PREDICTING THE CHURN IN TELECOM INDUSTRY

Prof S. Chandrasekhar B. Tech. M. Tech. ( IIT, Kanpur) Ph. D (USA)

Introduction

“Churn” is a common phenomenon that occurs in telecom Industry. By “Churn”

we mean those customers, who will be leaving us in near future. If we are able to
predict in advance, the attributes of customers whom we are going to loose in
near future one can take corrective action so that we can minimize this
phenomenon.

As we are aware that acquiring a new customer is more expensive than retaining
an existing customer. In fact it costs anywhere between 6 to 10 times more for
acquiring a new customer. Also one bad customer can in fact spoil the chances
of acquiring 8 – 10 good customers.

Initially when the industry growth is high, churn may not be a problem. The
focus at this stage is on customer acquisition. For each customer churning out
there may be many who will be joining. As the industry matures, the churn rate
rises and there will probably be a time when for every one customer leaving only
one new will be joining or there may not be even one joining.

When this stage is reached, it will affect the bottom line for the company.

So predicting the churn is a very important step for keeping the bottom line of
the company in a competitive environment.
Also in a mature market growth comes from:

Maximizing profit from existing customers (cross selling and up
selling)

Retention of customers

Stealing new customers from its competitors.
Some unique features of Telecom Industry

In many respects telecom Industry is like any other service Industry like
Financial, Insurance & utilities. But it has some peculiar features like high cost of
acquisition, not much of face – face contact, little customer mind care. Compared
to some other industries there is tremendous amount of data available as each
transaction of the customer is recorded and data is in digital Format.

The problem:

One of a well known mobile phone company operating in the country for the
past two years is interested to make investment in CRM technology. The market
is becoming more and more competitive because of deregulation and slowly the
industry will reach a maturity stage.

The maturing of the market and the increasing competition, the company wants
to focus on its existing customers, how to keep them and make them more
profitable.

What benefit such a model helps

One of the important benefits of such as model will be to marketing department
/call center(s). The list of customers who are likely to churn in the coming
month will be made available to the above places so that further action can be
taken. Data analyzed from back end or Analytical CRM System will go to front
end or operational part of CRM closing the CRM loop. Marketing department
can plan giving customers discounts, other promotions/events, other products of
other sister companies wherever applicable.

Predicting the churn also helps us to approximately know the life time value of
customers. If a group of customers have a 20% chance of churning this month,

then we would expect them to remain customers for 5 months. (1 month ÷ 20%).
If the churn were reduced to 1%, then we would expect the customers to remain
for 100 months.

The other application is for prioritizing customer segments. If a segment is more

likely to churn, perhaps, they should not get a high value gift. May be a discount
might encourage them to stay. The issue may not be clear cut, but having a
churn score will definitely help in better Decision making.

HOW DO WE GO ABOUT CHURN PREDICTION?

Data requirements for the Analysis

The basic requirements are:

Data from customer information file like age, sex, Zip code etc.

Data from service account file such as Pricing plan, activation data,
contract identification etc.

Data from billing system such as number of calls, airtime, fixed line
time, total amount spent, no. of times calls made to customer care
center, change in price plan etc.
ANALYZING THE DATA
To analyze the churn there are different techniques available. Some of the
common models used are:

Regression analysis

Decision Trees

Logistic Regression

Neural networks
Out of the above models, Decision Trees are a good choice since they provide the
Rules that Business users can understand. Other models like Neural Networks,
Logistic Regression reduce the understandability of the phenomenon even
though they may be little bit more accurate than decision trees.

Sometimes important derived variables like growth rate of the number of calls
over a period of time, calls proportions changes etc. are included. One rule of
thumb of including derived variables are those that will explain the phenomenon
of real world rather than including mere mathematical transformations.

Building a churn model:

Let us build a Churn Model for a Telecom company. Using the data provided in
the dataset Telecom.xls. Though real life example may have many more
variables and volume of data will be extremely large nevertheless this case
indicates the methodology of model development.

The variables in the data set are:

Contract id

Phone number

Age (since how long he is the customer)

Change in price plan

Air charge

MTNL charge

Total call duration

Total charges

No. of incoming calls

No. of outgoing calls

Distinct numbers called

Status (continuing/left)

The method of construction of data set form operational system is as follows:

Assume that we are building a model to predict those customers who are likely
to leave in the month of September 2002. We take the sample data set and this
set consists of those customers who are continuing in the month of September
2002 and those who have left in the month of September 2002. Those who have
left have been coded as “Churned” and those who have not left have been coded
as “continuing”, irrespective of which day of the month they have left. For each
of these customers, we go back three months in time period i.e. for August 2002,
July 2002, June 2002. Three months of data will be used to build the model. Each
month will be divided into two windows of 15 days size. So we will have six
windows which are called w1,w2,w3,w4,w5 and w6. w1 is window for June 1 –
15, w2 for June 16 – 30 etc.

Each of the above variables from operational database is extracted, summarized

for each contract number. In that sense, each contract number can be treated as
equivalent to a customer. So after summarization, there will be one record per
customer containing data for all the above fields you can see this by opening the
excel file “Telecom.xls”.
The concept is illustrated pictorially as follows:

June 02 July 02 August 02 September 02

W1 W2 W3 W4 W5 W6 Churn
Continuing

July 02
Data for October 02
Analysis August 02 September 02

For this illustration we will use status as target variable and all
others as input variables. (If you are sure that some variables are
not useful, you can leave them from the analysis).

Use FORESIGHT to build the tree. (up to the depth of 3).

The details of building the tree are given in the software manual.

Since contract _Number and phone–number are unique, they should not be
included to develop the model. First build a normal tree with default values.
Use 20% of the data to test the model.

INTERPRETATION OF RESULTS:

The training data set contains 3715 samples (80%). Balance is Test set (20%)
which will be used to validate the model. The entire training set contains 37%
churners.
The first most important variable which has highest impact on churners is
change in price plan (ch_plan). The tree has made two Groups; one group for
which change in plan is ≤ 2 (group no. 1) and for another > 2 (Group no. 2). The
first Group consists of 30% churners and 70%non-churners. While the second
group consists of 94% churners and 6% non churners. So it is important we
concentrate on customers in Group 2 and design an intervention program to
retain them.

Now let us concentrate on Group no. 2. By double clicking this group, this node
is split into Two group. One is the group where w6TOTAL CHARGES ≤ 1187
and another w6TOTAL CHARGES > 1187. 96% of the churners are in group
where w6TOTAL CHARGES are < 1187.
So the next important attribute of churners is total charges in the window
preceding the month of churn.

Now let us come to the segment where change in plan is less than 2. Drill down
this segment by double clicking. Now the segment is broken into two groups:
one Group where w5outcalls are less than <=0. 87% of the churners are in group
where w5outcalls are <0.
So we can create business rule like:
1. If change in plan <=2 and
W5outgoing calls <=0 then 87%
2. If change in plan >2 and
W6totalcharges < Rs 1187
Then the group has 96% churners
We can drill further down for details. Normally in many practical applications
up to 5 levels should be sufficient to draw meaningful conclusions.

SECOND ANALYSIS:
Let us build a model to analyze the churn in top 25% of segment.

We define top 25% of customers based on total bill amount for the month of June,
July and August 2002. to extract this data do back to the earlier excel sheet and
use autofilter option of excel to select top 25%. Save this file as telecom data top
25 percent. Use this data set for further analysis.

Using FORESIGHT open the file for analysis. Since we have segmented the data
set using total bill amount this can be excluded from the analysis, the other
field(s) that can be excluded from the analysis are Last2TOT, TOTAIR and
RATL2AIRC.
As usual target variable is STATUS. Select 20% test data. This time change the
minimum number of outcome to 50 and percent of pruning to 10. The above
parameters will have effect on the depth of tree. This ensures that sum of
samples in any parts of the tree is 50.

Results Interpretations:
For understanding of the results, let us analyze the tree to a depth of Two/Three
levels.

The first important variable which will help to predict the churn is change in
price plan. Whenever change in price plan is greater than TWO almost all the
customers in this segment have churned out. (Churn rate 100%).

The second important group is those whose change in plan is less than or equal
to Two and in Window6 (i.e. Fifteen days window prior to prediction period).
MTNL charges less than Rs. 7.20. This segment also has a churn rate of 100%.

As an exercise compare the churn Groups characterizes of all customers Vs those

in Top 25% based to total amount.

CONCLUSIONS
The above case study illustrates how an insight into the process of “Customer
Churn” can be gained using the Analytical CRM techniques. It also tells the
management what are the important variables that affect the churn. This will
help to put a churn management system in place so that early information about
those customers can be obtained.

The output of churn Management system can be sued by marketing department

to develop suitable strategies/ interventions.
 HANDS ON EXERCISES

1. Select bottom 25% of customers from the Telecom.xls files and

built a decision Tree to get an Insight into attributes of
“churners”.
2. How this “churners group” compare with those of “churners
Group” in Top 25%
3. What is the average age and average air charges of churners of
Top 25% and bottom 25% of customers a the 2nd level of
decision tree
4. The total number of calls made to the customer care center is
given in second file Customercare.xls
Merge this information to the original data file segment the data into two
groups.
I. Those who have not made any calls to the customer care center.
II. Those who have made calls
Build churn prediction model to each group separately
Discuss the results
5. What additional / derived variables would you think can be
collected to improve the accuracy of the model?