Predicting Customer Churn In Tesco: C&R Tree Vs RFM Method

Introduction and problem definition

When customers cease their relationship with an organization then such customers are referred to as customer Churn. Organizations generally refer to investigate such customers when time has elapsed between the last purchase of such customers. Customer churn is a financial drain on the organization. They not only lose the customer’s revenue but also the organization has to invest separately for marketing cost to build a new customer. Thus reducing customer churn is a very important aspect for organization (Dahiya and Bhatia 2015).   

For the survival of an industry the customer is very important. Further customer relationship is essential for growth of the business. Organizations invest hugely in building customer relationship. The maturity of a customer can be estimated by the number of customers it is able to retain (Guo and Qin 2015). In the present assignment we analyse the data set of Tesco PLC. The organization runs a multinational supermarket chain from Welwyn Garden City, Hertfordshire, England, United Kingdom. It operates in 12 countries. In 1995 Tesco launched customer loyalty scheme – Tesco Clubcard. Tesco Clubcard has been cited as being responsible for growth of the organization. Tesco Clubcard owners gain one point for every £1 spent. Thus, Tesco is able to collect information on customer behaviour. This also enables Tesco to retain customers through targeted campaigns.

In the present assignment information from 1^st January 31^st December 2015 has been collected. The assignment is an attempt to predict factors responsible for churn of the customers based on different factors. The churn model is predicted with the C&RT model. The decision tree model is predicted with the aid of C&RT model. The RFM model was further used to evaluate the prediction from C&RT models.

The growth of internet has fuelled the growth of customers. However, this growth has also given rise to information about business. Today a customer can get information of a particular product sold by different brands easily. Thus a customer unsatisfied with a particular service provider then he can easily jump providers. This processes of jumping service providers is defined as churn. One of the most important measure of lost number of customers is Customer Churn. The prediction of customer churn thus becomes essential in order to maintain loyalty of customers. The process also helps in improving customer relationship management (CRM). Various factors are responsible for customer churn. Hence the prediction of customer churn becomes important since evaluation of the factors would help to provide information on causes of Churn (Almana, Aksoy and Alzahrani 2014).

According to Coussement, Lessmann and Verstraeten (2017) the presence of a saturated market due to the fierce competition amongst organization and continuous revised offerings by organizations to customers has initiated the process of customer defection. Thus it has become imperative that a targeted program be developed by business houses to prevent defection. But with the presence of huge amount of data there is little scope that such information can be derived manually. Thus one has to take the help of advanced algorithms and computers to find factors responsible for preventing defection.  

Literature review: Use only academic journal articles. APA style should be used for referencing.

To predict Churn of Customers different prediction techniques may be used. Logistic regression technique is used to predict churn by connecting the factors affecting churn with probability of occurrence of customer churn.

The most popular predict model used to predict churn is the decision tree model. The decision tree provides a chart which represents the relationship of the variables. The decision tree is created by starting from the root node. The gain ratio is calculated from information gained. In order to build a decision tree, CART, CHAID and C5.0 may be used (Hassouna, Elyas and Abou Trab 2015). There are many advantages of using decision tree. The most important advantage is it provides easy visualisation of the information. Further it is a non-parametric approach to assessing customer churn. Moreover, no prior assumptions is required to calculate Churn. In addition, decision tree can process both numerical as well as categorical data (Song and Ying 2015). However, decision tree suffers from certain disadvantages also. The formation of a decision tree is dependent on interaction between variables and their attributes. Thus, the more complex the variable interaction the more complex the decision tree. In addition, complex decision trees are difficult to visualise. Furthermore, it suffers from robustness due to training and testing data sets (Priyama 2013).  

According to Patil et al., (2017) customer churn occurs when the concerned customer terminates his association with the organization. Moreover, with online businesses coming into vouge the organization can get real time information towards interaction of the customer with the website. Churn generates a loss of revenue to the organization in terms of revenue due from the organization according to the researchers. Moreover, there also is a loss in terms of marketing loss since the organization needs to find another customer. Prediction models for churn are thus essential to understand the behaviour of the customer. The records provide for attributes which are essential for predicting churn as well as time of churn of the customers. With finances being involved in customer churn hence the accuracy of prediction of churn become relevant. With higher accuracy in prediction of churn the organization can initiate proactive action to prevent the process.

According to Jahromi, Stakhovych and Ewing (2014) management of churn by organizations is a two pronged strategy. The classical theory on churn management proposes to target those customers who have the highest probability to defect. Thus identifying churners and trying to prevent the process through incentives or inducements is the classical process. The more modern theory suggests to target customers who have the highest potential for profit. The profit margin generated from the customers can be grouped on the basis of profits. Thus the overall return from such customers is selected and thus a retention campaign may be initiated. Further, according to the researcher’s prediction of churn is difficult in a B2B business since customers are non-contractual. Thus in the absence of any binding on the customer to be loyal to the organization the prediction of timing of churn becomes difficult.

Methodology and empirical study: this should include a discussion of your analytical techniques, your model evaluation metrics, your working data, and your model building process.

For the CART method the Gini index is used measuring impurity. The Entropy and Gini index is calculated as: (Sahin, Bulkan and Duman 2013)

where p_i is the relative frequency of class i.

In the making of decision tree using CART, nodes are created. All variables are checked for their suitability of split. For categorical variables multiple splits are possible. For numerical variables binary splits are done. The minimum misclassification cost of each class is calculated and the least misclassification variable is used to conduct the split. The misclassification of a node associated with class and probability for a numerical variable is given as:

Once the C&RT model is built the accuracy of the model needs to be predicted. This can be done through splitting the initial data set into training and testing set. The model is initially build on the training set. After the model is build the model the accuracy of prediction of the model is evaluated on the testing set. The accuracy and precision of the model provides information on the predictability of the model.  

RFM is the ideal mechanism of splitting customer on the basis of information. The information required by RFM method is recent purchase of the customer and the value of the purchase. The higher value of the purchase and the higher the frequency of purchase provides higher values of RFM (Sarvari et al., 2016).

The advantages of RFM are its combination with customer segmentation, it can predict customer churn in short terms thus preventing churn (Dursun and Caber 2016). However, it suffers from disadvantage since it uses only three criteria for calculation while other factors are not considered. Higher correlation between monetary value and frequency of purchase is only considered. Non-profit customer is rarely considered in the process.

According to Chen, Hu and Hsieh (2015) management of churn by organizations is essential to minimize loss by businesses. Churn also is helpful in retaining high-value customers; thus the organization is able to maximize loss. According to the 80/20 rule 20% of customers of an organization provides for 80% of the revenue to the organization. Thus retaining of high value customers who provide are loyal to the organization pays. The identification of factors of RFM varies according to the industry and business. Thus, identification of the variables representing RFM for each industry becomes a tedious process. Moreover, according to the researcher customer loyalty can be affected by the longevity of the association of the customer with the organization.  

The word RFM stands for Recency, Frequency and Monetary. Recency signifies how lately the customer has made a purchase. Frequency signifies the occurrence of purchase of customers. Monetary value signifies the amount of money spent by the customer.

Research by Zaki et al., (2016) recommends RFM model since it transforms the transactional information (purchase behaviour) of the customer to profit of the business. Thus information on customer’s behaviour is segregated on the basis of purchasing behaviour. In the initial stage the customers are grouped which is preceded by providing ranks to the customers. Customers with high recency values earn a low score. Similarly, customers with low recency values earn high scores. This process is followed since the more recent the purchase by a customer the more valuable he is to the organization. Similar steps are followed for frequency and monetary values. Thus is this process each customer receives a RFM score.   

Results: evaluate your analysis results to explain how the constructed models perform and also how they are positioned against a random model (random guessing).

The information for churn in the dataset was a result of automatic recording by computers. Thus all information was present in the dataset i.e., there was less chance of missing data. Thus at the pre-handling stage presence of missing information was manually observed for the present research.

Data transformation refers to the situation of converting data from one format to another format. This process becomes essential when information from one source is required for another source. In the present case prior information for churn as generated from the C&RT model was required to be transformed for suitability in RFM model.

The quantification of churn was carried out with the help of two different processes. In the first stage IBM SPSS modeller was used. C&ART model of IBM SPSS Modeler was used to model the churn of the client in light of the information collected by TESCO. To model churn of customers the training dataset provided for preliminary investigation was used. The dataset provided was imported into IBM SPSS Modeler. The data was post-processed with “type” filed operation. In this operation “churn” was set the role as the “target” variable and “ID” was given the role as “none.” Further the “C&AR” model was selected and integrated into the model building process. Thus, the relevant model for prediction of churn was generated from the dataset. After the initial model was build, the model was further analysed and graphs and charts were generated.

For the second stage the information as derived from IBM SPSS Modeler was validated. The validation of the C&AR model was done with the help of RFM model. The process to validate the model was done using MS EXCEL. Two step Clustering technique was used in the RFM model. This process was used since the requirement was to group the clients on the basis of churn.

The initial churn of the customers is build using the given model:

Figure 1: The initial C&RT Model

The above C&RT model provides information regarding the churn of the customers from the training dataset.

The model was subsequently changed to accommodate bot training and testing data set.

Figure 2: The Final C&RT Model

The above C&RT model provides information regarding the churn of the customers from both the training and testing dataset.

Figure 3: Prediction Importance

The above chart shows that importance of independent variables in prediction of churn. From the above figure it can be assumed that T.Last and Total profit has higher prediction importance in churn of customers as compared to other independent variables.

The importance of the predictors are provided in the following table.

Table 1: Predictor Importance

Nodes	Importance	Importance	V4	V5
BH.spent	0.0031	0.0031	BH.spent	0.0031
DL.spent	0.0031	0.0031	DL.spent	0.0031
FV.spent	0.0031	0.0031	FV.spent	0.0031
AP.spent	0.0031	0.0031	AP.spent	0.0031
MT.spent	0.0031	0.0031	MT.spent	0.0031
Socio.Economic	0.0031	0.0031	Socio.Economic	0.0031
GM.spent	0.0031	0.0031	GM.spent	0.0031
LQ.spent	0.0031	0.0031	LQ.spent	0.0031
Total Profit	0.2553	0.2553	Total Profit	0.2553
T.last	0.7139	0.7139	T.last	0.7139

From table 1 and figure 3 it is found that T.Last is the most important variable for the prediction of churn of customers. T.last represents the time difference between the first and last purchase of the customer. Thus it is found that the time difference between purchases is the most important factor for churn. The second most important factor is the profit of customers. The importance of profit is given a factor of 0.2553. All other independent variables are same importance to predict churn of customers.

Conclusion and Recommendations: This should consider the implications of your results and how they may contribute to customer retention and reduce marketing expenditure.

 

Figure 4 below represent C&AR model for churn of customers. The model also shows that the initial split in the dataset is due to T.last of the customers. The first split takes place for T.last less than or equal to 15.214 and above 15.214.

Similarly, the second split in the decision tree for the sub-branches T.last is again factor. For all T.last below 15.214 the split for T.last takes into account a value of 6.071. For all T.last above 15.214 the split for T.last takes into account a value of 22.929.

The above table presents the analysis of the prediction model as predicted from the use of both the training and testing data set.

Table 3: Confusion Matrix

Correct	15240	76.85%
Wrong	4590	23.15%
Total	19830	100%

Table 4: Accuracy of the training model

Correct	9548	95.48 %
Wrong	452	4.52%
Total	10000	100%

From table 3 it is found that the training model is 76.85% correct in predicting the model. Similarly, the accuracy of the model built through the C&RT process is found to be 95.48% accurate.

Figure 5: Lift Chart

The lift chart for training and testing data set is presented in figure 5.

For the RFM model the RFM scores of the customers were created. The gain and cumulative lift values were calculated.

Table 5: RFM Analysis

Decile	Number of Cases	RFM Score	Cumulative RFM score	% of events	Gain	Cumulative Lift
1	1000	3000	3000	3.33	3.33	0.33
2	1000	3003	6003	3.34	6.67	0.33
3	1000	6000	12003	6.67	13.33	0.44
4	1000	6003	18006	6.67	20.00	0.50
5	1000	9000	27006	10.00	30.00	0.60
6	1000	9003	36009	10.00	40.00	0.67
7	1000	12000	48009	13.33	53.33	0.76
8	1000	12003	60012	13.33	66.67	0.83
9	1000	15000	75012	16.66	83.33	0.93
10	1000	15003	90015	16.67	100.00	1.00
Total	10000	90015

Figure 6: Gain Chart from RFM analysis

Figure 7: Lift Chart from RFM analysis

Through the use of Lift chart in RFM model it is found that the model is 94.87% accurate. The percentage of accuracy is almost similar to the accuracy as generated by C&RT model in IBM SPSS modeller.

Table 6: Accuracy of RFM model

Correct	9487	94.87%
Wrong	513	5.13%
Total	10000

Conclusion

The discovery of information is essential for the growth of a business. Moreover, data mining processes are essential for mining the data. Since the volume of data is huge hence selecting the perfect data mining tool as well as mining method becomes essential. The use of a data mining tool provides refined information which in due course of time can be transformed to actionable information. In this case the prediction of factors responsible for churn of customers for TESCO plc. can provide the organization with means to prevent churn.

The management of TESCO plc. had initiated the process of customer loyalty scheme in order to retain customers. Towards the goal it had also introduced clubcard points which could be collected on every purchase by the customer. However, TESCO plc. was also concerned of the churn amongst the customers. Churn of customers acts as a loss to an organization. Thus to prevent customer churn it is essential that factors be introduced which would prevent churn. However, introduction of factors is a financial burden to an organization. Thus it becomes prudent to provide extra benefits to those customers who are likely to move away from the organizations. Moreover, factors which may be introduced to keep the loyalty of customers has been uncovered.

Thus, the present research concentrated on the factors which can be recognised for churn of customers. From the present research TESCO plc. can find that T.Last and profit of customers are responsible for predicting churn of customers. Thus from the information gained by the organization TESCO plc. can focus on those customers for whom T.last is least as well as n customers who provide good profit.   

Reference 

Almana, A.M., Aksoy, M.S. and Alzahrani, R., 2014. A survey on data mining techniques in customer churn analysis for telecom industry. International Journal of Engineering Research and Applications, 45, pp.165-171.

Chen, K., Hu, Y.H. and Hsieh, Y.C., 2015. Predicting customer churn from valuable B2B customers in the logistics industry: a case study. Information Systems and e-Business Management, 13(3), pp.475-494.

Coussement, K., Lessmann, S. and Verstraeten, G., 2017. A comparative analysis of data preparation algorithms for customer churn prediction: A case study in the telecommunication industry. Decision Support Systems, 95, pp.27-36.

Dahiya, K. and Bhatia, S., 2015, September. Customer churn analysis in telecom industry. In Reliability, Infocom Technologies and Optimization (ICRITO)(Trends and Future Directions), 2015 4th International Conference on (pp. 1-6). IEEE.

Dursun, A. and Caber, M., 2016. Using data mining techniques for profiling profitable hotel customers: An application of RFM analysis. Tourism management perspectives, 18, pp.153-160.

Guo, F. and Qin, H.L., 2015, November. The Analysis of Customer Churns in e-Commerce Based on Decision Tree. In Computer Science and Applications (CSA), 2015 International Conference on (pp. 199-203). IEEE.

Hassouna, M., Tarhini, A., Elyas, T. and AbouTrab, M.S., 2016. Customer Churn in Mobile Markets A Comparison of Techniques. International Business Research, 8(6), 224-237

Jahromi, A.T., Stakhovych, S. and Ewing, M., 2014. Managing B2B customer churn, retention and profitability. Industrial Marketing Management, 43(7), pp.1258-1268.

Patil, A.P., Deepshika, M.P., Mittal, S., Shetty, S., Hiremath, S.S. and Patil, Y.E., 2017, August. Customer churn prediction for retail business. In 2017 International Conference on Energy, Communication, Data Analytics and Soft Computing (ICECDS) (pp. 845-851). IEEE.

Priyama, A., Abhijeeta, R.G., Ratheeb, A. and Srivastavab, S., 2013. Comparative analysis of decision tree classification algorithms. International Journal of Current Engineering and Technology, 3(2), pp.334-337.

Sahin, Y., Bulkan, S. and Duman, E., 2013. A cost-sensitive decision tree approach for fraud detection. Expert Systems with Applications, 40(15), pp.5916-5923.

Sarvari, P.A., Ustundag, A. and Takci, H., 2016. Performance evaluation of different customer segmentation approaches based on RFM and demographics analysis. Kybernetes, 45(7), pp.1129-1157.

Song, Y.Y. and Ying, L.U., 2015. Decision tree methods: applications for classification and prediction. Shanghai archives of psychiatry, 27(2), p.130.

Zaki, M., Kandeil, D., Neely, A. and McColl-Kennedy, J.R., 2016. The fallacy of the net promoter score: Customer loyalty predictive model. Cambridge Service Alliance, pp.1-25.