See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/323671215

Kalman filter based time series prediction of cake factory daily sale

Conference Paper · October 2017

DOI: 10.1109/CISP-BMEI.2017.8302108

CITATIONS READS
2 3,475

5 authors, including:

Jionglong Su Fei Ma
Xi'an Jiaotong-Liverpool University Xi'an Jiaotong-Liverpool University
65 PUBLICATIONS   515 CITATIONS    54 PUBLICATIONS   286 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Image Recognition View project

Applications of Artificial Intelligence in Portfolio Management View project

All content following this page was uploaded by Jionglong Su on 27 July 2019.

The user has requested enhancement of the downloaded file.

 Kalman Filter Based Time Series Prediction of
Cake Factory Daily Sale
Jiaxuan Wu Qing Fang Yangying Xu Jionglong Su Fei Ma
Mathematical Sciences, Xi’an Jiaotong-Liverpool University, Suzhou, China
Telephone: (86) 512 8816 1633, Email: {jionglong.su, [email protected]}

Abstract—Accurate prediction of future daily sales is a crucial [13]. In the meantime, its time consuming learning process
step towards optimal management of daily production of a cake and poor generalization ability restricts its application in real
factory. In this study, an interacting multiple model integrated world [14].
kalman filter was used to predict the future daily sales of cake
products. Two years daily sale history of 108 cake products were In recent years, the state space model which embeds Box-
used to train and test the proposed method. Our experiments Jenkins models into Kalman Filter has been suggested ( [15],
show that 1) running interacting multiple models of different [16]). These Kalman filtering algorithms can implement the
orders in parallel is more effective than single classical interacting optimal estimation of the system states. It constructs optimal
multiple model; 2) when only daily sale data was used, the state estimations by methods such as maximum likelihood
proposed method predicted 33.54% of sales within ±10% of true
sales; 3) when more variables, including festival and weekend, estimation (MLE) or least square solution and is self-adaptive
were combined into the prediction, 34.38% of predicted sales in nature [17]. These improves the accuracy and efficiency of
were within ±10% of true sales. time series prediction.
The Interacting Multiple-Model (IMM) is a method which
Keywords-Kalman filter, interacting multiple models, time allows optimal solution to be obtained among several possible
series, daily sale prediction, cake factory models [18]. Each model can effectively describe changing
parameters and each solution to the problem is a combination
I. I NTRODUCTION of several models.
In this paper we combine Kalman filter with Interacting
A time series is a series of data indexed in time order. Time Multiple-Model for online prediction of daily sales of a
series forecasting focuses on the influence of the rules of a cake shop. Sales prediction is essential for operation of food
given phenomenon to predict future trends based on historic industry due to short shelf life of raw materials. Accurate
data [1]. Traditional Box-Jenkins regression models such as sales forecasting can result in reasonable resource allocation
Autoregressive (AR), and Autoregressive Integrated Moving and optimization in advance, As a result, this can effectively
Average (ARIMA) have been widely used in the past for reduce wastage or bottleneck. Moreover, in order to improve
stationary data analysis [2]. In recent years, the modeling and the predicting accuracy and efficiency, there are several im-
prediction of the non-stationary time series raises considerable provements applied in our model :
interest to the research community [3]. • Parallel IMM with different orders used in this new model
In the last few decades, artificial neural networks (ANN) has make the lag in AR model can be adapted in the process
been widely used to predict the nonlinear processes due to its of optimal selection.
map-approaching and self-learning abilities [4]. For example, • The new method to achieve self-updated observation
Karunanithi et al. [5], Hu et al. [6] used both classical and noise variance Rt has been applied in this model, more-
recurrent multilayer perceptron (MLP) neural networks for over, we set some constant choice for the value of
forecasting software reliability. Inspired by both the multilayer system noise variance Qt to simplify operations and avoid
perceptron (MLP) and wavelet decomposition, Wang et al. ( backtest overfitting.
[7], [8], [9]) proposed the wavelet packet MLP (WP-MLP) • The new method to achieve self-updated observation
and demonstrated that it can be a promising method for noise variance Rt has been applied in this model, more-
time series prediction. Xu et al. [10], Wang and Gupta [11] over, we set some constant choice for the value of
applied feed-forward MLP to conduct prediction. The method system noise variance Qt to simplify operations and avoid
based on a BiLinear Recurrent Neural Network proposed by backtest overfitting.
Ghazali et al. [12] has also been proven to have robust abilities
in modeling and predicting time series. However, there are II. M ETHODOLOGY
some flaws on these methods: though ANN does not build
on specific mathematic model, it requires a large number A. Autoregressive Model
of training samples. And noisy training samples can lead to The autoregressive model, which describes a time-varying
error learning results such as slow convergence parameters and process, specifies that the estimated variable only depends
suboptimality which will finally result in invalid prediction linearly on its own previous data and a stochastic term [19].
 The notation AR(p) is defined as: Pt|t−1 represents the error covariance matrix. The prior esti-
p
X mate x̂t|t−1 and Pt|t−1 are made for the measurement update.
yt = αi yt−i + t (1) As yt has been obtained at time t, this model can update the
i=1 conditional mean and covariance matrix of estimates according
where yt is the time series under investigation, αi denotes the to the three equations:
auto-regression coefficient, p represents the order of the model Kt = Pt|t−1 HtT [Ht Pt|t−1 HtT + Rt ]−1 (6)
which is generally very much smaller than the length of the
x̂t = x̂t|t−1 + Kt [yt − Ht x̂t|t−1 ] (7)
series, and t is always assumed to be the Gaussian white
noise. Pt|t = Pt|t−1 − Kt Ht Pt|t−1 (8)
Autoregressive model assumes that the series xt is sta- where Kt is the Kalman gain matrix that minimizes the à
tionary, which means the joint probability distribution of the posteriori error covariance, Pt|t denotes the updated error
stochastic process does not change when shifted in time. covariance matrix. Once this model obtains the outcome of
However, if we regard the series as a whole, the order of the next measurement, these estimates are updated based on
the whole series may not be the most appropriate for every a weighted average. The algorithm is recursive and does not
time step because the series may be nonstationary. Hence, we require any additional past information, only taking the present
consider that running the models which have different orders in input measurements and the previously calculated state and its
parallel and selecting the output based on estimated accuracy uncertainty matrix into consideration.
may be more efficient than classical means. In section 3,
1) Embedding Autoregressive Model into the KF: To solve
our model associated with Kalman filter and the interacting
the problem that the classical AR(p) model can only be fitted
multiple-model will be tested in detail.
using stationary data, the AR(p) model is embedded into the
Kalman filter, since the KF can accept non-stationary data.
B. Kalman filter(KF) Therefore, for the system equation of state space model, we
In classical filtering and prediction of dynamic system set At to be identity matrix so that xt follows a random walk.
by means of the state transition, one is subjected to many For observation equation, we rewrite the AR model as
limitations which may adversely impact on their practical
yt = Ht xt + vt (9)
applications [20]. To sidestep the difficulties which will curtail
the real-world usefulness, the Kalman filter provides an effi- where  
cient recursive means to estimate the state space and is very Ht = yt−1 yt−2 ··· yt−p (10)
powerful to solve linear filtering problems [21].
and
The Kalman filters, based on linear dynamical system  T
xt = α1 α2 ··· αp (11)
discretized in the time domain, can estimate the current state
with the estimated state from the previous time step and the The order of classical AR model is determined after obtain-
current measurement. In other words, this model assumes that ing all data. However, in the KF, we will not test the data first
the state of a system at a time t grounded on the prior state and we just do the prediction directly. Hence we can run the
at time t − 1 and gets an observation yt of the true state xt at model in real time.
time t according to the state space model:
2) Estimation of Rt : The process noise covariance Qt and
xt = At xt−1 + wt (system equation) (2) the measurement noise terms Rt are often assumed to be
yt = Ht xt + vt (observation equation) (3) unknown constants, which can simplify the KF. However,
the value of Qt and Rt should be time-variant since we are
where xt denotes the state vector at time step t, At represents estimating non-stationary time series. In this paper, we let
the state transition matrix based on the system state parameter Rt be self-adaptive and Qt be an unknown discrete random
at time step t − 1, wt is the process noise vector at time step variable. We give several possible values of Qt , which will be
t, yt denotes the vector of measurements, Ht represents the explained in detail in next section.
transformation matrix that maps the state vector parameters One method to estimate the measurement noise terms Rt is
into the measurement domain, vt is the measurement noise to define
term at time step t. Mt = yt − Ht x̂t|t−1 (12)
The Kalman filter consists of two stages: prediction and
measurement update. The equation related to prediction stage Since
are: yt = Ht xT + vt (13)

x̂t|t−1 = At x̂t−1|t−1 (4) Thus

Pt|t−1 = At Pt−1|t−1 AT


t + Qt (5) Mt = (Ht xt + vt ) − Ht x̂t|t−1 = Ht xt − x̂t|t−1 + vt (14)
where Qt denotes process noise covariance matrix associated Hence
with noisy control inputs, x̂t|t−1 is the predicted state mean, Mt = Ht êt|t−1 + vt (15)
where • The mixing weight
(i|j)
êt|t−1 = xt − x̂t|t−1 (16) To begin with, the mixing weight µt−1 could be gener-
ated from the probability of mode i, which is represented
By conditioning on all past observations, we can obtain (i)
as µt−1 and the transition probability from location i to
ij
V ar (Mt |all past observations) = Ht V ar êt|t−1 HtT + Rt


location j, denoted as πt−1 , which could also be obtained
(17) from time step t-1. The related equation is:
Because (i)
π ij µt−1
h i
Pt|t−1 = E êt|t−1 êT
t|t−1 (18) (i|j)
µt−1 = PN t−1(kj) (23)
(k)
k=1 πt−1 µt−1
Then
V ar (Mt ) = Ht Pt|t−1 HtT + Rt (19) where
(i)
Therefore, the measurement noise covariance Rt can be µt = P (rt = i|{yk }tk=1 ) (24)
estimated according to the equation: (ij)
πt−1 = P (rt = j|rt−1 = i, {yk }t−1
k=1 ) (25)
Rt = V ar (Mt ) − Ht Pt|t−1 HtT (20) • The mixed state and covariance [22]
However, Rt may become negative because of the subtrac- (i)
X (j) (i|j)
x̄t−1|t−1 = x̂t−1|t−1 µt−1
tion. Since Rt represents the measurement noise, it cannot be
j
zero or negative. When Rt equals zero, the measurement is X h (j)  
(i) (i)
perfect and the noise does not exist, which is impossible. As P̄t−1|t−1 = Pt−1|t−1 + x̄t−1|t−1 (26)
a result, we propose a heuristic solution: when Rt is negative, j
we set Rt to be 0.001.  T 
(i) (j) (i|j)
In the next section, multiple model algorithms are being x̄t−1|t−1 − x̂t−1|t−1 µt−1
considered, which can be used to run the Kalman filters in (j) (j)
parallel and switch among these models based on a transition where x̂t−1|t−1 and Pt−1|t−1 are the updated state and
probability matrix. covariance from the j th Kalman filter in time step (t−1).
C. The Interacting Multiple-Model (IMM)
In reality, time series is sophisticated to predict and it may 2) Implementation of the multiple models:
be modeled with only one single Kalman filter even though
the estimation is good. Furthermore, when the Kalman filter is • The predicted state mean and covariance
used for prediction, we generally regard the covariance Q, R in (i)
This step consists of a state prediction, x̂t|t−1 and an
the state space model as unknown constant, which means we (i)
estimate of the covariance of the prediction error, Pt|t−1 .
do not figure out a self-adaption method to update the value
of Q and R. (i)
x̂t|t−1 = Ax̄t−1|t−1
(i)
(27)
In order to solve this issue, we shall utilize the Interacting
(i) (i)
Multiple Model (IMM) in which several Kalman filters with Pt|t−1 = AP̄t−1|t−1 AT +Q (28)
different values of Q and R are taken into consideration.
where i represents the ith Kalman filter in the IMM
The IMM is an estimator which uses several Kalman filters
model.
running in parallel and each filter is utilized to describe
• The Kalman Gain
a different model of volatility. Meanwhile, Discrete-Time
Furthermore, the system state estimation could be ob-
Markov Jump linear systems are applied in order to achieve
tained by first calculating the Kalman gain.
the goal of switching among different modes. Suppose the
−1
discrete locations are represented by rt ∈ {1, ...N } where T
 T
(i) (i) (i)
Kt = Pt|t−1 H (i) H (i) Pt|t−1 H i + R (29)
N is the number of modes, hence the state space model of
Kalman filter can be transformed into: • The modified state mean and covariance
 
(i) (i) (i) (i) (i)
Xt+1 = At (rt )Xt + wt (rt ) (21) x̂t|t = x̂t|t−1 Pt|t + Kt y t − H (i) x̂t|t−1 (30)
yt = Ht (rt )Xt + vt (rt ) (22)
3) Update model probabilities [22]: In this step, we ma-
To apply Kalman filter in IMM, the estimation algorithm can jorly use Bayes Rule to compute the updated model probabil-
(i)
be expressed in terms of four steps: ities µt
(i) (i)
(i)
µ̂t|t−1 Lt
(i) (i) µt = P (31)
1) The Mixed state x̄t−1|t−1 and covariance P̄t−1|t−1 : To (j) ˆ (j)
(i|j) µ
j t|t−1 tL
calculate the mixed state and covariance, mixing weight µt−1
and transition probability from mode i to j πtij , need to be first where Lt is the conditional probability of mode i, represented
obtained. as model likelihoods. Which is given by
  T  −1   is applied here [24], which treats transition probabilities as
− 12 (i) (i) (i) random variables. In this case, the estimation function can be
e ỹt St ỹt
(i)
Lt =  1 (32) represented numerically by using the stocahstic gradient desent
(i)  2 algorithm:
det 2πSt 

n
(ij) (ij)
π̂t =P π̂t−1 +
(i) (i)
where ỹtis the measurement residual, and St is the residual (ij)

εt P (yt |rt = j, {yk }t−1 t−1
k=1 )π̂t−1 P (rt−1 = i|{yk }k=1 )
 (40)
covariance, which are calculated as: PN t−1 PN (ij) t−1 
T j=1 P (yt |rt = j, {yk }k=1 ) i=1 π̂t−1 P (rt−1 = i|{yk }k=1 )
(i) (i)
St = H (i) Pt|t−1 H (i) + R (33)
(i) (i) where εt is the learning rate, P{·} represents an operator in
ỹt = yt − H (i) x̂t|t−1 (34)
order to limit the updated probabilities so that:
4) Updated state mean and covariance: From the updated N
(i) (ij) (ij)
X
computed in the preceding paragraph, the updated state x̂t|t π̂t = 1 and π̂t ≥0 (41)
(i) j=1
and updated covariance Pt|t and they can be computed by the
following equations Then we need to simplify this formula to fit our model, then
n the estimation function can be written as:
(i) (i)
X
x̂t|t = µt x̂t|t (35) (j) (i)
Lt µt−1
(ij) (ij)
t−1 ω̂t = π̂t−1 + εt PN (j) PN (ij) (i)
(42)
n    T  j=1 Lt i=1 π̂t−1 µt−1
(i) (i) (i) (i)
X
Pt|t = Pt|t + x̂t|t − x̂t|t x̂t|t − x̂t|t µt (36) (ij)
(ij) ω̂
t−1 π̂t = PN t (ij)
(43)
j=1 ω̂t
5) Determine the transition probability matrix (TPM): In
(ij)
the IMM algorithm, the TPM, denoted by πt , is a very where Lt and µt are as defined in previous section.
(ij)
important component. Since the true value of πt is hard
6) Application in our model: Our IMM model will have 4
to define, we use an estimator to represent its value. In this
Kalman filters with different Qt and Rt , the estimated self-
report, we focus on a time-variant TPM because a constant
adaptive Rt is mentioned in previous section and the estimated
one is not accurate for the whole process as those parameters
value Qt will be discussed in the following.
in our model updated over time. Because of the popularity
of multiple-model, there are lots of methods in estimating the 7) Determine the value of Qt : The process noise covariance
TPM because it largely influences the result of multiple-model Qt cannot be measured since it cannot be observed, which
algorithm. Recursive minimum mean square error estimation is makes the estimation of it difficult. Hence we regard Qt as
one of the most popular methods and quasi-Bayesian estimator an unknown constant, we need to adjust its initial value to
is a typical example of it, which is defined as [23]: get a more accurate estimation. However, from our previous
(ij) experiments, there are enormous combinations of values that
(ij) θt
π̂t = (37) Qt could take. Furthermore, the model will be less feasible if
t+1
the users need to adjust Qt every time when they estimate
where a new time series. Hence, we decide to determine several
(ij) (ij) sets of combinations of Qt in order to reduce the frequency
(ij) (ij) θ γt
θt = θt−1 + PN t−1 (ij) (ij)
(38) of adjusting Q. The table I shows part of our predetermined
θt−1 γt value.
 j=1  T 
(i) (j) (i)
µt Lt − π̂t−1 Lt
(ij)
γt = 1 + (39)
µTt π̂t−1 Lt TABLE I
C OMBINATIONS OF Q
(ij)
In the equations, π̂t represents the estimated transition
probability from the ith Kalman filter to the jt Kalman Q1 Q2 Q3 Q4
(1) (2) (N ) 0.06 0.03 0.005 0.002
filter in IMM at time t, Lt = [Lt Lt · · · Lt ]T and
(1) (2) (N ) T 0.05 0.01 0.001 0.008
µt = [µt µt · · · µt ] denote model likelihoods and
probabilities in IMM for Kalman filtering algorithm, π̂ (i) = 0.1 0.3 0.01 0.08
(i1) (i2) (iN )
[π̂t π̂t · · · π̂t ]T represents the transition probability 0.06 0.03 0.001 0.008
vector of mode i and hence the TPM can be written as 0.1 0.03 0.01 0.08
(1) (2) (N )
π̂ = [π̂t π̂t · · · π̂t ]T . 0.1 0.01 0.001 50
However, the divergence between the TPM estimator and its 0.1 0.01 0.001 0.0001
true value still exists. In order to reduce this divergence, an-
other method called recursive Kullback-Leibler (RKL) method
 III. R ESULT & A NALYSIS TABLE III
O UR MODEL’ S RESULTS
A. Our Model
We focus on setting up a model which contains six different item number i0.1
IMM following AR models with order p from 4 to 9, where, 16110001 10
in each IMM, there are four different modes. Because we run 16110002 14
six different IMM simultaneously in one round, we will have 16110003 18
six different predicted values in each time step t. We shall use 16120001 8
the Mean Square Error(MSE) to determine which predicted y 16120005 10
value will be used . The formula of MSE is:
PN
(yt − ybt )2
M SE = N =n+1 (44) C. Multivariable
N −n+1
Besides sales volume, we also consider the influence of
To be more precise, we calculate the MSE of the prediction in other factors, such as weekend, weather and festival. However,
time step t − 2 to t − 1 and based on its value, we choose the during our experiments, we found out that the impact of
one with the smallest MSE from the predictions to become weather is trivial, thus, we only take festival and weekend into
final result in time step t. account. We regard them as binary variables, if true, value is
B. Parallel IMM 1, if not, value is 0.
In order to test the efficiency of multivariable, we decide to
In this part, we aim to compare the effectiveness between compare the results of our model under the circumstances of
our model and the traditional model. The traditional model is using multivariable and without using multivariable. To better
the single IMM following AR model with order p, with four illustrate our result, here we use Pi , which represents the
different KF in the IMM. percentage of i0.1 in 30 days, denoted as Pi = i0.1 /30.
We use 108 time series from a cake shops daily sales
of different products for the past two years. Initially, the D. Analysis
effectiveness of our classic models are tested. In this part,
In the table IV, we only put part of our results because of the
we only use the products previous sales volume to predict
large volume of data. From those output, we can figure out that
its future data, regardless other factors, such as weather and
even though there is a small proportion of products perform
holidays.
better without multivariable, the majority of products have a
Though many approaches can be used to test the effective-
more accurate result if considering multivariable. Moreover,
ness of the model, we choose the accuracy i0.1 to do the
the condition is same for the overall result. The total average
evaluation. The notation i0.1 denotes the number of outputs
of parallel IMM with multivariable is that i0.1 equals to 10.31
that fluctuate within 10% of the true sales in the last 30 days.
and P =34.38%, which is slightly bigger than single-variable
Since there are four different KF in the IMM, we provide
case, which is i0.1 equals to 10.06 and Pi =33.54%. The value
four possible values of Qt and set the same Qt for each time
of Pi increase by 2.5% after applying multivariable. Therefore,
series as the follows:
we can conclude that, according to our experiments, taking
Q 0.1 0.01 0.001 0.0001 festival and weekend into account will improve the overall
performance of our estimation.
Some related results are listed in the table II.
TABLE IV
MULTIVARIABLE & SINGLE VARIABLE
TABLE II
T RADITIONAL MODEL’ S RESULTS
multivariable single variable
i0.1 item number i0.1 Pi i0.1 Pi
item number p=4 p=5 p=6 p=7 p=8 p=9 16110001 11 36.67% 10 33.33%
16110001 4 2 1 5 6 8
16110002 7 6 6 10 11 10 16130001 23 76.67% 20 66.67%
16110003 11 10 11 12 11 10 16130013 13 43.33% 11 36.67%
16120001 3 4 8 6 4 6
16120005 9 5 4 6 7 8 16130017 20 66.67% 21 70.00%
16140003 17 56.67% 16 53.33%
Then, we run the 6 models with different p in parallel. The 16130014 10 33.33% 9 30.00%
table III shows the relevant results. 16130003 13 43.33% 12 40.00%
From the above two tests, we can see the second tests 16130004 9 30.00% 8 26.67%
results is better than the first one, which proves that running 16130005 7 23.33% 8 26.67%
these models of different orders in parallel is more effective 16130006 15 50.00% 14 46.67%
than traditional means. In the next part, we will consider 16170009 10 33.33% 10 33.33%
multivariable to improve our model.
E. Discussion In this case, Figure 2 is used to compare the efficiency of
using multivariable and without using multivariable, fig 2(a)
We use the criteria of i0.1 to compare the prediction shows the model with single variable and fig 2(b) depicts the
efficiency of our model with traditional model, the overall multivariable model. From these two pictures, the usage of
average will increase by about 60% to 250% according to multivariable will improve the effectiveness.
the equation (1) after applying our model, followed are part
of results: IV. C ONCLUSIONS
i0.1 (parallel IMM) − i0.1 (single IMM with order p) In this paper, we mainly use the methods of (1) Parallel
×100%
i0.1 (single IMM with order p) IMM with different orders (2) self-updated parameters (3)
(45) Multivariable to improve the accuracy and efficiency of time
series forecasting based on state space model. The prediction
used 2 year’s data of daily sales of a cake shop to evaluate
Fig. 1. Parallel IMM with updating orders VS Single IMM the performance of Parallel IMM and Multivariable which are
based on self-updated variance compared with traditional state
space model. it can be concluded from the results that both
three methods made the state space model showed relatively
better efficacy and accuracy when it comes to forecasting, as a
result, predicted values will be more reliable when conducting
under these improvements
ACKNOWLEDGMENT
(a)
This study is supported by the National Natural Science
Foundation of China (Grant No. 61501380).
R EFERENCES
[1] S. Kendall and K. Ord, Time Series. Oxford University Press, 1990.
[2] H. Akaike, “Markovian representation of stochastic processes and its
application to the analysis of autoregressive moving average processes,”
Annals of the Institute of Statistical Mathematics, vol. 26, no. 1, pp.
363–387, 1974.
(b) [3] C. K. . I. Waeto, S, “Forecasting time series movement direction with
hybrid methodology,” Journal Of Probability & Statistics, pp. 1–8, 2017.
[4] H. Yuan, J. Liu, H. Pu, J. Mao, and S. Gao, Prediction of Chaotic Fer-
roresonance Time Series Based on the Dynamic Fuzzy Neural Network.
Figure 1 uses the last 80 data of one product, fig 1(a) shows New York, NY: Springer New York, 2012.
[5] N. Karunanithi, D. Whitley, and Y. K. Malaiya, “Prediction of software
the model with single IMM and fig 1(b) depicts the parallel reliability using connectionist models,” IEEE Transactions on Software
model. From these two pictures, paralleled model perform Engineering, vol. 18, no. 7, pp. 563–574, 1992.
better than another one. [6] Q. Hu, M. Xie, S. Ng, and G. Levitin, “Robust recurrent neural
network modeling for software fault detection and correction prediction,”
Reliability Engineering & System Safety, vol. 92, no. 3, pp. 332 – 340,
Fig. 2. Multivariable VS Single Variable 2007.
[7] L. Wang, K. K. Teo, and Z. Lin, “Predicting time series with wavelet
packet neural networks,” in Neural Networks, 2001. Proceedings. IJCNN
’01. International Joint Conference on, vol. 3, 2001, pp. 1593–1597
vol.3.
[8] K. K. Teo, L. Wang, and Z. Lin, Wavelet Packet Multi-layer Perceptron
for Chaotic Time Series Prediction: Effects of Weight Initialization.
Berlin, Heidelberg: Springer Berlin Heidelberg, 2001, pp. 310–317.
[9] Y. Fang, K. Fataliyev, L. Wang, X. Fu, and Y. Wang, “Improving the
genetic-algorithm-optimized wavelet neural network for stock market
prediction,” in 2014 International Joint Conference on Neural Networks
(a) (IJCNN), 2014, pp. 3038–3042.
[10] K. Xu, M. Xie, L. Tang, and S. Ho, “Application of neural networks in
forecasting engine systems reliability,” Applied Soft Computing, vol. 2,
no. 4, pp. 255 – 268, 2003.
[11] S. Gupta and L. Wang, “Stock forecasting with feedforward neural net-
works and gradual data sub-sampling,” Austr. J. Intelligent Information
Processing Systems, vol. 11, no. 4, 2010.
[12] R. Ghazali, A. J. Hussain, D. Al-Jumeily, and P. Lisboa, “Time series
prediction using dynamic ridge polynomial neural networks,” in 2009
Second International Conference on Developments in eSystems Engi-
neering, 2009, pp. 354–363.
[13] R. S. Sexton and R. E. Dorsey, “Reliable classification using neural net-
(b) works: a genetic algorithm and backpropagation comparison,” Decision
Support Systems, vol. 30, no. 1, pp. 11 – 22, 2000.
[14] B. Safarinejadian, M. A. Tajeddini, and A. Ramezani, “Predict time
series using extended, unscented, and cubature kalman filters based
on feed-forward neural network algorithm,” in The 3rd International
Conference on Control, Instrumentation, and Automation, 2013, pp.
159–164.
[15] S. J. Julier and J. K. Uhlmann, “Unscented filtering and nonlinear
estimation,” Proceedings of the IEEE, vol. 92, no. 3, pp. 401–422, 2004.
[16] I. Arasaratnam and S. Haykin, “Cubature kalman filters,” IEEE Trans-
actions on Automatic Control, vol. 54, no. 6, pp. 1254–1269, 2009.
[17] G. Kitagawa and W. Gersch, “A smoothness priorsstate space modeling
of time series with trend and seasonality,” Journal of the American
Statistical Association, vol. 79, no. 386, pp. 378–389, 1984.
[18] X. Bi, J. Du, J. Gao, W. Wang, and Y. Gao, “The imm tracking algorithm
for maneuvering target with adaptive markov transition probability
matrix,” in 2015 IEEE 10th Conference on Industrial Electronics and
Applications (ICIEA), 2015, pp. 1279–1282.
[19] S. W. Qingling Wu, Time series analysis. Changsha: Hunan Science
and Technology Press, 1989.
[20] R. Kalman, “A new approach to linear filtering and prediction problems,”
Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45, 1960.
[21] B. Gary and W. Greg, “An introduction to the kalman filter,” Proc of
SIGGRAPH, Course, vol. 8, no. 27599–23175, p. 41, 2001.
[22] J. Svec and X. Katrak, “Forecasting volatility with interacting multiple
models,” Finance Research Letters, vol. 20, pp. 245–252, 2017.
[23] P.Vesselin, X.Jilkov, and L. Rong, “Online bayesian estimation of
transition probabilities for markovian jump systems.” IEEE Transactions
on Signal Processing, vol. 52, no. 6, pp. 1620–1630, 2004.
[24] R. Guo, M. Shen, D. Huang, X. Yin, and L. Xu, “Recursive estimation of
transition probabilities for jump markov linear systems under minimum
kullback;leibler divergence criterion,” IET Control Theory Applications,
vol. 9, no. 17, pp. 2491–2499, 2015.

View publication stats