Journal of Intelligent Learning Systems and Applications, 2011, 3, 90-102 doi:10.4236/jilsa.2011.32011 Published Online May 2011 (http://www.SciRP.org/journal/jilsa) Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a Retail Bank by a Recommender System Approach Michele Gorgoglione, Umberto Panniello Department of Mechanical and Management Engineering, Polytechnic of Bari, Bari, Italy. Email: [email protected], [email protected] Received July 10th, 2010; revised October 1st, 2010; accepted December 1st, 2010. ABSTRACT Customer churn may be a critical issue for banks. The extant literature on statistical and machine learning for customer churn focuses on the problem of correctly predicting that a customer is about to switch bank, while very rarely consid- ers the problem of generating personalized actions to improve the customer retention rate. However, these decisions are at least as critical as the correct identification of customers at risk. The decision of what actions to deliver to what customers is normally left to managers who can only rely upon their knowledge. By looking at the scientific literature on CRM and personalization, this research proposes a number of models which can be used to generate marketing ac- tions, and shows how to integrate them into a model embracing both the analytical prediction of customer churn and the generation of retention actions. The benefits and risks associated with each approach are discussed. The paper also describes a case of application of a predictive model of customer churn in a retail bank where the analysts have also generated a set of personalized actions to retain customers by using one of the approaches presented in the paper, namely by adapting a recommender system approach to the retention problem. Keywords: Customer Churn, Customer Retention, Personalization, Predictive Models, Recommender Systems 1. Introduction knowledge. However, these decisions are at least as cri- tical as the correct identification of customers at risk. A Retail banks often deal with customer churn. Among the customer churn prediction model should be integrated several issues addressed by Customer Relationship Ma- with a model to decide what personalized marketing ac- nagement (CRM), identifying the customers who are tion to deliver to customers; otherwise the benefits of about to quit the relationship with a company is one of using accurate predictive models would be lost. the most important in the financial services industry. The issue addressed by this research is to identify a set When competition becomes tougher, when laws decrease of approaches to generating actions to retain customers, either the barriers to entry or the customer’s switching once customers at risk are identified and profiled. We costs, or when a company aims at strengthening its posi- believe that this issue is important for both businesses tion in a new market, the issue of retaining customers and launching customer retention campaigns and researchers avoiding customer churn becomes even more crucial. aiming at building models of customer churn. From a In the last decade several computer science scholars business perspective, the problem is to support managers have tackled the problem of building accurate models to identify customers at risk in a bank by using statistical, in making the right decision on how to define personal- machine learning and data mining approaches [1]. How- ized retention actions. In fact, this decision has to con- ever, much literature only focuses on the problem of sider both organizational aspects, such as control and correctly predicting that a customer is about to switch. efficiency, and technological aspects, such as the accu- Very rarely the problem of generating personalized ac- racy of the algorithms. The challenge for managers is to tions to improve customer retention rates is considered. integrate the results of a predictive model of customer The decision of what actions to deliver to what customers churn with the decision of delivering certain marketing is normally left to managers who can only rely upon their actions. The problem is how to associate the right action Copyright © 2011 SciRes. JILSA Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 91 Retail Bank by a Recommender System Approach with the right customer. From a scientific perspective, a to review the approaches to deliver actions to individual comprehensive model of how to generate actions might customers. A review of studies on personalization can be be helpful to develop more effective data mining and found in [17]. The authors consider three main areas, statistical models to support CRM processes. namely computer science and information systems (CS/ By looking at both the scientific literature on CRM IS), marketing, management science and economics. It is and personalization, this paper proposes a number of possible to extract from that review the works directly approaches which can be used to generate personalized related to the problem of generating personalized actions. marketing actions, and discusses benefits and risks asso- In the CS/IS area the main streams dealing with ac- ciated with each model. Finally, a case of application is tions are Recommender Systems and Web Contents Per- described. A retail bank developed a predictive model of sonalization. In both cases the problem of generating customer churn. The model was integrated to a method to personalized actions is solved by processing individual generate personalized actions aimed at customer reten- customer profiles which describe personal preferences. In tion. The method is based to the approach followed by a recommender system the profile describes a customer’s Recommender Systems. The results of this model are preference by including either a set of ratings to certain compared to those obtained by the same company in a products or information on customers’ previous pur- formerly launched retention campaign. chases [18]. The action is the recommendation of a prod- uct and it is generated by analyzing the similarities be- 2. Prior Research tween the profiles of customers who bought or rated a The problem of customer churn has been analyzed by product and those of customers who did not. These sys- scholars in the fields of marketing and analytical CRM, tems are the most common approaches to personalization by analyzing several aspects. For instance, [2] analyzes in real industrial settings ever since Amazon adopted a the possible payoffs of retention and acquisition strate- Recommender System [19]. In the case of Web Content gies depending on the market structure. An analysis of Personalization the user’s profile includes statistics on the switching costs is provided by [3]. Customer churn the usage of the Web and the action is generated by asso- has also been studied by analyzing CRM-related prob- ciating a content with similarities with the user’s profile lems such as the effectiveness of loyalty programs [4] [20]. and customer satisfaction [5]. Scholars in marketing have dealt with the problem of Analytical models to predict customer churn have been generating personalized actions by studying competitive developed in several areas and industries as well. Mar- personalized promotions [21], customization [22], the keting models have been deployed to predict customers personalization of services [23], and recommender sys- at risk [6,7] as well as data mining models [8], especially tems. In all cases, except that of recommender systems in telecommunication-related industries [9-11]. Specific which was reviewed above, the problem is solved by prediction models have developed also in banking and using optimization models that allow the analyst to find finance [12,13]. the best price and/or product based on individual cus- This body of literature have mainly focused on the tomers response models. The response models are built problem of correctly predicting customer churn. Very by collecting data on customers’ previous behavior. rarely the problem of how to generate personalized ac- The research in management science [24] and eco- tions, once customer churn has been modeled, is consid- nomics has mainly faced the problem of generating per- ered. For instance, [14] states that a good customer churn sonalized actions by studying personalized pricing and model should also provide managers with a way to gen- mass customization [25]. The first stream falls in the erate personalized actions, although the research focuses marketing approaches reviewed above. The second str- only on predicting customer churn. The need to integrate eam mainly analyzes the benefits and costs of producing the analysis of customer behavior with the generation of as many products as the number of different customers’ actions in a comprehensive model is maintained by [15], individual preferences a company can identify in the while [16] suggests to include customers found at risk of market [26]. In these models the action is offering of a leaving the company in the organization of events in or- vast variety of products. Actions are generated by pro- der to strengthen their relationships with the company. cessing customers’ preferences in an aggregated way, not As the literature on customer churn does not directly individually, while the choice of the specific product is face the problem of how to generate actions, and in order left to customers. to achieve a more general understanding of the problem, A few research streams can be added to the review we have broadened the focus of the literature analysis to taken from [17] which do not directly deal with person- the research on personalization-related areas. The goal is alization, namely Direct Marketing and Knowledge Dis- Copyright © 2011 SciRes. JILSA 92 Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a Retail Bank by a Recommender System Approach covery in Databases (KDD). An extensive literature ex- decided a-priori or a-posteriori. In the first case the front ists in the area of direct marketing dealing with targeting office managers select the right customers to whom a problems. The largest body of knowledge is related to the certain action can be offered, while in the second actions process of segmentation, targeting and positioning. It are decided by the front office managers, who have to be includes several methods to define and select a target given the necessary degree of autonomy, based on their market and position an offer, i.e. define a specific mar- understanding of customers’ needs. keting mix (or any other marketing action) to that spe- 3. Approaches to the Definition of cific market. Several scholars in marketing have warned Personalized Actions about the fact that a segmentation should be done only aiming at a clear marketing goal [27], i.e. segments should By reviewing the literature, five main categories can be be built in order to deliver targeted actions. The research identified to classify the approaches to the generation of on direct marketing has produced models to define and personalized actions. Each category represents a set of deliver actions to more granular targets, even single indi- homogeneous approaches which can be used to decide viduals. For instance, “RFM models” represent custom- what action should be delivered to what customers in a ers by their Recency (i.e., the time since they have made CRM program. The approaches are reported in Table 1 the last purchase), Frequency (i.e., how often they pur- and discussed below. chase), and Monetary value (i.e., the average money they Computational approaches include all those approaches have spent) in order to compute the probability to which that build a complete model of customers’ behavior, ac- a customer will make the next purchase by a certain pe- tions, and customers’ reactions based on information riod of time. RFM models have quickly become common stored in a data set. These approaches can use both data in direct marketing and have been extended to give rise mining [32] and optimization models [21-23]. An exam- to more sophisticated models such as Automatic Interac- ple of applications in banking and finance is a retail bank tion Detection Models and Regression Scoring models which stores the data related to promotion of stocks, and [28]. In the industrial applications of these models, man- the reactions of customers who might have purchased agers often define a very small set of actions, even one those stocks or not. The fundamental condition that en- action (e.g. proposing a discounted subscription to a new ables the adoption of these approaches is the complete- service), and run the model in order to select a subset of ness of data. If the data sets do not include, for instance, individual customers who represent the optimal target. the reaction of each customer to a certain marketing Knowledge Discovery in Databases (KDD) is a pro- campaign, then neither a mining algorithm nor an opti- cess to find valid, new, understandable and useful pat- mization model can be run to generate personalized ac- terns in data [29]. This process can be seen as a possible tions. Computational approaches allow a company to way to target actions to customers. If the rules derivable fully automate the generation of actions based on cus- from KDD are “actionable”, i.e. present information on a tomers’ profiles, as no human decision is needed. A li- customer useful to decide “what to do” in order to mod- mitation is that only marketing actions already launched ify that customer’s behavior, then a manager can define before can be considered in such approach. The full cov- marketing actions tailored to that customer [30,31]. The erage of customers may be another problem because main limitation of these approaches is that either they are some customers’ reactions may remain unknown (for in- not efficient, due to the big amount of discovered rules stance, when a customer does not respond to a survey). which need human supervision, or actions are not per- Similarity-based approaches are those used by Re- sonal, because when rules are aggregated to improve effi- commender Systems [19] and Web content personaliza- ciency the target becomes aggregated as well. Starting tion methods [20]. This kind of approach assumes that from KDD, some attempts were done to define a general actions are related to customer preferences, preferences method to mining actionable patterns in databases and may be inferred by customer profiles, and that either thus generating personalized actions based on data de- similar customers behave similarly or similar actions scribing customers’ reactions to the company’s previous cause similar reactions. A “similarity-based” approach offers [32]. does not require to store as much information as a com- A last remark is useful to highlight that many compa- putational approach. Recording customers’ preferences is nies, especially in banking and finance, make use of front enough, because it is assumed that the unknown prefer- office personnel to generate actions tailored to the per- ences of a customer can be derived by identifying the sonal characteristics of customers. In this approaches, similarity with other customers. However, the twofold managers have to talk to customers, understand their pre- condition of applicability of such approaches is that cus- ferences and propose them specific offers. Actions can be tomers’ profiles have to represent preferences, and only Copyright © 2011 SciRes. JILSA Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 93 Retail Bank by a Recommender System Approach actions associated with those preferences can be gener- rather from customers’ choices. This makes customiza- ated. For instance, a customer who owns multiple credit tion different from other personalization approaches [26]. cards can be classified as a customer who “prefers” using The offer has to be granular enough in order to adopt this credit cards, whereas the fact that a customer has a mort- approach. For instance, a retail bank can propose cus- gage does not necessarily represent a “preference”. A “si- tomers to choose one among many discounted options milarity-based” approach is useful to automate the per- (e.g., a discount on bank transfers, credit cards, cash sonalization process. An example of this approach in cards, e-transactions, etc.). The association between ac- banking is given in Section 5. tions and profiles is performed by the customers instead Bottom-up approaches include the knowledge discov- of the company. For this reason, the control over the ery methods [31,33] and the use of front office personnel. process is low and the resulting actions can turn out to be These approaches consist of two separate steps: 1) pro- quite expensive. However, the targeting is expected to be filing customers, 2) deciding proper actions, where the quite effective. first step has to precede the second step (i.e., actions de- pend on profiles). They cannot be fully made automatic 4. Which Approach is Best? Discussion of because only the first step is performed by an algorithm. Benefits and Risks For this reason these approaches are typically not very The approaches identified above have characteristics that efficient. The condition of applicability is that the deci- make them beneficial in certain settings while risky in sion-making effort has not to exceed the company’s re- sources: either the number of customers is low or the others. The suitability of each approach should be thor- number of decision-makers is high. The advantage is that oughly discussed by managers based on the observation targeting can be very effective because each profile is of their benefits and risks compared to the company thoroughly analyzed before generating a proper action. goals and business conditions. Typically, managers have An example of bottom-up approach in banking is the to deal with trade-offs. The main characteristics which work of a financial advisor who manages a portfolio of can differentiate the five approaches are the following: customers. The advisor has periodic conversations with a  control on the decision-making process; customer, analyzes her needs and proposes tailored fi-  automation of the personalization process; nancial solutions.  effectiveness of targeting; Top-down approaches include the direct marketing  process efficiency; approaches [28]. They consist of the same two separate  scope of actions; steps typical in bottom-up approaches. However, in this  coverage of customers. case, the decision of what actions to deliver is made be- fore the definition of customers’ profiles and, hence, pro- 4.1. Control on the Decision-Making Process files depend on actions. For instance, a retail bank man- Companies normally want to keep as much control as agers may first decide to offer customers a discount on possible on the definitions of the actions to deliver. Pos- bank transfers, and then select the target customers by sible reasons are the need of controlling the costs related building appropriate profiles. For this reason top-down to marketing actions, and avoiding to leave too much approaches do not cover the whole customer base: only power to front office personnel. In this case, the decision customers falling in the target profiles are delivered the of what offering to customers should be centralized. As actions. In this approach, the control on the decision- an example, a retail bank may want to avoid that the front making process is very high, as the definition of actions office personnel makes the decision of what to offer to is made centrally (it is not delegated to front office per- the customers in their portfolio, because both costs may sonnel as in a bottom-up approach). Profiles can be de- increase and the bank could be exposed to opportunistic termined by either running algorithms or asking the front behavior. In this case, the suitable approaches to be used office personnel to identify suitable target customers. are “computational”, “similarity-based” and “top-down”. The condition of applicability is that actions can be de- In all these approaches, the decision of what action to fined before customers’ profiles. In some CRM problems, deliver is made centrally, by the support of either an al- such as reducing customer churn, this condition can be gorithm or the strategic orientation. The “bottom-up” and questionable. “customization” approaches may be risky because the Customization approaches are those which offer cus- first may involve the use of front office personnel, the tomers many different options and let the customers second requires the intervention of customers. choose the suitable one. This approach is typically adopt- 4.2. Automation of the Personalization Process ed in mass customization [25]. In this approach actions do not derive from processing customers’ profiles but The need of automating the process of profiling, defining Copyright © 2011 SciRes. JILSA 94 Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a Retail Bank by a Recommender System Approach personalized offers and delivering actions can come from involved in the process. The efficiency of a “customiza- several reasons, such as the need of keeping the cost low tion” approach seems to depend on the business settings, and the promptness of CRM processes high. For instance, as it can be easy to implement but the results can turn out companies in the financial services industry may want to to be expensive to the company. Efficiency can be meas- be prompt and fast in starting a retention program when ured by the number of people involved and/or by time the churn rate suddenly increases. In this case, automat- spent on the process. ing the process is critical. “Computational” and “similar- 4.5. Scope of Actions ity-based” approaches are the most suitable when the need of automation is high. In fact, they can be imple- Companies may want to be creative in the definition of mented into an algorithm which can be integrated in the the actions to deliver to customers. The proposed approa- company information system. The remaining approaches ches have different impacts on the capability of the com- (“bottom-up”, “top-down” and “customization”) can only pany to define a wide spectrum of actions. The only ap- be partially automated, as it was observed in the previous proaches that look suitable to this aim are those in the section. “bottom-up” category. In fact, it always requires the hu- man supervision, to either infer the needs from the pro- 4.3. Effectiveness of Targeting files and, in turn, the actions, or talk to each customer, Companies involved in CRM programs are often con- dig into his/her preferences and needs, and finally find cerned in maximizing targeting, i.e. to maximize the out the best offer to deliver. The scope of actions in probability that a customer receives an offer tailored to “computational” and “similarity-based” approaches are her own needs. For instance, a retail bank may want to limited by the nature of the data. In the former, the only make sure that each customer is offered exactly the pro- actions which can be offered to customers are those re- duct she needs. In these case, the best approaches to de- corded in the databases, for which customers’ reactions fine personalized actions are the “computational”, “bot- are already recorded and codified. In the latter, actions tom-up” and “customization” ones. In the first approach have to be related to the variables used to profile cus- actions are derived by knowing the customers’ response. tomers. In the “top-down” approach, the scope of the In the second, actions are defined by analyzing each pro- actions is usually small because companies define a file. In the third, customers themselves decide what ac- small set of offers. In the “customization” approach the tion they prefer. Targeting is less effective in the “simi- scope of actions depends on the spectrum of different larity-based” and “top-down” approaches because the choices customers are provided with. A potential meas- customers’ need are inferred by similarity or simply not ure for this variable can be the number of different ac- considered when the decision of actions is made. The tions. effectiveness of targeting can be evaluated by the “re- 4.6. Coverage of Customers demption rate” which measures how many customers react the expected way to a marketing campaign. Not all approaches allow a company to fully cover the customer base with personalized actions. The “similar- 4.4. Process Efficiency ity-based” and “bottom-up” approaches guarantee a full Whatever the goal of a CRM program, companies have coverage. In the first, actions can be defined by similarity. to keep the costs of CRM as low as possible. Companies Even customers for whom the amount of data is very may choose the best approach to define personalized little (e.g., new customers), actions can be derived by the actions by considering the level of resources they can preferences of similar customers. In the second, human allocate to that process. Some approaches requires less decision makers can rely on all their knowledge to define resources than others. Particularly, “similarity-based” actions starting by profiles. On the contrary, in a “com- and “top-down” approaches look suitable to this aim. The putational” approach the coverage is limited to customers algorithms required by the first approach are relatively for whom relevant information on actions and reactions simple, as it is witnessed by their quick diffusion in in- have been recorded. In a “top-down” approach, actions dustrial settings. The second approach makes the integra- are defined independently of profiles, with the conse- tion between the decision of actions and definition of quence that not all profiles can be associated with actions. profiles simpler than other. On the contrary, the algo- In a “customization” approach only customers who make rithms required by a “computational” approach are more an explicit choice are then offered the actions, those who complex. A “bottom-up” approach requires either a huge do not want to express any preference remain outside the effort in terms of supervision of the knowledge discovery coverage of the CRM program. The coverage of custom- process or a high level of resources in terms of people ers can be measured by the percentage of customers for Copyright © 2011 SciRes. JILSA Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 95 Retail Bank by a Recommender System Approach whom actions can be derived from profiles. They were ordered by the number of signals of attrition Table 1 represents each approach in terms of its char- in a ranking from high to medium risk. acteristic, conditions that must be true to adopt that ap- Once the customers were identified, the team had to proach, the benefits and risk that should be discussed in carefully face the problem of what actions should be part order to select the most appropriate approach. of a marketing campaign. A fairly straightforward choice would be leaving the decision to the personal financial 5. A Case of Application advisors, who are sales reps with a portfolio of about 150 Born as a small local bank, the company went through a (on average) customers to manage. Since they know each significant M&A process over the last few years. It ac- customer personally, they could decided the best offer to quired many new customers as a result of this process, retain each customer. However, the bank top manage- currently having about 300 000 customers in total and ment decided that the actions to deliver to customer had being a medium-sized Italian retail bank operating at a to be identified centrally. The reason was threefold. Fir- national level. In 2008 the bank managers had observed a stly, not all customers had a direct and strong relation- growing churn rate. Although the churn rate was much ship with their personal advisors. Some preferred to in- smaller than the acquisition rate, the cost for acquiring a teract with the bank via the home banking system or new customer was much higher than the cost of retaining other virtual channels. Secondly, the high number of a customer. Moreover, the cost for the re-acquisition of a personal advisors, about 2000, would make the process lost customer was even higher. These observations made quite expensive. The auditing process would be slow and the need of reducing customer churn one of the bank’s complex. Thirdly, the top managers wanted to keep a strategic priorities. strong control over the whole customer relationship At the end of 2008 a team of managers was involved management process. The problem was too crucial to be in a project aimed at both identifying customers at risk of left to the initiative of front office personnel. leaving the bank and launching a marketing campaign to Therefore, in order to generate appropriate retention retain those customers. By talking to financial advisors actions, the managers had several meetings with senior and senior managers, the team identified a number of managers and sales/marketing managers. A list of few events which normally signal that a customer is about to marketing actions was generated, such as offering the leave the company. The most important signals were a customer a credit card for free or a discount on the an- relevant decrease in the assets, the sale of more than one nual fee. The list was communicated to all the personal financial products formerly owned, the cancellation of all financial advisors. They were told to pick one or more the automatic outgoing payments and incoming credits. actions among those in the list for each customer identi- By analyzing the customer data set, the managers identi- fied at risk of leaving the bank, depending on the cus- fied the customers at risk by selecting those customers tomer’s profile. The actions had to be suitable to each who presented one or more signals of imminent attrition. specific customer and picked based on the advisors’ Table 1. Approaches to the definition of personalized retention actions. Approach Characteristics Conditions Benefits Risks The data set includes information on • Control Complete model of preferences, • Limited scope of actions Computational prior marketing actions and • Automation actions and reactions • Customer coverage customers’ responses • Targeting • Control Actions are related to customer Comparison between customer • Automation Similarity-based preferences, and preferences may be • Limited scope of actions profiles • Efficiency inferred by customer profiles • Coverage The definition of customers’ The supervision effort has to be • Targeting • Low efficiency Bottom-up profiles precedes the definition affordable (in terms of number of • Scope of actions • Lack of control of actions profiles and resources) • Coverage The definition of customers’ • Lack of targeting Actions can be defined independently • Control Top-down profiles follows the definition • Limited coverage of of customers’ profiles • Efficiency of actions customers Customers are free to choose • Low efficiency Customization The offer has to be granular enough • Targeting the appropriate action • Lack of control Copyright © 2011 SciRes. JILSA 96 Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a Retail Bank by a Recommender System Approach knowledge of the customer’s behavior. Secondly, the group decided not to include the data re- As a result, a redemption rate of 78.5% was observed, lated to customers in the branch offices most recently meaning that 1062 customers out of 1353 contacted cus- acquired from other banks, in order to keep the data ho- tomers responded positively to the marketing action mogeneous. Thirdly, the data would be referred to the proposed and did not leave the bank, thus reducing the last two years. The reason was that the bank had older overall churn rate. data only for some customers. The data available on cus- Although the results were not impressive, they were tomers coming from other banks recently acquired by fairly aligned with the managers’ expectations. In fact, M&A processes was not older than two years. Moreover, the manager themselves were aware of some flaws in the a technical constraint in the storage service made the first project. First of all, actions were independent of retrieval of data referred to years preceding the last two customers’ profiles. The decision of what to offer to cus- very slow. Finally, the data would include demographical tomers at risk was made upstream, based on previous information on customers, such as age and job, the his- experiences rather than a deep knowledge of each custo- tory of the relationship between customers and bank (e.g., mer. Therefore, a very small number of customers would when an account was open or when a financial product receive an offer tailored to their actual needs. Another was purchased and/or sold), the products owned by cus- flaw of the method was the very limited possibility to tomers (e.g., credit cards, cash cards, stocks, insurances, automate the process. While the identification of custo- etc.), the operations made on the account. Operations mers at risk could be implemented on an information were classified and divided into subcategories: 1) pay- system, the decision-making process through which to ments by card, bank transfers and cheques (incoming and identify retention actions could not. outgoing), 2) payments to utility services, 3) salaries and Encouraged by the results and the commitment of the similar incomes credited on the account, 4) other opera- bank executives, the team decided to launch a new, im- tions. Other information for each customer were also proved project of customer retention. The main goal included in the data set, such as the name of the cus- would be improving targeting and the possibility of au- tomer’s personal financial advisor, the branch office the tomating the process. customer belongs. Once the group was provided with a data set, a pre- 5.1. Predictive Model of Customer Churn liminary analysis was run in order to check the correct- The managers decided to collaborate with a group of ness of the data. The acquisition and analysis proceeded analysts, including this paper’s authors. The project fol- in an iterative way, until a complete and correct set of lowed the Cross Industry Standard Process for Data Mi- data was provided. However, the data on operations on ning (CrispDM) methodology [34]. The first step was customers’ accounts was available only for the last year. aimed to codify the managerial knowledge and beliefs, to For this reason that specific data was not used in the define organizational and business constraints and to subsequent modeling phase. The initial dataset included identify the sources of useful information. This was help- around 200 000 customer, 300 000 records and 100 va- ful to guide the analysts in the definition of the variables riables. that would be included in the model. For instance, in the The next step was aimed to clean and modify the data managers’ experience the more the products owned by a according to the analysis’ goals and method. Some re- customer the higher the probability that she would be cords were deleted while others were re-coded, and seve- loyal, while customers owning only one product had the ral variables were transformed starting from the initial highest probability of leaving the company. The products data set. The records referred to users older than 70 years which made customers loyal were typically financial were dropped from the database (customers older than 70 products such as loans, stocks and bonds. could leave the bank for “natural” reasons more than for In the next step, the analysts discussed with the man- a real choice). The records referred to customers with agers the nature of the data sets to use based on the first more than one account were merged into one overall re- step output. The data warehouse service was outsourced cord, in order to obtain a data set where each customer by the bank. Therefore, any data set had to be explicitly corresponded to one record. Several variables were asked to the outsourcer which managed the data storage. transformed. Some were discretized (e.g., length of the Firstly, the team decided to focus only on customers in relationship between customer and bank). For some va- the “basic” segment with at least a charged bank account. riables the analysts computed statistics such as average, The “basic” segment was the one with the highest num- maximum and minimum value (e.g., operations made on ber of customers and the highest churn rate. Customers in the account). Mistakes were cleaned (e.g., double codes higher segments shall not be included in the analysis. for the same value of a variable) and some records with Copyright © 2011 SciRes. JILSA Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 97 Retail Bank by a Recommender System Approach too many missing values were deleted. Table 2. Variables used in the modeling phase. The final number of customers in the data set was Variable Type 203 196. More than 200 variables were included in the new data set, although only a small subset would be used Length of customer-bank relationship numerical1 in the subsequent modeling phase. The depth of historical Place of residence nominal data owned by the bank was two years, namely 2007 and 2008. During the analysis which was performed in the Type of card owned nominal2 first months of 2009, the bank continued to gather similar Maximum value of a customer’s transaction numerical data in order to apply the model to a new set of data. Type of transaction corresponding to the maximum The core step of the project was aimed to build a pre- nominal value dictive model of customer churn and test it on actual data. After several experiments, a relatively small subset of 27 Maximum quantity of a product purchased numerical variables, taken from the initial set of 200, was selected Type of product corresponding to the maximum nominal to build the model. The set included the length of the quantity purchased relationship between customer and bank, the customer’s Type of account owned by the customer nominal3 place of residence, the type of card owned (e.g., cash, People sharing the same account with the customer binary credit or pre-paid card), the type of financial products and services owned (e.g., stocks, loans, insurances) and Products owned by the customer nominal4 their quantity and value, the type of account owned (cor- Automatic payment of utility services binary responding to a certain set of discounted options on the account), the amount of payments and incomes auto- Automatic credit of incomes on the account binary matically debited and credited, respectively, on the ac- 1All the numerical variables were discretized; 2split in 4 binary variables; count (see Table 2). 3split in 3 binary variables; 4split in 11 binary variables. After several trials, the algorithm used to build the predictive model was J4.8. The training and validation tive rate equal to 0.916. The second model selected 32 938 sets included all data referred to 2007 and 2008. Each potential churners among the active customers with an point was represented by the figure at the end of each accuracy equal to 67.34% and a true positive rate equal semester in 2007 and 2008, therefore 4 vectors of 27 va- to 0.994. The potential “churners” identified by the first riables for each customer were used at most to train the model were a subset of those selected by the second model. A ten-fold cross-validation method was used. model, except 282 customers (0.14% of the whole cus- This data included information about 11,000 customers tomer base). who had abandoned the bank between 2007 and 2008 The bank’s managers decided to select the second and about 180,000 customers who had remained loyal in model. The decision was based on the fact that the cost the same period. An additional test was performed by of a false negative was considered substantially higher applying the model, once trained and validated, to the than the cost of a false positive. The managers preferred data related to the first semester of 2009. This data set to contact more customers, even loyal customers mis- was gathered while the analysts were building the model classified as potential “churners”, than risking not to and the test consisted of applying the selected model, identify actual customers at risk. Moreover, the second with the selected settings, to the actual data in order to model performed significantly better than the first, once predict the churner users. tested on the first semester of 2009, as the second identi- Two predictive models were built by the analysts and fied 96.89% of customer who actually left the bank, provided to the bank managers who would select one, whereas the percentage of re-identification in the first based on some managerial implications. The models pre- model was 79.33%. dicted whether a customer would leave the company or 5.2. Selection of a Suitable Approach to the not, based on the data. In fact, because of the limited Generation of Personalized Actions temporal depth, it was not possible to train the model to predict customer churn within a specific period of time. The next problem was to integrate churn prediction with The output of the models were two lists of customers the decision of which marketing actions should be deliv- with a predicted binary value (“will leave” or “will be ered to which customer. The analysts proposed all the loyal”) and a corresponding score. The first model se- five approaches presented in the section above to gener- lected 11,070 potential churners among the active cus- ate personalized actions for those customers predicted as tomers with an accuracy equal to 87.35% and a true posi- potential “churners”. Firstly, the team discussed the con- Copyright © 2011 SciRes. JILSA 98 Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a Retail Bank by a Recommender System Approach ditions of applicability of each approach. Secondly, the paigns had been carried out by defining non-personalized main managerial issues were ranked by priority in order actions, and the goal was just to improve targeting. A to discuss benefits and risks. “top-down” approach would represent too small a step The only condition to discard, among those listed in towards personalization, while a “similarity-based” ap- Table 1, was the one related to the “computational” ap- proach would allow the bank to achieve a better targeting. proach. Prior actions and customers’ response data had Figure 1 reports a graphical representation of the process never been stored in the dataset by the bank. Given the followed to select the most suitable approach to the defi- absence of this kind of data, a “computational” approach nition of personalized actions. could not be adopted. All the other approaches could be 5.3. Generating Retention Actions by a potentially adopted. Actions could be inferred by cus- Recommender System tomer profiles by a “similarity-based” approach, at least partially, because they included some behavioral data The most common practical applications of the “similar- (products owned). The bank could afford the supervision ity-based” approach are Recommender Systems (RS). of effort in a “bottom-up” approach if the process was Given a set of customers and a set of products or items, a delegated to front office personnel. Actions could be de- RS predicts the unknown utility of an item for a customer. fined independently of customers’ profiles, in a “top- If item j is predicted to be of a high utility for customer i, down” approach, because managers had some experience then the system delivers the marketing action of recom- on prior campaigns. Finally, the offer was granular mending that customer to purchase that item. The action enough to let the customers choose in a relatively wide is personalized because the set of items recommended to set of alternatives, thus adopting a “customization” ap- customers is different for each customer. proach. More formally, a RS deals with two types of entities: The discussion of benefits and risks was crucial to se- users (e.g., customers) and items (e.g., products). In the lect the most appropriate approach. The two main issues transaction-based RSs, the utility of an item for a custo- raised by the managers were ranked as (1) keeping high mer is measured by a Boolean variable indicating if the control on the marketing process, (2) achieving better user owns a particular item or not, or with the purchasing targeting with respect to prior marketing campaigns. According to the first issue, the team excluded the adop- frequency of an item, or with the usage frequency of that tion of both a “bottom-up” and a “customization” ap- item [35,36]. Based on the known values of utility, a RS proaches. In fact, given the high number of customers tries to estimate the utility of the yet unseen items for predicted at risk, a “bottom-up” approach would be fea- each customer. In other words, a RS can be viewed as the sible only by delegating the task of revising each profile rating function R that maps each user/item pair to a par- and deciding the corresponding actions to the front office ticular utility value [18]: personnel. This would entail a substantial loss in control. R: Users × Items  Utilities (1) The same would happen by letting customers choose among different optional actions in a “customization” One of the main tasks of a RS is to make this function approach. The choice between a “similarity-based” and a total by estimating the unknown utilities. The estimation is “top-down” approach was eventually made based on the done by using a similarity function, which can be designed second issue. In fact, the bank’s prior marketing cam- by following one of the following three approaches [18]: Wh at No data on Exclude the conditions are actions and “computational” not true? reactions approach Exclude the Exclude the Select the What are the Control “bottom-up” and Targeting “top-down” “similarity-based” main issues? “customization” approach approach approaches Figure 1. Representation of the decision-making process. Copyright © 2011 SciRes. JILSA Beyond Customer Churn: Generating Personalized Actions to Retain Customers in a 99 Retail Bank by a Recommender System Approach  Content-based recommendations: the user is rec- bank, place of residence). ommended items similar to the ones the user pre- According to the standard collaborative filtering ap- ferred in the past; proach, the “neighborhood”, i.e. the set of customers z  Collaborative-filtering recommendations: the user with tastes and preferences similar to customer i, was is recommended items that people with similar formed by using the cosine similarity, which is given by tastes and preferences liked in the past; the following formula:  Hybrid approaches: they combine collaborative-  r r filtering and content-based methods. iz i,s z,s simi,zcosi,z  sSiz (2) This general approach had to be adapted to the prob- i  z  r2  r2 lem of generating retention actions to customers at risk of 2 2 i,s z,s sSiz sSiz churning. Three hypotheses were identified in order to consider the approach feasible: 1) recommending certain where r and r are the ratings of item s assigned by i,s z,s products to customers can improve retention; 2) the pro- user i and user z respectively. S ={s  Items|r ≠   iz i,s ducts able to improve customer retention are not the r ≠ } is the set of all items co-rated by both user i and z,s same for all customers and can be identified by looking user z, and iz denotes the dot-product between the at the behavior of loyal customers; 3) similar customers vectors i and z. In this specific case, the customer i be- behave similarly and have similar preferences. longs to the set of customers predicted as churner, While the third hypothesis has much support in the whereas the customer z belongs to the set of loyal cus- marketing literature, the first two were significantly tomers Z. Only the customers predicted as loyal with a supported by the management expertise and partially probability higher than 90% were considered in Z. The by empirical evidences. In fact, the bank’s managers strong hypothesis of this approach is that similar cus- were confident that certain products, such as bank’s tomers behave similarly, so a customer predicted to leave stocks, insurance products and loans, directly contrib- the company can be retained if she is recommended ute to improve customers’ loyalty. These beliefs had products characterizing the loyal customers with similar emerged in the first project step in the statements of tastes and preferences. several managers interviewed by the analysts. The Therefore the main difference between a standard re- managers were also persuaded that the products which commender system approach and the system used in this can improve retention are not the same for all custom- case is the way the neighborhood is formed. While in the ers, as each customer may have a different sensitivity former, the neighborhood include all the customers simi- to different products. These differences come from lar to the customer to whom an action has to be delivered, several causes, such has different personality traits and in the latter the customers included in the neighborhood different life cycle stages. Because of this variety of were taken from a subset of all customers, namely the behavior, the statistical analyses only partially vali- most loyal customers. The final output of the recom- dated these beliefs. Customers had been split into loyal mendation algorithm was a list of products that should be and those who had abandoned the bank, although some suggested to each customer at risk. differences could be observed between the two groups As a result, a redemption rate of 81.6% was observed, in the distribution of the products owned by customers, meaning that 1313 customers out of 1609 contacted cus- they were statistically significant only at a low prob- tomers, responded positively to the marketing action ability. proposed and did not leave the bank, thus reducing the As a consequence of these observations, the team of overall churn rate. analysts and managers decided to adopt a collaborative 5.4. Discussion of Results filtering approach to recommendations, where each customer at risk would be recommended products that The method used in the first project can be classified as a loyal customers with similar tastes and preferences hybrid of a “top-down” and a “bottom-up” approaches. liked in the past. In fact, on the one hand the team of managers decided Each user was defined as a vector containing both de- what actions should be delivered to customers independ- mographic and transactional data. That kind of informa- ently of the profiles of customers at risk. The decision tion was used because two users can be defined as simi- was made by talking to senior managers and was based lar only if they are similar in terms of preferences (e.g., on a general knowledge of the way customer reacted to products and services owned by customer, quantity and marketing campaigns in the past. On the other hand, the amount of transactions) and other behavioral and demo- team asked the front office personnel (the personal fi- graphical variables (e.g., length of relationship with the nancial advisors) to select the most appropriate action Copyright © 2011 SciRes. JILSA