Customer base management in a prepaid mobile market : Usage risk and usage opportunity model

One of the major challenges of mobile operators in a fierce competitive and multi-simming mobile market is how to gain share of wallet and grow both value and market share. This is because a prepaid customer makes a daily decision on when to recharge, which of the SIMs to recharge and what to do with the recharge. Hence, operators are faced with dwindling revenue as daily recharges and usages are impacted by customers’ behavior. In this paper, we develop a model which can identify irrational usage behavior within the customer base at an individual customer level. Our Usage Risk and Usage Opportunity (UR/UO) model captures all the usage related behavior within the customer base and enables an immediate intervention by the mobile operator to leverage on the opportunities to increase revenue and mitigate the usage risk. We applied the model on a robust sample size of prepaid customer base and model the customer’s usage along the grower, decliner, flat, stopper and new users of all products and services available in the offer portfolio. Opportunities are identified, quantified and prioritized across all service portfolios for a one to one marketing intervention.


INTRODUCTION
In mobile industry, one of the most challenging problems over the years is the customer base management.In order for an operator to ensure an optimal Average Revenue Per User (ARPU), the base needs to be properly engaged and managed.With increasing customer data at CDR level due to increasing customer base and usage triggered-events, understanding one to one customer behavior at granular level becomes a complex exercise for operators.Though, the most visible development amongst operators in the recent years is how they have successfully leveraged on the intelligence derived from customers' data to take timely and robust business decision.However, several behavior exists within the base that take a long time before they are picked up or identified and addressed due to large data set that is associated with the mobile industry.Hence, a need for a robust way of understanding the complex structures of customer behavior and how they impact the usage revenue and business performance.
However, growing numbers of CDRs of mobile customers have given rise to many research areas within the marketing science and analytics of customer base management in mobile industry.Most of these methodlogies and approaches are not easy to implement within the marketing setting of operators.Our approach is different in this regard.In this paper, we focus primarily on models that can be adopted, implemented and put to use by any operator regardless of the size of the customer base.
Research into customer base has grown especially in the area of model development.We have seen growth in the publication of prediction of customer lifetime value and development of models for customer-specific offers (Fader et al., 2005;Rust and Verhoef, 2005).Moreover, several review papers have been published (Villanueva and Hanssens, 2007;Rust and Chung, 2006;Gupta and Zeithaml, 2006;Neslin et al., 2006;Gupta et al., 2006; as cited in Verhoef et al., 2007a).Fader and Hardie provided a deeper coverage of modeling issues across noncontractual and contractual settings of mobile customer base (Fadie and Hardie, 2009).Our paper differs from these papers in several aspects (Verhoef et al., 2007a).Firstly, this paper provides a robust and comprehensive approach towards value optimization within the customer base.We identify how an operator can quickly identify customers that are spending somewhere else within a competitive environment (share of wallet).We therefore focus on the prepaid segment (non-contractual) where operator has no assured revenue from any individual customer.Managing this segment of the customer base is more difficult as customer usage can be erratic.A customer can use $30 this week and use only $5 the following week.The ability to quickly identify this erratic behavior of customer usage within the base for immediate intervention is another aspect that distinguished our approach to other publications.
Several approaches have been adopted in modeling behavior of customers in a non-contractual setting (prepaid).These models are more complex due to the nature of this segment.Mining the rich CDR is an important process that can be described as the bedrock of decision making process.It is also a process aimed at the discovery and consistent use of profitability knowledge from the enterprise data (Ling and Yen, 2001).Starting with (Ngai et al., 2009), data mining models generally include association, classification, clustering, forecasting regression, sequence discovery and visualization.In customer base management, a combination of data mining models is required to support an effective base management strategy.From churn management to usage optimization and to ensuring that a customer is on the right tariff plan -prepaid base management indeed involves a series of complex activities.As much as the operator wants to get the maximum share of wallet from every customer within the base, it is important that every customer also get value from the operator on every recharge that is made.Fader and Hardie (2009) considered a probability model for customer base with a mindset that observed behavior is the outcome of an underlying stochastic process.However, most models did not consider com-Dairo 43 petition effects.They have left operators with no quick insights into the impact of competition activities and other related variables on the usage and recharge behavior of the prepaid customer base which is the focus of this paper.One explanation for the missing of this important aspect in most models is that data on competition interactions are not available in customer data base (Verhoef et al., 2007b,c).However, some inferences can be deduced from interconnect data in the case of prepaid customer base.
In this paper, Usage Risk and Usage Opportunity (UR/UO) model was developed that can be used to understand critical usage behavior within the customer base.Our model reveals usage opportunities and risks that lie within the base.In the next section, we take a look at the dynamics of usage and recharge in a noncontractual setting and present the need for our model in base management.This is followed by the development process of the Analytical Data Store (ADS).We outline the assumptions and derived the usage model.The relative performance of the UR/UO model when applied to the customer data set is examined.Next, we quantify and prioritize the opportunities and risks under the model assumptions.We conclude with discussion on several additional issues that arise from this paper and how operators can leverage and implement this model to optimize usage revenue and performance.

Dynamics of customer usage and recharge
The nature of prepaid customers makes it difficult for the operator to be certain of value that can be realized from a customer over a period of time.This is because there is no notice on when a prepaid customer will increase usage, decrease usage or stop using a particular service.A prepaid customer can stop to perform revenue generating activities at any time and resume usage after few days.Prepaid customers are constantly making revenue impacting decisions by the way they use their phones -when they top-up, where they top-up, how much balance they keep and what number to call (off-net or on-net).It is possible for a customer to go dormant and churn from the network after a sudden stoppage of revenue generating events without giving any noticeable signal to the operator.
Due to this uncertainty of prepaid usage revenue, it becomes very important for operators to have a robust way of detecting usage associated risk of every customer at one to one level.We present a model that will not only alert the marketing team of the usage risk and associated opportunities of every customer week on week, but also reveal the specific behavioral change that is responsible for the change in the customer usage.
Recharges or top-up are characterized by a prepaid market.This is the only way a prepaid customer can have access to the mobile phone.For an operator, recharge behavior of a customer is a top priority.This is because customers would use whatever they have recharged.Analyses have also shown over the years that for an operator, the difference between total recharges of operator's prepaid customer base and total usage is ±1%.For an operator with a robust value and customer engagement, total usage can surpass total recharge in some months occasionally.This is as a result of customers that use their total recharges for the month and consume the carry over balance from previous month.However, there could be unused balance in the customer account by the time the customer goes into inactivity state.This however, depends on the reason for inactivity that could eventually lead to the churn of the customer.
Fluctuations in the usage and recharge behavior of customers present a difficult problem for the operator.Operator can witness millions of dollars decrease in revenue within a month if a sudden change in behavior of a group of customers is not quickly noticed and addressed within a short period of time.Conversely, operator can witness millions of dollars growth in overall monthly revenue if usage opportunities within the base are detected and optimized accordingly (Figure 1).The question is, how quickly can an operator identify the risk and opportunity within the customer base consisting of millions of customers?

Analytical Data Store (ADS)
The data come from analytic experiment of a prepaid base of a mobile operator in an emerging market.The firm exists within a very fierce competitive environment.We use this data to understand the usage behavior of prepaid segment across different product and service portfolios and assess the associated opportunities and risks across the customer base due to the customers' usage behavior.
The primary objective of this model is to be able to assign prioritized interventions on a one to one basis to custstomers by leveraging on the various usage opportunities and identified risks across the customer base.This underlying infrastructure for this model is the Analytical Data Store (ADS) (Dairo and Akinwumi, 2014).Operators require structured, reliable and insightful information at their disposal.This is a pre-requisite to execute a usage model and methodology that will create incremental revenue via targeted marketing actions.The information from this infrastructure can be used for the following: Automation of the usage opportunity and usage risk model; (ii) Use as the input for Measurement and Evaluation (M&E) of all marketing actions; (iii) Use as the basis for advanced analytics/reporting and other advanced models such as predictive modeling; The ADS is the central data mart that is created; this in turn feeds the other modules and become the "single source of the truth" for the marketing team.The ADS architecture in Figure 2 facilitates all marketing and value management activities, including ADS creation and ADS modeling creation which give all outputs that are needed for effective targeting and measurement.
ADS sources information from the operator source systems and data is collected from all subject areas.
1.All information is refreshed on a weekly basis using the oracle built in date function classified as to_char(Date,`WW`); 2. Account information -product adoption, status and activity/inactivity dates; 3. Telecom usage -minutes of use, SMS events, Data consumption, both MoC 1 and MtC 2 ; 4. Payment / recharge behavior -Denominations and frequency; 5. Usage pattern -Hours of day usage pattern, days of week usage pattern; For successful implementation and execution of ADS, we ensure the following: Underlying tables are available in the schema; (ii) Underlying tables are updated with the information required for the time period stipulated; (iii) Before the execution, information for the current week is added.Two types of table are defined:  Snapshot tables -these populated information as of the last day of the week e.g. if the week is 44 th week in 2012 then the last day will be Saturday the 17 th of November. Weekly tables -when populating cumulative tables e.g.usage revenue, then full week data need to be available in the base table before execution.

Model
Our objective is to develop a model of a typical prepaid customer in order to identify the opportunities and associated risks to the customer usage revenue.The approach is to identify the product portfolio where there are opportunities, risks and quantifies the revenue potential in order to give priority to what customers will be offered at one to one customer level.
To start, we identify all customers' billable events and classify them as follows: 1.
Data events: all data related events such as data bundle purchase and pay as you go data usage;

2.
Voice: we further classify voice events into different legs -on-net, off-net and international; 3. SMS: This is also segmented into on-net SMS, off-net SMS and international SMS.All associated usage related to data, voice and SMS are considered under this model.Secondly, we define our specific period for this model.While we have over six months data in our ADS, the model focuses more on the last four weeks.This is the most critical period for a marketing team to optimize customer usage when an observed behavior is identified.

Assumption 1
Consider the usage of a customer within the customer base of an operator in a defined period along a defined product or service, the customer can only be a New (N), an existing (E), or a Non-User (NU).

Assumption 2
When a customer within the customer base of an operator is an Existing user of a product, the customer can only be a Flat (F), a Dropper (D), a Grower (G) or a Stopper (S) in the usage of the product within a defined period.
A New user (N) is defined as customer who started the use of a product or service in the last period of consideration.In the case of our model, our last period is the last week of data in which the model is refreshed.An Existing user (E) is a customer that is using a particular product week on week while a Non-User (NU) is a customer that is yet to use a particular product in the past.For example, a customer that has never made an international call before is a non-user of international call.
A customer is said to be Flat (F) on the usage of a product if the customer usage on that product grows or declines within a tolerable and defined threshold.We described a Dropper (D) as a customer who declines in a particular usage leg in the period under consideration.A Grower (G) grows in the usage of a product while a Stopper (S) stops using a product within a period.

Derivation of model
Consider an active customer A with revenue generating events for the last four periods.Customer A is revenue generating at T=1, 2, 3 and 4, where T =1 represents the first week.T 1 ,T 2 ,T 3 , and T 4 will represent our weekly period throughout this paper.The objective of the model is to identify the usage opportunities and usage risk at individual customer level such that every customer can be targeted with appropriate offers accordingly at the beginning of T 5.
Let the total usage 3 of customer A at period t 1 be µt 1.To be part of the weekly ADS for the weekly refresh, µt 1 > 0, in this case, we have eliminated all customers that are inactive within this period.The ASPU 4 of customer A from week 1 to week 3 can be expressed as Let the week 4 usage be α, usage differential £ can be expressed with the following equation: The next step is to find the usage differential £ for customer A on every usage leg in the period where T =1 which will form the basis of the behavioral pattern of customer A. Suppose customer A performs on-net voice call, off-net voice call, international call, on-net SMS, offnet SMS, international SMS and data.This customer is regarded as a typical customer that performs revenue generating events on all legs at time T=1.However, this is not a general behavior of all customers in a particular week.
Using the differential usage £, we define customer A along the following usage behavior: Where ∂ is an arbitrary usage differential ratio that would 3 Usage here refers to billable events from the main account of the customer not usage from bonus accounts.

4
(ASPU) -Average Spend Per User be chosen depending on how aggressive we want to be.We chose 0.25.We modeled equation (3) on every customer along each product portfolio.A customer can be a new user (A N ) on international SMS and at the same time a Dropper (A D ) on data, Flat (A F ) on voice on-net call and Grower (A G ) on voice off-net.The UR/UOmodel shows clearly the opportunities and usage risk of every customer within the base.

Opportunities
We present expression for the quantification of opportunities after identification of the opportunities and risks at individual customer level.Suppose this model is refreshed on the first day of the week according to the operator's calendar, what we would find out is that there will be multiple opportunities and associated risks per customers.For marketing team, it is critical for them to understand who to target and with which offer along the order of potential incremental revenue that could be achieved.Also, marketing wants to know the customers at risk and what usage legs they are at risk and the associated revenue potential along every usage leg that is at risk (Table 1).
Our quantification of opportunities is based on the following assumptions.

Assumption 3
A user with up-sell potential is likely to accept an offer that leverages price elasticity with a positive effect upon acceptance.

Assumption 4
The cross-sell usage opportunity is calculated under the premise that a new service brings the ASPU of that service as an incremental of the current ARPU5 .
Using the value segmentation as acknowledged by Dairo and Akinwumi (2014), suppose customer A is a platinum6 customer in the last calendar month.Let the ASPU of all platinum on voice on-net service in T=4 be β voice on-net .Suppose customer A is a non-user of voice on-net service, then, Customer A potential revenue if cross-sell for voice on-net can be expressed as P R (A NU ) = β voice on-net (4)

GROW
Up-Sell • Identify users that are potential candidates to increase their usage of a product • Quantify the potential to up-sell a product by the difference between the average consumption and the maximum consumption observed during the analyzed period Cross-Sell

•
Identify users that are potential candidates to start using a product by matching non user profiles with current user profiles • The quantification of the individual opportunity is assumed to be the product ARPU Decliner Mitigation • Identify users that are decreasing the usage of a product or that have stopped using a product • The quantification of the decline mitigation is made by locking the current expenditure of the decliner and an assumption of a percentage of users to be recovered by reactivation campaigns KEEP Reactive Product Usage

•
Identify users that have churned only a specific product or service • Quantify the potential to motivate the return to usage of that revenue stream Churn Prediction

•
Identify users that have a high probability to churn in the next two months • The quantification of the individual opportunity is based on the lifetime extension of the current average ARPU Suppose A is Flat (A F ) on data usage, we can express the potential revenue of the customer if up-sell as P R (A F ) = ∂β data (5) Suppose customer A is a stopper (A S ) on international SMS service, the potential revenue of A if up-sell is Where β is is the ASPU leg of international SMS service for all platinum customers with international SMS usage in week T=4.Potential opportunity for customer A who is a dropper (A D ) say on on-net SMS service is expressed as Where £ is the usage differential of the average of week T=1 to T=3 and T=4 of the on-net SMS ASPU of platinum customers while α on-net_sms is the ASPU of platinum on-net users in week T=4.Every usage opportunity is calculated for every product or service available in the portfolio with the objective of having at least one marketing offer available to be executed (Figure 3).The combination of usage opportunity category and number of available services or products provide the full set of unique usage opportunities.
We therefore rank the usage opportunities of every customer along potential revenue in the form r 1, r 2,……………….r n where n is the number of usage opportunities.The ranking will help the operator to maximize returns by focusing on most promising opportunities at individual customer level (Tables 2 to 5).

Conclusion
We have developed a model that can be appropriately used in the effective management of customer base usage revenue in a prepaid mobile market.A market where customers make daily revenue impacting decisions by the way they use their phones -when they recharge (top-up), where they top-up, whom to call, when to call and how much balance they keep.By using 50,000 sample customer data contained in the four week CDR data of a prepaid base, we found through our model that operators can detect irrational usage behavior from the base on a weekly basis regardless of the size of the operator.
Our model identifies and segments the base into      Growers, Flat, Non-Users, Stoppers and New Users on all services and products that are available within offer portfolio.The UR/UO model identifies the usage leg where up-selling and cross-selling opportunities can be quickly leveraged and where risk of stoppers and flat users can be quickly mitigated with the appropriate offers.
It also identifies the customers with high potential to increase their usage of certain products.We quantified the potential to up-sell a product by the difference between the average consumption and the maximum consumption observed during the analyzed period.
In the case of cross-selling opportunities, we identified customers that are potential users of a product by matching non-user profiles with current users' profiles by assuming the quantification of the individual opportunity to be the service ARPU.To mitigate declining usage, users that are either deceasing or stopped the usage of a service are identified through the model.The quantification of potential revenue is made by locking the current spend of the decreasing customers through an arbitrary differential ratio assumption.
As we mentioned at the beginning of this paper, what UR/UO model set out to achieve is to allow the operator to understand the impact of every customer behavior on the usage revenue.Operator will know which segment of the base is impacting the revenue negatively and why.With this, no activity or change in behavior of customer can have a negative impact on the overall revenue more than a week before it is picked up by the operator if the model is refreshed at the beginning of the week.Due to several opportunities that would be identified for each customer since customer uses more that one service within the operator portfolio, the opportunities are quantified and ranked based on potential revenue as shown in Table 5.Since all opportunities cannot be addressed with marketing activities at the same time, operator will have to focus on the top ranking opportunities.
UR/UO methodology provides a platform to plan marketing related campaign activities but the execution needs to be extremely flexible to adapt to every user's action.Our result shows that the top 10 intervention/opportunities contribute 90% of the revenue potential.Up-sell/cross-sell contributes 76% of all these opportunities.Cross-sell/Up-sell off-net, on-net voice and data have the maximum revenue potential.For operator, it is pivotal to run pilots for up-sell and cross-sell campaigns that will be used to target every opportunity before the scaling up the offers to a larger base.
Implementation of this model implies a change of operational mindset from a vertical product or product silo approach to a horizontal one to one approach for best activity optimization.
The usage opportunity quantification methodology has a list of non-exhaustive value for operators amongst which are:

Figure 1 .
Figure 1.A typical operator usage and recharge trend.

Table 2 .
First ranking of UR/UO potential.

Table 3 .
Second ranking of UO/UR potential.

Table 4 .
Second ranking of UR/UO potential.

Table 5 .
Ranking of UR/UO potential and intervention mapping.
across product portfolio that may end up with a global answer for an operator (technical implementation, product revamp e.t.c).