Telecom Business Review

12 (1) 2019, 46-49

http://publishingindia.com/tbr/

Big Data Analytics for Customer Lifetime Value

Prediction
Aslekar Avinash*, Piyali Sahu**, Arunima Pahari***
*Associate Professor, Symbiosis Institute of Telecom Management, Constituent of Symbiosis International
(Deemed University), Maharashtra, India. Email: [email protected]
**Student (Analytics and Finance), Symbiosis Institute of Telecom Management, Constituent of Symbiosis
International (Deemed University), Maharashtra, India. Email: [email protected]
***Student (Analytics and Finance), Symbiosis Institute of Telecom Management, Constituent of Symbiosis
International (Deemed University), Maharashtra, India. Email: [email protected]

Abstract
Predicting all the values that a business can derive from its long-standing relationship with its customers is Customer Lifetime
Value. As it helps to create a sustainable relationship with selected customers, the importance of CLV is growing at a brisk pace,
generating higher revenue that in turn enhances business growth. With increasing competition, retaining existing customers is
more profitable than acquiring new customers. To manage and allocate resources efficiently for each and every customer, big
data analytics comes into play. Taking this into account, a large amount of data should be taken into consideration, such as
client's attrition, objectives, diverse products and services that they use, client's characteristics viz demographic, psychographic,
geographic etc. First step to this is data cleaning, pre-processing and data manipulation to achieve a meaningful outcome or
information from the raw data, followed by data analysis and visualization. Techniques that can be recommended for the data
analysis and visualization of CLV model can be Stepwise regression, Classification and regression trees (CART), Generalized
linear models (GLM). To determine the dynamic view of customer behavior, future marketing strategies and to foster brand loyalty,
prediction of a proper CLV model is much needed.
Keywords: CLV, Predictive Models, Pareto/NBD Model, Gamma-Gamma Model, Retail Industry CLV, Purchase Count, Lifetime
Value, Monetary Value

1. Introduction Data growth can be classified according to - amount of

input data (volume of data), data input type (variety of
data), faster ingest of data (velocity of data), data quality
The enhancement of faster processing through Pentium
or trustworthiness (veracity of data). Through big data
processors introduced back in the early 1990s marked the
analytics it is easier to find unknown patterns and market
beginning of the era of information economy. It improved
trends, correlations, association, customer preference and
the processing speed of the computers, the ability to run
other actionable insights. The analytical observations can
state-of-the-art software devices and store huge amount of
lead to better decision making, finding out more effective
data (Kudyba, 2002). Big Data is one of the most hyped
terms in today’s market and there is no consensus as how marketing strategies, new revenue opportunities, improve
to define it. It is not a specific application type but can be operational efficiency. It gives competitive advantage over
rather considered as a collection of trends observed from its competitors and provide with business benefits. Big
multiple application types. Big data can include data from data is used with related concepts of Business Intelligence
web server logs, internet clickstream data, social media (BI) and data mining. But the hyped term is different
content, social media activity reports, customer emails, from the other two when data complexity, data volume
surveys, CDRs, data from sensors connected to Internet and the huge data sources are taken into consideration.
of Things. Data warehouses may not be capable of handling the high
 Big Data Analytics for Customer Lifetime Value Prediction  47
processing demands faced due to huge volume of data sets transaction and how many transactions the customer has
which needs to be frequently updated or even in real-time. made. The demographic or the earlier transaction details
With the purpose of collecting, processing and analyzing are not considered.
of big data, a set of new technologies including Hadoop
and tools such as YARN, Spark, Hive are budding. These
technologies support the processing of vast and diverse
2.2 Gamma-Gamma Model: Extension to the
data across clustered systems. Pareto/NBD Model
With the increase in customer transaction data it has been
Pareto/NBD focuses only on the number of transactions/
quite an interest to estimate the value of customers or
purchase count throughout lifetime but does not focus on
assessment of customers. This is an important trend in the
the monetary value component whereas; Gamma-Gamma
disciplines such as - accounting, finance and especially
model focuses on the monetary value also.
marketing catering to various sectors (Petrison, Blattberg
& Wang, 1993). Gamma-Gamma model is based on three general
assumptions:
2. Probability Models ∑∑ The monetary value of a transaction randomly var-
2.1 Pareto/NBD Model ies around their average transaction value at cus-
tomer level.
Pareto/NBD Model is one of the most widely used methods
for calculating Customer Lifetime Value (Schmittlein, ∑∑ The average transaction value is independent of
Morrison and Colombo, 1987). It is used to predict the time for an individual but it varies across customers.
future activities of the customer. The order history is ∑∑ The distribution of average transaction does not de-
considered to be the primary input and takes into order pendent of the transaction process across customers.
the frequency and recency of the orders. The basic model
Hence, monetary value i.e. Gamma-Gamma model
stimulates two events - coin and dice. The coin is used
can be formed separately from the lifetime compo-
to determine the churn rate of the customers and dice to
determine how many items a customer will order. Pareto nents and purchase count of the model i.e. the Pareto
distribution is used to model the coin whereas negative model.
binomial distribution is used to model the dice.
The prediction done by this model rests on the following 3. Two Models Together: Estimation
assumptions (Charu, Trisha, Jarred & Edward, 2013): of CLV at Customer Level
∑∑ Consider the customer to be alive even after they
become permanently inactive. ∑∑ The expected number of purchases done by the cus-
∑∑ While alive, the customer’s purchasing behavior tomer in a forecasted period is computed by Pareto/
follows a Poisson process with a transaction rate to NBD model.
be considered. ∑∑ The monetary value to each of these future purchas-
∑∑ There is heterogeneity in transaction rates and drop- es is assigned by Gamma-Gamma model.
out rates across customers that follow a gamma
∑∑ To forecast CLV for each customer became simpli-
distribution.
fied tying these two models. Comparing CLVs be-
This model focuses only on the time length for which came easier considering the monetary value of life-
a customer is observed, time of the customer’s last time purchases.
 48   Telecom Business Review: SITM Journal Volume 12 Issue 1 September 2019

Fig. 1: Two Models: Pareto/NBD and Gamma-Gamma

4. Use Case for Retail Industry ∑∑ Prediction of CLV Using Probabilistic Models:
These models (Pareto/NBD model) work on a meth-
A non-contractual relationship exists between businesses od which predicts a customer’s expected purchases
and customers in e-Commerce or retail business. In the in the next period going by their transactional histo-
non-contractual world, churning of customers happens ry. Data should be predicted by the model and tested
silently. So, it is more complex to model customer and cross-validated by machine learning practices
lifetime value for non-contractual businesses and we need to achieve higher model accuracy for predicting
prediction for this. CLV. According to that trained model, the histori-
∑∑ Data to Analyze cal probability of customer being alive is predicted,
o Customer’s last transactions. given a customer’s transaction history. For predict-
ing monetary value, Gamma-Gamma model is ideal,
o Is the customer “dead”, or is the customer alive
since this model considers the economic component
but dormant?
of each transaction and then estimates customer’s
o What is the frequency of repeat purchases made probability to remain “alive” indicating proper cus-
by the customer? tomer’s CLV prediction in retail businesses.
o The age of the customer or the duration between
a customer’s first purchase and the end of the pe- 5. Value Addition of Pareto and
riod (under consideration). Gamma-Gamma Model for
o Recency or the age of the customer when they Predictive CLV Calculation
made their most recent purchases (Thus if only 1 Historical customer lifetime value is the simplest
purchase is made, the Recency is 0.). approach for calculating CLV. The historical methods
∑∑ RFM (Recency, Frequency, Monetary) Analysis: do not account for time. It is solely based on the past
transactions but doesn’t help in predicting the future
RFM analysis is used to quantitatively determine the
activities of the customer. This approach is valid if the
recency of purchase done by the customer, purchas-
customers have same behavioral pattern and over a same
ing frequency of a customer and the expense done time span. Among customers, there can be a good amount
by the customer (monetary) from this, probability of heterogeneity. Historical approach of CLV estimation
of customers who are surely alive, can be predicted. will apply recency of the last purchase, thereby filtering
Customers having very high frequency and very high out the criteria of segregation between the non-active and
recency are likely to be the best customers in future. the active users.
 Big Data Analytics for Customer Lifetime Value Prediction  49
The aim of a predictive model is identifying the References
purchasing behavior of customers in order to predict the
Custora, U. (n.d.). Pareto/NBD model. Retrieved from https:
future activities of a customer. In the non-contractual //university.custora.com/for-marketers/clv/advanced/
business context, Pareto model is a frequently applied pareto-nbd
probabilistic model. It focuses on three latent behaviors,
Fader, P. S., & Hardie, B. G. S. (2013). The gamma-gam-
the time period for which the customer is in relationship
ma model of monetary value. Retrieved from http://
with the company or defining customer lifetime behavior www.brucehardie.com/notes/025/gamma_gamma.pdf
and the number of purchases made by the customer over a
Kudyba, S. (2002). Pentium processors improved the pro-
given period or the purchase rate. The lifetime distribution
cessing speed of the computers. Retrieved from https:
follows an exponential distribution with slope μ and the
//www.modernanalytics.com/wp-content/uploads
purchase count follows a Poisson distribution with rate λ.
/2014/07/Chapter-1.pdf
The Pareto model is trained to find out the parameters over
MyCustomer. (n.d.). Customer lifetime value: How on-
a particular training period within a minimum time span
line retailers can measure and use CLV. Retrieved
that corresponds to thrice of the typical inter-purchase
from https://www.mycustomer.com/community/blogs/
time of the customers. Pareto model takes the historic or
silviya-dineva/customer-lifetime-value-how-online-
transaction data as primary input to predict for the future, retailers-can-measure-and-use-clv
thereby helping in business contexts.
Petrison, B., & Wang, P. (1993). During the past cen-
tury, database marketing techniques have become in-
6. Conclusion creasingly important. Retrieved from https://www.
Pareto/NBD model mainly focuses only on modelling scholars.northwestern.edu/en/publications/database
purchase count and lifetime but Gamma Gamma model -marketing-past-present-and-future
provides monetary aspect of the future purchases also. Schmittlein, D. C., Morrison, D. G., & Colombo, R.
Hence, Gamma Gamma model is basically the monetary (1987). Counting your customers: Who are they
value extension to the Pareto/NBD model. The expenditure and what will they do next? Management Science,
of a client is assumed to be independent of the transaction 33(1), 1-24. Retrieved from https://www.jstor.org/
behavior in case of Gamma-Gamma model. Models stable/2631608?seq=1#metadata_info_tab_contents
like Pareto/NBD only predict the transaction behavior, Towards Data Science Pages. (n.d.). Retrieved from https:
independent of the monetary value and this limitation is //towardsdatascience.com/whats-a-customer-worth
being nullified by Gamma-Gamma model. -8daf183f8a4f
Copyright of Telecom Business Review is the property of Publishing India Group and its
content may not be copied or emailed to multiple sites or posted to a listserv without the
copyright holder's express written permission. However, users may print, download, or email
articles for individual use.