World Academy of Science, Engineering and Technology

International Journal of Economics and Management Engineering

Vol:14, No:1, 2020

Predicting the Success of Bank Telemarketing Using

Artificial Neural Network
Mokrane Selma

 competitiveness. Indeed, AI, data mining or machine learning

Abstract—The shift towards decision making (DM) based on can solve complex problems within businesses. Based on data,
artificial intelligence (AI) techniques will change the way in which these tools are able to anticipate customer feedback and know
consumer markets and our societies function. Through AI, predictive their purchases in order to guide more accurately the
analytics is being used by businesses to identify these patterns and
marketing campaign. These actions help build customer
major trends with the objective to improve the DM and influence
future business outcomes. This paper proposes an Artificial Neural loyalty and optimize marketing actions. Ultimately, predictive
Open Science Index, Economics and Management Engineering Vol:14, No:1, 2020 publications.waset.org/10010974/pdf

Network (ANN) approach to predict the success of telemarketing analytics is a tool that is nowadays perfectly integrated into
calls for selling bank long-term deposits. To validate the proposed the life of businesses, to become progressively indispensable.
model, we uses the bank marketing data of 41188 phone calls. The An intelligent system for prediction of the success of bank
ANN attains 98.93% of accuracy which outperforms other telemarketing using a data-driven approach is suggested by
conventional classifiers and confirms that it is credible and valuable
[2]. Research on the success of bank telemarketing has
approach for telemarketing campaign managers.
deserved increasing interest in recent years. Using data mining
Keywords—Bank telemarketing, prediction, decision making, for bank, direct marketing was developed in [3]. In [4] authors
artificial intelligence, artificial neural network. used deep convolutional neural network for predicting the
success of bank telemarketing.
I. INTRODUCTION Neural networks are widely used tools for classification [5],
estimation [6], [7], and prediction [8], [9]. Artificial neural
T ELEMARKETING is a personalized and interactive
method for consumers to learn about products, services or
opportunities. When the consumer has limited time or is
technique has been applied in many areas. The back
propagation is one of the most learning algorithms used for the
training process of ANN. ANNs are becoming a powerful tool
immobile, telemarketing delivers products and services, and
for prediction. This method is used to explain and predict one
makes buying opportunities possible with local and distant
or more observable and effectively measured phenomena. The
vendors [1].
analysis is done on all variables of the database using a
Many consumers would be greatly inconvenienced without
supervised learning classification (the class is known) based
telephone access to such products and services as; travel
on automatic learning. This technique aims to extrapolate new
reservation, catalog purchases, subscription renewals, charity
information from hidden information (this is the case of
donations, banking services, insurance purchases, food,
scoring), in this case there is a target variable to be predicted.
medicine purchases and personal services.
In this study, the data used for Bank Telemarketing
With a broader market available to vendors, more choices
prediction is a dataset publicly available for research. The
available to consumers and sales costs frequently reduced
dataset used for prediction was collected from a Portuguese
through telemarketing, consumers may benefit from lower,
retail bank. The approach used in this study is the ANNs of the
long-term product and services prices.
type Multi-Layer Perceptron (MLP).
Through AI, many companies and businesses have
This paper is organized as follows. The description of the
benefited from the advantages of prediction analysis. This
dataset and ANN with its architecture and learning algorithm
technique makes it possible to anticipate an event, predict
are introduced in Section II. In Section III, results and
incidents, and even behaviors based on data set, after having
discussions are given in order to demonstrate the effectiveness
established a correlation between data variables, it tries to
of the proposed technique. Section IV gives the conclusions.
predict what will happen in the future. Anticipating actions in
companies and being able to control them helps to prevent
II. MATERIAL AND METHODS
problems and conflicts from occurring in the future, and
increasing productivity is an objective for many businesses. A. Dataset
Today, the terms AI, big data or predictive analysis are very The data used in this study are real data downloaded from
often used. They represent a new sector that is still unclear to UCI Machine Learning Repository [10]. It was gathered from
the public. a Portuguese retail bank [11] over five years, from May 2008
Businesses today are oriented toward the predictive to June 2013. The dataset was collected by marketing
analytics to solve internal problems and to improve their campaigns of a Portuguese banking institution, based on
phone calls. In order to access if the product (bank term
M. Selma is with the Department of Economics, Abdelhamid Ibn Badis
University of Mostaganem - 27000, Algeria (phone: +213(0)778757655; e-
deposit) would be subscribed “yes” or “no”, all clients were
mail: [email protected],). contacted. Dataset contains 41188 phone contacts composed

International Scholarly and Scientific Research & Innovation 14(1) 2020 1 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Economics and Management Engineering
Vol:14, No:1, 2020

of 21 attributes with a label attribute as shown in Table I.  The hidden layer: it is located between the input layer and
that of the output; a MLP may contain one or more hidden
TABLE I
ATTRIBUTE INFORMATION layers. The only peculiarity is that starting from the input
Num. Attribute Name Description Type layer toward that of output and by passing through the
1 Age It is age of client. Numeric hidden layers, all neurons of a layer are the entries of each
2 Job It is type of client’s job. Categorical neuron of the next layer (only the layer that comes just
3 Marital It is client’s marital status. Categorical after).
4 Education What is the highest education of client? Categorical
5 Default Does client has credit? Categorical
6 Housing Does client has housing loan? Categorical
7 Loan Does client has personal loan? Categorical
8 Contact What is a contact communication type Categorical
of client?
9 Month What is the last month of the year Categorical
contracting to the client?
10 Day of Week What is the last day of the week Categorical
Open Science Index, Economics and Management Engineering Vol:14, No:1, 2020 publications.waset.org/10010974/pdf

contracting to the client?

11 Duration How long does it contact to the client? Numeric
12 Campaign Number of contacts performed during Numeric
this campaign and for this client
13 Pdays Number of days that passed by after the Numeric
client was last contacted from a
previous campaign
14 Previous Number of contacts performed before Numeric
this campaign and for this client
15 Poutcome Outcome of the previous marketing Categorical
campaign Fig. 1 A MLP with one hidden layer
16 Emp.var.rate Employment variation rate Numeric
17 Cos.price.idx Consumer price index Numeric
Recall that the output of the elementary cell (linear
18 Cons.conf.idx Consumer confidence index Numeric
perceptron) considered for the development of a MLP is a
19 Euribor3m Euribor 3 month rate Numeric
20 Nr.employed Number of employees Numeric
differentiable function of real variables, to thereby a perfect
21 Label Has the client subscribed a term Categorical example of differentiable function of real variables is the
deposit? sigmoid function defined by:
B. ANNs
𝜎 𝑦 (1)
Neural networks [12] are tools widely used for
classification, estimation, prediction and segmentation. They
It can be noted that the derivative of the function σ is simple
are derived from biological models and are composed of
to calculate:
elementary units: neurons. They are organized according to an
architecture and they are well adapted to the problems
including continuous variables possibly too noisy. They get 𝜎̀ 𝑦 𝜎 𝑦 1 𝜎 𝑦 (2)
good performance.
It is essential that this calculation is simple because the
1. MLP
derivative of this function will be used in the update rule of
MLP is one of the most used neural networks for weight by the back-propagation algorithm [11].
approximation problems, classification and prediction. It is We can now define a MLP considered in the following of
usually composed of three layers of neurons totally connected. this paper with the elementary cells of perceptron type topreal
A MLP is a set of interconnected neurons. This constitutes entries x⃗ x , … , x is defined by the real synaptic weights
an extension of the perceptron model, with one or more hidden
w⃗ w , … , w , and the output y is calculated by:
layers between the input and the output. Each neuron in a
layer is connected to all the neurons of the previous layer and
the next layer. The activation functions used in such networks 𝑦 𝑥⃗ 𝑤ℎ𝑒𝑟𝑒 𝑦 𝑥⃗𝑤⃗ ∑ 𝑤𝑥 (3)
are mainly threshold or sigmoid functions. It can solve non-
2. Learning by Back Propagation Algorithm
linearly separable problems and more complicated logical
problems. In our case we will study the three-layer networks. The principle of the algorithm is to minimize an error
 The input layer: it is the first layer of the network where function E. It is then to calculate the contribution of each
no calculation is made. The input cell only receives in synaptic weight to this error. Indeed, each of the weight
influences the corresponding neuron, but the change to this
input the values of attributes: 𝑥 , 𝑥 , … , 𝑥
neuron will affect all neurons of the following layers. The
 The output layer: it is the last layer of the network that
error therefore measures the gap between the desired and
contains one or more decision cells.
calculated outputs on the full sample. The problem is to

International Scholarly and Scientific Research & Innovation 14(1) 2020 2 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Economics and Management Engineering
Vol:14, No:1, 2020

determine a vector w⃗ which minimizes E w⃗ . The error is for Polish coal mines. Experimental results show that the MLP
example: has good predictive effect and high accuracy.

𝐸 𝐸 ⃗, ⃗ 𝑤⃗ ∑ 𝑦 𝑦 (4)

The back-propagation algorithm is then written:

Input: a sample Ω of R *R
A MLP with an input layer𝐶 , q-1 hidden layers C , … , C - , and
an output layer𝐶 ,
Random initialization of weights in 0,1 for each neuron j of the
network.
Set the value of .
Repeat
Take an example x⃗, y ⃗ in Ω
Open Science Index, Economics and Management Engineering Vol:14, No:1, 2020 publications.waset.org/10010974/pdf

Calculate the output 𝑦⃗of MLP for input 𝑥⃗

---Calculation of 𝛿 back-propagation---
For all output cells 𝒋, calculate
𝛿 𝑦 1 𝑦 𝑦 𝑦
End For
For each layer from q-1 to
For each cell j of the current layer Fig. 2 MLP Prediction results
∀𝑖 ∈ 𝑆𝑢𝑐𝑐 𝑗 (Succ j is the set of cells which take as input the output
of the cell j) calculate Major advantage of using ANN over other statistical and
𝛿 𝑦 1 𝑦 ∑ 𝛿 ∗𝑤 numerical techniques lies on their capability to capture non-
End For linear relationship among concerned variables and
End For optimization can be done very fast, without mathematical form
--- Update weights --- of the relation between input and the output data is necessary.
For all weights Their ability to learn by example makes them very flexible
𝑤 ⟵𝑤 𝜀𝛿 𝑥 and powerful.
End For In our future work; it is recommended to improve the
End Repeat
effectiveness of the proposed model by several means. It
Output: a MLP defined by the initial chosen structure and weights
𝑤 would be important to optimize the ANN model by combining
it with other techniques such as genetic algorithms and
III. RESULTS AND DISCUSSIONS metaheuristics.
In this section; to show the contribution of the bank REFERENCES
telemarketing prediction by MLP, a model is developed; able
[1] P. Kotler, and K. L. Keller. “Marketing Management”.14th Edition, New
to predict the success of the bank telemarketing using real Jersey: Prentice Hall, 2012.
data. [2] S. Moro, P.Cortez, and P. Rita. “A data-driven approach to predict the
In this section, to show the contribution of the prediction by success of bank telemarketing”. Decision Support Systems, 2014, 62,
22–31. http://dx.doi.org/10.1016/j.dss.2014.03.001.
ANN approach, ANN technology has been approved on bank [3] S. Moro, R. M. S. Laureano and P. Cortez. "Using Data Mining for Bank
telemarketing. Direct Marketing: an Application of The CRISP-DM Methodology," in
Fig. 2 shows accuracy rate of ANN method that is quite Proceedings of the European Simulation and Modelling Conference -
ESM' 2011, Guimarães, 2011.
high; however, the proposed technique can provide a high rate [4] K. H. Kim, et al., “Predicting the success of bank telemarketing using
than other methods. This method has proven to be effective in deep convolutional neural network,” in Proc. 7th International
prediction the success of bank telemarketing with a high rate Conference of Soft Computing and Pattern Recognition (SoCPaR),
Fukuoka, 2015.
of precision. [5] F. J. L. Iturriaga, and I. P. Sanz. “Bankruptcy visualization and
prediction using neural networks: A study of US commercial banks”.
IV. CONCLUSION Expert Systems with Applications, 2015, 42 (6), p. 2857-2869.
[6] Kwon, H.-B., & Lee, J. “Two-stage production modeling of large U.S.
In this study, a non-traditional approach based on the ANN banks: A DEA-neural network approach”. Expert Systems with
technology has been developed in order to predict the bank Applications, 2015, 42(19), 6758–6766.
telemarketing. 98.93% accuracy rate has been achieved with http://dx.doi.org/10.1016/j.eswa.2015.04.062.
[7] L. G. Brock III, and L. B. Davis. “Estimating available supermarket
the proposed method. The obtained results have shown that commodities for food bank collection in the absence of information”.
the effectiveness of the presented approach is able to predict Expert Systems with Applications, 2015, 42(7), 3450-3461.
seismic hazard more accurately than methods routinely used in [8] C. Jeong, J. H. Min, and M. S. Kim. “A tuning method for the
architecture of neural network models incorporating GAM and GA as

International Scholarly and Scientific Research & Innovation 14(1) 2020 3 ISNI:0000000091950263
 World Academy of Science, Engineering and Technology
International Journal of Economics and Management Engineering
Vol:14, No:1, 2020

applied to bankruptcy prediction”. Expert Systems with Applications,

2012, 39 (3), pp. 3650-3658.
[9] P. Du Jardin. “Predicting bankruptcy using neural networks and other
classification methods: The influence of variable selection techniques on
model accuracy”. Neurocomputing, 2010, 73 (10), pp .2047-2060.
[10] A. Asuncion, and D. Newman, CA: University of California, School of
Information and Computer Science, UCI machine learning repository.
Irvine, 2012.
[11] S. Moro, P. Cortez and P. Rita. “A Data-Driven Approach to Predict the
Success of Bank Telemarketing”. Decision Support Systems, Elsevier,
2014, 62:22-31.
[12] S. Bajpai, K. Jain, and J. Neeti, “Artificial Neural Networks,” the
International Journal of Soft Computing and Engineering, vol. 1, no.
2011, pp. 2231-2307, 2011.
Open Science Index, Economics and Management Engineering Vol:14, No:1, 2020 publications.waset.org/10010974/pdf

International Scholarly and Scientific Research & Innovation 14(1) 2020 4 ISNI:0000000091950263