Contents Business Understanding: Introduction ......................................................................................................... 3 Business Objective ........................................................................................................................................ 3 Data Mining Objective .................................................................................................................................. 3 Data Set ......................................................................................................................................................... 3 Data Preparation ........................................................................................................................................... 3 Data Modeling............................................................................................................................................... 4 1.
Decision Tree (Binary) ....................................................................................................................... 5
2.
Decision Tree (Three-way tree) ........................................................................................................ 6
3.
Logistic Regression ............................................................................................................................ 6
4.
Logistic Regression with Transform Variables .................................................................................. 7
5.
Neural Networks ............................................................................................................................... 7
6.
Neural Networks after transform variables and variable selection .................................................. 8
Evaluation ..................................................................................................................................................... 9 Profitability of a Proactive Retention Plan.................................................................................................. 10 The key variables predicting churn: ............................................................................................................ 11 Possible Incentives Offered ........................................................................................................................ 11 Test Measures ............................................................................................................................................. 12 Profitability Matrix .................................................................................................................................. 12 Net Additions minus Existing Customers ................................................................................................ 12
1|Page
List of Figures Figure 1: Process flow Diagram..................................................................................................................... 4 Figure 2: 2-way decision tree that resulted from cell2cell data set ............................................................. 5 Figure 3: Variables in descending order of their importance helping in splits for 2-way Tree .................... 5 Figure 4: Result summary of 2-way decision tree......................................................................................... 5 Figure 5: Variables in descending order of their importance helping in splits for 3-way Tree .................... 6 Figure 6: Result summary of 3-way decision tree......................................................................................... 6 Figure 7: Result summary of Regression without transformation variables ................................................ 7 Figure 8: Result summary of Regression with transformation variables ...................................................... 7 Figure 9: Result summary of Neural Network without Variable transformation and selection ................... 8 Figure 10: Result summary of Network with Variable transformation and selection .................................. 8 Figure 11: Comparison of cumulative lift value for different techniques..................................................... 9
2|Page
Cell2Cell: The Churn Game Business Understanding: Introduction Cell2Cell is the 6th largest wireless company in the US, giving service to nearly 10 million subscribers, serving more than 210 metropolitan markets & 2900 cities (covering nearly all 50 states). The company is currently facing a major problem of customer churn. We are using SAS EM 4.3 to develop a model for predicting customer churn at Cell2Cell.
Business Objective
Reduce churn for the company Improve profitability Identifying incentives offered to the customers with high risk of churning
Data Mining Objective
To develop an accurate predictive churn model (Lift value of at least 1.75) To identify the factors that are important in driving subscribers churning
Data Set The given data set consists of 71,047 rows & containing a total of 78 variables (including a variable named “CHURN”, signifying whether the customer had left the company two months after observation). One of the variables named “CALIBRAT” was used to differentiate the validation dataset from training dataset. Training dataset contained data of 40,000 customers and validation dataset contained 31,047 customers.
Data Preparation The dataset was divided in training and validation datasets, using “CALIBRAT” as the partition variable (value of 1 was used training and value of 0 was used for validation). Variable “CHURN” was set as target variable and some other variables (those not related to business objective) were rejected.
3|Page
Variable No.
Original Variable
Changed Variable
22 26 30
Churn CSA Customer
Target Rejected Rejected
77 Calibrat 78 Churndep Table1. Showing variables that were rejected
Rejected Rejected
Data Modeling Total of 6 different models were used to predict the churn of customers. These models were:
Decision Tree (binary) Decision Tree (three way tree) Logistic Regression Logistic Regression with Transform Variables Neural Networks Neural Networks after transform variables and variable selection
SAS EM 4.3 was used to run these 6 models. Snapshot of the model is shown below.
Figure 1: Process flow Diagram
4|Page
1. Decision Tree (Binary) For both 2-way and 3-way tree gini-reduction method was used. The assessment criteria was set to be “Proportion of event in top 10%”.
Figure 2: 2-way decision tree that resulted from cell2cell data set
Figure 3: Variables in descending order of their importance helping in splits for 2-way Tree
As can be seen from above figure, EQPDAYS – Number of days of the current equipment (split at <302), MONTHS – Months in service (Split < 11 months) are most important variables that resulted in splits.
Figure 4: Result summary of 2-way decision tree
5|Page
As can be seen from figure above, with number of leaves greater than 34, no significant split happens.
2. Decision Tree (Three-way tree)
Figure 5: Variables in descending order of their importance helping in splits for 3-way Tree
The important variables are very similar to that used in 2-way decision tree. First four variables are same for 2-way and 3-way tree.
Figure 6: Result summary of 3-way decision tree
With number of leaves greater than 93, no significant split happens in 3-way tree.
3. Logistic Regression Here, no transformation of variable was carried out.
6|Page
Figure 7: Result summary of Regression without transformation variables
4. Logistic Regression with Transform Variables In this model, few variables were transformed. Details of transformation of variables are as below: 1. Variable MOU- Minutes of Usage was transformed using log transformation. The variable had high skewness earlier. 2. CUSTCARE- Mean number of customer care calls was transformed into a squared variable. As high number of customer care calls suggests high number of complaints, it can be a major cause for churn. Sqauring the variable helps increase its influence. 3. Using decision trees, the 2 major variables EQPDAYS and MONTHS were identified and transformed by creating 2 buckets. For EQPDAYS, the cut off value for bucket used is 301 days and for MONTHS, the cut off value used is 10 months. 4. Other important variables like CHANGEM and CHANGER were transformed but they didn’t improve performance.
Figure 8: Result summary of Regression with transformation variables
5. Neural Networks Here, no transformation and variable selection was done. 7|Page
Figure 9: Result summary of Neural Network without Variable transformation and selection
6. Neural Networks after transform variables and variable selection Here, the transformed variables were made to pass through variable selection process with default settings. Initially, transformed variable CUSTCARE was rejected by neural network model. But, the same was forced to be used. With this the performance of the neural network model improved.
Figure 10: Result summary of Network with Variable transformation and selection
As can be seen from figure above, the average error has come down a little bit and has less variation as compared to that of neural network model without transformation and variable selection.
8|Page
Evaluation The lift chart of all the techniques is shown below:
Figure 11: Comparison of cumulative lift value for different techniques
Legend: Tree : 2-way decision tree Tree-2 : 3-way decision tree Neural : Neural Network without transformations Neural-2 : Neural Network with transformed variables and variable selection Reg : Regression without variable transformation Reg-2: Regression with variable transformations Table below shows cumulative lift values for different techniques at different percentiles (deciles).
Percentile 2-way Tree 3-way Tree Regression 10 1.931 1.904 1.658 20 1.706 1.699 1.642 30 1.451 1.573 1.5 40 1.353 1.444 1.4 50 1.294 1.325 1.343 60 1.254 1.263 1.253 70 1.236 1.192 1.203 80 1.127 1.131 1.131 90 1.056 1.07 1.067 100 1 1 1
Neural Networks Regression with Neural with Transformation Networks Transformation 1.511 2.102 1.691 1.42 1.79 1.593 1.336 1.604 1.489 1.318 1.457 1.416 1.297 1.411 1.346 1.21 1.314 1.284 1.159 1.217 1.194 1.127 1.133 1.131 1.076 1.071 1.062 1 1 1
Table 2. Showing cumulative lift values at different deciles for different techniques As can be seen from table, the performance of “Regression with transformed variables” is best among the different techniques used. The lift value at first deciles with Regression with transformed variables is
9|Page
2.102 which is highest. It is also evident from table that “Regression with transformed variables” performs best at most deciles and significantly better till top 3 deciles. The performance of Neural Network technique is worse among the 3 techniques. Transformation and variable selection does help in improving the performance of Neural Network technique. 3-way decision tree, though underperforms when compared to 2-way decision tree at first deciles, performs better than 2-way decision tree at other deciles. So, based on the table above, the best technique is “Regression with transformed variables” which gives a lift value of 2.102 which is well above the target of 1.75.
Profitability of a Proactive Retention Plan Using regression model with transformed variables as inputs, the following values are calculated. Assumption: Subscriber in the 1st deciles is targeted. β = Base line churn rate= 1.96 % λ= Lift = 2.102 γ = Success rate = λ-1= 1.102 LVC= Lifetime value of customer C= Cost of incentive Profit = Probability of Churner* Success Rate* LVC-C = β* λ* γ* LVC –C LVC = Monthly Revenues * (1+r)/(1+r-Retention Rate) r= discount rate= 10% Total average monthly churn rate= 2% Retention Rate (annual) = (1-(.04*12))= 0.76 Average Monthly revenues per customer = 58.8528 Average LVC per customer = 58.8528*(1+.1)/(1+.1-.76) = 190.406 Thus Cell2Cell can spend a maximum of 190.406 on a customer. Total number of customers = 10,000,000 subscribers Using 1st deciles, who have highest probability of churning: 10 | P a g e
Average incremental revenue
= Number of customers contacted* β* λ* γ*(Average LVC) = (10% of 10 million)*1.96*2.102*1.102*190.406 = 864.472 million
Thus Cell2Cell can afford a maximum incentive cost of 864.472million.
Using other monthly revenues we have: Monthly revenues
Lifetime Value of Customer
Maximum Incentive Cost
30 50 70 90 110 130 150
97.05876356 161.7646059 226.4704483 291.1762907 355.8821331 420.5879754 485.2938178
440.6614469 734.4357448 1028.210043 1321.984341 1615.758639 1909.532936 2203.307234
The key variables predicting churn:
EQPDAYS: Number of days of the current equipment MONTHS: Months in service CAHNGEM: % Change in minutes of usage MOU: Mean monthly minutes of use. RECCHRGE: Mean Total recurring charge CHANGER: % Change in revenues RETCALL: Customer has made call to retention team.
Variables like months, MOU, RETCALL, and Change in Revenues are Key performance indicator of telecom businesses. It indicates longer the time a customer remains with a Cell2Cell lesser will be its churn. Also number of calls made to the retention team has direct effect on the churn rate. Variable like MOU, RETCALL, RECCHARGE are actionable. Cell2Cell can offer incentive plans to improve minutes of usage as this will result in lower churn rate. Also Retention team can play an active role in improving EQPDAYS, MONTHS and RECCHARGE through loyalty and service programs.
Possible Incentives Offered Based on above derived important variables, the following incentives plan can be offered to the customers to reduce the possible churn 11 | P a g e
From the model we got EQPDAYS as one of the primary factors for churn prediction. It makes business sense, as a customer who changes his old cell phone is likely to churn, because many mobile service providers gives new cell connection bundled with cell phone. From decision tree we can see that after 302 days customers change their handset. o So the company can come up with a plan of offering customers new cell phone (at a price slightly higher than the cost price) without changing their connection. This offered at 280 days mark. This offer will allow the company to keep hold of their customers (ensuring future revenue) without any incremental cost. From the decision tree it is visible that the probability of churn increases if EQPDAYS is more than 301 and CHANGEM value is less than -131% (i.e. decrease in minutes by 131%). Hence the company should try to increase the usage minutes by the customers having old cell phones. o This can be done by providing discount plan to these customers. Customers with CHANGEM value less than -131% along with MOU value less than 416mins and the customers with CHANGEM value more than -131% along with MOU value less than 1875mins have a very high tendency of churn. o Such customers can also be offered with discounted calling plans to increase their minutes of usage. In case of LCV customers, the company should offer little more incentives to such customers. This is because even if in the short term the company incurs more cost, but retaining such customers will increase cash flow to the company in the long run.
Test Measures Profitability Matrix Profitability in this case can be represented as shown below: Decision (from Predictive Modeling) Churn Not Churn Churn
True Positive (Profit = LCV - Incentive Cost)
False Negative
Not Churn
False Positive (Loss = Incentive Cost)
True Negative
Actual
Net Additions minus Existing Customers This is another measure that can be used to measure the success of the program. As negative net additions minus existing customers is a concern for any company in the industry in question, a positive number in mature market would suggest successful implementation of data analytics.
12 | P a g e
