You are on page 1of 18

Pre-paid Customer Churn Prediction Using SPSS

Sanket Jain
GBS Business Analytics and Optimization Center of Competence, CMS Analytics India

Date of writing: November 15 2010

ABSTRACT Given the dynamic nature of pre-paid mobile phone subscribers and the ease with
which they can stop using their phone services without giving any notice, combined with the
increasing influence of their group of close friends/family/peers, the task of managing churn has
become of prime importance to CSPs. Here, I used SPSS to predict churn. The model takes as
input data of customer demographics, social network behavior, call usage and tariff plan. The
model has been compared with FOAK (First Of A Kind) assets that make use of social
networking based highly sophisticated churn prediction models, and it yielded 95% accuracy
with C5.0 modeling technique. Reasons have been explored for the superior performance of C5.0.
Then, additional input variables that can further increase the model accuracy and explain more
its variability have been enlisted. Later, some campaigns and offers for the pre-paid subscribers
have been discussed. Offering pre-paid plan as Next Best Offer to those who are currently using
post-paid connections but of late have been exhibiting behavior that more fits pre-paid plan
package - came out as one of the future research areas. Another conclusion is to offer such plan
packages as flex-pay package that can tie post-paid kind of services to relevant pre-paid
customers so as to increase their stickiness with their service provider. This paper is dedicated to
TM Forum Organization.

Keywords ARPU; MOU; Recharge; C5.0; MNP; Loyalty; Next Best Offer.

Mobile phones are now fast becoming a commodity. Most cellular circles in growing countries
like India now have as many as 5-7 offerings such as GSM and CDMA being provided by service
providers. In such a competitive market, it becomes very easy to switch your cellular service
provider merely at the drop of a hat. Rapid advancement in next-generation services has drawn
the attention of teenagers and rural subscribers by features like free SMS. It has also enabled the
service providers with tools to attract working professional by offering them free STD minutes,
bundling and VAS. Tumbling ARPU figures, an ever increasing MOU (Minutes of Usage),
together with the fact that ARPUs rate of decline has outpaced the MOUs rate of increase, there
is an increasing pressure on service providers to maintain their margins. (ARPU is declining
because more and more subscribers are signing up and a constant decline in rate plans and an
upsurge in attractive tariff plans.) Soon, such a commoditization will cause the market price of
mobile telephone services to fall to the marginal cost of lowest-cost volume producer.

This is causing churn. The problem of churn originated from European countries where the
matured markets gave an incentive to the operators to try to attract customers from competitors.
From a global perspective, churn of mobile operators led them to loss of ~$100 billon USD per
year (Berson et al., 2000). It has assumed alarming proportions in growth markets like India as
well. According to Gartner research, Indias churn rate is a high 3.5 6% a month, aggregating to
~40-50% every year. This fact combined with the high installation and marketing costs, makes it
5-10 times more expensive to acquire a new customer than to retain an existing one (Ruta et al.,
2006). All this shows that the churn of an existing customer, especially in pre-paid category,
could hit the bottom-line of these operators. Also, today, most of Chinese and South African
service providers are selling handsets along with the tariff plans, which means that handset in
itself is giving the customers an incentive to churn.

These factors make the task of predicting churn as one of high priority for service providers. To
stay competitive in this market, they must be able to correctly predict risky subscribers on whom
the subsequent retention efforts should be focused. On top of that, MNP (Mobile Number
Portability) is looming as a threat to service providers. MNP can provide flexibility to subscriber
by letting him change the service provider at will. It can also fuel a hard-nosed battle between
providers. The only flip side to MNP for the customer is that the operators will charge
maintenance and monthly fees from them, and the time taken to port the number from one
provider to another will cause some inconvenience to the customer. So, MNP will cause churn.

Unlike post-paid subscribers, pre-paid subscribers can annul their service without giving any
prior indication because they are not bound by any contract. They might churn if their current
needs charge or if they get influenced by their social network of family and friends. This dynamic
situation makes the task of predicting the likelihood (and timing) of churn very important in the
context of pre-paid segment. So, in order to survive competition, telecommunications service
providers must detect the main reasons for both the expected churn and the churn that happens
after the event has taken place in pre-paid category because this information can help them to
customize their offers. It can be a tool to effectively anticipate the demands of their key
customers who have the highest churn propensity, fully knowing that retention can have a huge
impact on LTV.

The churn propensity is calculated on a contract basis rather than a per customer basis
CION.pdf). The main reason for this is that many important predictor variables such as LOS
(Length Of Service) are associated with contracts rather than customers. Also, if a subscriber has
multiple contracts, then each of these contracts contributes to revenue for the service provider. On
the other hand, the mailings and other follow-up actions target customers, not contracts, thereby
shifting the focus of analysis at customer-level rather than at contract-level. To resolve this
problem, some post-analysis processing must be carried out to summarize the predictions for
customers as individuals. We are currently trying to find a solution to this.

The customer attributes typically considered in a churn analysis can be categorized into:
Demographics (age, socio-professional category),
Contractual data (subscriptions and cancellation of services),
Quality data (number of dropped calls, number of complaints related calls),
Billing data (recharge amount, amounts withdrawn for the services, current tariff plan,
net present value of subscriber), and
Usage data (total numbers of calls, LOS, percentage share of Local/STD/International
calls, peak call consumption and average consumption).

An enterprise can increase its profits by 25-95% by reducing its churn by just 5% (Reichheld
et al., 1990), which shows the impact of doing analytics. In order to reduce the losses caused by
churn, operators have to find the most valuable customers who are inclined to churn, and then
carry out retention policies for them.

Here is a formula to correlate churn problem with the ultimate goal of achieving loyalty:
Higher the churn, lower the chances of being loyal. So, Churn% ~= 100% Loyalty%.

Now that we know the reasons for churn, lets look at the typical category of customers who
prefer pre-paid phones to post-paid phones in the first place. Some of these users are:
Price-sensitive user (especially in countries like India),
Those who know that their usage is going to be too low to justify investing in a high
MOU post-paid plan,
People who want to stay anonymous,
People with poor credit history, and
Customers who want to try different networks.

It is very important to know the business goal along with the data mining goal. For e.g., the
business goal could be to reduce churn rate by 15% in next 6 months. Whereas, the data mining
goal could be to achieve ~95% accuracy in prediction with a Lift of > 2 being captured in top 20
percentile users (where accuracy is defined as the ratio of predicted churn to actual churn). Also,
it is the proportion of correct churn predictions, not the number of absolute correct predictions,
which should be more important for the business to analyze. According to Burez et al. (2009),
Lift is defined as ratio of precision to overall churn rate. According to
, only a small fraction of the subscriber base can be contacted at any given time, and the
subscribers with the highest churn scores are assigned top priority. So, a churn prediction system
should be measured by its ability to identify churners within its top predictions. Performance is
measured using lift. For any given fraction 0 < T < 1, lift is the ratio of the number of churners
among the fraction of T subscribers that are ranked highest by the proposed system, to the
expected number of churners in a random sample from the general subscribers pool of equal size.
For e.g., Lift of 5 at a fraction T = 0.01 means that if we contact the 1% of subscribers ranked
highest by the proposed system, we expect to see five times more people who planned to churn in
this population than in a 0.01 fraction random sample of the population. The performance of a
churn prediction system is completely characterized by its derived lift curve, which maps each
fraction 0 < T < 1 (horizontal axis) to the lift (vertical axis) that is obtained by the system. In
general, the lift curve is monotonically decreasing, since it is usually harder to provide a
substantial lift for larger fractions.

A Call Detail Record (CDR) is the computer record produced by a telephone exchange containing
details of a call that passed through it ( It is the
automated equivalent of the paper toll tickets that were written and timed by operators for long
distance calls in a manual telephone exchange.

There were 31,769 records of pre-paid customers that were available, out of which the Partition
node in SPSS did the split (a 70:30 split ratio was chosen), thereby causing 22,179 records in
training data set and 9,590 records in validation data set.

In a typical scenario, churn occurs only 2-3% of the times. This causes imbalanced data. This
problem is crucial in churn prediction because we need to maximize the instances of recognizing
the minority class. Two methods for dealing with class imbalances are: oversampling and
downsizing. Oversampling consists of re-sampling the small class at random until it contains as
many examples as the other class. Downsizing consists of the randomly removed samples from
the majority class population until the minority class becomes some specific percentage of the
majority class. Using one of these two techniques, the churn population was made 46.3%.


Existing churn prediction methods like decision tree (DT) classify customers as churners or non-
churners while ignoring the timing of churn event. DT builds interpretable models that show the
patterns discovered. It is more on classifying the customers into two groups. Hence, it can take
into account both churn and non-churn; whereas ANN and regression have been trained to make
calculations to decide if customers are churners. Accuracy depends mainly on the weights for the
neural networks and coefficients for the regression. Also, the churn prediction models like
decision trees, logistic regression, etc. can give the probability to churn for a customer. Because
of such fierce competition in the pre-paid phone market over the last few years, we assume that
higher the churn probability score, sooner he will churn (typically, no activity in past 20-30 days).
This assumption simplifies our task of building accurate prediction models because it obviates the
need to build survival analysis models that take into account the customer life cycle. Also, this
definition removes those cases that are involuntarily churned by the CSP.

3.1 Defining the framework for churn prediction model
The general framework of the churn predicting model is shown in Figure 1. It first samples
subscribers for training of the predictor and then pre-processes it for missing data, outliers, etc.
After Feature Selection (FS), data can be used for the predictor training and testing (not for
validation). Models based on expertise first conduct FS before other preprocessing methods.
During FS, the variables that got excluded were for national call cost and international call

Social networks have also been addressed by IBM Research. They try to answer whether the
decision of a subscriber to churn is dependent on existing members of community with whom he
is related. SNAzzy shows that diffusion models built on call graphs have superior performance to
baseline model. TABI, another asset of IBM, finds Groups and subsequently Group Leaders.
However, I feel that, although their approach and results might be very good, we should try to
build our own data model that is separate from both TABI and SNAzzy, and we must consider
including data from handset, account balance and inactivity.

The data used for this study was CDR, Social Network, Tariff and Customer Information. It was
only synthetic data. We are still trying to get some real data. The time horizon selected was six
months. The data used was having six months of information. Total of twenty five variables were
given. The data had already been cleansed for missing values, outliers, etc. The churn population
had also been boosted from 2-3% to 46.3% so as to allow for meaningful prediction of churn.

The data given social network, CDR, tariff and customer profile was read into SPSS and then
merged by customer id. It was then clustered using Auto Cluster, which chose Two Step
clustering method to other clustering methods (see Figure 1 below).
Here, I have applied a new technology (SPSS Modeler) for doing churn prediction modeling.
Hence, these significant predictor variables will constitute a part of the Information Framework
(SID) of TM Forum.

Figure 1. Auto Cluster Output Summary

Chief reasons for using Two Step Cluster analysis are:
File size is big enough (31,769 customers in the CDR data) to prefer Two Step to either
hierarchical or k-means.
If you have both continuous and nominal predictors, only two step method will work.
k-means being based on Euclidean distance would suffer because it depends on the units
of measurement for the variables used. The outliers will be selected as the initial clusters,
resulting in outliers forming clusters with few cases. Hence, it is required to remove the
outliers before doing k-means clustering.

However, I chose k-means method because it is simple to interpret and one can control the
number of clusters that he desires. Figure 2 below shows the k-Means method summary. Cluster
quality was Fair using Silhouette measure. The outcome shown below using 4 clusters and 14
variables was the best possible outcome after several trials.

Figure 2: k-means Clustering Output

PCA/Factor analysis was also done using this data. Using PCA, it was found that five
components explained only ~59% of the variability in the model. This suggests that the
representation done using PCA was not satisfactory. However, I have used PCA analysis only for
illustration purpose. PCA should not be used to assess multicollinearity. Instead, we should
follow a correlation matrix and then compare the VIF (Variance Inflation Factor) values.

This data was then partitioned (70:30), which was then used to create logistic regression and other
classification models. LR (Logistic Regression) model yields continuous probabilities that are
discriminated into churners vs. non-churners by using certain threshold value (0.5).

Models that you build (train) must be assessed with separate testing data that was not used to
create the model. The training and testing data should be created randomly from the original data
file. They can be created with either a Derive or Sample node, but the Partition node allows
greater flexibility. With Partition node, SPSS Modeler can directly create a field that can split
records between training, testing (and validation) data files. Partition nodes generate a partition
field that splits the data into separate subsets or samples for the training and testing stages of
model building. When using all three subsets, the model is built with the training data, refined
with the testing data, and then tested with the validation data. The Partition node creates a
categorical field with the role automatically set to Partition. The set field will either have two
values (corresponding to the training and testing files), or three values (training, testing, and

Results from LR (Logistic Regression) model:

Pseudo R-Square Table
Cox and Snell 0.509
Nagelkerke 0.68
McFadden 0.515

Table 1a. A list of Pseudo R-square values

Before inputting the variables into the regression and classification models, following variabes
were removed for obvious reasons: Customer_id, pay method, churn, connect_date and
record_count. Backward elimination was chosen for this exercise because churn is a binomial
target variable. Log likelihood function was chosen as the goodness of fit test because it is
superior to Walds test. Using this, following variables were removed:
a. Variable(s) removed on step 2: LeaderNonCarrierCalls.
b. Variable(s) removed on step 3: AveNonCarrierCalls.
c. Variable(s) removed on step 4: Min_R.
d. Variable(s) removed on step 5: Ratio_In_Network_Calls.
e. Variable(s) removed on step 6: %LeaderOutGroupCalls.
f. Variable(s) removed on step 7: %AveINGroupCalls.

Using forward selection, following variables were included.
a. Variable(s) entered on step 1: Handset.
b. Variable(s) entered on step 2: LeaderVoiceCalls.
c. Variable(s) entered on step 3: tariff.
d. Variable(s) entered on step 4: Dropped_Calls.
e. Variable(s) entered on step 5: Gender.
g. Variable(s) entered on step 7: OffPeak_mins_Mean.
h. Variable(s) entered on step 8: Peak_mins_Mean.
j. Variable(s) entered on step 10: Weekend_calls_Sum.
k. Variable(s) entered on step 11: Age.
l. Variable(s) entered on step 12: L_O_S.
m. Variable(s) entered on step 13: Min_R (this variable was removed later).
n. Variable(s) entered on step 14: SocialGroupSize.
o. Variable(s) entered on step 15: Peak_calls_Mean.
p. Variable(s) entered on step 16: Ratio1 (all)_transformed.
q. Variable(s) entered on step 17: #In net (all)_transformed.
r. Variable(s) entered on step 18: OffPeak_calls_Sum.
s. Variable(s) entered on step 19: Max_R.
*t. Variable(s) entered on step 21: Ratio1_transformed.
u. Variable(s) entered on step 22: Weekend_mins_Sum.
v. Variable(s) entered on step 23: SocialGroupRank.

*Note: Step 20 was removed manually because it was later dropped in Feature Selection.

Finally, logistic regression model was running with the Enter option instead of Forward or
Backward. Surprisingly, Max_R was found to be a significant variable, whereas Min_R was
not. (Note that TABI algorithm proves that Min_R is a significant variable, whereas Max_R is
not.). This could be further explored.

Modeling Technique Accuracy Lift Percentile
Logistic Regression 86.56% 2.02
C5.0 94.65% 2.05
QUEST 84% 2.04
CHAID 88% 2.04
C&RT 88.70% 2
Analysis (DA)
80.60% 1.55 61
Neural Network - -

Bayes - -

Table 1b. A Comparison of models using Lift and Accuracy

Note: The accuracy figures above represent the True Positive rate from the confusion matrix.

From Table 1b above, we can conclude that C5.0 should be the preferred model to Logistic
Regression. However, CHAID works the best if Percentile is also taken into consideration. It
could be because C5.0 incorporates variable misclassification costs. It allows a separate cost to be
defined for each predicted/actual class pair. C5.0 then constructs classifiers to minimize expected
misclassification costs rather than error rates. The cases themselves may also be of unequal
importance. In pre-paid churn case, the importance of each case may vary with the value
associated with the subscriber. C5.0 has provision for a case weight attribute that quantifies the
importance of each case; if this appears, C5.0 attempts to minimize the weighted predictive error
rate. It can automatically winnow the attributes before a classifier is constructed, discarding those
that appear to be only marginally relevant. For high-dimensional applications, winnowing can
lead to smaller classifiers and higher predictive accuracy, and can even reduce the time required
to generate rule sets.

Note that the model accuracy of C5.0 model increased by 1% only when the variable
Dropped_Calls is taken as a Weight Field.

The effect of derived variables on the accuracy of the model was also studied. From the data
provided, only two derived variables could be formed: average peak call duration, and average
off-peak call duration. Incorporation of these raised the accuracy of C5.0 model from 94% to
94.2%. However, the need to check for multi-collinearity was obviated because PCA gave 5
components with 62.3% variability in scenario B and 59.3% in scenario A (A = After, B =
Before). But, nearly all models including CHAID did not give favorable result when the model
was evaluated on Profit (see Figure 3 below), indicating that the model may not be a good
candidate for measuring profit.

Figure 3. Profit Output of Training vs. Test using CHAID

Using Web Graph (see Figure 4), the churn variable was analyzed along with gender, handset and
tariff. It turned out that Handset model ASAD90 and CAS30 are significant in predicting churn
when compared to their ability to predict Active, i.e., non-churners. Hence, we can conclude that
handset, and hence price, is the leading reason for churn to happen. Also, we can test the
following hypothesis: Are the owners of these handsets "ASAD90 and CAS30 relatively more
down-market in their economic or demographic profile?

Figure 4. Web Graph Output (circle layout)

Using CHAID, Leader voice calls, handset and LOS variables are used as classifiers to split the
data set into >=2 partitions.

Before launching the campaign, we should verify the accuracy of the model by
comparing the predictions with actual churn instances of the most recent month. Doing
this will help us see the results of the model on LIVE Test (or Unseen) data.
Sometimes, it could become somewhat difficult to convince the management to think of a
solution beyond regression. If it is ok to go ahead with C5.0 instead of Logistic
Regression, then higher predictive accuracy can be attained. This is because of its
capabilities mentioned above: Boosting, Differential misclassification costs, Winnowing,
Support for cross-validation trials and sampling, and Case weights.
It needs to be known whether a lift of ~2 would be acceptable or not.
There are a lot of misconceptions about the use of R-square. Some think that it needs to
be >0.95. However, this may not always be wise to reject a model whose R-squared value
is less than 0.95. We should not ignore pseudo R-squared value of 0.68, because it could
still help in revealing the most significant predictors.

After getting all the input variables that can improve the predictive capability of our model, the
next step would be deploy effective retention related activities that can be used as a strategic
roadmap for the telecom service providers. Here are a few retention related ideas:
1. Consider a scenario where the users International call time for a month is 60 minutes
and his IDD bill type is not premium. This would tantamount to rather high churn
likelihood. To fine tune your retention efforts, it may be worth calling him and suggesting
a change of billing type to one that would suit his high level of international calling.
2. Build a framework for capturing his most recent requirements. In this regard, the telecom
providers can attempt to create an innovative solution that predicts the customers who are
most likely to switch from post-paid to pre-paid connection if they have been exhibiting
pre-paid behavior of late. Later, they can offer a tailored pre-paid plan as next best offer
for each of such subscribers. A focus on pre-paid plan can pay off the operators for
following reasons:
It has larger segment of subscriber base and favored by youngsters. This natural
segmentation makes targeted marketing campaigns easier and more effective.
Pre-paid market revenue has been enjoying good growth over the past few years.
There are no risks of late payment or bad debt.
Rate changes can be applied instantly by the operator, thereby affording the operators
to be flexible and more reactive to competitive conditions.

3. Set up a model to intelligently detect the latest competitor offerings in real time.
4. If he has been found to be high risk, and if he is valuable with a business professional
profile, currently traveling to a different location from his home location, then offer him
five free SMS as soon as he reaches the away location.
5. Upon his completion of say, 6 months of service, offer the following: (a) Points for
outgoing calls, file transfer, SMS, etc. (b) Tiered points for high monthly recharge
amount, and points for every top-up. (c) Tiered points for 1/2/3 years of sign up. (d)
Special incentives for choosing lifetime service. (e) More points for referring a friend
who then becomes your subscriber.
6. Self-serviced kiosks for providing convenient e-service as a differentiator.
7. Accept credit cards as mode of payment.
8. Develop a contact strategy and promote a product like service upgrade during the call.
9. Engage them in communication and thank them for their service every month. Set thank
you messages on IVR to thank them upon every monthly recharge!
10. Get involved in content partnerships to offer VAS to key users. For e.g., while offering
him SMS service, offer account balance inquiry, cheque book request, balance transfer
capability among authorized accounts only all using FREE SMS.
11. New products and services development such as VAS are now necessary to retain him.
Voice mail and Alert features can be bundled and offered to those who cant take a call
because (a) their phone is switched off, or (b) they may be travelling outside their
networks, or (c) to those who need to know the folks who tried to reach him while they
were busy. Another example of VAS could be an offering that sends an SMS to the user
with a list of all the missed calls in case his phone was switched off.
12. Factor the following in your pre-paid networks structure: International roaming, IN, etc.
13. Offer him free SMS for next month if he tops up by a preset amount this month.
14. Operator branded self-service kiosks with data, voice, wireless, etc.
15. If both the scope and the infrastructure allow, then, to deal with the churn threat of MNP
in India, you can offer a family pack with bundled services (pre-paid mobile telephone
service with Broadband) to the user.
16. Educate a new prospect of your unique selling propositions to avoid him from accepting
another offer from a competitor.
17. Companies like Videocon in India have started offering cable service along with their
TVs. So, companies like IDEA Cellular should offer bundled service of cable service, TV
and mobile phone so that they can offer a combo-package to its most valuable customers
so that it can gain a strategic position in this ever-competitive environment.
18. There are many people especially in India where a young working professional settles
down at his home city after working for some duration in cosmopolitan cities like
Bangalore. So, if we can find the average duration (say, x days) for which he goes to a
big city for working (e.g., Bangalore), where he has been using the pre-paid service from
a provider P, and let us say that he returns to his hometown H and spends at least a
month there (say, y days, where y is greater than thirty days). Then, we can predict
both x and y. This is because if he has already spent y days at H, then it is quite
likely that he would continue living there for a longer time. This predictive information
can have good value to the provider P by giving him an offer that is more relevant to
the location H, assuming that H falls under one of its circles. Another takeaway from
this is that the provider P can focus on what areas it needs a large presence. This can be
done if they come to know the top 3 cities where the customers leave Bangalore to settle
down at their home.

Here are the additional variables required to increase our models predictive power:
1. Migration data (from say, Active status to Grace period, Grace to Active, Grace
to Churn, etc. Hypothesis: If there is a pattern of such migration, then it can be a
predictor of future behavior.
2. Number of days between inactivity and recharge. Hypothesis: If there is a pattern of say,
10 days of inactivity before which he generally recharges, and if he has been inactive
recently for more than 10 days, then we can use it as a churn indicator.
3. Remaining balance: Hypothesis: Lower the remaining account balance, higher his churn
likelihood. Also, it could be important to analyze the ratio of balance to top-up. In other
words, how much of his balance goes unutilized when his validity period gets expired.
Another pattern worth exploring could be to explore how many customers opt for
recharge just one or two days within the expiration of their validity period.
4. Location data. Hypothesis: Voluntary churn could be higher in Kolkata than in
Bangalore because of higher social connectivity of customers in Kolkata.
5. Assigning weights to a service provider. For e.g. Sprint may be having more coverage in
USA than T-Mobile. So, Sprint should be assigned more weight than T-Mobile.
6. A high standard deviation of calls received from customers using other GSM (or CDMA,
depending on which country you are based out of) providers can impact high churn.
Likewise, another hypothesis could be: Low standard deviation of incoming calls from
other providers (away-net providers) can impact high churn.
7. Roaming data.
8. If he has made several complaints throughout the year, we can create a new variable that
records his frequency of raising complains. Hypothesis: Someone with an average of six
complaints every month may differ from a customer with the same number of complaints
but that were either made in a very short duration of time (less than 2 weeks) or those
that were staggered across a very long timeframe (12 months).
9. Watching churn events more closely. Hypothesis: If a negative event is followed by
another negative event (e.g., filing a complaint twice), it can lead to churn.
10. Brand data. Hypothesis: Brand can impact a customers decision of churning.
11. If he has a history of high billing mainly due to STD calls from Airtel and later it is found
that he is calling a number from Airtel itself but in another location, then offer him an
Airtel calling pack. OR offer him a tailored tariff plan.
12. Handset cost. Hypothesis: Low handset rates can increase churn, especially in countries
like China and South Africa.
13. National (NAT) Call Costs and International (INT) Call Costs data. Using this data, it
would be worthwhile to look at the ratio of NAT calls to INT calls.
14. Billing dispute data. Hypothesis: High number of billing disputes and billing fraud is one
of the chief reasons for churn to happen.
15. Blocked calls data (blockage due to demand for that network area exceeding the supply
capacity). Hypothesis: Other than dropped calls, blocked calls can also be a reason for
customer leaving your service.
16. We have got 6 months of data here. We can split the data into two halves of three months
each and then monitor the % relative decrease in top-up frequency, top-up amount,
incoming calls, outgoing calls, etc. Hypothesis: It is enough to use 3 months to do a
robust and reliable churn prediction because of the dynamic nature of this business. So, if
there is any non-sequential pattern in the 6-month data, then it could be discovered and
thereby help in making a more accurate churn prediction.
17. Data for virtual churners (incidental churn). There are many subscribers who are no
longer active simply because they have changed their phone number but neither their
tariff plan nor their service provider. This phenomenon of virtual churn can lead to a loss
of data of their previous phone number. This data loss can affect out analysis too. To deal
with this situation, we can suggest the operators to keep an active database of Nationality
and Date of Birth for each customer. Doing this can almost always guarantee the chances
of locating a unique customer (by looking for his full name, nationality and date of birth).
The introduction of NID (National Identification) can also help in avoiding this data loss.
18. The number of unique individuals called by your subscriber.
19. The number of people who churned and also featured in prior months top 10 frequently
called list of your subscriber.
20. The number of people from your customers close contact list (the top 10 list) who, 3-4
months ago, were not using your service (say you are Cingular), but who churned from
their service provider (say AT&T) and then joined your service. Hypothesis: From these
Top 10 contacts list, greater the number of people who that subscriber is able to
influence and bring them to you as your customer, higher the chances that he would be
an influencer for you, and hence, more attention should be paid to retain him. In fact,
a special weightage can also be assigned to this kind of variable. We can consider only
those people in his contact list who have at least one instance of two-way call with our
customer. A similar variable to be considered could be Top 10 contacts with highest
21. Separate churn prediction models for 1-month to be and 3-month to be churners
should be made. This is because there could be cases with only 0.4 churn propensity
score for the next month, but with a high 0.8 score for churning after three months.

Once these variables have been obtained, they need to be prioritized such that only the top 5-10
that would be retained as the key predictors.

Here is an analytical approach that can act as a blueprint for telecom providers:
How to treat those who will most likely not churn soon: A customers churn score
generally shows a gradual increase over the time. Lets say that at one point of time,
it will cross 0.5 score. If there is a sizeable proportion of customers, say, 70%, who
had churned three months after they had attained a 0.5 score, then we can conclude
with 70% confidence that my customer will not churn in next ninety days if he has
got a score of say, 0.1. Such an approach can help the telecom operator by telling
them the customers who are currently not at any significant risk of churning.
How to treat those have an in between churn score, i.e., who will neither churn
soon nor remain for a long time (with say, a score of 0.2 to 0.7): Explore a 2-way
prediction approach where subscriber base is first analyzed for non-churners using
one model and then analyzed for churners using the other model. This would lead to
certain customers appearing in both the churn and non-churn groups, therefore a third
group of fuzzy customers would be created. An analysis of this idea can lead to
make a decision of whether these customers should belong to the non-churner group,
the churner group or remain in their own group as customers requiring no immediate
How to treat those who will most likely churn soon (as indicated by a high SPSS
based churn score): Offer special incentives to those who meet all these conditions:
(1) they made at least one outgoing activity in the last 3 days, (2) their first activity
date after taking connection was at least 90 days before, and (3) they have been
recharging their balance every month since the day they got activated.

For prepaid users, there must be some users who have reasonably large validity period of say, six
months but consume all their minutes in a very short span of time, say two weeks. It would be
interesting to know the proportion of such users. Then, we can possibly design some special
offers for them. These customers would either be very rich or very irrational. They would be
irrational if they are young students with low income, if any. Whereas rational users (i.e., the rich
ones) would be those who have their income greater than the average income of that particular
country and/or the standard deviation of their income is close to zero (standard deviation is taken
to take care of the fact that they are not new to earning). We can call these categories of users as
"A1" and "A2" respectively. On the contrary, if they consistently miss their validity period and
consistently stretch their usage to grace period, then they can be treated opposite to those in
categories A1 and A2. We can call this category of users as "B. All other "in between" customers
would fall in category "C". Doing this will form a unique and novel way of customer

Next, we can explore hypothesis such as the one below:
These early-out customers belonging to categories A1 and A2 could be significantly more
susceptible to churn as compared to the late-out customers who belong to category B. In
addition to churn, this type of analysis could have some impact on network capacity usage
contribution from the pre-paid users. Also, if there is a sizeable proportion of such users (say
20%) who constrain the network in those couple of weeks, then it is worth sending them offers so
that they can spread their usage across six months. Offers could be designed such as six months
free talk time if they spread their consumption of their current minutes to at least three months.

Once it is known that these customers are pre-paid - we can do this in two ways: (1) their phone
number will indicate whether it is a post-paid connection or pre-paid, and (2) The IBMs master
data management would contain the information whether a customer is carrying a post-paid
connection or pre-paid we can look at certain descriptive, preventive and predictive metrics for
checking network capacity usage contribution from the pre-paid users. Metrics like bandwidth
traffic and bandwidth utilization are very seasoned metrics, though you will have to further
differentiate by wire-line and wire-less technologies; and then within wireless by 2G, EDGE, 3G,
4G for the current technologies. However, both these metrics - traffic and utilization - are "after-
the-fact" metrics. Real value-add would come from finding "preventive" metrics like bandwidth
optimization, bandwidth balancing, bandwidth switching, etc. Also, some predictive metrics
could be explored like bandwidth capacity planning, bandwidth disaster recovery, optimal point
at which to switch from 2G to 3G and vice versa, optimal point at which to switch from 3G to 4G
and vice versa, and point at which to switch from EDGE (2.75G) to 3G and vice versa.

Since churn prediction and prevention are linked to the companys retention activities, so this
entire exercise falls under the Loyalty and Retention part of the TM Forums eTOM (a business
process framework). The website below demonstrates the role of Loyalty and Retention in the life
cycle of the business process framework of TM Forum. Note that this typically starts with your
customer calling the call center representative to complain about poor quality of service, dropped
calls, incorrect bills, etc.

We can work out some focused and creative promotions with targeted offers that could help the
service providers to move from simply doing churn management to the loyalty game. Some of
these promotions could be launched as follows:

Attracting Early Adopters
Social media awareness campaign: Provision to upload your photos directly to your
Facebook account, and also grant free unlimited access to Facebook, Flickr, etc.
Nex-Gen campaign: Offer strong brand device by offering next gen services such as e-
browsing, GPS navigation, email services and 3G/3.5G/4G services (GPRS) like video,
broadband, MMS, data download.
Smart campaign: When your subscriber reaches the threshold for a bonus, you can
automatically provision that reward for the subscriber and then notify him in real-time of
this reward
paid_campaign_lifecycle_toolkit.pdf). Also, send him Thank you SMS every time he
does a topup beyond a certain minimum amount say 50 INR.
Higher away-network usage campaign: The higher score on "away-net usage" signifies
that he has called very frequently to other networks (
Industry&id=2598262). A targeted campaign can be performed with the price plan
beneficial to call other networks. A further analysis of the called away-net numbers can
result in identifying frequently called off-net numbers which can be targeted by
campaigns as a candidate of acquisition.

Retention Offers
Short talk campaign: If your customer typically makes a call of 2-5 minutes duration
only, then send him with an automated SMS with 5 additional minutes of top-up.
First five minutes campaign: If the customer is making calls that do not even last for 5
minutes, then encourage him to talk more by offering him to pay only for 1
5 minutes,
then remaining minutes upto 30 minutes would be free of charge.
Night users campaign: Pay less for evenings if you predominantly call during nights.
Student campaign: For student segment, who like to play with iPOD and other gadgets,
offer free iPod devices to those who have been shopping for at least 500 INR every
month since the last 6 months.

Cross-sell Offers
FlexPay offer campaign: monthly fee of 50 INR for 200 minutes of usage, and then
topup if need arises for filling up a few days of outage (http://www.pre-paid- You can also give post-paid services like free
voice minutes to qualified pre-paid customers. This could increase their stickiness with
you because post-paid customers are typically more loyal than pre-paid users.
Bundle your services to build a large subscriber base by offering pre-paid services
combined with other products/services that you may have, combined with free caller
tunes for a month, free voice mail, and free traffic alerts. For instance, according to
strategies-and-minimising-churn, Virgin Mobile UK managed to build a large subscriber
base by offering pre-paid services combined with other products/services from the Virgin
Group (i.e. airline-ticket contests, DVDs, etc.).

Up-sell Offer
Customer experience campaign: On March 8 (Womens day for women only), send a
promotion via SMS (or email) for recharge of > 30 INR (Indian Rupees) this week before
March 15 and you will be eligible for lucky draw where 1000 free SMS or 1000 free
outgoing calls will be offered to 1000 lucky winners. Then, on March 15, activate the
trigger, and send SMS saying: If you are eligible, then reply to this text with your name
and phone number. A reminder SMS needs to be sent and then the user will enter the
lucky draw (random draw). If he qualifies as one of the 1000 winners, then call center
will send him SMS on April 1 saying Congratulations! 1000 free SMS have got
activated for this month for you!

In this paper a customer churn analysis was presented for pre-paid mobile phones. The analysis
focused on churn prediction based on logistic regression and other classification techniques using
SPSS Modeler (formerly Clementine). The different models predicted the actual churners with
greater than 85% accuracy, which was quite good. A low of 68% model variability being
explained by the chosen predictors in logistic regression could be due to the dynamic nature of
the churning customer profile. We can do better by considering more variables. The findings of
this study indicate that the user should update the logistic regression model to be able to produce
predictions with high accuracy. Also, the C5.0 model can be the model of choice because it
proved to be more accurate than any other model, including logistic regression. The effect of
derived variables on the accuracy of the model was also studied. From the data provided, only
two derived variables could be formed: average peak call duration, and average off-peak call
duration. Incorporation of these further raised the accuracy of the C5.0 model. We could further
explore derived variables of Outgoing call duration as: Avg_Outgoing_Calls_
Greater_Than_Five_Min, and Standard_Deviation_Outgoing_Calls_Greater_Than_Five_Min
(where standard deviation is measured month over month variability in usage behavior).

CSPs can attempt to create an innovative solution that predicts the customers who are most likely
to switch from post-paid to pre-paid connection if they have been exhibiting pre-paid behavior of
late. Later, they can offer a tailored pre-paid plan as next best offer for each of such subscribers.

If the need be, then we can also calculate the aggregated proportion of subscribers who are about
to churn. Although individual churn score would be the most important to look at, yet the CSPs
can get a sense of the extent to which churn is hurting them if they look at aggregate proportion
of those who are most likely to defect. Other variables worth exploring would be:
Is he with or without discount package? (Discount package is one of the top reasons to
churn, because it is closely linked to price, which is actually the biggest reason for
customers to churn. In some countries, as much as 40-50% people churn primarily
because of price. Other significant reasons for churn are extent of coverage of a
providers service area, Quality of Service, the kind of advertisement carried out by it to
attract prospective customers, whether he is carrying only a pre-paid connection, and the
customers curiosity to switch to competitor.)
Average number (and duration) of outgoing calls made to GSM operators.
Does he have a second number?
Standard deviation (and average) of calls more than 5 minutes.
Average of maximum (and minimum) calls.

However, there is one caveat while using any modeling tool such as SPSS Modeler. If the
customer has been showing an increasing trend of his phones usage, then our SPSS churn
prediction model would tend to suggest that the CSP should focus its efforts on sending him a
higher rate plan that can eventually generate potentially higher revenues. However, he may have
increased his recent usage only because of a battery problem that resulted in frequent email
synchronization. So, at times, the churn prediction itself could be misleading. If we can detect
such patterns in data that can give meaningful predictions, then it could help enhance the value of
next offer for him. This is akin to the false alert, and needs to be studied carefully.

A. Berson, S. Smith, and K. Thearling, Building data mining applications for CRM, New York:
McGraw-Hill (2000).
F. F. Reichheld, W. E. Jr. Sasser, Zero defections: Quality comes to services, Harvard Business
Review, Vol.68, 1990, pp. 105-111 (1990).
Behara, R. S., Fisher, W. W. and Lemmink, J. Modelling and Evaluating Service Quality
Measurement Using Neural Networks. International journal of operations and production
management, 22, 10, 1162-1185 (2002).
Sherali, D., Hanif, Hobeika, G., Antoine and Jeenanunta, Chawalit. An Optimal Constrained
Pruning Strategy for Decision Trees (2007).
Hadden, J., Tiwari, A., Roy, R., Ruta D.: Churn prediction: Does technology matter (2006).
Hong X., Zigang, Z., Yishi, Zhang. Churn Prediction in Telecom Using a Hybrid Two-phase
Feature Selection Method. Third International Symposium on Intelligent Information Technology
Application (2009).
Ruta, D., Nauck, D., Azvine, B.: K nearest sequence method and its application to churn
prediction. In: Corchado, E., Yin, H., Botti, V., Fyfe, C. (eds.) IDEAL 2006. LNCS, vol. 4224,
pp. 207215. Springer, Heidelberg (2006).
Eastwood, M., Gabrys, B.: A Non-sequential Representation of Sequential Data for Churn
Prediction. Computational Intelligence Research Group, School of Design, Engineering and
Computing, Bournemouth University (2009).
Rosset S., Abe N.: Data Analytics for Marketing Decision Support, IBM T.J. Watson Research
Center (2006).
S. Rosset, C. Perlich, B. Zadrozny, S. Merugu, S. Weiss and R. Lawrence, Customer Wallet
Estimation. 1st NYU workshop on CRM and Data Mining (2005).
S. Merugu, S. Rosset and C. Perlich, A New Multi-View Regression Method with an Application
to Customer Wallet Estimation. The Twelfth ACM SIGKDD International Conference on
Knowledge Discovery and Data Mining, August (2006).
R. Koenker, Quantile Regression. Econometric Society Monograph Series, Cambridge University
Press (2005).
Archaux, C., Laanaya, H., Martin, A., Khenchaf, A.: An SVM based Churn Detector in Pre-paid
Mobile Telephony (2004).
Dasgupta, K., Singh, R., Viswanathan, B., Chakraborty, D., Mukherjea, S., & Nanavati, A.
A.: Social Ties and their Relevance to Churn in Mobile Telecom Networks. Proceedings of
the 11th international conference on Extending database technology pp. 668677 (2008).
S. Y. Hung, D. C. Yen and H. Y. Wang, Applying data mining to telecom churn management,
Expert Systems with Applications, Vol.31, pp. 515524 (2006).
Ascarza, E., Hardiey, B. Modeling Churn and Usage Behavior in Contractual Settings, March
Bauer H., Hammerschmidt M., Braechler M.:The customer lifetime value concept and its
contribution to corporate valuation. Yearbook of Marketing and Consumer Research, vol. 1
Buckinx W., Van den Poel D.: Customer base analysis: partial detection of behaviorally loyal
clients in a noncontractual FMCG retail setting. European, Journal of Operational Research 164
(2005) 252268.
Buckinx W., Verstraeten G., Van den Poel D.: Predicting customer loyalty using the internal
transactional database. Expert Systems with Applications xxx (2005).
Hwang H., Jung T., Suh E.: An LTV model and customer segmentation based on customer value:
a case study on the wireless telecommunication industry. Expert Systems with Applications 26
(2004) 181188.
Grsoy, Umman Tuba imek: Customer churn analysis in telecommunication sector,
Department of Quantitative Methods, School of Business Administration, Istanbul University,
Istanbul, Turkey, Istanbul University Journal of the School of Business Administration
Cilt/Vol:39 (2010), 35-49.
Burez, J., Van den Poel, D.: Handling class imbalance in customer churn prediction, Expert
Systems with Applications 36 (2009) 46264636.
TM Forum Best Practices and Standards: