A Modified ARIMA Model Based on Extreme
Value for Time Series Modelling
Yaohui Bai
Cloud Computation and Big Data Research Center
Jiangxi University of Finance and Economics
Nanchang, China
e-mail: byhnpu@163.com
Benting Wan
School of Software & Communication Engineering
Jiangxi University of Finance and Economics
Nanchang, China
e-mail: ren_btw@163.com
Xixia Zong
Xi'an Aerospace Propulsion Test Institute
Xi'an, China
e-mail:zxx_0126@163.com
Wenyuan Rao
School of Software & Communication Engineering
Jiangxi University of Finance and Economics
Nanchang, China
e-mail: raowy2002@gmail.com
Abstract—The ARIMA model is an important method and is
widely used in time series modelling. The model relies
heavily on autocorrelation patterns in the data, and doesn't
consider other factors. However, in most cases, the extreme
value of series has an influence on the subsequent behaviour
of series. But this information isn't considered in the original
ARIMA model. To solve this problem, We proposes a
modified ARIMA model based on the past maximum and
minimum value of series to solve the modelling tasks which
includes the factor of past extreme value in the model. The
modified model is tested on USD-EUR exchange rate time
series. The experimental results show that it is possible to
improve the performance by considering extreme value for
time series modelling compared to the original ARIMA
model.
Keywords- Time Series; modelling; ARIMA; extreme value;
I. INTRODUCTION
A time series data is a sequence of numerical
observations naturally ordered in time. Recently, the rapid
development of information technologies have led to the
situation in which huge amounts of information data
accumulate in quick speed and in fact constitute various
time series. The modelling of such time series is extremely
important and vital, and has been attracting the attention of
both practitioners and researchers. However, it is also
considered a rather difficult problem, due to the many
complex features frequently present in time series, such as
irregularities, volatility, trends and noise, and so on. A
number of techniques have been developed in an attempt
to model time series based on their present and past
behaviour.
Traditional time series modelling technologies, such as
autoregressive integrated moving average(ARIMA)[1],
exponential smoothing[2][3], decomposition[4], etc., have
been widely and successfully used. More recently a
number of machine learning techniques, such as neural
networks[5~7], fuzzy systems[8], genetic algorithm[9],
and SVM[10] are becoming promising directions in this
fields. Some showed improvement compared to traditional
models. Although a number of modern intelligent
technologies can be available in machine learning and
pattern recognition, it has been rarely used in temporal
aspects of time series modelling for the discrete values
output.
Among them, ARIMA model is considered the most
common choices for the time series modelling, which was
popularized by George Box and Gwilym Jenkins in the
early 1970s. The model only use the information in the
series itself to model, and doesn't include other
independent variables. However, the model is linear, and
can be used to model stationary as well as non-stationary
time series. With the widely application of the ARIMA
models, there are a number of variations on the ARIMA
model. For example, for the multiple time series, a
VARIMA model may be appropriate, and for the seasonal
effect, it is generally better to use a SARIMA.
Although the ARIMA model has been widely applied
in time series modelling, the model has some shortcomings.
The model rely heavily on autocorrelation patterns in the
data, and doesn't consider other factors. However, in most
cases, the extreme value of series has an influence on the
subsequent behaviour of series. Therefore, in these cases,
when we consider the model , we should include the
extreme value of series into the model. This work proposes
a modified ARIMA model based on the past maximum
and minimum value of series to solve modelling tasks.
II. METHODOLOGY
A. Time Series
A time series is a set of data points, measured typically
at successive times spaced at uniform time intervals, and
defined by ,
1
,
2, ,
tt
X x R t N
International Conference on Mechatronics, Electronic, Industrial and Control Engineering (MEIC 2015)
© 2015. The authors - Published by Atlantis Press