Review Popular Techniques for Managing and Understanding Customer Churn in Banking and Finance

Abstract

Customer churn has been evolving as one of the major problems for financial organizations. The incessant competitions in the market and high cost of acquiring new customers have made organizations to drive their focus towards more effective customer retention strategies. Though banking and finance sectors exhibit low customer churn rates as compared to other sectors, the impact on profitability by losing a customer is comparatively high. Customer churn management plays a vital role for an organization to enhance long term profitability. Hence much research has been going on for understanding customer’s switching behavior and determining its determinants. There also has been a significant development of churn management techniques for effectively modeling customer churn in banking and finance sectors. The focus of this paper is to review most popular techniques and methodologies that have been identified from the literature for managing and understanding customer churn in banking as well as finance sectors. The benefits and limitations of the identified techniques are discussed and directions for future research are offered.

Keywords: Customer switching behavior, Bank, Data mining, Churn Management

Introduction

The effect of global melt down had a fierce impact on economies of both developed and developing nations. The crisis in the US sub-prime mortgage market brought tremors in financial markets all over the world resulting into intense competition for survival. This caused banks and financial companies to focus more on retaining customers rather than investing in acquisition of new customers. Moreover, acquiring new customers can cost five times more than satisfying and retaining existing customers . Hence the need for customer churn management in banks and finance sectors became inevitable. The paper provides a holistic view of the current practices related to customer churn management in banking and finance verticals. The most prominent techniques and methodologies identified from the literature for managing churn are also reviewed. The paper also identifies some of the major gaps in the existing literature and provides areas of future research that still needs to be paid attention for effective management of customer churn. This would help practioners to rectify their current practices and also look at other techniques and dimensions of customer churn management.

Section 2 provides the methodology used for reviewing the existing literature. Section 3 provides definition of churn and churn management as well as types of churn have also been discussed. The section also describes about existing frameworks for churn management. A discussion on data requirement and gathering has been provided in section 4. Issues related to data sampling have been also discussed. Section 5 discusses about various concepts and techniques related to customer churn management. Section 6 provides discussion on limitations of the existing studies and areas of future research.

Methodology

The paper is based on the methodology suggested by for reviewing the literature. The relevant literature on customer churn management was found by exploring several journal databases like ebsco, science direct, citeseer, emerald insight, jstor and so on. A majority of the papers are from the journal Expert Systems with Application, followed by European Journal of Operational Research and papers from some international conferences. A concept-centric approach was followed while reviewing all the papers. This resulted in a concept matrix containing all the prominent concepts found across all the papers. The concepts were then isolated and studied by taking unit of analysis as a group, an organization or an individual. It was seen that most of the studies identified from the literature were carried out at an organizational level only. We have discussed each of the identified concepts related to customer churn management in banking and finance domains latter in this paper.

Churn Management

Churn is defined as the movement of customers from one provider to another and churn management is the process for holding profitable customers with the company through appropriate marketing campaign and retention strategies. There are majorly two categories of customer churn, voluntary churn and non-voluntary churn. Non-voluntary churn is initiated by the company in which a company withdraws its service from a customer. On the other hand voluntary churn is initiated by the customer when he/she decides to terminate his/her service from the provider, hence voluntary churn is more difficult to determine. Voluntary churn has been a challenge for all the companies and it constitutes of major portion of company’s total churn. A voluntary churn can be either incidental or deliberate. Incidental churn happens due to circumstances which prevents customer from continuing his service with the provider. Deliberate churn occurs when a customer decides to switch to another service provider. Some reasons for this type churn include bad service quality or attractive and low priced offers by competitors .

Customer churn has become a massive problem that affects other aspects of Customer Relationship Management (CRM) . For banks and financial organizations maintaining relationship with the customer is of highest priority. Although these sectors exhibit a low churn rate, the impact of losing a single potential customer can have a drastic effect on company’s profitability. Hence it’s essential for companies to efficiently manage customer churn for long term profitability and survival in the market.

There have not been many studies on forming a churn management framework apart from and recently . Figure 1 shows the churn management framework proposed by which captures the common requirements essential for an efficient churn management system.

Fig. 1. CMF Structure

The framework in comparison to the ordinary churn prediction program, provide insights to customer churn behavior which can help managers for planning effective marketing strategies.

Considering all the methodologies identified from the literature for predicting churn in banking and finance domains, most of them relate closely to the methodology suggested in above discussed frameworks. These frameworks have been applied and tested for telecom sector; however their application in banking and finance sector is yet to be investigated.

Data for Customer Churn Management

The existing studies suggest that selection of data can be based on the type of analysis to be done as different combinations of the data hold different analytical capabilities hence it is essential to identify the data that best suits the type of analysis to be performed . In some of the cases it might happen that required data is not available from company’s database, in such cases customer surveys can be conducted to obtain the required data .

The event of customer churn is often rare in case of banking and financial service providers. It has been observed from previous studies that the percentage of churners in a data would be comparatively less than that of non-churners. One of the most common approaches to deal with this problem has been to oversample the data with significant data of churners. This problem of class imbalance has been addressed by and studied the use of random sampling, advanced under sampling, boosting and cost sensitive learner techniques for handling the problem of class imbalance.

Customer Churn Management

There are many concepts to be considered within customer churn management. This section provides discussion on some of the most prominent concepts of customer churn management related to banking and finance sector.

Customer Satisfaction

Several studies have been carried out for measuring customer satisfaction and identifying determinants of customer satisfaction. One of such studies had been done by for determining critical satisfaction dimensions for Greek private bank sector to evaluate performance of the bank and identify distinct groups of customers. The study illustrates the implementation of a preference disaggregation methodology in the private bank sector using Multi criteria Satisfaction Analysis (MUSA) for identifying and understanding behavior of various customer groups. The practical implication of this study can help in improving the performance among all the branches of the bank by establishing an internal benchmarking system based on customer satisfaction evaluation of each of the branches. Moreover, considering the recent sub-prime mortgage crisis, a strong image of a bank or any finance company does not influence customer satisfaction but can be an important driver of customer dissatisfaction in case of organization’s poor image . This behavior is more related to interpersonal relationship between the customers and the employees of the bank and their attitude towards the customers. proposed a three phase approach for increasing customer satisfaction according to which different strategic actions can be taken for different levels of customer satisfaction. This phase wise approach can help managers in re-designing the products/services and re-engineering of various processes according to customer needs.

Customer potential and Life Time Value

Due to increasing availability of customer data, banks and finance companies are moving towards more customer-centered approach for marketing which leads to a strong need for estimating Customer’s Lifetime Value . The CLV for a customer is defined as the present value of future cash flows yielded by the customer’s product usage, without taking into account previously spent costs . In situations where the duration and nature of customer relationship is uncertain, CLV can serve as important tool for measuring customer loyalty and their profitability to the organization. CLV can be used as effective measure for detecting customer churn by estimating CLV for all customers . CLV has also been used as a significant indicator in predicting customer churn and consequently planning proactive customer retention strategies for the high-risk customers . Such studies can also help banks and finance companies in estimating acquisition cost for a customer by predicting CLV of a customer beforehand.

The potential value of a customer is defined as the profit or value delivered by a customer if this customer behaves ideally, i.e., the customer purchases all products or services he currently buys in the market at full prices at the focal company . The information regarding the potential value of a customer depends on customer’s purchasing behavior data. has presented a framework that provides insight into potential value of customers of a multi-service industry. The potential value of the customers enables organizations to plan their investments in customers that are potentially valuable and also minimize their investments in non-valuable customers.

Switching Behavior

The changes in customer’s switching behavior have not been paid much attention in the previous studies. The understanding of customer’s switching behavior would significantly help organizations in identifying factors leading to churn. identified seven key factors that influence switching behavior of customers of Chinese bank. These factors are switching costs, price, service quality, organization’s reputation, effective advertising competition, distance in terms of reachability and involuntary switching factors. Moreover, the situations when customers intend to switch persists at each stage of the CRM. The findings from show that loan negotiations are one of the most sensitive situations where great care should be taken while dealing with the customers. In addition, stiff competition in the market also leads a customer to switch for better deal. studied effects of various types of relationship bonds on customer switching behavior. For a bank, the study emphasizes on the importance of understanding the relationship bonds between different segments of customers in order to differentiate themselves from the competitors. The implications derived from indicate that personality traits and emotional intelligence of an insurance industry customer highly influence his switching behavior. Hence to increase customer loyalty, providing public aids for emergencies and disasters can have a positive impact on customer’s emotional intelligence and consequently on personal traits.

Feature Selection

In terms of retaining the accuracy of churn prediction, the process of reducing the number of the features is called feature selection . Feature selection is the most critical part of predictive churn modeling due to high dimensional characteristics of the churn data. used PCA for reducing the feature set to predict customer churn in credit card users. A study carried out by using six different datasets indicates that considering all the information available reduces the predictive power of detecting churn as some of the information can be misleading when training the model. A framework for feature selection specific to customer churn prediction has also been proposed by .The framework is divided into two categories of data – structured and unstructured. However the framework has only been tested for structure data.

Predictive Modeling

The most critical part of churn management framework is to the have ability to predict future. This section is further divided into following two sub sections:

Modeling Techniques

There have been wide ranges of techniques used for predicting customer churn in banking and financial sector. Table 1 gives an overview of the techniques used among all the papers along with the goal of prediction.

Decision trees are one of the most commonly used techniques for predictive modeling. Some of the most widely used decision tree algorithms include Classification and Regression tree (CART), C5.0, and ID3. shows the use of decision trees in predicting customer churn.

Support Vector Machine (SVM) uses a nonlinear mapping to transform the original data into a higher dimension. It then separates the data of one class from other by a decision boundary called as hyper plane. used a version of SVM called the Least Square Support Vector Machine (LS-SVM) and Rough Set Theory (RST) for predicting churn in credit card customers. For predicting customer churn in banks, SVM was found to better than other algorithms like Neural Networks, logistic regression, Naïve Bayesian and decision trees predicted customer’s potential value by using a linear regression probit, a multi regression probit and Naïve’s model. However the average performance of the probit models was found higher than that of Naïve’s model. used quantile regression technique for predicting CLV of credit card customers.

The use of robust techniques like Random Forests (RM) for predicting customer is increasing in case of banks and financial services. It can be seen from Table 2 that RM is the most widely used technique for predicting customer churn. Random forests are a combination of tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all trees in the forest . used Random Forests to predict customer retention for a financial firm. The prediction was done considering customer’s likelihood of buying new products and the decision to cancel the existing product. proposed an extension of the original Random Forest Method called Improved Balanced Random Forest (IBRF) technique by combining the sampling technique of balanced random forests (bRF) and cost sensitive learning capability of weighted random forests (wRF). The study by also indicates better performance of wRF while predicting churn in case of banks.

Table 1. Techniques used in all papers

Customer churn

There also have been other methods for predicting churn apart from the traditional modeling techniques. Moreover simple methods like using financial metrics on data extracted through OLAP have been also been used .

Model Evaluation Metrics

There are several metrics for evaluating and measuring performance of the classification models. We would be discussing all of the model evaluation metrics used for evaluating performance of the models shown in Table 1.

• Accuracy/Error percentage – It is most simple and basic criteria for evaluating model performance. It measures the percentage of observations correctly classified from all the observations.
• Lift – For calculating the Lift, the dataset is divided into several portion known as deciles then the ration of correct instances in a specific decile is divided by the average ratio of correct instances in the entire population . A lift chart is usually drawn for comparing performance of different classification models, the higher curve the better the performance of the classification model.
• Receiver Operating Characteristic (ROC) Curve – ROC curve shows difference between true positive rate (the proportion of true observations that are correctly identi¬ed) and false positive rate (the proportion of false observations that are incorrectly identi¬ed as positive) for a given model .
• Area Under ROC Curve – has suggested AUC to be a good evaluation metrics in comparison error/accuracy for evaluating performance of the model. AUC of 1 is perfect prediction indicating zero error. AUC of 0.5 is random prediction indicating no relationship between predicted and actual value .
• Mean Absolute Prediction Error (MAPE) – It measures how predicted values deviate from the actual value on an average. used MAPE for evaluating performance of the model.

Apart from the above discussed model evaluation metrics, the life span of a model should also be taken into consideration. discussed about the life span of the churn models. The study explains the variation in the model by considering multicollinearity in the data, omitted variables, and actual changes in the situation in which the model was build.

Customer Segmentation and Profiling

The process of segmentation analysis describes the characteristics of groups of customers within the data, and putting customers into segments according to their affinities or similar characteristics . A study by emphasizes on cross selling the most convenient products to the identified segments on the basis of their characteristics in order to maximize profit and to retain customers. A simpler way of grouping credit card customers is by Usage Segment Code (USC) which clusters existing customers into distinct groups based on their spending pattern, delinquent status and history and so on. segmented the customers of an insurance company into a two by two segmentation matrix on the basis of customer’s potential value an discussed strategies for each segment. The matrix used is shown in fig. 2.

Fig. 2. Segmentation with current value and customer potential .

proposed a segmentation matrix based on two aspects of future value, the CLV prediction and uncertainty around the prediction. Such segmentation schemes will allow decision maker to consider the risk of a customer segment not being profitable as expected or an opportunity that a segment is more profitable. In order to overcome the issues of cross sectional data, performed clustering of bank customers based on longitudinal data set using a clustering technique based on traditional DBSCAN algorithm. A panel or longitudinal data provides of multiple observations for each individual over time this helps in revealing hidden information in data in contrast to cross sectional data.

Retention Strategy

The most important part in customer churn management is planning of effective and durable retention strategies. did association rule mining to discover rules between the segments of customers and their demographic and geographic characteristics for devising better bank marketing strategies. used sequential pattern to detect early warning signs before losing a valuable customer for devising more effective and customer centric retention strategies. explained a customer retention process based on its customer value wherein the customers at risk of churning were offered series of promotions starting with the least attractive offer to the most attractive offer for persuading customer to stay.

Discussion and Directions for Future Research

Although, significant research has been undertaken for understanding customer churn behavior and consequently managing churn in banking and financial organizations, there certain gaps in the existing studies which still need strong attention. Almost all the studies on understanding customer switching behavior adopted a cross sectional analysis of the data due to which much of the hidden information was not considered. The findings from some of the studies do not claim their significance on their external validity . The existing studies focus only customer’s past and present behavior which limits the study in identifying potential factors that may influence a customer to churn . There has not been much focus on selection of appropriate sample of population to be examined. This can lead to much biased results due to which the results obtained cannot be generalized for other regions or domains. Whereas for predicting customer churn, the use of feature selection for selecting optimum inputs to model have not been considered in many studies which can significantly degrade the performance and accuracy of the models.

In addition to above discussed gaps in existing studies, future research can be done for identifying churn determinants for banks and financial organizations that can be generalized across the globe. Considering the nature of the business pertaining to these verticals are into, much research is essential for developing a standard churn management framework specific to this domains. This would help in understanding customer churn in these domains more closely and would consequently lead to planning of more effective marketing strategies. Apart from traditional algorithms the use of more sophisticated algorithms like Rule Induction and Ensemble methods should be examined for improving accuracy in predicting churn. The use of RMF (Recency, Monetary, and Frequency) variables can also be used for better analyzing the customer base. Moreover, the use of behavioral churn determinants as inputs to the predictive model or pattern mining techniques can be examined for improving accuracy of the techniques.