You are on page 1of 6

Group 8 DMBD Assignment

Question 1. What technique do you think is most suitable predicting probabilities that
customers will leave given the data at hand and why?

Objective of chosen model:


Primary: Identify the factors contributing to the non-renewal of contracts and estimate the
probability of occurrence of non-renewal of contract (Probability of churn within a given
time-period)
Secondary: Enhance customer experience & retain customers (preventing churn) without
spending too much on customer retention through price cuts/discounts
Note: objective is on targeting 100 customers in 2 months because of resource constraints
Probable techniques that can be used
To predict whether a customer will be a churner or a non-churner, a number of predictive
techniques can be used like neural networks, decision trees, logistics regression, binomial
logit model, discriminant analysis etc., with each having their own pros and cons.

We have used CHAID (Chi-square Automatic Interaction Detector) to predict the customer
churn based on its overall accuracy (94.9%), highest area under the curve (.748) and a higher
Lift of top 30% (2.045). The screenshot of the same is as given below.

Question 2. What factors contribute the most to the likelihood that customers will leave?

The top 3 drivers (based on predictor importance) that come out based on the analysis on
the given data come to be as follows:

Based on the output above QWE Inc. should concentrate upon (top 3 predictors)
● Customer Age in months,
● Customer Happiness Score 0-1
● Customer Happiness Score Month 0(Current Month)
Group 8 DMBD Assignment

Question 3. Provide the list of 100 customers with highest churn probabilities and top three
drivers of churn for each customers

Given below is the list of 100 customers with their respective probabilities.

Rule Set Used For Classification:


Group 8 DMBD Assignment
Group 8 DMBD Assignment

4. If you are using any predictive modeling technique, please explain how good model is?
For Training Data (Coincidence Matrix and Gains Chart)

For Testing Data (Coincidence Matrix and Gains Chart)

The model correctly predicts 95.04% times whether a customer will leave in case of training data.
AUC for training data is around .779
The accuracy with testing data is also good at 94.5%. Both these indicate that this is a good model.
Group 8 DMBD Assignment

AUC for testing data is around .725

Using the above result, the organization can design relevant marketing programs for those identified
set of 100 customers based on the most important factors. This saves huge amount of expenditure
without significantly affecting price cuts/ discounts.

Appendix: Given below is the decision tree for reference—


Group 8 DMBD Assignment

You might also like