Technological Forecasting & Social Change 202 (2024) 123250

Contents lists available at ScienceDirect

Technological Forecasting & Social Change

journal homepage: www.elsevier.com/locate/techfore

A neural network-based predictive decision model for customer retention in

the telecommunication sector
Rahul Thangeda a, Niraj Kumar b, *, Ritanjali Majhi c
a
National Institute of Technology, Warangal, India
b
Essex Business School, University of Essex, Southend-on-Sea SS1 1LW, United Kingdom
c
National Institute of Technology Karnataka, Surathkal, India

A R T I C L E I N F O A B S T R A C T

Keywords: Acquiring a new customer is far more expensive than retaining a customer. Hence, customer retention is a key
Customer retention aspect of business for a firm to maintain and improve on its market share and profit. The paper analyses customer
Churn prediction retention strategies by employing an artificial neural network-based decision model to a real-life dataset
Artificial neural network
collected from 311 mobile service users in India. Seven linear and non-linear adaptive models are developed
Telecommunication sector
using features related to customer dissatisfaction (DSF), customer disloyalty (DLF) and customer churn (CF).
Findings of this study suggest that non-linear models are most efficient in predicting customer churn, and both
DSF and DLF variables significantly affect the retention strategy. Three groups of customers are discussed in this
study in the order of least likelihood of churning to most likelihood. Finally, a priority matrix based on key
performance indicators is proposed to help service providers target potential customers to retain.

1. Introduction (2019) and Kumar (2022) also highlighted that a 5 % decrease in churn
rate could increase the profit of a firm by 25 to 85 %. Hence, it is
The world of business is rapidly growing over time. There is intense important to hold and retain existing customers in the highly competi­
competition in every sector, and it has become critical to not only focus tive market in the telecommunication sector (Babu et al., 2014).
on acquiring new customers but also to retain them (Rust et al., 2004; Customer churn is one of the major problems in developing countries
Jiang et al., 2023). The telecommunication sector is no exception to this. like India, particularly due to the unprecedented increase in mobile
The service price and switching costs in this sector have significantly phone users and the competitive pricing strategies and add-on services
reduced over the years (Ferreira et al., 2019), and therefore customer provided by the telecommunication companies. India has the second
retention has been a major concern (Óskarsdóttir et al., 2017; Amin largest number of mobile phone users in the world after China. The
et al., 2019). exponential growth in this sector in the last few years has primarily been
There are several reasons for retaining existing customers. Firstly, driven by a wider availability of services, affordable tariffs, the porta­
the market is saturated to a point where the untapped customer base is bility of mobile numbers, the expansion of the 3G/4G network, evolving
rapidly shrinking (Kostić et al., 2020). Secondly, the cost of attracting consumption patterns of customers and supportive regulatory initiatives
new customers is substantially higher, given the scale of advertising and (APEDB, 2023). There is a huge penetration of mobile phones in urban
promotion required (Ullah et al., 2019; Pakurár et al., 2019; Sivadas and parts of the country, but gradually this is increasing to semi-urban and
Baker-Prewitt, 2000; Leong et al., 2022). It costs five to six times more to rural parts as well. This has resulted in intense competition between
acquire new customers than to retain existing ones (Verbeke et al., 2011; service providers to acquire and retain their customers. Customers now
Kumar, 2022). It is reported that an average 25–30 % annual churn rate have much wider options to choose from among service providers, and
is witnessed in the telecommunication sector. Past literature such as therefore firms are increasingly attempting to understand churn be­
Ahmad et al. (2019) and Saleh and Saha (2023) suggested that the churn haviours of customers and devising new strategies to retain their
rate of customers has a huge impact on the long-term value of a business, customer bases (Venkatesan and Kumar, 2004). In this context, the
because it affects the length of service and future revenue. Ullah et al. paper primarily focuses on developing an intelligent decision framework

* Corresponding author.
E-mail addresses: [email protected] (R. Thangeda), [email protected] (N. Kumar), [email protected] (R. Majhi).

https://doi.org/10.1016/j.techfore.2024.123250
Received 12 December 2022; Received in revised form 25 January 2024; Accepted 26 January 2024
Available online 26 February 2024
0040-1625/© 2024 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

for customer churn prediction and a retention strategy. The paper is 2.1. Churn management theories
guided by two key research questions: (a) How could customer churn
behaviour be predicted more accurately in practice, using neural network- Churn management is primarily about assessing which customers are
based models considering different sets of inputs? and (b) How could a going to leave a firm and evaluating the most effective insights needed in
customer retention strategy be shaped by prioritizing the possible churners retaining those customers (Hung et al., 2006). In the current competitive
into different levels? These two key questions can assist companies to business environment, customers are subjected to several choices and
reach out to their important customers first, who could possibly churn. attractive offers when it comes to deciding on their service providers.
Developing a robust decision framework for churn prediction de­ However, when customers become dissatisfied with the service, they
mands moving away from conventional techniques based on statistical tend to churn, and as a result this impacts the profitability of the firm
models like regression frameworks, decision trees, etc., which could be (Anderson, 1994; Villanueva et al., 2008). The problem of churn has
more appropriate for problems with linear relationships (Ahmad et al., increasingly become critical in the era of digital- and subscription-based
2019; Lalwani et al., 2022). Non-linear predictive models must be used service offerings, reducing the switching costs (Hochstein et al., 2020).
for predicting churn behaviour for greater accuracy of results; however, Most literature (Winer, 2001; Burez and Van den Poel, 2007; Ascarza
most non-linear models have also been criticised for being computa­ et al., 2016) discussed two types of approaches to manage churn: reac­
tionally expensive and taking more time for training and testing pur­ tive and proactive. Firms following the reactive approach tend to wait
poses (Majhi et al., 2009; Basaran et al., 2014). The future usability of until customers request the termination of their relationship with the
some of these models is also very limited; many of the conventional service provider, and then offer an incentive to retain them. However,
models cannot be reused for new datasets. Therefore, in this paper, an the proactive approach suggests identifying and engaging customers in
adaptive learning model based on neural networks has been developed advance of them deciding to churn. Firms following the proactive
to predict the churning behaviour of customers. The proposed model is approach tend to offer targeted provisions or incentives to keep their
equipped with a feedback loop to learn from the errors made in pre­ customers before churning. It involves understanding the factors
diction and accordingly train for variation in customers’ behaviour. The affecting the churn rate (such as demographics, customer attitude,
impact of various satisfaction and loyalty factors on churn behaviour has length of stay, perception of fairness) and analysing customers’ potential
also been considered in the proposed predictive decision model. The of churning based on those factors (Bolton, 1998; Bolton and Lemon,
model takes less computational time without compromising the accu­ 1999; Neslin et al., 2006). Past studies (Lewis, 2005; Ascarza et al.,
racy of the results. It is comparatively simpler to train and test in 2018) explained the advantages and challenges of both approaches in
practice. Further, the models discussed in prior studies do not provide an detail. For instance, the reactive approach could be simpler, because a
appropriate classification of the churn level of customers for decision firm does not need to engage with customers in advance to identify who
makers. The proposed model in this paper also considers this problem is at higher risk of churning and can save the resources and efforts
and improves the classification accuracy of different churn levels. The needed to do so. However, this could be more challenging to manage in
adaptive model classifies the churn level of the consumers from low to terms of rescuing all customers who have decided to leave, threatening
high based on the performance criteria of the service providers. This the long-term sustainability of a firm. As a result, the incentive cost to
enables the service providers to devise custom-made strategies for win back customers is typically high, as compared to the proactive
different levels of churners. This study, therefore, contributes to approach (Lemmens and Gupta, 2020). On the other hand, the proactive
bringing together two critical aspects of effective churn management, by approach needs systematic engagement with customers starting from
efficiently predicting customers’ churn behaviour and by classifying identifying who may potentially churn and an investment in advanced
different churners based on their likelihood of churning to aid service analytics to accurately predict the customers’ churn. This approach also
providers to take necessary actions to retain them. carefully considers the incentive cost and value of retaining those
The paper is structured as follows. It begins with outlining the identified customers.
problem context and explaining the importance of understanding cus­ Several past studies (Keaveney, 1995; Kim et al., 2004; Gustafsson
tomers’ churning behaviour in the telecommunication sector. Later, a et al., 2005; Sohn and Lee, 2008; Al-Refaie et al., 2018) attempted to
detailed discussion of churn management theories and empirical studies understand the reasons behind churn and investigated the relationships
on predicting customer retention is provided. This is followed by a among these factors to retain customers. Customer satisfaction and
discussion of the data collection approach and the survey instrument. loyalty have prominently featured in literature on churn behaviour.
Subsequently, the proposed model is discussed, and findings are pre­ However, there has been an equal number of criticisms in literature
sented. The paper ends with a comprehensive discussion of findings and relating to loyalty and churn (Bolton et al., 2000). For instance, Reinartz
the contributions of the study to literature and practice. and Kumar (2002) argued that the relation between loyalty and profit­
ability is weak, and they found that long-term customers pay lower
2. Literature review prices compared to short-term customers. Ascarza et al. (2018) also
discussed that firms need to be aware of a trade-off in engaging with a
Several studies in the past (Reichheld and Sasser, 1990; Reichheld, customer too late or too early. Time to engage with a customer is also
2003; Adebiyi et al., 2016; Ascarza et al., 2018; Hochstein et al., 2020) crucial because customers may not prefer to stay with a firm if offers
discussed different aspects of customer churn management in a variety come too late, and it becomes too costly for the company. However,
of industries. However, a comprehensive analysis of past literature connecting with customers too early may look irrelevant to those cus­
identifies two key clusters of studies: one group of literature focuses on tomers and may even influence them to start thinking about churning.
the theoretical aspect of churn management, whereas a second group Therefore, analytical tools based on artificial intelligence could play a
discusses the empirical methods to predict customer churn behaviour. In greater role in facilitating decision makers to develop an optimal system
this section, both clusters of literature are critically reviewed to discuss to predict churn and develop their retention strategies.
the gaps and frame the narratives to explain the significance of this Targeted retention schemes have the potential to reduce churn rates
study. This section has been divided into two sub-sections. The first sub- and have lower incentive costs, but these schemes would not be useful if
section primarily discusses the literature on proactive and reactive ap­ predictions of customer churn are wrong and do not reflect real-life
proaches of customer churn management and their fundamental prin­ scenarios. Due to predictive inaccuracies, firms may target the wrong
ciples. Later, the second sub-section focuses on empirical studies and customers or engage with them too early or too late (Ascarza et al.,
models developed to effectively manage the churn. The research gaps 2016). Therefore, the need of developing an effective churn prediction
identified will help to build the rationale to develop adaptive artificial model is crucial for an efficient churn management system (Adebiyi
neural network-based models to predict customer churn behaviour. et al., 2016). Thus, this study largely focuses on the proactive approach

2
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

to develop a robust predictive model for churn management. The next call details, complaint data, service bills and payment information
section provides a review of studies of empirical models of predictive (Hung et al., 2006; Sudharsan and Ganesh, 2022). Although, Gerpott
churn management approaches. et al. (2001) argued that customer satisfaction and loyalty values are
expected to affect churn prediction, we still found very few studies that
2.2. Empirical models of predicting churn focused on considering satisfaction and loyalty as two important attri­
butes to predict customer retention. Studies discussing the combinative
Most past literature on predictive modelling for churn management impact of satisfaction and loyalty factors on non-linear adaptive models
discussed quantitative modelling approaches; however, a limited num­ for predicting churn behaviour are also scarce. Therefore, in the present
ber of studies could also be found on qualitative or mixed approaches study, an attempt has been made to study the impact of different satis­
(Baeke and Van den Poel, 2010). Traditional approaches of customer faction and loyalty factors on churn behaviour.
service and support may not yield great benefits to firms, and therefore Furthermore, several studies highlighted the importance of linear/
new artificial intelligence-based innovative approaches are needed to non-linear and simulation methods in developing a predictive model
support proactive churn management. In this regard, this paper intends for churn. For instance, Shieh et al. (2014) developed a three-layer hi­
to support the proactive churn management approach and develops an erarchy for factors that affect the adoption of mobile services. A fuzzy
innovative predictive model for churn management. This will help analytic hierarchy process (FAHP) was employed to understand the
better understand the needs of customers to predict churn and enhance weight of each factor, thereby, to study their significance. Similarly,
the optimal development of marketing campaigns. Coussement et al. (2017) studied the impact of data preparation on
Recognising that customer churn is a non-linear phenomenon (which customer churn prediction performance, whereas Ahmed and Mahes­
implies that contributing factors are non-linearly related to the level of wari (2017) proposed a hybrid firefly classifier model to predict
churn), and with an increased availability of data and understanding of customer churn in the telecommunications industry. Kim and Yoon
features causing churn, several researchers employed statistical models (2004) developed a binomial logit model to estimate the churn and
to predict churn. For instance, Neslin et al. (2006) were the first to study loyalty of Korean Telecom customers, whereas Keramati and Ardabili
data-based algorithms to predict churn, but later several other models (2011) proposed a churn model for Iranian mobile operators using
like logistic regression, decision trees (McCarty and Hastak, 2007; Lal­ binomial logistic regression. Lalwani et al. (2022) used machine
wani et al., 2022), random forest technique (Lariviere and Van den Poel, learning models to predict churn after pre-processing the data. Further,
2005; Lalwani et al., 2022), neural network-based models (Zahavi and Ahmad et al. (2019) developed a churn prediction model to help tele­
Levin, 1997; Khodabandehlou and Zivari Rahman, 2017; Yu et al., 2018) communication operators predict customers who were likely to leave.
and support vector machines (Shin and Cho, 2006; Keramati et al., Social network analysis (SNA) was used as a feature to develop this
2014) appeared. Moreover, Verbeke et al. (2011) proposed a rule in­ model. Amin et al. (2019) observed that different evaluation metrics
duction technique-based customer churn prediction model with were also used to understand the significance of the model developed.
improved performance. They applied two data-mining algorithms (lo­ Additionally, some of the past studies focused on understanding the
gistic regression and decision trees) to build a churn prediction model relevant factors responsible for churn. For instance, Basaran et al. (2014)
using credit card data collected from a Chinese bank. Guidolin and studied important factors for deciding on an operator in the Turkish
Guseo (2015) developed an innovative model that evaluated the effect mobile market. Similarly, Jaiswal et al. (2018) studied the impact of
of competition between competing products both on the dynamics factors like acquisition, length of relationship, service communication,
within the products and cross-product word of mouth for pre-recorded product return activity and type of products purchased in online markets
music in the US market, contributing to identifying the social and on customer retention. Also, Shin and Kim (2008) identified the factors
market factors affecting churn. Lamrhari et al. (2022) also used a social that influence a customer to switch from their current service provider.
customer relationship management (CRM) analytic framework as a Recently, Kim et al. (2019) indicated that non-monetary switching costs
feature to improve customer retention. Further, Huang et al. (2012) have a moderating effect on customer retention.
applied new features and a window-based technique to predict churn in This review of past literature highlights that in the last two decades,
the landline telecommunication sector in China and showed its superi­ several literatures (Basaran et al., 2014; Ahmad et al., 2019; Lalwani
ority in yielding better churn prediction performance in landline tele­ et al., 2022) have focused on developing numerous approaches to esti­
communication services. A hybrid model based on a genetic algorithm mate customer churn behaviour. These approaches can be broadly
and a neural network has been proposed for predicting customer divided into two groups: statistical and soft computing. The statistical
retention in cellular wireless network services (Huang et al., 2012). It is techniques are mostly linear methods, whereas the churn prediction
reported that this new approach outperforms the statistical z-score problem is a non-linear one. Therefore, the linear methods provide a less
model on all performance criteria. accurate prediction of churn. Subsequently, random forest techniques
Among other sectors, studies in the telecommunication sector on and support vector machines have been applied to predict churn. These
churn behaviour have also featured in past literature. For instance, Idris soft computing-based methods are non-linear in nature and have been
et al. (2012) used a particle swarm optimization (PSO)-based data- reported to provide improved estimation of churn compared to statis­
balancing method for churn prediction in this sector and argued that tical approaches. However, they are also not free from shortcomings.
this approach performed better compared to other approaches for pre­ Accurate classification of churners plays an important role in customer
dicting churners. Praseeda and Shivakumar (2021) also proposed a behaviour, as it is directly related to the growth of business in the
fuzzy PSO-based feature-selection method to develop a clustering al­ organisation. The available methods in the literature do not provide an
gorithm for predicting churn. Similarly, Verbeke et al. (2012) developed appropriate classification level of customers for decision makers. Hence,
a profit-driven data-mining approach for churn prediction, and Chen this study identifies this as a research gap and attempts to improve the
et al. (2018) employed a data-mining approach to identify mobile classification accuracy by developing efficient non-linear models.
opinion leaders. However, several literatures (Hadden et al., 2007; Bin Ascarza (2018) argued that it is not wise to only target customers based
et al., 2007; Ismail et al., 2015; Sudharsan and Ganesh, 2022) still on risk of churn but that it should also be based on who might need
argued that limited data-mining techniques have been employed to interventions. The proposed model not only predicts the churn but at the
tackle the problem of customer retention in the telecommunication end recommends who the customers to be targeted are as a priority.
sector. The challenges with existing non-linear-based churn prediction
Generally, the features used for churn prediction in the telecom­ methods are that they are computationally expensive and take more
munication industry include recency, frequency and monetary vari­ time for training and testing operations. As a result, these techniques
ables, customer demographics, contractual data, customer service logs, may not be useful for online prediction purposes. Therefore, in this

3
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

paper, a novel approach is proposed to address this problem of high customers are two important determinants for estimating the level of
computational time without sacrificing the accuracy. Also, the future churn. Therefore, in this study, satisfaction and loyalty are used as in­
usability of the models discussed in past literature is very limited; many puts to the proposed model. The questionnaire consists of 25 questions
of the conventional techniques cannot be reused for new datasets in the on satisfaction and 14 questions on loyalty. Respondents were asked to
future. However, the proposed adaptive technique develops the pre­ record their responses on a scale of one to five on a Likert scale.
diction model by training it through old datasets and subsequently de­ All responses received were subjected to factor analysis based on the
velops a model suitable for predicting churning behaviour based on principal component analysis (PCA) method. As a result of factor anal­
future datasets. Periodic training of the model may be necessary in the ysis, five factors (Adaptability, Customer Service, Connectivity, Skills of
future due to variation in the data and outcomes. Employees, Advertisement) pertaining to ‘satisfaction’ and three factors
(Reliability, Repurchase Intention, Continuity in Usage) pertaining to
3. Data collection and sampling ‘loyalty’ were identified. Total variance explained by the variables for
both satisfaction and loyalty are presented in Appendix B. For customer
This study focuses on proposing a neural network-based non-linear satisfaction, five factors have eigen values >1 and express a cumulative
adaptive model to efficiently predict the churn level of mobile phone variance of 56.47 %, whereas for customer loyalty, three factors have
users in India. Primary data for this study were collected through a eigen values >1 and express a cumulative variance of 54.78 %. The
structured questionnaire sent to mobile phone users from the Andhra summary results of factor analysis are presented in Table 1.
Pradesh (AP) telecom circle, one of the highest tele-density telecom Data were normalized to ensure that each input lay between 0 and 1
circles (licensed service areas – LSAs) in India. The AP circle included for developing the proposed prediction model. It is known that higher
mobile phone subscribers from the Andhra Pradesh and Telangana states levels of dissatisfaction and disloyalty in customers lead to higher
in India, and it was one of the first circles to introduce mobile phone probabilities of churning (Bolton et al., 2000; Gerpott et al., 2001;
services in India in 1997 (Mukhopadhyay et al., 2023). The tele density Gustafsson et al., 2005). Hence, these two values were obtained by
in the AP circle is about 93 % and it provides on average the highest subtracting the corresponding satisfaction and loyalty levels of the re­
download speed (TRAI, 2023). Further, the AP circle also has the highest spondents from unity and are used as inputs for model development.
number of 5G smartphone models, representing 14 % of all 5G-capable Provision was also made in the questionnaire to collect the responses in
smartphones in India (APEDB, 2023). The AP circle is also among the top terms of the churning behaviour of customers. The magnitude of
five LSAs with the highest number of mobile number portability (MNP) dissatisfaction, disloyalty and four normalized responses to the ques­
requests in India, reaching 55.73 million in November 2022 (TRAI, tions pertaining to churn are used as inputs to the model and the cor­
2023). This clearly highlights the high frequency of change in mobile responding churn response of the customers is used as a target or desired
phone service providers and reflects a high rate of churn. The compe­ value of the proposed model.
tition among service providers is very intense in this circle. Therefore,
the AP circle was the most appropriate to conduct this study to develop 5. Basis of development of different prediction models
an efficient predictive model to estimate customer churn in the tele­
communication sector. Customers residing in cities generally get the As churn management becomes such a complex problem, where
best possible service from service providers, and wider opportunities to several factors are involved in taking an efficient decision, the funda­
switch between service providers. Therefore, in this study, data were mental principles of system dynamics would help to bring different
collected from the highest tele-density cities in the AP circle, like viewpoints to overcome the increasing complexity of the issue (Azadeh
Hyderabad, Warangal, Vijayawada and Vishakhapatnam (Khatri, 2022). et al., 2014). The telecommunication industry is a complex system
A random sampling method was used to distribute the questionnaire involving customers, service providers, mobile phone manufacturers,
to mobile service customers in the AP circle. Random sampling helps to retailers and policymakers. Churn management could be considered as
avoid any general bias in collected data. The survey was administered one of numerous subsystems, which needs to optimise several variables
through an online platform by using Google Forms. It was targeted to to reach an optimal outcome. The market dynamics and competitions
adult customers of >18 years of age without any further restriction on among service providers significantly affect the cost structure in this
gender and age. It was initially sent to 400 potential respondents with sector. Hence, it is necessary to understand all the factors affecting churn
the intention of receiving a sizable data sample to implement proposed behaviour before developing the neural network-based predictive
predictive models based on neural networks (Hatcher and O’Rourke, model. In this paper, customer-related variables affecting their churn
2013). A similar approach for sampling was also followed by several past were identified by directly connecting with them through a survey in­
studies, such as Babu et al. (2014) and Adebiyi et al. (2016). Data were strument. Later, once the set of key variables were identified, then an
only collected at one point in time based on the survey instrument adaptive neural network-based model was proposed with a feedback
described in the next section. 340 responses were received from the loop to learn from the errors made in the prediction. The dynamic
customers, but 29 did not respond to all the questions and were therefore simulation proposed in the predictive model with the feedback loop
not considered for further analysis. Finally, a sample of 311 responses would be beneficial to reach to an optimal result (Sterman et al., 2015).
were considered for further analysis in this study. In this study, seven different non-linear models have been developed
based on different combinations of inputs. Factor scores (F) obtained
4. Survey instrument from factor analysis of the variables and the direct value (V) estimated
from the responses received from the respondent are two different sets of
A structured questionnaire (Appendix A) was developed for data inputs. The basis of selection of these models is explained in the
collection to test the efficiency of the proposed neural network-based following. Four typical responses pertaining to churn have been
prediction model. The questionnaire was divided into four sections. collected from the customers and then normalized to serve as inputs
The first section consisted of questions on general usage of mobile
phones. The responses of customers pertaining to four questions in this
Table 1
section were used as inputs to the model. These questions were (i) the
Results of factor analysis for satisfaction and loyalty using PCA.
number of years they had been using a mobile phone, (ii) their current
Key Factors
service provider, (iii) the expenditure on their mobile phones per month,
and (iv) the number of service providers they had used in past. Several Satisfaction Adaptability, Customer service, Connectivity, Skills of Employees,
past literatures (Bolton et al., 2000; Gerpott et al., 2001; Gustafsson Advertisement
Loyalty Reliability, Repurchase intention, Continuity in usage
et al., 2005) suggested that the satisfaction and loyalty values of

4
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

(denoted as FC) to the proposed prediction model. The corresponding i - Number of experiments conducted.
prediction model is termed as Model-1. The findings of past literature A unity bias input is also applied through a weight wb (i) to achieve
(Bolton et al., 2000; Gerpott et al., 2001; Gustafsson et al., 2005) sug­ improved prediction performance. The predicted churn output ̂c (r, i) of
gested that factors relating to dissatisfaction as well as loyalty strongly the model is obtained as:
influence the churn level of a customer. Similarly, customer dissatis­
∑
N
faction can reduce the customer base of a firm and impact its reputation c (r, i) =
̂ xn (r).wn (i) + wb (i) (1)
(Levesque and McDougall, 1996). Customers who are not satisfied could n=1
tell others about their dissatisfaction and might churn (Lejeune, 2001).
where the symbols ̂c (r, i), c(r, i) and e(r, i) denote the estimated
Lee et al. (2001) also highlighted a high relationship between the
churning level, the customer’s response on churn. The error term be­
switching costs, customer satisfaction and loyalty. Accordingly, from the
tween the actual churn value and its estimated one is given by:
responses collected, the dissatisfaction factor (DSF) and the disloyalty
factor (DLF) of each customer was estimated. e(r, i) = c(r, i) − ̂c (r, i) (2)
In Model-2, the DSF, DLF and FC values have been used as the inputs
During the training period, when the attributes relating to a
and the predicted churn level of the customer as the corresponding
customer churn are applied, the model produces an estimated churn
output of the model. To assess the effect of the DLF, another model is
level which is then compared with the churn value provided by the
proposed (Model-3) in which the DLF and the FC are chosen as inputs.
corresponding respondent to produce the error term defined in (2).
Similarly, in Model-4, the DSF and the FC are employed as inputs with an
Using the input value corresponding to each connecting path and the
objective to study the effect of dissatisfaction factors on churning.
error value given in (2), the change in weight value in each nth path is
In Models 2, 3 and 4 the magnitudes of key factors pertaining to churn
computed as:
as well as the DSF and the DLF computed from responses of the cus­
tomers are used as inputs. In developing the remaining models, a new dwn (r, i) = 2 × μ × xn (r, i) × e(r, i) (3)
strategy is adopted in which the dissatisfaction values (DSV) and/or
For the bias path, the change in weight is given by:
disloyalty values (DLV) obtained directly from the respondents along
with the factors relating to churning are used as inputs instead of using dwb (r, i) = 2 × μ × e(r, i) (4)
the values of different factors. Accordingly, Models 5, 6 and 7 are
where μ = convergence coefficient which controls the convergence
developed with different combinations of the above-mentioned attri­
rate, and its value lies between 0 and 1.
butes. The details of different combinations of inputs used are termed as
The same process is repeated for all R respondents in a single
Types 1–7, and the corresponding Models 1–7 are listed in Table 2.
experiment. At the end of an experiment, each path accumulates R
The accuracy of churn prediction achieved from these models is
number of weight changes. The average weight change in each con­
obtained using different combinations of real-life data of customers as
necting path after an experiment is obtained as:
inputs and is compared with the proposed models. The development of
the proposed models and a discussion on the results are in the next 1 ∑R

section. adwn (i) = ⌊ dwn (r, i) ⌋ (5)

R r=1

6. Deployment of novel adaptive models Where adwn (i) = average change in weight value in nth path during
the ith experiment.
In this section, both linear and non-linear adaptive models are pro­ Similarly, the average change in weight in the bias path is given by:
posed for prediction of churn and their performances are compared
using collected data. All models proposed in this section are adaptive in 1 ∑R
adwb (i) = ⌊ dwb (r, i) ⌋ (6)
nature. Such models have the advantage of reusability using new/dy­ R r=1
namic data of customers and hence can also be used for prediction of
Finally, the updated weight value in each nth path is given as:
churn levels of future customers.
wn (i + 1) = wn (i) + adwn (i) (7)

6.1. Adaptive linear prediction model The bias weight is updated as:
wb (i + 1) = wb (i) + adwb (i) (8)
A simple adaptive model, as shown in Fig. 1, is developed to predict
churn. where, wn (i + 1) represents the updated weight value at the begin­
In Fig. 1, xn (r, i) and wn (i) represent the input of nth attribute of rth ning of (i + 1)th experiment corresponding to nth path for rth respondent.
respondent and connecting weights corresponding to nth attribute dur­ Subsequently, the mean square error E(i) during ith experiment is
ing ith experiment respectively. In model development, the following computed using (9) to plot the training or convergence characteristics of
terminologies are used: the churn prediction model:
N- Total number of input attributes pertaining to churning. ∑R 2
e (r, i)
R- Total number of respondents whose data are used for training E(i) = r=1 (9)
R
purposes.
where, R= total number of respondents.
Table 2
The experiments are continued until E(i) attains the lowest possible
Description of models and their inputs. magnitude. The convergence plot exhibits the relation between E(i) and
the number of experiments i. This plot indicates (i) whether the model
Model Inputs Output
has attained convergence (when the E(i) value attains the lowest
Model − 1 Type − 1 (FC) possible constant value), (ii) the number of experiments required for
Model − 2 Type − 2 (DSF, DLF, FC)
training the model, and (iii) the accuracy of prediction that is achievable
Model − 3 Type − 3 (DLF, FC)
Model − 4 Type − 4 (DSF,FC) Predicted churning level/churning value from the trained model.
Model − 5 Type − 5(DSV,DLV,FC) After completion of training, the weights attained by each path
Model − 6 Type − 6(DSV,FC) represent the final weights of the proposed model. At this stage, the
Model − 7 Type − 7(DLV,FC) model is complete, and it is ready for validation or testing purposes.

5
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

6.2. Adaptive non-linear prediction models corresponding to N attributes are represented as:

Zk (r) = [ x1 (r) , sinπx1 (r) , cosπx1 (r) , sin3πx1 (r) , cos3π x1 (r) , sin5πx1 (r) , cos5πx1 (r)………xN (r) , , sinπ xN (r) , cosπxN (r) , sin3πxN (r) , cos3πxN (r) ,
(11)
sin5πxN (r) , cos5πxN (r) ]

In this section, three different non-linear prediction models are The corresponding weight vector during ith experiment is repre­
developed and analysed. These are structurally simple, easy to imple­ sented as, Wk (i) (1 ≤ k ≤ K). The predicted churn of rth respondent
ment and, performance-wise, superior to the results provided by previ­ during kth experiment is given by:
ously reported linear techniques such as the regression or adaptive
linear combiner model. c (r, i) = ZKT (r)Wk (i) + wb (i)
̂ (12)
These models are almost similar to each other structurally except in
where wb (i) represents the weight value corresponding to the bias
the functional expansion block. In this section, models are presented in
(unity) input during ith experiment.
two different ways using trigonometric and Legendre-based expansion
The error term e(r, i) is then computed using (2). The expressions for
schemes. The details of various expansion schemes are illustrated in
change in weight, the change in bias weight, the average change in
Table 3. The objective behind using such non-linear expansion is to map
weight, the average change in bias weight, the weight update equation,
the input data to non-linear values and then employ the non-linearly
the bias update equation and the mean square error are identical to (3)
transformed values as inputs to efficiently estimate the churn level.
to (9) respectively. The only exception is that wherever the subscript ‘n’
Adoption of such a strategy, though simple, makes the prediction more
appears in these expressions, it would be replaced by ‘k’.
efficient. The remaining of these non-linear models such as the con­
necting weights, the computation of the output, the error values, the
weight update schemes, the training and testing methodologies are 6.3. Comparison of different models
identical to that of the adaptive linear combiner model dealt with in
Section 6.1. As mentioned in the previous sub-section (Table 2), seven prediction
The complete scheme of development of the trigonometric,
expansion-based, non-linear churn prediction model is shown in Fig. 2. Table 3
The Legendre expansion-based models are similar to that of Fig. 2 and Expanded values for different non-linear schemes for a single input case.
are shown in Fig. 3. Input Model Expansion
In Fig. 2, only odd-numbered sine/cosine expansions are used, as Single input Linear x
such a combination provides better prediction performance. Further, for (x) Trigonometric x, sinπx, cosπx, sin3πx, cos3πx, sin5πx, cos5πx)
the present application, each input is expanded to seven terms, out of (odd numbered expansion)
In general sin(2n − 1)πx; cos(2n − 1)πx
which all except the first one are related to the input in a non-linear
Legendre L1 (x) = x
manner. Such choice of odd-numbered trigonometric expansions is 1( )
L2 (x) = 3x2 − 1
based on trial and error, keeping in view the accuracy in the prediction 2
of the model. M represents the number of each sine and cosine terms. 1 ( )
L3 (x) = 5x3 − 3x
2
Thus, each single input is transformed into L = (2M + 1) terms. If the 1( )
L4 (x) = 35x4 − 30x2 + 3
number of inputs of the model is N, then the total number of terms after 8
1(
trigonometric expansion becomes:
)
L5 (x) = 63x5 − 70x3 + 15x
8
1 (
(10)
)
K = N(2M + 1) L6 (x) = 231x6 − 315x4 + 105x2 − 5
16
1 ( )
kth (1 ≤ k ≤ K) number of expanded terms for rth respondent L7 (x) =
16
429x7 − 693x5 + 315x3 − 35x

Fig. 1. Adaptive Linear Combiner Model for Churn Prediction.

6
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

Fig. 2. Trigonometric expansion-based non-linear adaptive model for customer churn prediction.

models have been proposed based on the use of Type 1 to Type 7 inputs. problem of churn-level estimation is inherently a non-linear one because
The block diagrams in Fig. 4 explain the way the models are built based the churn level of a customer is related to its input factors in a non-linear
on input schemes. These models are simulated using MATLAB software manner. Out of the two non-linear models, the performance ranking
and real-life data obtained from the customers through the survey. 80 % based on accuracy of prediction is observed to be the trigonometric and
of the collected data is used for training the model, and the remaining Legendre models in that order.
20 % is employed for validation purposes. The comparison of convergence characteristics among the best per­
Following the procedure discussed in Section 6.1, the training forming six trigonometric, expansion-based, churn prediction models is
operation is carried out and the convergence characteristics obtained for shown in Fig. 6. A systematic comparison of the accuracy of prediction
different models corresponding to different input conditions, as listed in under identical input conditions is shown in Table 5. It shows that
Table 2. The comparison of the training characteristics obtained for Model-2, which refers to Type-2 inputs (DSF, DLF and FC), performs the
Model-1 for three (one linear and two non-linear) expansions is shown in best, whereas Model-1, which takes the input FC, performs the worst.
Fig. 5. Similar models were developed for the rest of the six models and In Figs. 1 to 4, the approach of studying the relationship between
convergence characteristics. The comparison reveals that the rate of factors and customer churn incorporates the feedback loop through
convergence of the trigonometric, expansion-based model provides the which the weights are updated so that the model converges, as shown in
best prediction accuracy among all the proposed seven models. The Fig. 5.
comparison of prediction performance in terms of percentage of errors Using Table 4, the first five best models based on the accuracy of
during training and testing has been separately obtained through churn prediction and the corresponding inputs used during training and
simulation studies for all the seven models and for all combinations of testing phases are separately obtained from the simulation results and
input expansion schemes. The corresponding results are shown in are listed in Table 5.
Table 4. In-depth examination of these results demonstrates that in
terms of prediction accuracy, the trigonometric, expansion-based 7. Discussion
models outperform the other expansion-based models. Further, it is
observed that the linear model provides the worst performance, as the Customer retention is critical for the success of a company.

7
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

Fig. 3. Legendre expansion based non-linear adaptive model for churn prediction.

Fig. 4. Block diagram for prediction of churn level of customers using type − 1 inputs in model-1.

8
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

proactive approach to manage customers’ churning behaviour (Ascarza

et al., 2018). To further aid the service providers, this study also assists
to categorise customers based on their churn level (Table 6) and key
performance indicators like monthly revenue and longevity of the
customer. Some past studies (Hwang et al., 2004; Sharaf Addin et al.,
2022) have also segmented customers into different groups based on
various criteria; however, the combinative effect of satisfaction and
loyalty factors on predicting churn behaviour has not been extensively
discussed.
After completion of training of all seven churn prediction models, it
is observed that in all cases, the trigonometric, expansion-based models
perform the best. Further, out of all input combinations, the DSF, DLF
and FC inputs corresponding to Model-2 produce the highest accuracy of
prediction. This indicates that factors related to satisfaction, loyalty and
churn behaviour contribute to predicting customer churn. These factors
were also considered important in past studies (such as Gerpott et al.,
Fig. 5. Comparative performance of the convergence of model-1 using different 2001; Gijón et al., 2013; Díaz, 2017; Pasape, 2022). Based on the
linear/nonlinear expansions. aforesaid explanation, the best grouping of 63 customers under testing is
given as 8, 37 and 18 for Groups A, B and C respectively, as shown in
Table 4 Table 7. From these findings, it is evident that 18 out of 63 customers,
Model wise comparison of the accuracy of churn prediction. that is 35 %, belonging to Group C would most likely churn. The primary
S. Model Type of % error during % error during focus of the service provider should be given to this group (C) as there is
No. No. polynomial training testing a high chance that these customers might churn in the near future.
Trigonometric 9.3138 10.9867
Group B should be given the next focus, as they are next most likely
Legendre 17.3689 18.096 group of churners.
1 Model-1 Linear Combiner 18.2015 23.522 The present investigation has provided segments of customers who
Trigonometric 5.412 5.6786 might churn in the near future, and service providers must work on
Legendre 12.4357 15.06
improving the satisfaction level of customers of Groups B and C so that
2 Model-2 Linear Combiner 15.6138 20.1772
Trigonometric 7.8104 8.2752 customer retention increases and thereby the service providers’ growth
Legendre 15.7735 16.1618 rate improves. Further grouping of customers is done based on key
3 Model-3 Linear Combiner 17.1105 22.6989 performance indicators like monthly revenue and longevity (age). This
Trigonometric 6.7511 7.6842 enables service providers to find out who they should be targeting or
Legendre 13.4733 14.0658
4 Model-4 Linear Combiner 16.0881 21.0761
prioritizing. A matrix is developed considering the significant customers
Trigonometric 9.104 10.5479 that fall into Group C and then taking into account the longevity of using
Legendre 16.5188 17.5761 the service and the revenue they are contributing on a monthly basis.
5 Model-5 Linear Combiner 17.6577 23.0034 The reasons for considering these two metrics are that revenue repre­
Trigonometric 9.5059 10.8914
sents the contribution the customer is making, and age represents the
Legendre 16.963 17.7909
6 Model-6 Linear Combiner 17.9247 23.3608 loyalty of the customer. This matrix further helps to filter the profitable
Trigonometric 8.9763 10.6212 and loyal customers, as shown in Fig. 7.
Legendre 16.7568 17.928 The customers who generate revenue for the service providers and
7 Model-7 Linear Combiner 23.1504 17.9157 who have stayed for a long time and been loyal must be first reached out
to by the service providers to understand their pain areas and retain
them. This will enable revenue generation and the retaining of loyal
customers. In Fig. 7, it can be observed that there are around four cus­
tomers in the quadrant who are generating INR 350 of monthly revenue
and have stayed for more than two and a half years with the service
provider. In essence, the accurate estimation of churn level of mobile
phone customers greatly helps the service providers to adjust their ser­
vice strategy to reduce the number of churners and thereby help in the
growth of service providers.

8. Conclusion

This paper analyses the dynamics of customer churn, which is an

important factor in a firm’s strategy of customer retention. Customer
satisfaction and customer loyalty are assumed to determine customers’
tendency to stay with a firm or to move away. Since customer churn is a
Fig. 6. Comparative convergence performance among the best performing
trigonometric expansion based models. non-linear function of both customer satisfaction and customer loyalty,
the typical linear regression methods have been ineffective in the pre­
diction of customer churn. Therefore, this paper developed the non-
Understanding churn behaviour is important for the mobile service
linear adaptive models which are far more accurate in studying
provider to orient their marketing strategy and to enhance customer
customer behaviour. Using real-life data collected from customers of
satisfaction level (Amin et al., 2019). In addition, if the service providers
mobile phone service providers, the performance of proposed models
are able to identify possible churning customers, then they are able to
during training and testing has been evaluated. The paper concludes that
devise strategies to retain them (Babu et al., 2014). The models devel­
the second model using trigonometric expansion and inputs such as the
oped in this paper help to predict the churn level, supporting the
DSF, the DLF and the FC has performed the best. This indicates that

9
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

Table 5
List of five best models along with the inputs based on the accuracy of churn prediction.
Training phase Testing Phase

Rank Based Performance Model No. Types of Expansions Inputs used Model No. Types of expansions Inputs used

1 2 Trigonometric DSF,DLF,FC 2 Trigonometric DSF,DLF,FC

2 4 Trigonometric DSF,FC 4 Trigonometric DSF,FC
3 3 Trigonometric DLF,FC 3 Trigonometric DLF,FC
4 7 Trigonometric DLV,FC 5 Trigonometric DSV,DLV,FC
5 5 Trigonometric DSF,DLV,FC 7 Trigonometric DLV,FC

Table 6 approaches discussed in the literature tended to focus more on only

Grouping of the customers based on churning levels. predicting the churn in general and not specifying different categories of
S. No. Churning level range Group churn levels. The paper provides a comprehensive approach for service
providers to understand and predict the churn rate of their customers
1. <0.33 A
(Less likelihood of being churners) and guides them to devise innovative retention strategies according to
2. 0.34–0.66 B varying churn levels.
(Medium likelihood being churners)
3. 0.67–0.99 C
(High likelihood of being churners)
8.1. Contributions to literature

The paper significantly contributes to the body of knowledge in

Table 7 churn management. Although in the last two decades several literatures
Grouping of Customers Using Different Models and Types of Expansions. (Ahmad et al., 2019; Lalwani et al., 2022) have applied numerous
S.No. Model No. Trigonometric Legendre Linear Model methods to estimate customer churn, most of these suffer from either
Expansion Expansion (No expansion) inaccurate prediction due to a mismatch of the nature of the problem
No. of customers No. of customers No. of and the model used (e.g. statistical models) or inappropriate classifica­
customers tion of churners, which does not provide sufficient insights for decision
A B C A B C A B C makers. This study has attempted to improve the accuracy of prediction
and provided appropriate classification categories of customers with the
1. Model-1 10 31 22 7 31 25 0 60 3
2. Model − 2 8 37 18 10 37 16 10 37 16
help of customer data and an adaptive neural network-based prediction
3. Model − 3 9 29 25 7 32 24 0 60 3 model with a feedback mechanism. The proposed model not only pre­
4. Model − 4 6 36 21 7 32 24 1 59 3 dicts the churn but at the end recommends who the customers to be
5. Model-5 7 30 26 5 34 24 0 60 3 targeted are as a priority. This paves the way for developing the right
6. Model-6 9 26 28 6 32 25 0 60 3
strategy for customer retention by service providers. Further, existing
7. Model-7 7 30 26 5 33 25 0 59 4
non-linear-based churn prediction methods are computationally
expensive and take more time for training and testing operations. As a
result, these techniques may not be useful for online prediction pur­
poses. The proposed model in this study also addresses this problem of
high computational time without sacrificing the accuracy. Moreover, the
future usability of the models discussed in the literature is very limited.
Many of the conventional techniques cannot be reused for new datasets
in the future. However, the proposed adaptive technique develops the
prediction model by training it through old datasets and subsequently
develops the model suitable for predicting the churn behaviour in the
future. Periodic training of the model may be necessary in the future due
to variation in the data and outcomes.

8.2. Managerial implications

The predictive decision model proposed in this study will greatly

benefit managers to accurately understand the churn behaviour of cus­
tomers, which will return in better retention of those customers. Using
the best model and range of churn level, the customers under testing
Fig. 7. Possible churners prioritisation matrix.
were grouped into three categories, namely no churners (Group A),
churners (Group B) and highly possible to churn (Group C). For any
dissatisfaction and disloyalty factors greatly influence the churn level of business, it would be important to devise effective retention strategies
customers. for different groups of customers and prioritize those who have a higher
The paper also highlights the novelty of the proposed model, which is chance of leaving. The model could help practitioners to reduce the
adaptive in nature and uses a feedback loop that can minimize training incentive cost by not targeting low-risk customers. Also, in a large
errors and provide more accurate results. It also helps in categorising organisation, if a similar model was run, there would be a possibility of a
customers based on their likelihood of being churners and on certain key large set of possible churners. Hence, to further prioritize customers to
performance indicators. The ability of the proposed model to predict target, a two-by-two matrix considering the revenue and longevity of a
customers’ churn behaviour and their potential likelihood of leaving customer has been developed in this study. Service providers could also
their current service providers will be beneficial for service providers to consider any other performance indicators relevant to them to classify
develop tailor-made strategies to retain different customer groups. Past potential customers to target. Consequently, the service provider can
focus first on those groups that have higher importance and ensure that

10
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

their profit-creating and loyal customers do not leave. The proposed system model to enhance industrial intelligence: the case of a large broiler industry.
Syst. Res. Behav. Sci. 31 (2), 236–257.
model in the paper follows the proactive churn management approach
Babu, S., Ananthanarayanan, D.N., Ramesh, V., 2014. A survey on factors impacting
and provides a research-driven guide to managers in their quest for the churn in telecommunication using data mining techniques. International Journal of
optimal development of marketing strategies to retain valued customers. Engineering Research & Technology 3 (3), 1745–1748.
Baeke, P., Van den Poel, D., 2010. Improving purchasing behaviour predictions by data
augmentation with situational variables. International Journal of Information
8.3. Potential limitations and future scope Technology & Decision Making 9 (6), 853–872.
Basaran, A.A., Cetinkaya, M., Bagdadioglu, N., 2014. Operator choice in the mobile
telecommunications market: evidence from Turkish urban population.
The model in this study is developed to run in real time. If the fea­ Telecommunications Policy 38 (1), 1–13.
tures are to be modified, then it needs to be retrained. It is built to Bin, L., Shao, P., Liu, J., 2007. Customer churn prediction based on the decision tree in
personal handy phone system service. Proceedings of the IEEE International
address the computational time issues that neural network models conference on service systems and service management 1–5.
usually have. However, similar models can be developed to assess Bolton, R.N., 1998. A dynamic model of the duration of the customer’s relationship with
customer churn with different variables and can also be connected to a continuous service provider: the role of satisfaction. Mark. Sci. 17 (1), 45–65.
Bolton, R.N., Lemon, K.N., 1999. A dynamic model of customers’ usage of services: usage
CRM databases to run in real time. Considering the limited resources and
as an antecedent and consequence of satisfaction. J. Market. Res. 36 (2), 171–186.
time, only a generic sample of respondents was used for this study, Bolton, R.N., Kannan, P.K., Bramlett, M.D., 2000. Implications of loyalty program
which does not provide a detailed analysis of gender-, age- and income- membership and service experiences for customer retention and value. J. Acad.
level factors and their relative impacts on churn behaviour. However, a Mark. Sci. 28 (1), 95–108.
Burez, J., Van den Poel, D., 2007. CRM at a pay-TV company: using analytical models to
study with a large dataset in the future could test these different effects. reduce customer attrition by targeted marketing for subscription services. Expert
Moreover, the validation of the models was only compared using the Systems with Applications 32 (2), 277–288.
data collected at one point in time and was not tested with longitudinal Chen, C.P., Weng, J.Y., Yang, C.S., Tseng, F.M., 2018. Employing a data mining approach
for identification of mobile opinion leaders and their content usage patterns in large
datasets. However, future research could make contributions in this telecommunications datasets. Technological Forecasting and Social Change 130,
direction. The validation of the proposed model can be done on CRM 88–98.
databases to improve its validity and reliability. Coussement, K., Lessmann, S., Verstraeten, G., 2017. A comparative analysis of data
preparation algorithms for customer churn prediction: a case study in the
telecommunication industry. Decis. Support. Syst. 95, 27–36.
CRediT authorship contribution statement Díaz, G.R., 2017. The influence of satisfaction on customer retention in mobile phone
market. J. Retail. Consum. Serv. 36, 75–85.
Ferreira, P., Telang, R., De Matos, M.G., 2019. Effect of friends’ churn on consumer
Rahul Thangeda: Writing – review & editing, Writing – original behavior in mobile networks. J. Manag. Inf. Syst. 36 (2), 355–390.
draft, Methodology, Formal analysis, Data curation, Conceptualization. Gerpott, T.J., Rams, W., Schindler, A., 2001. Customer retention, loyalty, and satisfaction
Niraj Kumar: Writing – review & editing, Writing – original draft, Su­ in the German mobile cellular telecommunications market. Telecommun. Policy 25
(4), 249–269.
pervision, Methodology, Formal analysis, Data curation, Conceptuali­ Gijón, C., Garín-Muñoz, T., Pérez-Amaral, T., López-Zorzano, R., 2013. Satisfaction of
zation. Ritanjali Majhi: Writing – review & editing, Writing – original individual mobile phone users in Spain. Telecommun. Policy 37 (10), 940–954.
draft, Supervision, Formal analysis, Data curation, Conceptualization, Guidolin, M., Guseo, R., 2015. Technological change in the US music industry: within-
product, cross-product and churn effects between competing blockbusters.
Methodology. Technological Forecasting and Social Change 99, 35–46.
Gustafsson, A., Johnson, M.D., Roos, I., 2005. The effects of customer satisfaction,
Data availability relationship commitment dimensions, and triggers on customer retention. Journal of
marketing 69 (4), 210–218.
Hadden, J., Ashutosh, T., Rajkumar, R., Dymitr, R., 2007. Computer assisted customer
Data will be made available on request. churn management: state-of the art and future trends. Comput. Oper. Res. 34,
2902–2917.
Hatcher, L., O’Rourke, N., 2013. A Step-by-Step Approach to Using SAS for Factor
Appendices. Supplementary data Analysis and Structural Equation Modeling. SAS Institute, Inc, Cary, N.C.
Hochstein, B., Rangarajan, D., Mehta, N., Kocher, D., 2020. An industry/academic
Supplementary data to this article can be found online at https://doi. perspective on customer success management. J. Serv. Res. 23 (1), 3–7.
Huang, B., Mohand, T.K., Brian, B., 2012. Customer churn prediction in
org/10.1016/j.techfore.2024.123250.
telecommunications. Expert Syst. Appl. 39 (1), 1414–1425.
Hung, S.Y., Yen, D.C., Wang, H.-Y., 2006. Applying data mining to telecom churn
References management. Expert Systems with Applications 31, 515–524.
Hwang, H., Jung, T., Suh, E., 2004. An LTV model and customer segmentation based on
customer value: a case study on the wireless telecommunication industry. Expert
Adebiyi, S.O., Oyatoye, E.O., Amole, B.B., 2016. Relevant drivers for Customers’ churn
Systems with Applications 26 (2), 181–188.
and retention decision in the Nigerian Mobile telecommunication industry. Journal
Idris, A., Muhammad, R., Khan, A., 2012. Churn prediction in telecom using random
of Competitiveness 8 (3), 1–12.
Forest and PSO based data balancing in combination with various feature selection
Ahmad, A.K., Jafar, A., Aljoumaa, K., 2019. Customer churn prediction in telecom using
strategies. Computers & Electrical Engineering 38 (6), 1808–1819.
machine learning in big data platform. Journal of Big Data 6 (1), 1–24.
Ismail, M.R., Awang, M.K., Rahman, M.N.A., Makhtar, M., 2015. A multi-layer
Ahmed, A.A., Maheswari, D., 2017. Churn prediction on huge telecom data using hybrid
perceptron approach for customer churn prediction. International Journal of
firefly based classification. Egyptian Informatics Journal 18 (3), 215–220.
Multimedia and Ubiquitous Engineering 10 (7), 213–222.
Al-Refaie, A., Al-Tarawneh, M., Bata, N., 2018. Study of customer churn in the telecom
Jaiswal, A.K., Niraj, R., Park, C.H., Agarwal, M.K., 2018. The effect of relationship and
industry using structural equation modelling. Economy & Business Journal 12 (1),
transactional characteristics on customer retention in emerging online markets.
393–411.
J. Bus. Res. 92, 25–35.
Amin, A., Al-Obeidat, F., Shah, B., Adnan, A., Loo, J., Anwar, S., 2019. Customer churn
Jiang, P., Liu, Z., Zhang, L., Wang, J., 2023. Hybrid model for profit-driven churn
prediction in telecommunication industry using data certainty. J. Bus. Res. 94,
prediction based on cost minimization and return maximization. Expert Systems
290–301.
with Applications 228, 120354.
Anderson, E.W., 1994. Cross-category variation in customer satisfaction and retention.
Keaveney, S.M., 1995. Customer switching behavior in service industries: An exploratory
Mark. Lett. 5, 19–30.
study. J. Mark. 59 (2), 71–82.
APEDB, 2023. India overview of Telecom Sector. Available at: https://apedb.gov.in/tel
Keramati, A., Ardabili, S.M., 2011. Churn analysis for an Iranian mobile operator.
ecom.html.
Telecommunications Policy 35 (4), 344–356.
Ascarza, E., 2018. Retention futility: targeting high-risk customers might be ineffective.
Keramati, A., Jafari-Marandi, R., Aliannejadi, M., Ahmadian, I., Mozaffari, M.,
J. Market. Res. 55 (1), 80–98.
Abbasi, U., 2014. Improved churn prediction in telecommunication industry using
Ascarza, E., Iyengar, R., Schleicher, M., 2016. The perils of proactive churn prevention
data mining techniques. Appl. Soft Comput. 24, 994–1012.
using plan recommendations: evidence from a field experiment. J. Market. Res. 53
Khatri, H. 2022. India already has a large addressable market for 5G. Available at: https
(1), 46–60.
://www.opensignal.com/2022/09/07/india-already-has-a-large-addressable-mark
Ascarza, E., Neslin, S.A., Netzer, O., Anderson, Z., Fader, P.S., Gupta, S., Hardie, B.G.,
et-for-5g.
Lemmens, A., Libai, B., Neal, D., Provost, F., 2018. In pursuit of enhanced customer
Khodabandehlou, S., Zivari Rahman, M., 2017. Comparison of supervised machine
retention management: review, key issues, and future directions. Cust. Needs Solut.
learning techniques for customer churn prediction based on analysis of customer
5, 65–81.
behavior. J. Syst. Inf. Technol. 19 (1/2), 65–93.
Azadeh, A., Darivandi Shoushtari, K., Saberi, M., Teimoury, E., 2014. An integrated
artificial neural network and system dynamics approach in support of the viable

11
R. Thangeda et al. Technological Forecasting & Social Change 202 (2024) 123250

Kim, H.S., Yoon, C.H., 2004. Determinants of subscriber churn and customer loyalty in Sharaf Addin, E.H., Admodisastro, N., Mohd Ashri, S.N.S., Kamaruddin, A., Chong, Y.C.,
the Korean mobile telephony market. Telecommunications policy 28 (9–10), 2022. Customer mobile behavioral segmentation and analysis in telecom using
751–765. machine learning. Appl. Artif. Intell. 36 (1), 2009223.
Kim, M.K., Park, M.C., Jeong, D.H., 2004. The effects of customer satisfaction and Shieh, L.F., Chang, T.H., Fu, H.P., Lin, S.W., Chen, Y.Y., 2014. Analyzing the factors that
switching barrier on customer loyalty in Korean mobile telecommunication services. affect the adoption of mobile services in Taiwan. Technol. Forecast. Soc. Chang. 87,
Telecommunications policy 28 (2), 145–159. 80–88.
Kim, M.K., Park, M.C., Lee, D.H., Park, J.H., 2019. Determinants of subscriptions to Shin, D.H., Kim, W.Y., 2008. Forecasting customer switching intention in mobile service:
communications service bundles and their effects on customer retention in Korea. An exploratory study of predictive factors in mobile number portability. Technol.
Telecommunications Policy 43 (9), 101792. Forecast. Soc. Chang. 75 (6), 854–874.
Kostić, S.M., Simić, M.I., Kostić, M.V., 2020. Social network analysis and churn Shin, H.J., Cho, S., 2006. Response modelling with support vector machines. Expert
prediction in telecommunications using graph theory. Entropy 22 (7), 753. Systems with Applications. 30 (4), 746–760.
Kumar, S., 2022. Customer Retention Versus Customer Acquisition. Forbes. Available at: Sivadas, E., Baker-Prewitt, J.L., 2000. An examination of the relationship between
https://www.forbes.com/sites/forbesbusinesscouncil/2022/12/12/customer-retent service quality, customer satisfaction, and store loyalty. International Journal of
ion-versus-customer-acquisition/?sh=dac10a11c7d6. Retail and Distribution Management 28 (2), 73–82.
Lalwani, P., Mishra, M.K., Chadha, J.S., Sethi, P., 2022. Customer churn prediction Sohn, S.Y., Lee, J.K., 2008. Competing risk model for mobile phone service.
system: a machine learning approach. Computing 1–24. Technological Forecasting and Social Change 75 (9), 1416–1422.
Lamrhari, S., El Ghazi, H., Oubrich, M., El Faker, A., 2022. A social CRM analytic Sterman, J., Oliva, R., Linderman, K.W., Bendoly, E., 2015. System dynamics
framework for improving customer retention, acquisition, and conversion. perspectives and modeling opportunities for research in operations management.
Technological Forecasting and Social Change 174, 121275. J. Oper. Manag. 39, 40.
Lariviere, B., Van den Poel, D., 2005. Predicting customer retention and profitability by Sudharsan, R., Ganesh, E.N., 2022. A swish RNN based customer churn prediction for the
using random forests and regression forests techniques. Expert Syst. Appl. 29 (2), telecom industry with a novel feature selection strategy. Connect. Sci. 34 (1),
472–484. 1855–1876.
Lee, J., Lee, J., Feick, L., 2001. The impact of switching costs on the customer TRAI, 2023. Highlights of Telecom Subscription Data. Available at: https://www.trai.
satisfaction-loyalty link: mobile phone service in France. J. Serv. Mark. 15 (1), gov.in/sites/default/files/PR_No.05of2023_0.pdf.
35–48. Ullah, I., Raza, B., Malik, A.K., Imran, M., Islam, S.U., Kim, S.W., 2019. A churn
Lejeune, M.A., 2001. Measuring the impact of data mining on churn management. prediction model using random forest: analysis of machine learning techniques for
Internet Res. 11 (5), 375–387. churn prediction and factor identification in telecom sector. IEEE access 7,
Lemmens, A., Gupta, S., 2020. Managing churn to maximize profits. Mark. Sci. 39 (5), 60134–60149.
956–973. Venkatesan, R., Kumar, V., 2004. A customer lifetime value framework for customer
Leong, V.S., Ahady, D.M.H., Muhamad, N., 2022. Corporate image as an enabler of selection and resource allocation strategy. Journal of marketing 68 (4), 106–125.
customer retention. Int. J. Qual. Serv. Sci. 14 (3), 486–503. Verbeke, W., Martens, D., Mues, C., Baesens, B., 2011. Building comprehensible customer
Levesque, T., McDougall, G.H., 1996. Determinants of customer satisfaction in retail churn prediction models with advanced rule induction techniques. Expert Syst. Appl.
banking. Int. J. Bank Mark. 14 (7), 12–20. 38 (3), 2354–2364.
Lewis, M., 2005. Incorporating strategic consumer behavior into customer valuation. Verbeke, W., Dejaeger, K., Martens, D., Hur, J., Baesens, B., 2012. New insights into
J. Mark. 69 (4), 230–238. churn prediction in the telecommunication sector: a profit driven data mining
Majhi, R., Panda, G., Sahoo, G., 2009. Development and performance evaluation of approach. Eur. J. Oper. Res. 218 (1), 211–229.
FLANN based model for forecasting of stock markets. Expert systems with Villanueva, J., Yoo, S., Hanssens, D.M., 2008. The impact of marketing-induced versus
applications 36 (3), 6800–6808. word-of-mouth customer acquisition on customer equity growth. J. Market. Res. 45
McCarty, J.A., Hastak, M., 2007. Segmentation approaches in data-mining: a comparison (1), 48–59.
of RFM, CHAID, and logistic regression. J. Bus. Res. 60 (6), 656–662. Winer, R.S., 2001. A framework for customer relationship management. Calif. Manage.
Mukhopadhyay, A., Bagchi, K.K., Udo, G.J., 2023. Exploring the main factors affecting Rev. 43 (4), 89–105.
mobile phone growth rates in Indian states. J. Knowl. Econ. 1–23. Yu, R., An, X., Jin, B., Shi, J., Move, O.A., Liu, Y., 2018. Particle classification
Neslin, S.A., Gupta, S., Kamakura, W., Lu, J.X., Mason, C.H., 2006. Defection detection: optimization-based BP network for telecommunication customer churn prediction.
measuring and understanding the predictive accuracy of customer churn models. Neural Comput. & Applic. 29, 707–720.
J. Market. Res. 43 (2), 204–211. Zahavi, J., Levin, N., 1997. Applying neural computing to target marketing. J. Direct
Óskarsdóttir, M., Bravo, C., Verbeke, W., Sarraute, C., Baesens, B., Vanthienen, J., 2017. Mark. 11 (1), 5–22.
Social network analytics for churn prediction in telco: model building, evaluation
and network architecture. Expert Systems with Applications 85, 204–220.
Dr Rahul Thangeda is currently working as Assistant Professor at School of Management,
Pakurár, M., Haddad, H., Nagy, J., Popp, J., Oláh, J., 2019. The service quality
National Institute of Technology Warangal, India. His research interests include solving
dimensions that affect customer satisfaction in the Jordanian banking sector.
consumer behaviour related problems using machine learning and Neural Network based
Sustainability 11 (4), 1113.
models.
Pasape, L., 2022. Enhancement of customer retention of telecommunication companies
in Tanzania through customer satisfaction. The Strategic Journal of Business &
Change Management 9 (3), 218–231. Professor Niraj Kumar (> 90 publications) is Professor of Operations Management and
Praseeda, C.K., Shivakumar, B.L., 2021. Fuzzy particle swarm optimization (FPSO) based the Head of the Strategy, Operations and Entrepreneurship group at Essex Business School,
feature selection and hybrid kernel distance based possibilistic fuzzy local University of Essex. His research interests are primarily in the area of operations man­
information C-means (HKD-PFLICM) clustering for churn prediction in telecom agement, big data practices and decision support systems. He has led and contributed to
industry. SN Applied Sciences 3, 1–18. several research projects funded by EU-Horizon 2020, Innovate UK, Newton Fund, British
Reichheld, F.F., 2003. The one number you need to grow. Harv. Bus. Rev. 81 (12), 46–55. Academy, British Council and UKIERI.
Reichheld, F.F., Sasser, W.E., 1990. Zero defections: quality comes to services. Harv. Bus.
Rev. 68 (5), 105–111.
Professor Ritanjali Majhi is currently working as Professor at School of Management,
Reinartz, W., Kumar, V., 2002. The mismanagement of customer loyalty. Harv. Bus. Rev.
National Institute of Technology Karnataka, India. She has published several papers in
80 (7), 86–94.
peer reviewed journals. Her research areas are applying soft and evolutionary computing
Rust, R.T., Lemon, K.N., Zeithaml, V.A., 2004. Return on marketing: using customer
models to marketing problems.
equity to focus marketing strategy. Journal of marketing 68 (1), 109–127.
Saleh, S., Saha, S., 2023. Customer retention and churn prediction in the
telecommunication industry: a case study on a Danish university. SN Applied
Sciences 5 (7), 173.

12