usual settings, however, we assume that we have no informa-
tion on the physical sampling interval τ 0 or rate 1/τ 0 although we may know those data are sampled on a daily, weekly,
monthly, or annually basis. Those conventional sampling frequency means neither sampling interval nor rate because the
sampling interval or rate is determined by the unit time scale which is usually unknown to experimenters. Therefore, we
have to estimate not only the parameter vector but also the sampling interval or rate.
In the following discussion, we focus one-dimensional diffusion processes for simplicity. Those arguments, however, can
be easily extended to multi-variate diffusion processes.

2.1. Maximized log-likelihood

Here, consider the maximum likelihood (ML) method for example. Let l(θ ) be the logarithm of the likelihood function
for given {Xtk }1≤k≤n . We get the ML estimate θˆ by maximizing l(θ ) with respect to θ when the sampling interval is known.
In our settings, however, the true sampling interval τ 0 is unknown. So, first we tentatively assign a possible value to the
sampling interval, denoted by τ , and then carry out the maximum likelihood estimation. The estimate obtained in such a
way depends on τ used for estimation, and thus the estimate should be expressed as a function of τ , or θˆ (τ ).
In terms of the ML method, the greater l (θˆ (τ )) is more desirable. Thus, it seems that we have only to choose τ which
gives the largest l (θˆ (τ )) among possible sampling intervals. But, this straightforward approach fails. Consider the differential
 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453 447

of l (θˆ (τ )) with respect to τ , which is given by,


dl (θˆ (τ )) 
m
∂ l dθk 
=
dτ
k=1
∂θk dτ θ =θˆ (τ ).
But, this differential is always zero because ∂ l/∂ θk (θˆ ) = 0 due to the property of the ML estimate. Hence, l (θˆ (τ )) is constant
for any τ and thus we are unable to distinguish l (θˆ (τ )) from l (θˆ (τ  )) for any different τ and τ  .
All is the same even when the least squares (LS) method is applied. The LS estimate θˆLS is given as the minimum point
for an objective function f(θ ). Hence, ∂ f /∂ θk (θˆLS ) = 0, and thereby df (θˆ (τ ))/dτ = 0. Hence, the optimized value of f is
constant for any τ , and thus this approach is also useless for our purpose.

2.2. Mean log-likelihood

Given time series {Xtk }1≤k≤n of a diffusion process starting at x0 , its log-likelihood function is denoted by
log( p(Xt1 , · · · , Xtn ; θ |Xt0 = x0 )). For simplicity, without specifying this condition, we denote,
l (θ ) = log ( p(Xt1 , · · · , Xtn ; θ ) ). (2)
According to the Kullback–Leibler divergence,
  
p(Xt1 , · · · , Xtn ; θ0 )
E0 log ≥ 0, (3)
p(Xt1 , · · · , Xtn ; θ )
where θ 0 represents the true parameter vector and E0 [] means the expected value computed by the probability density
p(Xt1 , · · · , Xtn ; θ0 ) for θ 0 . Thus E0 [l(θ 0 )] ≥ E0 [l(θ )], implying the greater E0 [l(θ )] is more desirable for given θ , and the equality
holds if and only if the two joint probability density functions are equal, or θ = θ0 . See [3] for example.
In our settings, since the log-likelihood function l depends on τ as well as θ , this inequality should be expressed as
E0 [l(θ 0 , τ 0 )] ≥ E0 [l(θ , τ )]. Here, suppose a function φ (θ , τ ), defined by φ (θ , τ ) ≡ E0 [l(θ , τ )]. Then, putting the ML estimate
θˆ (τ ) into θ of φ , we get,
φ (θ0 , τ0 ) ≥ φ (θˆ (τ ), τ ).
Here note φ (θˆ (τ ), τ ) is different from E0 [l (θˆ (τ ), τ )]; in the former θˆ (τ ) is treated as a fixed value, but in the latter it is to
be evaluated as a random variable by the expectation operator E[ · ]; φ (θˆ (τ ), τ ) is given by simply replacing θ in E0 [l(θ , τ )]
with θˆ , whereas E0 [l (θˆ (τ ), τ )] is given as the expected value of l (θˆ (τ ), τ ).
According to this inequality, we establish a new criterion on θˆ (τ ), which implies that the greater φ (θˆ (τ ), τ ) is more
desirable. Hence, we look for τ ∗ such that φ (θˆ (τ ∗ ), τ ∗ ) attains the maximum. In general, however, since it is difficult to
compute φ (θˆ (τ ), τ ), we approximate it by its sample mean. Suppose that the ML estimate θˆ (τ ) is obtained from a sample
path {Xt(0 ) }1≤k≤n . In addition to that, we have another N sample paths {Xt(i ) }1≤k≤n (1 ≤ i ≤ N), each of which is sampled inde-
k k
pendently of each other and has the same sampling interval as that of {Xt(0 ) }1≤k≤n . Hence, φ (θˆ (τ ), τ ) can be approximated
k
as follows:

1
N

φ (θˆ (τ ), τ ) ≈ log p(Xt(1i ) , · · · , Xt(ni ) ; θˆ (τ ), τ ) (4)

N
i=1

≡ l¯(θˆ (τ )). (5)

This arithmetical average l¯(θˆ (τ )) is hereafter referred to the mean log-likelihood of θˆ (τ ), and is used in the following
experiments instead of φ (θˆ (τ ), τ ).

3. Simulation studies

Suppose diffusion processes starting at x0 , which satisfy the following SDEs:

dXt = β (α − Xt )dt + σ dBt (6)


dXt = β (α − Xt )dt + σ Xt dBt , (7)
where α , β , and σ are constant, simply denoting these three tuple by the parameter vector θ . These SDEs are often used as
benchmark models of the interest rate for empirical or numerical experiments. See [1,2] and [4] for example. Discrete data
of size n are sampled with the sampling interval τ 0 , which is unknown to experimenters. An experimenter tries to estimate
τ 0 by looking for the sampling interval τ at which l¯(θˆ (τ )) attains the maximum.
In the experiments, two cases for τ 0 and three cases for n are considered; that is, τ0 = 1/50 and 1/100, and n = 10 0 0,
50 0 0, and 10,0 0 0. But the others are fixed through all the experiments by putting x0 = 1, α = β = 1, σ = 0.1, and N = 100.
448 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453

3.1. Quasi-maximum likelihood estimation

Given time series {Xtk }1≤k≤n , the log-likelihood function l(θ ) is given as,

l (θ ) = log ( p(Xt1 , · · · , Xtn |Xt0 = x0 ))

 

n  
= log p Xtk Xtk−1
k=1


n  
= log p Xtk Xtk−1 , (8)
k=1

where p(Xtk |Xtk−1 ) is the transition probability density function.

Although this transition probability density function is given as a solution to the Fokker–Planck equation, we do not gen-
erally expect to have its explicit formula except for special cases. Hence, we apply the quasi-ML method in our experiments.
In the quasi-ML method, the transition probability density function is given by assuming that the discrete-time process Xtk
follows conditionally a normal distribution, and thus we have only to compute its mean and variance; that is, we need to
specify μtk and σt2 appearing in,
k
 
1 (Xtk − μtk )2
p(Xtk |Xtk−1 ) =  exp − . (9)
2π σt2k 2σt2k

Here two methods of computing these moments are considered. One is to use the Euler method for discretization and
the other is to compute conditional moments directly. The latter method can produce the exact conditional moments for
our examples of SDEs and thus there is no approximation error in computing those moments.
For an SDE as in (1), the Euler method produces,

μtk = Xtk−1 + μ(Xtk−1 )τ (10)

σt2k = σ 2 (Xtk−1 )τ . (11)

Here note τ means the sampling interval assumed by the experimenter. The Euler method is simple but it is often the case
that the method generally produces neither exact moments nor accurate parameter estimates.
On the other hand, the computational method of conditional moments, the CM method shortly, produces the exact mo-
ments in such SDEs as in (6) and (7). Here, we give a brief explanation about the CM method but the details are discussed
in [10]. Let tk be a 2-dimensional vector whose components consist of the first and second conditional moments around
Xtk−1 ; that is,

tk = (Etk−1 [Xtk − Xtk−1 ], Etk−1 [(Xtk − Xtk−1 )2 ] ), (12)

where Etk−1 [· · · ] represents the expectation operator conditional of σ -algebra F tk−1 . According to the CM method,
 τ
tk = exp(Atk−1 t )btk−1 dt, (13)
0

where,
 
μ (Xtk−1 ) 1
μ (Xtk−1 )
Atk−1 = 2
2μ(Xtk−1 ) + (σ 2 ) (Xtk−1 ) 2μ (Xtk−1 ) + 12 (σ 2 ) (Xtk−1 )
 
μ(Xtk−1 )
btk−1 = .
σ 2 (Xtk−1 )
If Atk−1 is nonsingular,

tk = At−1
k−1
(exp(Atk−1 τ ) − I )btk−1 . (14)
Using this formula, the CM method gives,

μtk = ψtk ,1 + Xtk−1 (15)

σt2k = ψtk ,2 − (μtk − Xtk−1 )2 , (16)

where ψtk ,i is the ith component of tk .
 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453 449

-7.535 -7.53

-7.5355
-7.532
-7.536

-7.5365 -7.534
0 20 40 60 80 100 0 20 40 60 80 100
-7.525 -7.528

-7.53 -7.53

-7.535 -7.532

-7.54 -7.534
0 20 40 60 80 100 0 20 40 60 80 100
-7.525 -7.5332

-7.53 -7.5334

-7.535 -7.5336

-7.54 -7.5338
0 20 40 60 80 100 0 20 40 60 80 100
Fig. 1. Mean log-likelihood of SDE (6) when τ0 = 1/50: The left/right panel displays the results by the Euler discretization/the computational method of
conditional moments. The horizontal axis represents sampling rates.

3.2. Evaluation by the mean log-likelihood

For convenience, −l¯(θˆ (τ )) is considered instead of l¯(θˆ (τ )) itself in the followings, and thus this leads us to find the
minimum instead of the maximum. To find the minimum, we consider the collection of sampling rates over integer Z,
denoted by R, as possible sampling rates. More specifically, we put R = {1/τ ∈ Z |1 ≤ 1/τ ≤ 100} for 1/τ0 = 50 and R =
{1/τ ∈ Z |51 ≤ 1/τ ≤ 150} for 1/τ0 = 100, except for SDE (6) of 1/τ0 = 100 and n = 10 0 0, where we put R = {1/τ ∈ Z |1 ≤
1/τ ≤ 150}.
Results are graphically displayed by Figs. 1–4. The results of SDE (6) are shown in Figs. 1 and 2 and those of SDE (7) in
Figs. 3 and 4. Figs. 1 and 3 display the results when 1/τ0 = 50 and Figs. 2 and 4 for 1/τ0 = 100. In each figure, top, middle,
and bottom panels show change of −l¯(θˆ (τ )) to different sampling rates for n = 10 0 0, 50 0 0, and 10,0 0 0. The results of the
Euler method is given in the left-hand side of each panel and those of the CM method in the right-hand side.
Looking at Fig. 1, top panels show sharp decline at 1/τ = 32 in the Euler method and 33 in the CM method. These are
close to the true sampling rate 50. In the middle and bottom panels for the Euler method, however, the graphs do not show
clearly what the lowest point is, but the lowest points are 2 and 36 when n = 50 0 0 and n = 10, 0 0 0. By contrast, when the
CM method is used, it can be easily seen that the lowest points are 55 and 58 in the case of n = 50 0 0 and n = 10, 0 0 0,
which are very close to the true sampling rate. Considering that the performance of the quasi-ML method depends upon
the conditional moments used for the likelihood function, the results of estimation by the CM method may be more reliable
because it can produce the exact conditional moments while the Euler method produces approximate ones.
In Fig. 2, when the true sampling rate 1/τ0 = 100, the lowest points can be easily seen for each method, but they vary
considerably. In the Euler method, the lowest points are 131, 106, and 68 when n = 10 0 0, 50 0 0, and 10,0 0 0. The second
case is close to the true sampling rate. In the CM method, the lowest points are 51, 122, and 84. Except for n = 10 0 0, those
are around the true sampling rate.
As for SDE (7), we have similar results to SDE (6). Looking at Fig. 3 (1/τ0 = 50), the Euler method and the CM method
produce almost the same results for n = 10 0 0 and n = 50 0 0; the lowest points for n = 10 0 0 and n = 50 0 0 are 32 and 51
for the Euler method and 33 and 56 for the CM method. For n = 10, 0 0 0, the lowest points are 36 for the Euler method
and 58 for the CM method. Similar results to SDE (6) is also observed when 1/τ0 = 100; Fig. 4 shows that the lowest points
for n = 10 0 0, 50 0 0, and 10, 0 0 0 are 67, 107, and 69 for the Euler method, on the other hand, 97, 122, and 84 for the CM
method.
Additionally, we conducted another experiment for relatively small sample size because small sample was used for the
following empirical analysis, where data of size 70 is used and the exact conditional mean and variance can be computed.
450 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453

-8.2161
-8.212
-8.2161
-8.214
-8.2162
-8.216
-8.2162 -8.218
50 100 150 0 50 100 150
-8.217
-8.2196
-8.2198 -8.218
-8.22
-8.2202 -8.219
50 100 150 50 100 150
-8.2195 -8.2182
-8.2195
-8.2184
-8.2195
-8.2186
-8.2196
-8.2196 -8.2188
50 100 150 50 100 150
Fig. 2. Mean log-likelihood of SDE (6) when τ0 = 1/100: The left/right panel displays the results by the Euler discretization/the computational method of
conditional moments. The horizontal axis represents sampling rates.

-7.538 -7.532

-7.5385 -7.534

-7.539 -7.536

-7.5395 -7.538
0 20 40 60 80 100 0 20 40 60 80 100
-7.5392 -7.53

-7.5394 -7.535

-7.5396 -7.54
0 20 40 60 80 100 0 20 40 60 80 100
-7.538 -7.5354

-7.5356
-7.53805
-7.5358

-7.5381 -7.536
0 20 40 60 80 100 0 20 40 60 80 100
Fig. 3. Mean log-likelihood of SDE (7) when τ0 = 1/50: The left/right panel displays the results by the Euler discretization/the computational method of
conditional moments. The horizontal axis represents sampling rates.
 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453 451

-8.2189 -8.2168

-8.21895 -8.21685

-8.219 -8.2169

-8.21905 -8.21695
50 100 150 50 100 150
-8.223 -8.221

-8.2232 -8.2215

-8.2234 -8.222

-8.2236 -8.2225
50 100 150 50 100 150
-8.2215 -8.2202

-8.2216 -8.2204

-8.2217 -8.2206

-8.2218 -8.2208
50 100 150 50 100 150
Fig. 4. Mean log-likelihood of SDE (7) when τ0 = 1/100: The left/right panel displays the results by the Euler discretization/the computational method of
conditional moments. The horizontal axis represents sampling rates.

So, we consider the same number of observations, 70, as in the empirical analysis, and we used the CM method for the
quasi-ML estimation. Although there is no graph, the results for SDE (6) show that −l¯(θˆ (τ )) attains the lowest at 68 and
105 in the case of 1/τ0 = 50 and 1/τ0 = 100, respectively. As for SDE (7), the lowest points are 43 and 83 for 1/τ0 = 50 and
1/τ0 = 100, respectively. These results suggest the proposed method works even for small sample.

4. Empirical application

We apply the detection method to the time series of the Japanese stock price index, called TOPIX, which are sampled on
a daily basis from 2007 to 2012 (1471 observations). We are now interested in what sampling rate the time series has. It is
often the case that the sampling rate (interval) of daily data is blindly assumed to be 365 (1/365), but there is no reason
for assuming it. The aim of this section is to estimate the sampling rate from real data.
As in the option pricing theory, we assume the process of TOPIX, Xt , to follow the geometric Brownian motion:
dXt = μXt dt + σ Xt dBt ,
where μ and σ are constant. The logarithm of the original process, Yt = log(Xt ), satisfies the following SDE:
 
σ2
dYt = μ− dt + σ dBt .
2
Thus, given {Xtk }1≤k≤n ,
 
σ2
Ytk − Ytk−1 = μ− (tk − tk−1 ) + σ (Btk − Btk−1 ).
2
Here, Ytk − Ytk−1 represents the rate of return on TOPIX, denoted by Rtk , which is mutually independent because of mutual
independence of {Btk − Btk−1 }1≤k≤n . Thus, the joint density function of {Rtk }1≤k≤n is easily given by,

n
p(Rt1 , · · · , Rtn ) = p(Rtk ),
k=1

where
 
1 {Rtk − (μ − σ 2 /2 )τ }2
p(Rtk ) = √ exp − .
2π σ 2 τ 2σ 2 τ
452 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453

-6
x 10
-4.1

-4.2

-4.3

-4.4

-4.5

-4.6

-4.7

-4.8
100 150 200 250 300 350 400 450 500
Fig. 5. Mean log-likelihood (plus 6.8367) of TOPIX data: The horizontal axis represents sampling rates.

As in the previous section, first we estimate μ and σ and then estimate the sampling rate (1/τ 0 ) using several sets of
time series data independent of data for estimation. Because of limited amount of data, we used relatively small samples,
each of which consists of 70 observations, by dividing the total time span into 21 sub time spans. We estimate the parame-
ters from the first sample and then we compute l¯(θˆ (τ )) from other twenty samples, which are mutually independent due
to the assumption of the process Xt .
The result is displayed in Fig. 5. The sharp decline is observed at 294, implying the sampling rate of the daily TOPIX
data may be around it. This is closer to the sampling rate of 250 business days per year than that of 365 days. In terms
of the sampling interval, it is 1/294 ≈ 0.0034. On the other hand, the sampling rate of 250 days per year corresponding to
1/250 = 0.0040 while that of 365 days per year is 1/365 ≈ 0.002740. Interestingly, the result shows that the sampling rate
implied by daily data is closer to the sampling rate of 250 business days per year than that of 365 days.

5. Concluding remarks

When making statistical inference on discrete time series of diffusion processes, we need to pay our attention to the
sampling rate even if we know that they are sampled on daily, weekly, monthly, or annually base. These sampling infor-
mation has nothing to do with the physical number corresponding to the sampling interval or rate, which is necessary for
statistical inference such as parameter estimation. This paper proposed a method to detect the sampling rate hidden behind
observations.
The paper first shows that simple applications of the maximum likelihood estimation or the least squares estimation
do not work for the detection. Alternatively, on the basis of the Kullback–Leibler divergence, we proposed the mean log-
likelihood as criteria for detecting the sampling rate. In order to check its numerical performance, simulation studies were
carried out by using well known stochastic differential equations with several combinations of the sampling rate and the
number of observations. The results show that the proposed method can detect the sampling rate through those discrete
observations. In addition, modeling time series of the Japanese stock price index (TOPIX) by the geometric Brownian motion,
we apply the proposed method to TOPIX’s daily data and found its sampling rate around 294 which is closer to the rate of
250 business days per year than that of 365 days.

Acknowledgments

We are grateful to the editor and to anonymous referees for their helpful comments and suggestions.
 I. Shoji / Commun Nonlinear Sci Numer Simulat 62 (2018) 445–453 453

References

[1] Beskos A, Papaspiliopoulos O, Roberts GO, Fearnhead P. Exact and computationally efficient likelihood-based estimation for discretely observed diffu-
sion processes. J R Stat Soc B 2006;68:333–82.
[2] Chan KC, Karolyi GA, Longstaff FA, Sanders AS. An empirical comparison of alternative models of the short-term interest rate. J Finance
1992;47:1209–27.
[3] Cover TM, Thomas JA. Elements of information theory. 2nd. John Wiley & Sons; 2006.
[4] Durham GB, Gallant AR. Numerical techniques for maximum likelihood estimation of continuous-time diffusion processes. J Bus Econ Stat
2002;20:297–338.
[5] Gallant A, Long J. Estimating stochastic differential equations efficiently by minimum chi-square. Biometrika 1997;84:124–41.
[6] Kessler M. Estimation of an Ergodic diffusion from discrete observations. Scand J Stat 1997;24:119–211.
[7] Milstein GN. Numerical integration of stochastic differential equations. Kluwer Academic Publishers; 1988.
[8] Phillips PCB, Yu J. Maximum likelihood and Gaussian estimation of continuous time models in finance. In: Handbook of financial time series; 2009.
p. 497–530.
[9] Prakasa Rao B. Statistical inference for diffusion type processes. London: Arnold; 1999.
[10] Shoji I. Approximation of conditional moments of diffusion processes. Int J Comput Numer Anal Applic 2002;1:163–90.
[11] Sørensen M. Parametric inference for discretely sampled stochastic differential equations. In: Handbook of financial time series; 2009. p. 531–54.