Jump to the main content block

Marketing Analytics

Manage Customer Lifetime Value (CLV) by Big Data Analytics

National Sun Yat-sen Univ. Business Management Dept.

 


Data Driven Customer Value Management Framework

 

 

Abstract:  Customer centricity and value orientation has been the main themes of marketing, and customer lifetime value (CLV) is the core concept in these trends. In the literature, CLV used to be estimated by multi-variate, econometric or hierarchical Bayesian methods. Basically these methods predict CLV from a pre-determined, limited set of independent variables. Since the advent of digital marketing, when companies gather customer information from every digital marketing media (company we site, ecommerce sites, Google AdWords/Analytics, etc.), they usually have tens of thousands of variables (extracted from site visits, page views and click streams) per customers, which cannot properly be analyzed by the traditional methods. In this project we plan to incorporate predictive modeling (an advanced analytics technologies evolved from AI and machine learning) into customer value management framework (Kumar 2008); and use case studies to demonstrate how companies can make use of the ‘big’ data from the web, to improve their performance in digital marketing. Our objective is: to propose a managerial framework that applies big data analytics to customer (value) management; thereby the marginal effect of every marketing effort for every individual customer can be evaluated; thereby every marketing practice can be fully customized for optimal performance. Leveraging machine learning methods and predictive modeling techniques, we start by forecasting each individual’s CLV, and plotting the distribution of CLV for the customer base. We then divide the customer base into value-distinctive segments, estimate transition flows among the segments, and evaluate their average CLV’s separately. For each customer acquisition, development and retention instruments, we then build predictive models that forecast the responding probability and expected revenue for each customer; thereby managers can allocate and execute their marketing budget effectively. In the pioneer stage of this project, we have validated our procedure in a case study of churn management. With very limited CRM data, our model predicted churns correctively (Predicting AUC = 92.94%), select retention targets properly, and is expected to improve the performance of retention significantly (from –536,813 to +722,703). In the subsequent stages of this project, we plan to validate and refine our procedure for the other industries (telecom, e-commerce, etc.) Our final goal is to develop a common framework that generally applies to most B2C businesses, and may supplement it by specialized guidelines for certain specific industries.

Keywords: customer management, customer lifetime value, CLV, big data analytics, predictive modeling 

 

中文摘要

當今的行銷管理主要強調『顧客中心』和『價值導向』,其核心概念是『顧客終生價值 (customer lifetime value, CLV』,而以CLV為中心的行銷策略與實務,便統稱為『顧客價值管理(customer value management)』。過去這一方面的文獻主要以多變量、計量經濟、層級貝氏模型來推估CLV,這一些方法基本上都是以有限數目的、已知的自變數(以及其機率分布)來推估應變數。然而在『大數據』時代,公司常會透過各種數位行銷媒介(如官網、電商網站、社群網站、APP等等收集資料,針對每一位顧客通常可以整理出成千上萬個屬性變數(如針某一網頁的造訪次數、瀏覽流程、點擊順序等等)。由於變數的數量龐大,傳統的CLV方法並不能完全利用這些資訊。本案之中,我們計畫將近期由機器學習領域發展出來的『預測性模型 (predictive modeling)』,引進到『顧客價值管理­架構(CLV management framework, Kumar 2008) 』裡面,並且使用實際案例,演示如何在顧客價值管理流程裡面,全面性的、有效的利用數位行銷收集到的顧客資料。

本案的目標是提出一套顧客價值管理­架構,運用大數據分析技術做預測,使得每一個行銷手法應用在每一個顧客上的效果都能被估算,進而讓每一個行銷方案都能根據顧客的特性量身定制以達到最佳的效果。以機器學習和預測性模型為基礎方法,本計劃首先將估計每一位顧客之CLV,並描繪出整個顧客群的CLV分佈。接著,根據價值高低之差異將顧客群做區隔,估計每一位顧客在不同區隔之間轉換的機率,再分別計算他們的平均CLV。最後,從顧客獲取(acquisition)、顧客發展(development)和顧客保留(retention) 三個不同面向,我們會分別建立對應的預測性模型;用來預測每一位顧客、對每一個行銷方案的回應機率和預期收益,讓管理者可以運用數據來制定策略,更有效率地分配、執行行銷經費。

為了驗證我們的初步構想,我們我們已經將上述的方法和流程實際運用到一個的前導案例(有線電視與寬頻上網運營商)之中。初步的結果顯示,本計劃所建構之模型能夠正確的辨識出即將流失的顧客 (split sample AUC = 95.88%; predicting AUC = 92.94%),挑選出最值得保留的顧客,並能顯著改善顧客保留方案的績效(獲利從−536,813上升至+722,703)。本案進行過程之中,我們將與更多廠商(行動電話、電商平台)合作,持續的驗證、改進我們的方法和管理流程,建構一個大多數本土廠商都可以引用的、數據導向的顧客價值管理架構。

關鍵字:顧客管理顧客終身價值, CLV, 大數據分析預測性模型建構

 

研究計畫之背景。請詳述本研究計畫所要探討或解決的問題、重要性、預期影響性及國內外有關本計畫之研究情況、重要參考文獻之評述等。如為連續性計畫應說明上年度研究進度。

 

1. Introduction

Customer centricity and value orientation have been the main themes of marketing (Shah et al. 2006; Kumar & Rajan 2012), and customer lifetime value (CLV) is the core concept in these trend (Blattberg et al. 2009; Gupta et al. 2004; Kumar et al. 2006; Kumar et al. 2008). In the literature, CLV used to be estimated by multi-variate, econometric or hierarchical Bayesian methods. Basically these methods predict CLV from a pre-determined, limited set of independent variables. Since the advent of digital marketing, when companies gather customer information from all digital marketing media (company we site, ecommerce sites, Google AdWords/Analytics, etc.), they usually have tens of thousands of variables (extracted from site visits, page views and click streams) per customers, which cannot properly be analyzed by the traditional methods. In this project we plan to incorporate predictive modeling (an advanced analytics technologies evolved from AI and machine learning) into customer value management framework (Kumar 2008); and use case studies to demonstrate how companies can make use of the ‘big’ data from the web, to improve their performance in digital marketing.

When we contact the local industries, we find that most companies are enthusiastic in Big Data; many of them had invested a considerable amount of resource in chasing the zeitgeist; but few of them have ever captured the true potential of digital marketing and web analytics, less realizing substantial benefit. Some of them simply make aggregated statistics out of their existing CRM data; some are paralyzed by the intimidating amount of data generated from the WWW; others stay behind the sidelines, keep waiting for “enough” of data and wondering how “big” the machine should be, before they could take action. Actually, to leverage big data analytics, it doesn’t always take big data and large machine. For most of the SMB’s, the predictive modeling technique (a key part of big data analytics) can be executed on regular desktop computers, and could make good predictions out of even the simplest transaction logs and CRM data.

In order to help the local industry in harnessing these powerful techniques, we plan to develop a managerial framework that elaborates how companies can apply big data analytics to improve the efficiency of customer managements. Our objective is: to propose a framework procedure that applies data analytics to customer (value) management; thereby the marginal effect of every marketing effort for every individual customer can be evaluated; thereby every marketing practice can be fully customized for optimal performance. Leveraging machine learning methods and predictive modeling techniques, we start by evaluating each individual’s CLV, and plotting the distribution of CLV for the customer base. We then divide the customer base into value-distinctive segments, estimate transition flows among the segments, evaluate their average CLV’s separately, and forecast how the CLV’s might evolve overtime. For each customer acquisition, development and retention instruments, we then build predictive models that forecast the responding probability for each customer. Shortly speaking, our models would help to select target customers and estimate the expected return for every specific marketing tool; thereby managers can allocate the execute their marketing budget effectively.

V. Kumar could be the most influential scholar in the realm of customer management (Kumar 2008, 2011; &George 2007; &Rajin 2012; &Shah 2009; &Umashankar 2012; et al. 2006, 2008, 2010, and so on.) While the others mostly focused on certain specific aspects of CLV (evaluation, modeling, optimizing, statistical/analytical methods, etc., to be elaborated separately in section 2), Venkatesan & Kumar (2004) proposed a practical framework that systematically incorporated all CLV-relevant methods in making strategic decisions. They then applied this framework procedure in a real business case (with IBM), which resulted in a prized article in marketing science (Kumar et al. 2008). By elaborating the end to end process – from CLV evaluation/prediction, customer segmentation/selection, to the practices of customer acquisition, development and retention, to the strategies of budge and resource allocation, Kumar (2008) integrated the relevant theories and methods, and practically put them to work. With this pragmatic spirit, he successfully positioned CLV at the core CRM, and erected ‘customer centricity’ as one of the main themes in management science.

On the other hand, big-data analytics is a pragmatic science as well (James et al. 2014). Contrasting to the traditional methods primarily built for testing theoretical hypotheses, the techniques of predictive modeling are fully specialized in making predictions (Pentland 2014). Less concerning theoretical parsimony, these techniques can fully utilize the richness of big data, and usually lead to better predictions (Kuhn & Johnson 2013; Lu 2010; Trevor et al. 2009). Recently some researchers have started to use the big data analytics in customers management (e.g., Xiao et al. 2014); but instead of taking a holistic aspect (as in Kumar’s 2008), they mostly concentrated on specific algorithms and methods. With this project we aim to incorporate the pragmatic science of data analytics with the practical framework of customer management. Procedure-wise, we’d extend Kumar’s (2008) customer management framework by incorporating web analytics (Peterson 2004; Kaushik 2009) and digital marketing (Edelman 2010; Chaffey 2012); method-wise, we’d leverage on the leading edge technologies of cloud computing and predictive modeling (James et al. 2014; Kuhn & Johnson 2013; Trevor et al. 2009) to improve the performance of customer management.  

Inspired by Kumar’s (and followers’ such as Payne & Frow 2005; Holm et al. 2012) framework and case studies (Kumar 2008), we plan to validate, refine and extent our preliminary framework with a series of real business cases. In the pioneer stage of this project, we have tried our procedure in a business case of churn management. Fitted by extreme gradient boosting algorithm (a.k.a. ‘xgboost’, see Chen & He 2015), our model predicted churns correctively (Split sample AUC = 95.88%; Predicting AUC = 92.94%), helped to identify a priority list of targeted customer (1,116 out of 13,307 subscribers), and is expected to improve the performance of retention significantly (from –536,813 to +722,703). In the subsequent stages of this project, we also plan to test and adjust our procedure for mobile network and e-commerce platform operators. Our final goal is to develop a common managerial framework that generally applies to most B2C businesses, and supplement it with specialized guidelines for specific industries or business circumstances.

This project is expected to create value in three directions. First of all, the targeted framework procedure might serve as a start-up kit for local companies to jump start their big data initiatives. Rather than waiting for a “mature technological infrastructure” which might never be realized, local companies can start with basic transaction logs and CRM data, experience the power of big data, and enjoy its benefit right away. Second, in this project we will bring together and blend managerial, statistical and computational talents. Cross-disciplinary skill (mind) sets have repeatedly been regarded as the largest barrier in big data applications (Kuhn & Johnson 2016; Pentland 2014). Our project would develop cross-disciplinary talents in business school and our assisting companies, and help to relieve the human resource shortage. Third, while most of the recent studies focus on the technological details in applying specific analytics technique to specific tasks, we take a pragmatic perspective, impose managerial mind set, and emphasize on applying various analytical techniques on the entire cycle of customer management – from evaluating CLV, clustering/segmentation, to customer acquisition, customer development and customer retention, to marketing instrument selection and resource allocation. One the one hand, it is a systematic yet easy-to-get-started guideline for practitioners to introduce data analytics to their companies. On the other hand, it might serve as a preliminary blueprint, upon which a comprehensive framework for data-driven customer value management can be developed in the future.

 

2. Theoretical Background & Research Objectives

2.1 Trends in Marketing Management

Consumer Centricity & Value Orientation

Besides the traditional topics of market (target) segments, brand position and product oriented marketing mix (the 4 P’s, see Kotler & Armstrong 1996), customer centricity should be regarded as the most influential concept in marketing management today (Kumar 2011; Shah 2006). Prestige business schools, such as Harvard Business School, Wharton Business School (in U. Pennsylvania) and Sloan School of Management (in MIT), all place customer centricity at the core of their marketing curriculum (Gupta 2014). Basically, it emphasizes that the value of companies always come from their customers (Graf & Maas 2014; Kumar et al. 2006). Every marketing strategy should always starts from a company’s value proposition to its customers (Kumar et al. 2010). Furthermore, contrasting to companies’ value to their customers, customers’ value to the companies might be even more accountable in finance (Koosha & Albudvi 2015; Kumar et al. 2009), and more manageable in the practice of marketing. In this light, the concept of customer lifetime value (originally coined by Kolter in 1974) has become a main theme in the realms of marketing and management.

Consumer Lifetime Value (CLV)

In the aspect of (managerial) accounting, a customer’s value to a company (CLV) can be estimated by the sum of discounted cash flows (DCF) in her lifetime. In this perspective, the average CLV of a customer segment (or the entire customer base ) used to be estimated as:

where  represents the average margin (net profit), retention rate, interest rate and  growth rate of the segment (Rust et al. 2004; see Kumar & George 2007 for a Markov version).

        In addition to the basic (and arguably the most popular) form in equation (1), CLV could be evaluated in many other ways (see EsmaeiliGookeh & Tarokh 2013; Gupta et al. 2006 for reviews).  In one way or the other, CLV is always calculated from forecasted probability and amount of future purchases (Jain & Singh 2002). In order to improve the accuracy of forecasts, in the past decades several advanced statistical techniques have been adopted, including Pareto/NBD model (Schmittlein et al. 1987, improved by Reinartz & Kumar 2000 to accommodate non-contractual settings), BG/NBD model a.k.a. “Buy Till You Die” method (Fader & Hardie 2009; Jerath et al. 2011), Survival Analysis (Lu and Park 2003) and Hierarchical Bayesian models (Abe 2009; Karvenen et al. 2014) fitted by MCMC (Markov Chain Monte Carlo, a.k.a. Gibbs Sampling, see Andrieu et al. 2003) algorithm. Berger & Nasr (1998) proposed to apply Markov chain on the transition flows among customer segments. It supplemented the basic model with better flexibility (Mzoughia & Limam 2014) and make it possible to incorporate metrics from customer relationship management (Pfeifer & Carraway 2000; Romero et al. 2013).

Contrasting to these advanced (and somewhat complicate) methods, Fader at el. (2005) proposed to evaluate CLV in a “easy way”. Instead of working on the entire sequence, they transform each customer’s transactions into a triplet – Recency, Frequency and Monetary (RFM, see Gupta et al. 2006) before fitting the BG/NBD model. Perhaps due to its convenience, this approach is still widely adopted today (Nikumanesh & Albadvi 2014; Zhang at el. 2014). Similar to the concept of brand equity (Aaker 2009; Keller 1993), customer is also treated as companies’  equity (Blattberg & Deighton 1996; Kumar & Shah 2009). Recently even financial engineering methods are adopted in evaluation Customer Equity (Kumar & Umashankar 2012; Koosha & Albadvi 2015; Petersen & Kumar 2015).

Usually customers are different from each other; each contributes different revenue, incurs different cost and their net values (to the company) varies widely. In a case study, Kaplan (1989) found that, the first 20% of customers contributed 200% of a company’s total profit, and only a half of customers were profitable. If we sorted the customers by their values and align them on a horizontal line, when we plotted their cumulated profit over the line, we would see the famous whale curve (Kaplan & Atkinson 2015; Pfeifer et al. 2005). The hump of the curve indicates that companies may increase their profits by selectively foregoing some of their customers (Shin et al. 2012). In other words, for better financial performance, managing the composition of a company’s customer base is nonetheless important than increasing its size (Johnson & Selnes 2004; Schweidel et al. 2011).

Customer Management

Against the somehow outdated mottos of “customer is always right” and “the more customers the better”, proponents of customer management (Kaplan & Narayanan 2001; Bowman & Narayandas 2004) believe that companies should actively screen for ‘good’ customers and forego the ‘bad’ to maintain a lucrative customer portfolio (Schweidel et al. 2011; Tarasi et al. 2011). Customers come and go. To maintain and forester a sizable and profitable customer base, company need to pay attention in three tasks continually. First of all, they have to identify and acquire the most valuable customers. Second, they develop and transform the lower value customers into high value ones. Third, they retain their high value customers, and forego the customers with low or negative values. In other words, companies need to conduct customer acquisition (Lhoest-Snoeck et al. 2014), customer development (Dong et al. 2011) and customer retention (Aflaki & Popescu 2013; Mathur & Kumar 2013; Polo et al. 2011) in a ‘selective’ manner. Moreover, the effectiveness and efficiency of customer management depend on how well a company can make such ‘selections’ (Venkatesan & Kumar 2004).

2.2 Big Data Analytics

Big data helps to make good selection; but nothing is taken for granted. To make good selections out of big data, it involves new concepts, new analytical tools, new technologies and new processes, which could be unfamiliar to traditional analysts and researchers. Hereafter, we will refer these new techniques jointly as data analytics.    

From Multi-Variate Analysis to Bayesian Data Analytics

Basically, we build model to predict, and then use predictions to select. As we have mentioned in subsection 2.1, the modeling methods in predicting CLV have evolved from traditional multi-variate to Bayesian econometrics in the pasted two decades. These methods work for data extracted from CRM databases, but might become inadequate in analyzing ‘big data’ – the vast volume and variety of data gathered from web analytics, digital marketing and IoT (Hui 2014; McAfee et al. 2012; Yen & Ya0 2015).

Multi-variate analysis is still widely adopted in traditional market research. Under the paradigm of hypothesize and test, most of the multi-variate methods test for significant relationship between dependent and independent variables (Hayes 2013). However, such a concept of ‘significance’ does not really apply to big data. “When the sample is large enough, almost everything is significant” said Alex Pentland (2014), the founder of MIT media lab.

Most of those frequentist (a.k.a. Fisher-ian) methods were developed by 1980. In order to make meaningful inferences with limited computing resource, they made strong assumptions (on the distribution of random variables, see Hayes 2013). Since the 80’s, mixed-effects (Pinheiro & Bates 2006) and Bayesian (Bolstad 2013) methods had started to relax such assumptions; so that the resulting models had better flexibility to reflect the ever-growing complexity in human society (Berger 2013; Bolstad 2013). With multi-core CPU, the MCMC algorithm now can be executed on regular PCs; and that further promote the popularity of hierarchical Bayesian methods. Bayesian method is good at revealing latent insights with ‘limited’ amount of data (Berger 2013; Kruschke 2010); therefore it is widely adopted in machine (deep) learning and engineering applications, such as natural language processing, pattern recognition, etc. On the contrary, in today’s business environment where data is ‘big’ rather than ‘limited’, a new paradigm of data analysis – predictive modeling (explained below), has emerged.

Predictive Modeling & Selective Customer Management

Since the advent of e-commerce, companies gathered data before knowing what to do with it; and such accumulation of data is expected to go even faster under the trends of ‘big data’ and ‘internet of thing (IoT)’. Nowadays companies contact consumers in many digital channels – company web site, affiliate/referral sites, conventional and key-word ads, reviews and comments site such as Yelp, social network such as Facebook, content marketing channels such as YouTube, apps for mobile phones and so on. When companies record consumers’ every clicks on every digital marketing channels, soon they’d have tens of thousands of data fields per customer – an amount that cannot be properly analyzed with the aforementioned methods.

To cope with the volume, variety, velocity and veracity of such big data (Gupta et al. 2014), a new set (200+ methods, see Kuhn & Johnson 2013) of analytics methods – jointly named as predictive modeling (Kuhn & Johnson 2013; James et al. 2014), has become popular in the past few years. These methods basically evolved from the disciplines of AI and machine learning. Like their traditional counterparts, they also make use of analytical models, whereas their major purpose is to predict the future, instead of validating theories or testing hypotheses. Fueled by parallel computing technologies, predictive models can accommodate hundreds of millions (even billions) of records, with tens of thousands (even millions) of predictors per record.

Predictive modeling is widely adopted in leading digital companies: Google and Facebook to predict click through rates; Amazon and Netflix for product and movie recommendations; mobile network operators in predicting churn, marketing research firms in predicting response rate, and so on. Backed by cloud computing technologies, they build predictive models out of trillion of records, with millions of variables per record.

Notably, companies don’t need to be ‘big’ to utilize and take advantage of this new technology. Companies of any size, from thousands to millions of customers, with hundreds to tens of thousands of variables in CRM, can build predictive models simply by open source software and regular desktop PC’s. Furthermore, predictive models are very useful in customer management. Specifically, it helps to predict CLVs, risk of churn/foreclosure, and the response/conversion rates of recruiting, developing and retention campaigns. Predictions facilitate selections; thereby predictive models facilitate selective customer management. With predictions made by predictive models, modern companies can optimize their customer acquisition, development and retention programs, down to the per customer per campaign level.

That said, applying data analytics to customer management is not a trivial task. As will be depicted in section 3, it is a complicate iterative process that demands cross-disciplinary skill sets. Although there’s no barrier in equipment (usually a cluster of desktop PCs is good enough) or technology (most of the analytic tools are open source), most local companies still feel it difficult to get started with. To help the local companies, we see the need to develop a managerial framework that guilds the data-driven customer management process.

2.3 Objectives of this Project

On the one hand, the focus of marketing management has shifted from product-oriented marketing mix toward value oriented customer management; on the other hand, the ever advancing data analytics techniques are transforming the practice of marketing research. To help the local industry harnessing these advanced concept and techniques, in this project we plan to propose a managerial framework that applies data analytics technologies to customer management practices. To be pragmatic, we’d actually practice our process in local industries. These case studies would help to validate, refine and specialize our process for each specific industry. All participants in this project, teachers and students in our schools, engineers and managers in our assisting companies, would learn the related concepts and technologies, and have opportunities to practice them in real business cases. In summary, the major objectives of this projects are:

l   Applying data analytics technologies to the practice of customer (CLV) management

l   Propose a data-driven customer management framework for companies who wish to get started with (big) data analytics and digital marketing

l   Actually applying the process in local industries. Validate the proposed framework and get hand-on experience with real case studies.

l   Develop a pragmatic framework, whereby new theories of customer analytics and new data analytics technologies can be blended, and be brought to practice systematically

Talent has been repeatedly regarded as the largest obstacle in adopting data analytics and digital marketing; therefore the most important objective of this project might be: 

l   Train MBA/managers/marketers with basic data analytics skills, and equip IT/MIS specialists with marketing and business mind-sets

 

So that the local industry can leverage on the power of big data and IoT, and substantially be benefited from them.

 

 

 

研究方法、進行步驟及執行進度。請分年列述:1.本計畫採用之研究方法與原因。2.預計可能遭遇之困難及解決途徑。3.重要儀器之配合使用情形。4.如為須赴國外或大陸地區研究,請詳述其必要性以及預期效益等。

 

3. Research Methodology – Platform, Model & Process

Methodology in this project is elaborated in three levels, namely computation platform, predictive modeling and managerial framework. First, for scalability we’d adopt the open source cloud computation platform (3.1). Second, to leverage the largest and the latest libraries of analytic methods we’d use R as our major tool for predictive modeling (3.2). Third, on top of the aforementioned two elements, we develop a managerial framework, whereby companies can derive their customer management strategies by predictive modeling techniques (3.3). Limitation, difficulty and potential extensions of this project are then stated in sub-section 3.4.

3.1 Computational Platform & General Analytics Procedure

To cope with the vast volume, variety and veracity of big data, we plan to build our analytics tools on top of a cluster computing platform. As illustrated in Figure 1, we’d use this platform to perform four type of analytic tasks – E.T.L. (Extract, Transform & Load), Visualization, Prototyping and Full-Scale Modeling.

1.      ETL – Extract, Transform & Load

Big data analytics depends on parallel computation. Before analysis, data need to be extracted from its original sources, transformed into proper format and loaded into a distributed data store. In this project we use Apache Hadoop Distribute File System and its associated utilities, including Sqoop for ETL, Cassendra for SQL query, YARN for parallel programming and Presto, Hive and Pig for various types of scripting. After ETL, these tools also help to prepare (clean up, consolidate, normalize) the data for further examination and analysis.

2.      Visualization

Visualization helps to identify issues before, and present results after detail analysis. In this project we adopt the most popular big data visualization tool in the market – Tableau. If budget allowed, we’d use Tableau Server for concurrent accesses, otherwise Tableau Desktop. Either ways, the software package equips with comprehensive connectors that link to the underlying SQL servers and access the outputs generated from the other scripting/pre-processing tools.

3.      Prototyping

Initially we plan to start with a 20-data-note file system, which should be good enough for modeling most sub-terabyte datasets. Although it is possible to build full models in one stage, for flexibility and agility we choose to conduct two-stage modeling. Before building full scale models, we’d extract sub-samples (upto 4Gbytes each, bounded by the memory of the working machine) and build prototype models for concept proving. After feeding the sub-samples into the working PC, we can leverage R (RStudio) and its comprehensive library to make prototype models. In addition to proof of concept, some prototypes might also sever as segment-specialized models that make predictions for customers with similar characteristics. 

4.      Full-Scale Models

In order to model the full dataset, we make use of Apache Spark (Sankar & Karau 2015) – an in-memory parallel computing platform originated from UC Berkeley. It’s repeatedly proven that, in comparing to mere Hadoop, Spark’s RDD (resilient distributed dataset) technology can reduce the model building times in a factor of tens (Pentreath 2014; Frampton 2015). In addition to its core module, we’d also incorporate its built-in languages – Python and Scala, and its machine learning libraries (MLlibs). On top of Spark’s native interface, we’d also install interactive shells (Jupyter/iPython and/or H2O) to facilitate interactive analysis.

3.2 Predictive Modeling Process (R/Caret package)

Comparing to the traditional approaches, building predictive models from big data need to overcome two more problems. First, in the library there are hundreds of methods, and tens of new methods are developed every year. Different methods perform differently in different datasets; and usually we couldn’t tell which method was better before models were built. Second, to deal with the overfitting (Babyak 2004; Reunanen 2003) problem, every method is equipped with a few modeling parameters, which usually can only be fine-tuned by trial and error.

To overcome these problems, machine learning scholars has developed an iterative process (a.k.a. machine learning pipeline). First of all, based on the characteristics of the dataset, analysts identify a set of candidate methods from the method library. For example, in Table 1, Kuhn & Johnson (2014) provide a reference list of popular methods for general purposes. The iterative tuning process shown in Figure 2 is then applied to ‘each’ of the candidate methods. Briefly, the process consists of two cascaded loops. The outer loops for every combination of modeling parameters; and the inner iterates for every folds of cross-validation. At its end, this process produces an average (cross-validated) score per method, per parameter combination. The best model is then identified by comparing the scores.

As aforementioned, in order to identify the best one, hundreds (if not thousands) of models need to be created during the process. Even backed by parallel computing, the process is quite lengthy and error prone. In this project we’d adopt R’s caret package (Alfaro 2013; Kuhn 2008) to automize our data (pre)processing and model tuning process. The caret package can drive near 200 methods todays, and new methods has being incorporated frequently. It’s feature selection and model ensembling modules might also be helpful in this project.

3.3 Managerial Framework for Customer Management

At the center of this project is the data driven customer management framework. Actually the managerial framework itself is one of the major outputs of this project. Below we will elaborate on our ideas about the framework and its design principles. In Figure 3, we present a tentative version of this process. Notably, it is just our preliminary idea, and will continuously be concretized, refined, generalized and specialized during the project.

The Design Principles – Customer Centric, Value Oriented & Data Driven

The major objective of this project is to apply advanced data analytics techniques to the practice of customer management. To be compatible with the recent trends, the entire process is designed with three fundamental principles. First, customer centric: instead of product (marketing mix) and brand, we position customer at the center of marketing strategies. Second, value oriented: value is the key metric; and more than the company’s value to its customers, we focus on customers’ value to the company (i.e., CLV). Third, data driven: whenever possible, strategies should be derived from predictions made by predictive models.

Data & Source of Data

CRM databases should be the major data sources. Primary input includes (but not limited to): transaction logs, customer profiles and detail records of marketing efforts. In principle, we should be able to join all data by customers, including their transactions, site/store visits, and responses to previous promotions and advertisements. The CRM database may connect to web analytics data repositories (such as google analytics) and/or digital marketing platforms (such personalized web pages) in the latter stage of this project, as will be explained in 3.4.

CLV Evaluation

At the conjunction of the customer centric and value oriented principles, CLV is the central metric in this process. After loading the data, we’d always start by estimating CLV for each customer. CLV is the sum of discounted cash flows, which are the net of revenue minus cost. Usually the revenue part can be estimated by converting each customers transaction logs into recency, frequency and average amount of purchase. The difficult part is the cost. If cost could be evenly allocated to all customers (or by customer segment), estimating each individuals’ CLV is quite straight forward (to prevent duplication, we will explain the details of cost accounting in the case studies, see 4.1). Otherwise, complicate accounting methods, such as usage or activity based cost accounting, might have to be involved (see 4.2).

Segmentation & Transition Analysis

Customers can be segmented in many different ways (Konuş et al. 2008). Following the value oriented principle, we’d start with value based segmentation, with which customers are segmented by their estimated CLV. Notably, customers come and go, and their CLVs change with time. Low value customers might turn (be developed) into high value ones, and vice versa. One major tasks of customer management is to watch the in, out and transition flows amongst customer segments. The trends of these flows can be used to forecast the sizes of each segments, which in turn can project the population’s (and each segment’s) CLV for the long term (Kumar & George 2007; Kumar & Shah 2009). The transitions between segments can also be used in assessing performance, and setting goals for customer management campaigns. In fact, every customer management campaigns can be regarded as a kind of interference designed to alter the flows among segments. In this light, the overall objective of customer management can be defined by the size and composition of the customer base, and the goal for each campaign can be defined by its impact to the affected segments.  

Selective Acquisition

After knowing the CLVs of their existing customers, companies may wish to target their acquisition campaign to high value consumers. The problem is, how can we estimate the value of a ‘consumer’ before she ever becomes our ‘customer’? To predict the potential values of prospect consumers, we need to make use of observables variables of our existing customers. Using these observable variables as predictors, we can build a model to predict the value of consumers before they ever buy from us (Pfeifer et al. 2005).

Selective Retention

Attrition is a common nature of customer populations; therefore retention mechanisms have been developed to mitigate churn (Ascarza & Hardie 2013; Braun & Schweidel 2011; Kim & Moon 2012; Lemmens & Gupta 2013). Predictive models help to improve the effectiveness of customer retention in two major ways. First, we can build a model to predict the probability of churn for every customer. Alone with each individual’s CLV, these predictions help to identity and prioritize the customers to be retained. Second, for each retention campaign, we can build a model to predict its rate of successful retention for each customer. Thereby, the net effect of every retention campaign can be evaluated on the per customer basis. With these predictions, every retention tools can be optimized for its maximum expected return (Dong et al. 2011; Lhoest-Snoeck 2014).

Selective Development

Besides the in and out flows, the transitions among segments may also have significant impacts on the overall CLV. Broadly speaking, all efforts that aim to influence the transitions among segments can be regarded as customer development. Notably, deterring high to low transitions is nonetheless important than facilitating low to high flows. Again, predictive models help to forecast the probabilities of transition (a.k.a. conversion) down to the per customer, per instrument level; so that managers can generate a prioritized list of targeted customer, and evaluate the expected return of every development tool before it is deployed. 

Justification & Instrument Selection

As mentioned in the previous paragraphs (and illustrated in Figure 3), via the common process of modeling, prediction, selection, optimization and evaluation, managers can estimate the expected return of customer management tools before they were deployed. These ROI information should help customers in selecting and prioritizing among alternative campaigns. 

3.4 Limitations & Future Extensions

As a pioneer effort, we first put our focus on the structure and applicability of the framework. The aforementioned processes should have covered most of the major aspect of customer management. But we know it is still far from comprehensive. One of the limitation is on the dimension of segmentation. To be concise we choose to start with value based segmentation (which is compatible with the value-oriented principle.) In reality, managers may wish to adopt the other segmenting factors (ex., demographic, preferential or behavioral segmentation); sometimes it may also be beneficial to use multiple segmentation dimensions in parallel. Incorporating multi-dimension segments from the beginning could be challenging, especially when we do not know which dimensions should be included in advance. Recognizing its potential and importance, we thus plan to conduct multi-dimension segments as an extension in case study I (4.2).  

Making use of web analytics data, such as web server logs, click streams and Google analytics data streams, may also be a potential extension. To be concise, in the main project we primarily focus on transaction logs and CRM data. If we could allocate enough of resource in the latter stage, we might make use of web analytics data, as stated in case study II (4.3).

Contrasting to the traditional, broadcasting media, most of today’s digital marketing channels are interactive and bi-directional (Chaffey et al. 2012). Therefore, in addition to taking web analytics metrics as input, the predictions from the framework may also feedback to digital marketing platforms. After all, all of the marketing efforts are closely related to one another. Incorporating customer management and digital marketing should be very fruitful extension in the future (see 4.3).

 

4. Sub-Projects & Research Plan

        As a pragmatic project, we plan to apply the data driven customer management framework to real business cases. So far, we’d connected to a few local companies, including a CATV and ISP service provider (4.1), a mobile phone service operator (4.2) and an ecommerce platform operator (4.3). Progresses and plan for each of these cases are stated in the following sub-sections. While conducting these cases, we’d adjust and refine our framework and processes. The goal is to produce a data driven customer management framework that generally apply to most of the B2C companies, and a set of specialized guidelines for certain specific industries (4.4). The overall project plan is presented in Table 2, and explained in 4.5. 

4.1 Pilot Case: local CATV Operator

In order to validate and concretize our initial ideas, we’d cooperated with a local CATV/ISP operator. This operator has 100K+ CATV subscribers; about 40% of them also subscribe to broadband internet service. Our cooperation started with a churn management project targeting the broadband subscribers. In the past 5 years (Y2010 ~2014), the number of broadband subscribers have grown from 28K to 48K, and the churn rates were 13.1%, 15.1%, 13.3 and 16.1%, Y-to-Y. 

Segmentation & CLV Estimation

Broadband service basically charges by monthly fee, which is priced by data throughput. Based on the monthly fee we divide the population into 6 segments. Because of the fixed monthly fee, CLV for each segment is simply estimated by inserting the segment averages into equation (1).

Predicting Churn

The operator has a quite limited CRM database. There’re 85 variables per subscriber, but only 51 of them are applicable for churn prediction (including referral, acquiring media and channel, promotion events and packages, etc.) To reflect the real business situation, we right censor the data at each year end, and build a classification model to predict each individual’s probability of ‘churn in the upcoming year’. For example, first we erase all data latter than 2013/06/30; then for every existing customers at that time, we build a model to predict the probability of churn between  2013/07/01 and 2014/06/30. The model’s AUC on split sample is 95.9%; AUC against the actual (Y2013~Y2014) data is 92.4%. That is to say, randomly given a churn and a non-churn customers, our model can correctly distinguish them, 92.4% of the times. 

Selective Retention: Selection, Optimization & Evaluation

Given the estimations of CLV and probability of churn, we then simulate a case of selective retention with a hypothetic retention tool. Assuming the cost of retention is , and successful retention rate , for each segment , we can determine a cutoff probability , above which the expected return of retention is positive. With this criteria we selected 1,116 customers out of the 44,356 existing customers (at YE2013). When audited against the real data, we found that: if we had retained the selected customers in YE2103, the total expected return would have been ; but if we had randomly retained the same amount of customers, the total expected return would have drop to  Thereby the effectiveness of our selective retention process is proven.

Selective Development & Selective Acquisition

Based on the aforementioned results, we are expanding our cooperation toward customer development. The operator wish to convert its CATV subscribers to subscribe their broadband service. To this end, we are deploying our selective development process with the operator. So far, selective acquisition project has not been identified yet.

4.2 Case Study-I: Mobile Network Operator

        For mobile network operators, customer management is one of the most important and most challenging task (Lu & Park 2003). Taking the varieties in mobile phone bundles, different tariff plans, promotion events/packages and discriminated prices into concern, estimating the CLVs of mobile subscribers is a big challenge; and this could be a good opportunity for us to validate our managerial framework in real business environment. If we could have a successful track record in this challenging business circumstance, we’d have a good confidence to apply the framework elsewhere. In this light, we’re contacting mobile service providers and trying to initiate a case study.

CLV Estimation, Average Cost Allocation

For mobile service operators, cost accounting is a big challenge by itself. For simplicity, we’d start with average cost allocation, which means we evenly allocate all cost items to all subscribers. On the other hand, the calculation of revenue is complicated by “mobile phone bundle” – a popular acquisition and retention instrument in the market of mobile service. Operators subsidize the cost of handset to attract subscribers, in exchange of subscription contracts for 2 or 3 years. A typical cash flow in an n-month subsidy cycle looks like:

where  is the revenue in the n-th month, and  and  represent the initial and the subsequent subsidies. Taking monthly discount (d), risk of foreclosure (f), rick of churn (c) and average cost (k) into consideration, the discount cash flow could be expressed as: (based on our initial discussion with the operator, to be adjusted latter)

Based on our initially discussion with the operator, we plan to evaluate the CLV by a full subsidize cycle , plus the subsequent subsidy . For benchmarking, we’d project the subsequent revenue and subsidy streams from those of the first cycle. Letter on, we’ll pull in CRM data, and build models to predict revenue streams for individual customers. Present value for the first cycle is simply the sum of discounted cash flows. Present values for the subsequent cycles are adjusted by the risks of churn and foreclosure, and the potential of growth. Initially, the risk and growth factors might simply be segment averages. In the latter stage, they may also be predicted by models at the individual level. 

CLV Estimation, Usage Cost Allocation

If we adopt usage cost, we’d have to distinguish fix and variable cost, and the latter need to be accounted for in-net-voice, off-net-voice and data usage separately. Other than that, the estimation of CLV should be similar to the case of average cost allocation. Initially CLV can be evaluated by (to be adjusted):

In latter stage, we would link usage cost to revenue prediction. To do so, we also need to build models to predict in-net, off-net and data revenue streams separately.

Selective Customer Management

The selective acquisition, development and retention practices described in 3.3 should apply to this case without major modifications.

Extension: Multi-Dimensional Segmentation

In addition to value (CLV) based segmentation, we’ve also considered behavior (lifestyle), demographics and/or preference based segments. For example, in the market of mobile service, it might be useful to categorize customers into:

l   Listeners who receive but rarely make calls,

l   Phone Chasers who sign up for new phones,

l   Data Gluttons who use a lot of data bandwidth,

l   Nomads who churn at the end of every contract, etc.

Different segments may have different characteristics in many aspects. For example, gluttons and nomads may contribute different amounts of revenue and respond to customer management campaigns differently. Acquisition or retention programs that perform well in one segment might not work in the other. Therefore marketers usually need to develop different campaigns to address different segments. We understand that combining CLV with the other segmenting factors may significantly improve the flexibility of the framework. However, every industry has its unique structure, and except value, it’s very difficult to find another unified segmentation scheme (Forsyth 1999). Therefore we choose to focus on value based segmentation in the basic framework, and position multi-dimensional segmentation as an industrial specific extension in this case study.

4.3 Case Study-II: E-Commerce Web Site Operator

In this digital era, digital media have become the major channel to engage customers, and web analytics matrices (site visiting records, page views, click streams, etc.) are the major source of customer data. When a consumer actively search for product information on the web, she makes contacts with a company from many different media for many times, before she’d finally make a purchase (Elzinga et al. 2009). For example, a consumer may have:

l   heard a brand from one of her friends in Facebook,

l   saw a keyword ads in the second day,

l   been impressed by a banner ads two week later,

l   clicked and been directed to the company home page where she signed up for free membership and download a free mobile app.

Then, a month later:

l   when on a bus, she saw a promotion message in her mobile, connected to the promotion web page, and made a bookmark,

l   when she came home, she followed the book mark to the product page, and finally made a purchase.

It’d be the 45th days, since she first heard the brand in Facebook. In such a Consumer Decision Journey (CDJ, see Edelman 2010), the company engage with the consumer in many different media. These media have different effects and incur different costs. Some media are good in creating sales leads; others are good in promoting the brand or closing deals. In this era of digital marketing, the performance of digital companies by and large depend on whether they could leverage the strength of these digital media effectively and efficiently (Chaffey 2012).

        Data analytics and predictive modeling have been widely adopted among modern companies. For examples, Google and Facebook to predict click through rates, Amazon and Netflix for product and movie recommendations, and so on. These use cases had been published in engineering books, such as Raza (2014), Pentreath (2014) and Nicolas (2014), but not seen in the literature for customer management framework yet. Therefore we believe that extending our managerial framework to cover digital marketing could be a fruitful extension. Basically this objective can be achieved by adding two elements into the basic framework: (1) incorporating web analytic metrics, and (2) use predictive model in digital marketing.

Incorporating Web Analytics Metrics

Companies nowadays can (and most of modern digital companies do) capture and record every consumer engagements in every digital media (Peterson 2004; Kaushik 2009). Most of these web analytics matrices are customer-centric, and incorporating them into predictive models can improve the accuracy of predictions. However, owning to their vast volume and variety, we might not be able to handle them in traditional way. Tentatively, we plan to develop some connector modules, so that we can access various data repositories form web analytics platforms (such as google analytics, or ecommerce web servers).

Use Predictive Models in Digital Marketing

To their ends, acquiring one customer from an agent/dealer is not much different from acquiring one from Facebook Fan Page or Google AdWords. In this aspect, any digital marketing media might be treated as just another type of acquisition, development or retention tools. One key difference is: digital media are programmable (Chaffey 2012). During the case study. we plan find some ways to feed the predictions, or incorporate the predictive models themselves into various digital marketing platforms. It might take some efforts in the beginning; but once it’s done, ideally the platform would automatically generate personalized content for every individual consumer, which might significantly improve the effectiveness of the media.

Cooperating with Ecommerce Company

E-commerce platform operators should have the richest web analytics dataset. Therefore, we currently are negotiating with a medium size e-commerce platform operator for a MOU. Although project details is yet to be determined, the work items might include, tentatively:

l   evaluate every individual customer’s lifetime value,

l   assess the distribution of CLV,

l   perform value based segmentation,

l   observer the average probabilities of transitions among segments,

l   estimate the average CLV for every segments,

l   predict each individual customer’s probabilities of transition or churn, and

l   predict each marketing tool’s expected conversion rate for each individual customer (i.e., the probability of successfully transferring the targeted customer to the designated segment.).

With this estimation/predictions, we’d be able to:

l   generate an prioritized target list for every media,

l   estimate the expected (optimal) return for every media, and

l   justify and make selection from the alternative list of digital marketing media.

During the process, web analytics dataset would be utilized in all of the predictive models, and the experience would be documented as a case study.

4.4 Generalization & Specialization

        Generally, the common framework (see 3.3 & Figure 3) guilds the segmentation, modeling and optimization processes for customer management. First of all, we determine the segmentation by investigating the distribution of CLV. We then

        for every segments, estimate:

n   the Average Value of customers, and

n   the Average Probabilities of Transitions among Segments;

for every customers, predict:

n   the Probability of Purchasing in the upcoming period,

n   the Expected Revenue Contribution in the upcoming period,

n   her Lifetime Value, and

n   the Probabilities of Transferring (to the other segments);

for every marketing tools, predict (or generate):

n   the Expected Conversion Rate for every customers,

n   a Prioritized List of Target Customers, and

n   the Expected (Optimized) ROI.

On top of this common process, every case study also has its unique goal. In the case of mobile network operator, we aim to validate the applicability of our managerial framework in complicate business environment, and practice multi-dimensional segmentation; whereas in the case of e-commerce company, we’d blend CRM with Web Analytics data and make use of predictive models in digital marketing campaigns. In together, we aim to propose a general framework process that applies to most of B2C industries, and a set of specialized guidelines for certain industries –  such as telecommunications and ecommerce, and business circumstances – such as multi-dimensional segmentation and digital marketing.

 

 

預期完成之工作項目及成果。請分年列述:1.預期完成之工作項目。2.對於參與之工作人員,預期可獲之訓練。3.預期完成之研究成果(如期刊論文、研討會論文、專書、技術報告、專利或技術移轉等質與量之預期成果)。4.學術研究、國家發展及其他應用方面預期之貢獻。

 

5. Project Plan and Expected Impacts & Outcomes

5.1 Project & Publication Plans

We plan to deploy this project in 3 years; the overall plan is summarize in Table 2. We start by setting up the computing platform (3.1) and analysis tools (3.2) for the project and drafting the framework process for customer management (3.3). We then validate and adjust the framework and processes in the pilot case (4.1). Considering the loading constrains, Case Study I (4.2 – Mobile Network Operator) and Case Study II (4.3 – Ecommerce Web Site) would be started in the second and fifth quarters separately. In each case, deploying and validating the basic framework should be completed within a year, whilst the extensions might take longer. During the case studies, the framework process would be refined and finalized. Together with the case studies, we plan to publish the results of this project in three articles, tentatively named:

l   A Managerial Framework for Data Driven Customer Value Management,

l   Practicing Multi-Dimensional Segmentation in Data Driven Customer Value Management  (Case Study: a Mobile Network Operator), and

l   Applying Data Driven Customer Value Management in Digital Marketing (Case Study: an Ecommerce Web Site).

5.2 Analytics Training & Technology Transferring

Applied analytics is cross disciplinary. Therefore, the major objective of this project is to erase the boundaries amongst the silos, bringing in and blending talents from academia and industries, across the disciplines of managerial, statistics and computer sciences. Participant in this project includes:

l   Teachers, researchers, graduates & under-grad students from the departments of Business Administration, MBA and Information Management, and

l   Executives, Managers and Engineers from the departments of Marketing and MIS (in Telecommunications and Ecommerce sectors.)

would have an opportunity to cooperate and learn from each other and

l   experience the leading-edge technologies of cloud computing and data analytics,

l   practice customer-centric, value-oriented and data-driven strategic planning

l   get hand-on experience in applying data analytics to customer value management

5.3 Impacts & Significance

Many local companies have been enthusiastic with the promises of big data, yet paralyzed by its intimidating complexity. Confronting such a complicate, cross-disciplinary subject, hand on case studies guided by a systematic framework should be a reasonable approach. For the industries, the results of this project would be a practical reference for

l   Jumping start big data initiatives, and

l   Practicing data driven and value oriented customer management within their companies

While the leading digital companies (Google, Amazon, Apple, Netflix, etc.) are embracing data analytics and enjoying outstanding outcomes, it should be desirable to incorporate data analytics techniques into management science and marketing research. In the academic perspective, this project is also expected to:

l   Incorporate data analytics in management and marketing researches

l   Propose a managerial framework for applying data analytics techniques to customer value management

l   Serve as a conceptual yet practical framework, whereby further marketing practices and analytics methods can be integrated systematically

 

 

 

 

 


 

References:

 

1.           Aaker, D. A. (2009). Managing brand equity. Simon and Schuster.

2.           Abe, M. (2009). “Counting your customers” one by one: A hierarchical Bayes extension to the Pareto/NBD model. Marketing Science, 28(3), 541-553.

3.           Aflaki, S., & Popescu, I. (2013). Managing retention in service relationships. Management Science, 60(2), 415-433.

4.           Alfaro, E., Gámez, M., & Garcia, N. (2013). Adabag: An R package for classification with boosting and bagging. Journal of Statistical Software, 54(2), 1-35.

5.           Andrieu, C., De Freitas, N., Doucet, A., & Jordan, M. I. (2003). An introduction to MCMC for machine learning. Machine Learning, 50(1-2), 5-43.

6.           Ascarza, E., & Hardie, B. G. (2013). A joint model of usage and churn in contractual settings. Marketing Science, 32(4), 570-590.

7.           Babyak, M. A. (2004). What you see may not be what you get: a brief, nontechnical introduction to overfitting in regression-type models. Psychosomatic medicine, 66(3), 411-421.

8.           Berger, J. O. (2013). Statistical decision theory and Bayesian analysis. Springer Science & Business Media.

9.           Berger, P.D., & Nasr, N.I. (1998). Customer Lifetime Value: Marketing Models and Applications. Journal of Interactive Marketing, 12 (Winter), 17–30.

10.        Blattberg, R. C., Malthouse, E. C., & Neslin, S. A. (2009). Customer lifetime value: Empirical generalizations and some conceptual questions. Journal of Interactive Marketing, 23(2), 157-168.

11.        Blattberg, R.C., & Deighton, J. (1996). Manage Marketing by the Customer Equity Test. Harvard Business Review, 1996(Jul), 136–144.

12.        Bolstad, W. M. (2013). Introduction to Bayesian statistics. John Wiley & Sons.

13.        Bowman, D., & Narayandas, D. (2004). Linking customer management effort to customer profitability in business markets. Journal of Marketing Research, 41(4), 433-447.

14.        Braun, M., & Schweidel, D. A. (2011). Modeling customer lifetimes with multiple causes of churn. Marketing Science, 30(5), 881-902.

15.        Chaffey, D., Smith, P. R., & Smith, P. R. (2012). eMarketing eXcellence: Planning and optimizing your digital marketing. Routledge.

16.        Chen, T., & He, T. (2015). Higgs boson discovery with boosted trees. In Cowan et al., editor, JMLR: Workshop and Conference Proceedings (No. 42).

17.        Dong, Y., Yao, Y., & Cui, T. H. (2011). When acquisition spoils retention: Direct selling vs. Delegation under CRM. Management Science, 57(7), 1288-1299.

18.        Edelman, D. (2010). Branding in the Digital Age: You’re Spending Your Money in All the Wrong Places. Harvard Business Review. 2010(Dec).

19.        Elzinga, D., Mulder, S., & Vetvik, O. J. (2009). The consumer decision journey. McKinsey Quarterly, 2009(3), 96-107.

20.        EsmaeiliGookeh, M., & Tarokh, M. J. (2013). Customer Lifetime Value Models: A literature Survey. International Journal of Industrial Engineering, 24(4), 317-336.

21.        Fader, P. S., & Hardie, B. G. (2009). Probability models for customer-base analysis. Journal of Interactive Marketing, 23(1), 61-69.

22.        Fader, P. S., Hardie, B. G., & Lee, K. L. (2005). “Counting your customers” the easy way: An alternative to the Pareto/NBD model. Marketing Science, 24(2), 275-284.

23.        Forsyth, J., Gupta, S., Haldar, S., Kaul, A., & Kettle, K. (1999). A segmentation you can act on. The McKinsey Quarterly, 7.

24.        Frampton, M. (2015). Mastering Apache Spark. Packt Publishing

25.        Graf, A., & Maas, P. (2014). Customer value from a customer perspective–a comprehensive review. Springer Fachmedien Wiesbaden.

26.        Gupta, R., Gupta, H., & Mohania, M. (2012). Cloud computing and big data analytics: what is new from databases perspective?. In Big Data Analytics (pp. 42-61). Springer Berlin Heidelberg.

27.        Gupta, S. (2014). Core Curriculum Marketing. Harvard Business Publishing

28.        Gupta, S., Hanssens, D., Hardie, B., Kahn, W., Kumar, V., Lin, N., ... & Sriram, S. (2006). Modeling customer lifetime value. Journal of Service Research, 9(2), 139-155.

29.        Gupta, S., Lehmann, D. R., & Stuart, J. A. (2004). Valuing customers. Journal of Marketing Research, 41(1), 7-18.

30.        Hayes, A. F. (2013). Introduction to mediation, moderation, and conditional process analysis: A regression-based approach. Guilford Press.

31.        Holm, M., Kumar, V., & Rohde, C. (2012). Measuring customer profitability in complex environments: an interdisciplinary contingency framework. Journal of the Academy of Marketing Science, 40(3), 387-401.

32.        Hui, G. (2014). How the Internet of Things Changes Business Models. Harvard Business Review, 2014(Jul)

33.        Jain, D., & Singh, S. S. (2002). Customer lifetime value research in marketing: A review and future directions. Journal of Interactive Marketing, 16(2), 34-46.

34.        James, G., Witten, D., & Hastie, T. (2014). An Introduction to Statistical Learning: With Applications in R. Springer.

35.        Jerath, K., Fader, P. S., & Hardie, B. G. (2011). New perspectives on customer “death” using a generalization of the pareto/nbd model. Marketing Science, 30(5), 866-880.

36.        Johnson, M. D., & Selnes, F. (2004). Customer portfolio management: toward a dynamic theory of exchange relationships. Journal of Marketing, 68(2), 1-17.

37.        Kaplan, R. S. (1989). Kanthal (A). Case no.: 9-190-02, Harvard Business Cases

38.        Kaplan, R. S., & Atkinson, A. A. (2015). Advanced management accounting. PHI Learning.

39.        Kaplan, R. S., & Narayanan, V. G. (2001). Customer profitability measurement and management. Harvard Business School.

40.        Karvanen, J., Rantanen, A., & Luoma, L. (2014). Survey data and Bayesian analysis: a cost-efficient way to estimate customer equity. Quantitative Marketing and Economics, 12(3), 305-329.

41.        Kaushik, A. (2009). Web Analytics 2.0: The Art of Online Accountability and Science of Customer Centricity. Sybex, 1 edition (October 26, 2009)

42.        Keller, K. L. (1993). Conceptualizing, measuring, and managing customer-based brand equity. Journal of Marketing, 57(1), 1-22.

43.        Kim, Y. S., & Moon, S. (2012). Measuring the success of retention management models built on churn probability, retention probability, and expected yearly revenues. Expert Systems with Applications, 39(14), 11718-11727.

44.        Konuş, U., Verhoef, P. C., & Neslin, S. A. (2008). Multichannel shopper segments and their covariates. Journal of Retailing, 84(4), 398-413.

45.        Koosha, H., & Albadvi, A. (2015). Customer lifetime valuation using real options analysis. Journal of Marketing Analytics, 3(3), 122-134.

46.        Kotler, P. (1974) Marketing During Periods of Shortage. Journal of Marketing 38(3), 20–29.

47.        Kotler, P., Armstrong, G. (1996). Principles of Marketing, 7Th ed., Englewood Cliffs, NJ: Prentice-Hall

48.        Kruschke, J. (2010). Doing Bayesian data analysis: A tutorial introduction with R. Academic Press.

49.        Kuhn, M. (2008). Building predictive models in R using the caret package. Journal of Statistical Software, 28(5), 1-26.

50.        Kuhn, M., & Johnson, K. (2013). Applied predictive modeling. New York: Springer.

51.        Kumar, V. (2008). Managing customers for profit: Strategies to increase profits and build loyalty. Prentice Hall Professional.

52.        Kumar, V. (2011). Looking Through the Marketing Lens: My Journey so Far…. Review of Marketing Research, 8, 121-157.

53.        Kumar, V., & George, M. (2007). Measuring and maximizing customer equity: a critical analysis. Journal of the Academy of Marketing Science, 35(2), 157-171.

54.        Kumar, V., & Rajan, B. (2012). Customer lifetime value management: strategies to measure and maximize customer profitability. Handbook of Marketing Strategy, Edward Elgar Publishing, Northhampton, 107-134.

55.        Kumar, V., & Shah, D. (2009). Expanding the role of marketing: from customer equity to market capitalization. Journal of Marketing, 73(6), 119-136.

56.        Kumar, V., & Umashankar, N. (2012). 1 Enhancing financial performance: the power of customer metrics. Handbook of Marketing and Finance, 9.

57.        Kumar, V., Aksoy, L., Donkers, B., Venkatesan, R., Wiesel, T., & Tillmanns, S. (2010). Undervalued or overvalued customers: capturing total customer engagement value. Journal of Service Research, 13(3), 297-310.

58.        Kumar, V., Lemon, K. N., & Parasuraman, A. (2006). Managing Customers for Value An Overview and Research Agenda. Journal of Service Research, 9(2), 87-94.

59.        Kumar, V., Venkatesan, R., Bohling, T., & Beckmann, D. (2008). The power of CLV: Managing customer lifetime value at IBM. Marketing Science, 27(4), 585-599.

60.        Lemmens, A., & Gupta, S. (2013). Managing churn to maximize profits. Harvard Business School.

61.        Lhoest-Snoeck, S., van Nierop, E., & Verhoef, P. C. (2014). For New Customers Only: A Study on the Effect of Acquisition Campaigns on a Service Company's Existing Customers' CLV. Journal of Interactive Marketing, 28(3), 210-224.

62.        Lu, J. (2010). The elements of statistical learning: data mining, inference, and prediction. Journal of the Royal Statistical Society: Series A (Statistics in Society), 173(3), 693-694.

63.        Lu, J., & Park, O. (2003). Modeling customer lifetime value using survival analysis—an application in the telecommunications industry. Data Mining Techniques, 120-128.

64.        Ma, S. H., & Liu, J. L. (2007, August). The MCMC approach for solving the Pareto/NBD model and possible extensions. In Natural Computation, 2007. ICNC 2007. Third International Conference on (Vol. 2, pp. 505-512). IEEE.

65.        Mathur, M., & Kumar, S. (2013). Customer Retention Through Prioritization: Integrating Time-Dependent Context of Relationship Dynamics. Journal of International Consumer Marketing, 25(5), 332-343.

66.        McAfee, A., Brynjolfsson, E., Davenport, T. H., Patil, D. J., & Barton, D. (2012). Big data. The management revolution. Harvard Business Review, 2012(Oct), 61-67.

67.        Mzoughia, M. B., & Limam, M. (2014). An improved customer lifetime value model based on Markov chain. Applied Stochastic Models in Business and Industry. 2014.

68.        Nicolas, P. R. (2014). Scala for machine learning. Packt Publishing

69.        Nikumanesh, E., & Albadvi, A. (2014). Customer's life–time value using the RFM model in the banking industry: a case study. International Journal of Electronic Customer Relationship Management, 8(1-3), 15-30.

70.        Payne, A., & Frow, P. (2005). A strategic framework for customer relationship management. Journal of marketing, 69(4), 167-176.

71.        Pentland, A. (2014). Social Physics: How Good Ideas Spread-The Lessons from a New Science. Penguin.

72.        Pentreath, N. (2014). Machine learning with Spark - Tackle big data with powerful Spark machine. Packt Publishing

73.        Petersen, J. A., & Kumar, V. (2015). Perceived risk, product returns, and optimal resource allocation: evidence from a field experiment. Journal of Marketing Research, 52(2), 268-285.

74.        Peterson, E. T. (2004). Web analytics demystified: a marketer's guide to understanding how your web site affects your business. Ingram.

75.        Pfeifer, P. E., & Carraway, R. L. (2000). Modeling customer relationships as Markov chains. Journal of interactive marketing, 14(2), 43-55.

76.        Pfeifer, P. E., Haskins, M. E., & Conroy, R. M. (2005). Customer lifetime value, customer profitability, and the treatment of acquisition spending. Journal of Managerial Issues, 11-25.

77.        Pinheiro, J., & Bates, D. (2006). Mixed-effects models in S and S-PLUS. Springer Science & Business Media.

78.        Polo, Y., Sese, F. J., & Verhoef, P. C. (2011). The effect of pricing and advertising on customer retention in a liberalizing market. Journal of Interactive Marketing, 25(4), 201-214.

79.        Reinartz, W. J., & Kumar, V. (2000). On the profitability of long-life customers in a noncontractual setting: An empirical investigation and implications for marketing. Journal of Marketing, 64(4), 17-35.

80.        Reunanen, J. (2003). Overfitting in making comparisons between variable selection methods. The Journal of Machine Learning Research, 3, 1371-1382.

81.        Romero, J., Van der Lans, R., & Wierenga, B. (2013). A partially hidden Markov model of customer dynamics for clv measurement. Journal of Interactive Marketing, 27(3), 185-208.

82.        Rust, R. T., Lemon, K. N., & Zeithaml, V. A. (2004). Return on marketing: Using customer equity to focus marketing strategy. Journal of Marketing, 68(1), 109-127.

83.        Ryza, S. (2014). Advanced analytics with Spark. Ed. by Ann Spencer. O’Reilly.

84.        Sankar, K., & Karau, H. (2015). Fast data processing with Spark. Packt Publishing

85.        Schmittlein, D. C., Morrison, D. G., & Colombo, R. (1987). Counting Your Customers: Who-Are They and What Will They Do Next?. Management Science, 33(1), 1-24.

86.        Schweidel, D. A., Bradlow, E. T., & Fader, P. S. (2011). Portfolio dynamics for customers of a multiservice provider. Management Science, 57(3), 471-486.

87.        Shah, D., Rust, R. T., Parasuraman, A., Staelin, R., & Day, G. S. (2006). The path to customer centricity. Journal of service research, 9(2), 113-124.

88.        Shin, J., Sudhir, K., & Yoon, D. H. (2012). When to “fire” customers: Customer cost-based pricing. Management Science, 58(5), 932-947.

89.        Tarasi, C. O., Bolton, R. N., Hutt, M. D., & Walker, B. A. (2011). Balancing risk and return in a customer portfolio. Journal of Marketing, 75(3), 1-17.

90.        Trevor J.. Hastie, Tibshirani, R. J., & Friedman, J. H. (2009). The elements of statistical learning: data mining, inference, and prediction. Springer.

91.        Venkatesan, R., & Kumar, V. (2004). A customer lifetime value framework for customer selection and resource allocation strategy. Journal of Marketing, 68(4), 106-125.

92.        Xiao, J., Zhu, B., Teng, G., He, C., & Liu, D. (2014). One-Step Dynamic Classifier Ensemble Model for Customer Value Segmentation with Missing Values. Mathematical Problems in Engineering, 2014.

93.        Yen, B. & Yao, Y. (2015). The Internet of Things (IoT): Shaping the Future of e-Commerce. Harvard Business Review, 2015(Sep)

94.        Zhang, Y., Bradlow, E. T., & Small, D. S. (2014). Predicting customer value using clumpiness: From RFM to RFMC. Marketing Science, 34(2), 195-208.