Page 301 - Contributed Paper Session (CPS) - Volume 7
P. 301
CPS2105 Hermansah et al.
past and present values of variables to produce accurate short-term
forecasting, very well used to look at past patterns and then represent future
patterns for forecasting. ARIMA is a stochastic method that is very useful for
generating time series data where each event is correlated [3].
Developing of next is the Neural Network model. Neural Network (NN) is
an information processing system that has characteristics similar to biological
neural networks [4]. The NN forecasting method for its application is the Feed-
Forward Neural Network (FFNN) or often called the Back-propagation Neural
Network (BPNN). BPNN is the NN method with a network that uses errors to
change the value of its weights in the backward direction. To get this error, the
forward propagation stage must be done first. This network has an activation
function in two layers, namely the hidden layer and output layer. The
procedure for establishing BPNN begins with the selection of input variables
by looking at a plot of the Autocorrelation Function (ACF) or Partial
Autocorrelation Function (PACF) that is significant and determines the target
variable. The second stage is data sharing to training data and testing data.
Next is the determination of learning parameters. The formed network is
selected from the results of training and testing by looking at the smallest
Mean Squared Error (MSE) and the smallest Mean Absolute Percentage Error
(MAPE) [5].
In the past two decades, other techniques that are widely used are
wavelets or wavelet transformations. The use of wavelet transformation as an
analytical tool is caused partly because it has advantages in the process of
denoising, data compression, and multiresolution [6]. In this study the wavelet
method to be used is the Maximal Overlap Discrete Wavelet Transform
(MODWT) with the Daubechies 4 wavelet type. MODWT is considered more
suitable for time series data because in each decomposition level there are
wavelet coefficients and scale coefficients as many data lengths. This
advantage reduces the weakness of filtering with Discrete Wavelet Transform
(DWT) which cannot be done in any sample size [7]. Determination of
decomposition levels and coefficients used as input models using multi-scale
decomposition. The results of MODWT will be obtained smooth coefficients
and detail coefficients then smooth coefficients and detail coefficients will be
used to obtain the value of Multiscale Autoregressive (MAR) that will be used
for forecasting.
The purpose of this study is to compare the accuracy of forecasting using
the ARIMA, NN and Wavelet models. The case study used close of Indonesia
Sharia Stock Index (ISSI) data from September 4, 2017 to September 19, 2018.
The selection of the right model is very necessary to predict ISSI, so that an
action can be taken or a decision is made. The next three models will be used
to forecast the ISSI with the smallest MSE and MAPE values showing the best
performance.
288 | I S I W S C 2 0 1 9