Humanomics

Treating return of mudharabah time deposit as investment instrument: A utilization of

artificial neural networks (ANNs)
Saiful Anwar, Dadang Romansyah, Sigit Pramono, Kenji Watanabe,
Article information:
To cite this document:
Saiful Anwar, Dadang Romansyah, Sigit Pramono, Kenji Watanabe, (2010) "Treating return of mudharabah
time deposit as investment instrument: A utilization of artificial neural networks (ANNs)", Humanomics, Vol.
26 Issue: 4, pp.296-309, https://doi.org/10.1108/08288661011090893
Permanent link to this document:
https://doi.org/10.1108/08288661011090893
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

Downloaded on: 25 May 2018, At: 00:42 (PT)

References: this document contains references to 24 other documents.
To copy this document: permissions@emeraldinsight.com
The fulltext of this document has been downloaded 1456 times since 2010*
Users who downloaded this article also downloaded:
(2011),"A comparative study of the returns on Mudhārabah deposit and on equity in Islamic banks",
Humanomics, Vol. 27 Iss 4 pp. 229-242 <a href="https://doi.org/10.1108/08288661111181288">https://
doi.org/10.1108/08288661111181288</a>
(2012),"A critique on accounting for the mudarabah contract", Journal of Islamic Accounting and
Business Research, Vol. 3 Iss 1 pp. 7-19 <a
href="https://doi.org/10.1108/17590811211216032">https:// doi.org/10.1108/17590811211216032</a>

Access to this document was granted through an Emerald subscription provided by emerald-srm:600462 []
For Authors
If you would like to write for this, or any other Emerald publication, then please use our Emerald for
Authors service information about how to choose which publication to write for and submission guidelines
are available for all. Please visit www.emeraldinsight.com/authors for more information.
About Emerald www.emeraldinsight.com
Emerald is a global publisher linking research and practice to the benefit of society. The company
manages a portfolio of more than 290 journals and over 2,350 books and book series volumes, as well as
providing an extensive range of online products and additional customer resources and services.
Emerald is both COUNTER 4 and TRANSFER compliant. The organization is a partner of the Committee
on Publication Ethics (COPE) and also works with Portico and the LOCKSS initiative for digital archive
preservation.

*Related content and download information correct at time of download.

The current issue and full text archive of this journal is available at
www.emeraldinsight.com/0828-8666.htm
 H
26,4 Treating return of mudharabah
time deposit as investment
instrument
296 A utilization of artificial neural
networks (ANNs)
Saiful Anwar
Department of Management and Information System Science,
Nagaoka University of Technology, Niigata, Japan, and
School of Finance and Banking, STIE Ahmad Dahlan, Ciputat, Indonesia
Dadang Romansyah and Sigit Pramono
Department of Islamic Accounting, School of Islamic Economic (SEBI),
STIE Ahmad Dahlan, Ciputat, Indonesia, and
Kenji Watanabe
Department of Disaster Management and Safety,
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

Nagoya Institute of Technology, Nagoya, Japan

Abstract
Purpose – The purpose of this paper is to propose the development of return forecasting model for
mudharabah time deposit product in Islamic bank based on artificial neural networks (ANNs).
Design/methodology/approach – The analysis consists of two main elements. First element is
the identification and selection of significant macroeconomic variables that determine return
volatility of mudharabah time deposit in Indonesian Islamic bank industry. Second element is the
implementation of appropriate ANNs model according to neural networks properties, and model
evaluation based on simulated return predictions of mudharabah time deposit product in Bank
Syariah Mandiri (RR-BSM). Findings – It is shown that monthly changes of return can be predicted
quite well. The model provides a satisfactory result in forecasting RR-BSM for 12 months ahead
with 95.22 per cent accuracy. These results suggest that the ANNs can be applied as an adequate tool
to help depositors in predicting future return of mudharabah time deposit product.
Originality/value – There is believed to be no other empirical study of Islamic banks that
exclusively examines the utilization of ANNs to forecast time deposit return as well as return
from other investment instruments.
Keywords Islam, Banks, Indonesia, Macroeconomics, Return on investment, Neural nets
Paper type Research paper

1. Introduction
Accumulating studies have been reported that the motive of depositors in patronizing
the Islamic bank are similar with motive of depositors in patronizing the conventional
bank. Erol and El-Bdour (1989) have reported that religion is not the main reason
for depositors to select Islamic banks in Sudan and Turkey. The same finding was
confirmed also by Haron and Ahmad (2000) that most depositors in Malaysia are
motivated more by profit than religion. In the same view, Mangkuto (2004) in
Humanomics
Vol. 26 No. 4,
Indonesia has found positive correlation between deposit return and level of deposit
2010 pp. 296-309 in Bank Muamalat Indonesia. In accordance with other studies, applying vector
# Emerald Group Publishing
Limited
autoregressive
0828-8666 and impulse response function analysis, Kasri and Kassim (2009) found that
DOI 10.1108/08288661011090893
Indonesian Islamic banks deposit is significantly correlated with higher rate of return
 and lower interest rate which indicates an existence of profit macroecono e. M
motive. In brief, the motive is confirmed by Keynes theory (1936) mic Furtherm D
which said that return from deposit that he called interest is the condition ore, A
most dominant motive for people to save their money, since (Ahmad Dewaelhe ’
saving is not only intended to meet future consumption but also and Haron, yns and s
as an investment (Kasri and Kassim, 2009). 1998). This van Hulle a
Unlike interest that provides fix regular revenue, the return is (2007) c
of mudharabah time deposit which represented by monthly rate supported and c
of return (RR), is uncertain. In addition, the Islamic banks publish by Bashir Erdogan u
the RR report to assist depositors comparing the current RR (2000) who (2008) r
with current market interest rate, which the rate is written in confirmed used a
interest rate equivalent. The RR will vary among Islamic banks that GDP distribute c
since it depends on the bank’s profitability and pre-agreed profit and d lag y
and loss sharing (PLS) ratio. inflation model i
Our study attempts to fully reveal the application of artificial rate and logit s
neural networks (ANNs) on the future RR prediction as well as determine to i
on return prediction from other investment instruments such as financial forecast n
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

stock, bonds, etc. It will help depositor to anticipate loss performance probabilit t
probably happened due to bad financial performance of the of Islamic y of h
Islamic bank. In order to do so, this banks in default in e
study examines the return prediction accuracy of time Middle East conventio r
deposit product called countries. nal bank. a
mudharabah investment account in Bank Syariah Mandiri Many In n
(BSM) as a case study. methodolog accordanc g
We justify the importance of our study based on the ies have e with the e
following reasons. First, the fund of time deposit product been above o
contributes approximately 55.2 percent from the total depositor utilized in studies, f
fund in Indonesian Islamic bank during March 2008 to financial Atiya 5
February 2010. Moreover, the one-month time deposit which performanc (2001) 9
offers the highest rate, contributes the largest portion in the e analysis summariz .
total fund of time deposit for approximately 66.03 percent and es paper 3
according to Bank of Indonesia. In accordance with prior studies, prediction. of which -
these reports suggest that the most important reason for people For compares 7
in patronizing Islamic bank in Indonesia is the expectancy in example, forecastin 0
receiving better return. Second, Islamic bank offers mudharabah Meyer and g power .
or Pifer (1970) between 4
PLS contract for time deposit product. Consequently, the RR used linear ANNs p
given to depositor will be based on the bank profitability. probability and MDA e
Explicitly, if any loss happened to the bank, it will be borne model to in r
proportionally to depositor. predict bank predicting c
failure. Al- bankruptc e
2. Osaimy and y of 128 n
Li Bamahrama firms. In t.
ter h (2004) fact, B
at and Cihak ANNs e
ur and Hesse achieve s
e (2008) used correct i
multi- classificat d
re
discriminan ion e
vi
t analysis accuracy s
ew in the ,
2.1 Previous methods used in financial (MDA) to
analyze range of A
performance prediction
Islamic 77.8-81.5 z
Under PLS principle, the Islamic bank delivers return to
bank percent, i
depositors based on the pre- agreed PLS ratio. It means that the
performanc while s
return on deposit is fluctuating depends on the profit and
and Dar (2004) review 89 studies on corporate bankruptcy Mudhara
prediction. They confirmed that ANNs achieve 88 percent
accuracy rate in average while traditional statistical models bah
achieve 84 percent accuracy rate. time
deposit

2
9
7
 H Althoug predictive variables (Ravi et al., 2008). Moreover, Moody
h neural (1995) has reported that
2 networks macroeconomic variables are characterized as nonlinear time
6 have series data that violate the above assumptions. Therefore, in
, Figure 1.
demonstrat our recent study, Alyuda NeuroIntelligence
4 A model ed some software is employed to develop
of neuron successes neural networks model.
in this area, Actually, ANN is a part of artificial intelligence that is
only a few powerful to solve problem
studies especially with regard to pattern classification and recognition.
such as Al- It is a computational model which the structure and function
2 Osaimy mimic biological neuron in the human brain.
9 (1998) and Moreover, ANNs consist of a group of artificial neurons, which
Hall et al. are interconnected. Every single neuron processes information
8 (2008) (receiving input and delivering output) using a special algorithm
already function.
employed Serju (2002) mentioned two advantages of ANNs in modeling
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

ANNs to do the relationship between independent and dependent variables.

forecasting First, ANNs are universal approximators in that
in Islamic they can approximate whatever functional form to characterize the
banking time series data. It means ANNs perform as data driven rather than
research. model driven (Argyrou, 2006). Therefore,
they are very suitable for problems which data are available but
2 the underlying theoretical model is not known (Zhang et al., 1998).
. Second, unlike traditional statistical methods which
2 need a prior assumption regarding normality of data distribution
(Zhang et al., 1998), ANNs are superior to traditional statistical
methods due to their ability to deal with nonlinear data and multi-
A dimensional aspect. As a result, ANNs have been proven for better
N forecasting in
N long-term periods, but as good as traditional statistical
s methods in shorter periods.
Technically, the ANN process is shown in Figure 1. In the
m figure, there is a neuron j
o which has a certain number of inputs (x1, x2, x3, . . . , xj) and
d single output (yj). Each input has a weight (w1j, w2j, w3j, . . . , wij)
e as an indicator of importance of the incoming
l
Most of
statistical
technique
such as
logistic and
regression
analysis
relies on
strict
assumption
of linear
separability
,
multivariate
normality
and
independen
ce of
 signal to neuron ( j). The net value (uj) of the neuron is then calculated with the sum of Mudharabah
all the input values that multiplied by their specific weight.
Further, with reference to threshold value (tj) and activation function, the time deposit
neuron ( j) determines output value (yj) will be sent as output to other neuron. Each
neuron has its own unique threshold value (tj). If the net value (uj) is greater than the
threshold (tj), the neuron ( j) will send output (yj) to other neurons. In addition,
activation function is a function used to transform the activation level of a unit
(neuron) to an output signal. Currently, sigmoid and logistic are the most popular 299
activation functions used.
However, a single neuron is not useful to solve the problems. The research needs to
combine some neurons into multilayer structured which so called neural networks to
have the power for answering pattern classification and recognition problems. For
these purposes, this research employs a multi-layer feed-forward network which is the
most common type of neural network currently in use. Moreover, Figure 2 shows
(4-3-1)
neural networks with four-three-one network architecture (in abbreviated form, N )
that consists of one input layer with four neurodes, one hidden layer with three
neurodes and one output layer with one neurode. Every neuron in the layer works
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

with the same way as explained before.

Specifically, the input layer is a layer that is directly connected to outside
information. All data in the input layer will be feed-forward to hidden layer as the
next layer. Meanwhile, the hidden layer functions as feature detectors of input signals
and releases them to output layer. Finally, the output layer is considered as a
collector of the features detected and producer of the response. In the
networks, output from output layer is the function of the linear combination of
hidden unit’s activation; the hidden unit’s activation function is a nonlinear function
of the weighted sum of inputs.
Azadeh et al. (2007) explained the mathematical model of ANNs as in the following:

y ¼ f ðx; Þ þ " ð1Þ

where x is the vector of explanatory variables, u is weights vector (parameters) and

" is the random error component. Then, Equation (2) is the unknown function for

Figure 2.
A multi-layer
feed-forward network
 H estimation and prediction from the available data. As such, the model can be
26,4 formulated as:
" ! #
X
m nX
Y ¼ f v0 þ h jþ xi wij ð2Þ
vj
j¼1 i¼1

300 where Y is the network output, f the output layer activation function, v0 the output
bias, m the number of hidden units, h the hidden layer activation function, j the
hidden unit biases ( j ¼ 1, . . . , m), n the number of input units, xi the inputs vector (i
¼ 1, . . . , n), wij the weight from input unit i to hidden unit j, and vj the weights from
hidden unit j to output ( j ¼ 1, . . . , m).

3. Methodology
3.1 Data
In data collection, this research follows the bank failure theory issued by Gambs
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

(1977) which said, ‘‘Extremely bad management may not prove fatal to a bank until
economic condition leads to unexpected capital outflows or loan losses.’’ Therefore,
this research uses only macroeconomic variables to examine the volatility of RR.
Many researchers utilize macroeconomic variables in their empirical studies to
investigate financial performance of the firm. However, there are many types of
macroeconomic variables and difference number of the variables used in the
researches. Meanwhile, there is no consensus on the issue regarding the type and
number of variables should be considered when making judgments. Different
researcher produces different model using different set of macroeconomic variables.
Accordingly, the research performs two examinations as shown in Figure 3. The
first examination is carried out by selecting some of the eight macroeconomic
variables as the significant variables in affecting volatility of mudharabah RR of one-
month time deposit in Indonesian Islamic bank (RR-ALL). The eight
macroeconomic variables consist of exchange rate USD to Indonesian Rupiah
(EXCH), Jakarta stock indices (STIN), money supply as measured by M1, inflation
rate (INFR), average of one-month time deposit interest rate (INTR) of Indonesian
conventional bank, Central Bank
interest rate (BIRT), international gold price in USD per ounce based on London fix
price (GOLD) and average prices of Brent crude oil and West Texas Intermediate
crude oil in USD per barrel (OIL). These macroeconomic variables, except for GOLD
and OIL are collected from Indonesian central bank. Meanwhile, the GOLD and OIL
data are

Figure 3.
Research framework
 collected from World Gold Council and US Energy Information and Administration. Mudharabah
For this purpose, all data are collected during the period of January 2001 until
December 2008. time deposit
The second examination is conducted to predict future RR-BSM based on selected
macroeconomic variables resulted from the prior examination. Then, prediction
evaluation is carried out, based on appropriate model resulted from the networks
using new data which are collected during the period of January 2009 until December
2009. 301
In this step, the evaluation is calculated on the basis of standard statistical measures
like percentage errors in the
following:
jRR-actual ði Þ RR-predictionði Þj
%Error of prediction ðiÞ ¼ 100% ð3Þ
RR-actualðiÞ

where i ¼ 1, 2, . . . , n, with n being the number of testing data

points.
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

Finally, the accuracy power of ANNs is evaluated using the following formula:

Accuracy power ¼ 100% % Error of prediction ð4Þ

3.2 Using Alyuda Neurointelligence

The followings are the step-by-step software implementation based on the work of
Argyrou (2006). Actually, Alyuda has six main steps to build robust ANNs models as
follows:
(1) analyzing data;
(2) preprocessing data;
(3) designing the architecture of networks;
(4) training the networks;
(5) testing the robustness of the networks; and
(6) querying new data for out of sample prediction.
The first examination in this research is intended to construct ANNs as selector model
of significant macroeconomic variables in affecting RR-All. Therefore, the
examination will utilize only the first five steps. Moreover, the second examination is
constructed to develop ANNs as predictor model of future RR-BSM, so the
examination will completely employ all steps.
3.2.1 ANNs model as
selector
3.2.1.1 Data analysis. The input to Alyuda has similar form with a spreadsheet. At
the beginning, all data must be uncontaminated from data anomalies, because data
anomalies will give negative impact on neural network performance. According to
user manual, data anomalies are categorized into two conditions namely missing
values and outliers. In particular, missing values are unidentified values that
reflected as blank
cells in the input columns. Meanwhile, outliers are values that lie outside the range.
This analysis uses 108 time series which all of them are accepted for neural network
process.
After that, the column with RR data is set up as output, and the respective variables
are selected as input. Next, all data are partitioned into training, validation and testing
set using random method. As a result, the data are partitioned into: 74 data as training
set, 17 data as validation set and 17 data as test set.
 H 3.2.1.2 Data pre-processing. The next step is to normalizing the input data to make
26,4 it suitable for networks processing. The process simply converts the input data into
a new version before neural networks are trained. Bishop (1995) mentioned three
reasons for the necessity of input data normalization. First, to ensure the size of
input data reflects their importance level on determination of output required. Second,
to facilitate the random initialization of weights before training the network and
third, different variables possibly have different measurements unit; hence, their
302 typical values will be different significantly. In this step, all the input columns are
normalized in the same way since all data are in numerical values.
3.2.1.3 Designing the architecture of networks. With Alyuda, the research runs the
exhaustive search method to find the best possible architecture under following
limitations to avoid over fitting, namely r-squared as fitness criteria and 20,000
iterations.
Actually, this process is time consuming because Alyuda searches the best network
architecture among all possible alternatives in the specified range based on the
following criteria:
. The networks have the smallest training error.
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

. The networks have the smallest test error.

. The networks have the smallest difference between training and testing error.
. The networks have the simplest structure.
(10-3-1)
In this step, Alyuda choose N as networks architecture to select appropriate
macroeconomic variables with the value of r-squared is 0.9515 and the process stops
after 20,000 iterations. To be more specific, the chosen networks architecture as shown
in Figure 4 consist of one input layer with eight active neurons and two neurons as
date which plays no role in training or testing the neural networks, one hidden layer
with three neurodes and one output layer with one neurod.
3.2.1.4 Training the networks. In the training process, the research initially sets up
three configurations as shown in Figure 5. First, the logistic function is selected for all
the neuron in spite of the layer on which they are appeared. Second, the sum-of-
squared
errors are selected to minimize the output errors. The sum-of-squared errors are

Figure 4.
(10-3-1)
N network
architecture
 Mudharabah
time deposit

303
Figure 5.
Hidden layer activation
function, output error and
output activation function
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

summation of squared differences between the actual values and model’s output.
For completeness, the networks outputs are set up between 0 and 1 due to logistic
activation function used.
Then, the model is trained with specific treatment to avoid over fitting as the
following:
. choosing back propagation algorithm as learning algorithm;
. setting the learning and momentum rates at 0.1; and
. stopping the training process when the model’s mean squared error reduces by
less than 0.000001 or the model completes 20,000 iterations, whichever
condition occurs first.
As a result, the networks will generate information about the importance level of
each macroeconomic variable in affecting RR-ALL fluctuation, and at last, the model
completely performs as selector.
In Figure 6, it is shown that INTR, BIRT and M1 rank first, second and
third, respectively, in regard to the highest contribution which have affected RR-ALL
fluctuation.
This result is consistent with prior research conducted by Ahmad and Haron (1998).
They examined 15 Islamic banks from ten Muslim countries using regression analysis
and found
that interest rate, money supply and inflation are the first, second and third significant
variables which affect the return on capital. In principle, the volatility of return on
capital represents the profitability of Islamic bank which also gives impact to depositor
return.
Moreover, Alyuda also provides the importance level information in the form of
percentage as shown in Table I. This information suggests that OIL and GOLD give

Figure 6.
Importance graph of
independent variables
 H very small contribution with level of importance are 0.22 and 0.19 percent,
26,4 respectively. It means that only EXCH, STIN, M1, INFR, INTR and BIRT are selected
to be predictor variables in the second examination.
3.2.1.5 Testing the robustness of the networks. Before going to the second
(10-3-1)
examination, the research needs to investigate robustness of the N as
2
selector. The quality will be represented through value of correlation (r), R , mean of
304 absolute error (AE) and mean of absolute relative error (ARE) of all data sets as shown
in Table II. As a result, the networks have performed very well in selecting
macroeconomic variables.
2
According to Alyuda manual, the value of correlation (r) and R are the indicator of
multiple correlations between independent or predictor variables and dependent or
predicted variable. The coefficients of r can range from 100 to þ100 percent when
the
closer r is to 100 percent, the stronger the positive linear relationship between both
variables. In contrast, the closer r is to 100 percent, the stronger the negative linear
relationship will be. Besides, when r is near 0, it indicates no linear relationship
between independent or predictor and dependent or predicted variables. Meanwhile,
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

2
R is a statistical ratio that compares model forecasting accuracy with accuracy of
the simplest model that just uses mean of all target values as the forecast for all
records. The closer this ratio to 100 percent the better the model is. Moreover, small
positive values near zero indicate poor model. On the other hand, negative values
indicate models that are worse than the simple mean-based model.
Next, the robustness of the networks can also be examined through deviation value
between the predicted output and the desired one. This examination is based on mean
value of AE and ARE, because AE (in absolute value) and ARE are error values that
show the quality of in-sample prediction of the model. It means that the smaller the
error is the better quality of the networks will be.
Another way to test robustness of the networks is through an examination on
prediction vs actual graph based on in-sample data as shown in Figure 7. The figure
suggests that the performance of the model is very good. The quality is shown from
the
location of predicted line where is located very close to the actual
line.

Input column name Importance (%)

INTR 49.713013
BIRT 17.499574
M1 15.681459
STIN 9.735225
EXCH 5.747679
Table I. INFR 1.206882
Importance of OIL 0.221001
independent variables GOLD 0.195167

Data set Correlation (r) (%) R2 (%) Mean of AE (%) Mean of ARE (%)

Table II. Training data set 93.64 85.66 0.361526 0.044

Testing parameter Validation data set 90.91 56.08 0.612538 0.072
for the N (10-3-1) network Testing data set 88.74 58.28 0.6667 0.083
architecture All data set 91.62 77.29 0.449074 0.055
 Mudharabah
time deposit

305

Figure 7.
Prediction vs actual graph
of RR-ALL based on
in-sample data
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

3.2.2 ANNs model as predictor. In general, the second examination has the same
treatment as in the previous one. In this examination, the research uses six chosen
variables as predictor variables resulted from the previous examination namely
EXCH, STIN, M1, INFR, INTR and BIRT. Meanwhile, RR-BSM is used as predicted
variable.
3.2.2.1 Designing the architecture of networks. Using the same treatment for
(8-4-1)
designing the network as used in the first examination, Alyuda provide N
network architecture as shown in Figure 8 to perform accurate prediction. The value of
r-squared in this step is 0.920775 and the searching process stops after 20,000
iterations.
3.2.2.2 Training the networks. After designing the architecture, the networks are
trained with the same treatment as used in the prior examination. In this step, the
information about the importance level of predictor variables on affecting RR-BSM
volatility is generated as shown in Table III. This table shows that INTR, M1 and
STIN are the most significant variables in affecting RR-BSM.
3.2.2.3 Testing the networks. The next step is to investigate the robustness of
networks. Using the same treatment as used in the first examination, the research

Figure 8.
N (8-4-1) network
architecture
 H found that quality of networks is so reliable through the value of correlation (r), R ,
2

26,4 mean value of AE and mean value of ARE as shown in Table IV.
Moreover, the actual vs prediction graph based on in-sample data indicates the same
finding. The reliability is indicated by the similar pattern between predicted line and
actual line as shown in Figure 9.
3.2.2.4 Out of sample prediction. In the last step, the research investigates the
306 ability of the network to perform prediction based on out-of-sample data. For this
purpose, Alyuda provides a query feature to import out of sample data for doing
prediction
based on the same networks and learning experience. The new data used in this
process are the same predictor variables used for the period of January 2009 until
December 2009.
The results are shown in Table V. The table shows the comparison between actual
value of RR-BSM (RR-BSM actual) and the prediction value of RR-BSM (RR-BSM
prediction) resulted from the out of sample prediction.
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

Input column name Importance (%)

INTR 59.124384
M1 17.839376
STIN 16.91302
Table III. BIRT 4.20217
Importance level of INFR 1.386724
predictor variables EXCH 0.534326

Data set Correlation (r) (%) R2 (%) Mean of AE Mean of ARE (%)

Table IV. Training data set 93.89 86.94 0.564248 0.067

Testing parameter for Validation data set 92.94 79.64 0.661872 0.084
the N(8-4-1) network Testing data set 90.32 78.29 0.716035 0.079
architecture All data set 0.06 84.51 0.603629 0.071

Figure 9.
Prediction vs actual
graph of RR-BSM
based on in-sample
data
 Period RR-BSM actual RR-BSM prediction Difference (%) Mudharabah
time deposit
Jan-09 7.20 7.03 2.41
Feb-09 7.18 6.99 2.65
Mar-09 6.76 6.94 2.72
Apr-09 7.63 6.85 10.28
May-09 7.18 6.83 4.81
Jun-09 7.12 6.80 4.46 307
Jul-09 7.67 6.82 11.09
Aug-09 7.42 6.79 8.55
Sep-09 6.92 6.81 1.56
Oct-09 7.08 6.75 4.70 Table
Nov-09 6.81 6.73 1.21 V. The prediction
Dec-09 6.55 6.74 2.94 results based on out of
Difference in 4.78 sample data
average
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

Using formula (4), the research investigates the difference between RR-BSM actual
(8-4-1)
and RR-BSM prediction resulted from the process. As a final result, ANN with N
network architecture successfully predicts RR-BSM from January to December 2009
with accuracy of 95.22 percent in average according to formula (5).

4. Conclusion
This paper has empirically investigated the accuracy power of ANNs model to
perform prediction on RR of BSM. Accordingly, this paper proposes a utilization
of neural
networks to predict regularly one-month ahead of mudharabah time deposit return
since the model is able to provide acceptable accuracy. Hopefully, this effort
will
support the development of a comprehensive understanding regarding the nature of
investment on mudharabah time deposit product.
Further investigation on mudharabah time deposit product from all Islamic banks
in Indonesia and continuous development using web-based application will be
conducted. Accordingly, this future research will be able to maximize its primary aim
by providing information about RR prediction from all Islamic banks. The application
can be utilized by depositors to treat the product as pure investment instrument with
risk-return trade off attached to it. Finally, from time to time, depositors will consider
Islamic bank that gives better return with lesser risk considering macroeconomic
condition.
Besides giving benefit for depositors and Islamic banks, the application will
promote the expansion of Islamic bank industry. Because, the prediction information
will give depositors more options to find other Islamic banks will probably provide
better return. As a result, it will keep depositor to stay longer in Islamic bank industry
before shifting the money to conventional bank.

References
Ahmad, N.H. and Haron, S. (1998), ‘‘The existence of conventional banking profitability
theories in the Islamic banking system’’, Analisis, Vol. 5 Nos 1/2, pp. 89-97.
Al-Osaimy, M.H. (1998), ‘‘A neural networks system for predicting Islamic banks performance’’,
Journal of King Abdul Azis University: Islamic Economic & Administration, Vol. 11 No. 1,
pp. 33-46.
 ba Muamalat Indonesia)’’, unpublished Master thesis, University
H Al- nk
of Indonesia, Jakarta.
2 Argyr M
ua Meyer, P.A. and Pifer, H.W. (1970), ‘‘Prediction of bank failure’’,
Atiya, m
ala Journal of Finance, Vol. 25, pp. 853-68.
Azad t
Moody, J. (1995), ‘‘Economic forecasting: challenges and neural network
Azis, In
solutions’’, International Symposium on Artificial Neural
Bashi do
Networks, Hsinchu, 18-20 December, available at: http:// neuron-
3 Bisho ne
ai.tuke.sk/~daliman/Esej/2/moody95economic.pdf (accessed 20
sia
Cihak, December 2009).
’’
Dewa (‘‘
Erdog Th
Erol, e
G eff
Cit ect
y, s
M of
ont co
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

hly nv
Re en
vie tio
w, na
Vo l
l. de
62, po
Ju sit
ne, int
pp er
. est
10- rat
20. e
Hall, an
Haron d
M
Kasri,
ud
K ar
pe
nd ab
ap a
ata
n de
de po
po sit
sit
o yi
M el
ud
ha d
ra to
ba th
ter
ha e
da de
p po
pe
rtu sit
m gr
bu
ha o
n wt
de h
po
sit at
o ba
di nk
 Ravi, V., Kurniawan, H., Kok Thai, P.N. and Kumar, R.V. (2008), ‘‘Soft computing system for
bank performance prediction’’, Applied Soft Computing, Vol. 8, pp. 305-15.
Mudharabah
Serju, P. (2002), ‘‘Monetary conditions and core inflation: an application of neural networks’’, time deposit
working paper, Research Services Department, Research and Economic, Programming
Division, Bank of Jamaica, July.
Zhang, G., Patuwo, B. and Hu, M. (1998), ‘‘Forecasting with artificial neural networks: the state
of the art’’, International Journal of Forecasting, Vol. 14 No. 1, pp. 35-62.
309
Corresponding author
Saiful Anwar can be contacted at: olieanwar@gmail.com
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)

To purchase reprints of this article please e-mail: reprints@emeraldinsight.com

Or visit our web site for further details: www.emeraldinsight.com/reprints
 This article has been cited by:

1. Rashid Ameer, Radiah Othman, Nurmazilah Mahzan. 2012. Information asymmetry and regulatory
shortcomings in profit sharing investment accounts. International Journal of Islamic and Middle
Eastern Finance and Management 5:4, 371-387. [Abstract] [Full Text] [PDF]
Downloaded by Universitas Brawijaya At 00:42 25 May 2018 (PT)