Aim of the project
The aim of the project is to enable customer value based proactive retention in prepaid segment to target those customers who are in churn danger. The perception is that much less customer data is available for operators to analyze. Although this is true in comparison to the depth of data to post-paid segment where customer identification is easier and lifetime is longer, transactional behavior and community information mined from CDRs and billing system can enable data miners to derive up to 100+ meaningful variables that predict churn and help defining customer value. Coming up with an accurate prepaid churn model and customer value based retention can be a game changer for an operator in a high growth market where prepaid segment is the dominant profit pool, customer acquisition and retention costs are rocketing because price wars dominate the competition.
- Joint agreement developed between sponsors and other stakeholders with regards to definition of stages of churn. Our expert team will help define the best definition based on context, keeping in mind that definition shall enable actionable prevention. E.g. once last stages of churn are concluded based on customer top-up behavior, that is usually too late to act on.
- The project team focuses on prepaid churn prediction. Estimation of churn for the coming 3 months based on data analysis of past year, either from predefined tables in data warehouse, or data aggregated from core systems for analytic purposes. Creatively derived variables will be analyzed to predict future churn. This includes but not limited to top up frequency, time lapse since last top up, number of calls to internal/external networks, handset purchase, monthly spending, community scores, viral affinity, etc. The efficiency of the churn prediction model is defined in so called LIFT in the data mining jargon. For a short overview of LIFT, let’s assume that a telecom’s monthly churn is around 3% of prepaid portfolio and it has 100.000 customers. This means that the number of the churners in the population is 3000. Also, let’s assume that Telecom tries to reach out to all of these 3% with retention offer. If we pick 3000 customers by random sampling, only 90 churners will be found, which is the 0,09% of the whole population and 3% of the churners. In contrast, if Telecom developed a well functioning churn model, 540 (18%) to 900 (30%) of the churners can be found. Telecom’s LIFT number will be the ratio between nominator achieved by churn prediction model-, and denominator- random sampling. In this case the LIFT is 6 – 10. See the example below.
Assumption 3% monthly churn, 100.000 customers
||Churners found (%)
||540 - 900
||18% - 30%
- Once telecom has a well working churn model, customer value based retention is the next step. Basis for customer value can be revenue based or profit based. Profit based segments shall consider actual pricing discounts, interconnect fees from incoming calls, etc. Depending on budget available for retention, best customers in churn danger will need to be part of proactive retention campaigns. CLV alone can be a complex data mining project, but as a start a simple model is already a great advancement, and prepaid cost/revenue structure is usually easier to follow.
- To help calculating the ROI for this project, the project team will help translating profit increase achieved by each 1 unit increase in LIFT for different value based segments. The model can be also used to predict sustainability of portfolio, ie how many people from current active customer base will get to close to churn in the next 2-3 months, represented in a waterfall model. This will estimate new customer acquisition needs to keep or grow prepaid market share.
- Optional, not in scope of same project: Once operator identified who is likely to churn and who is most valuable to retain , product affinity scores need to be developed for for need and/or transactional behavior based segments, so that response rates for retention campaigns will be high.
Figure: Churn model performance
In the figure above, along X axis, we can see the percentiles based on ordered churn possibilities.
The main findings of prepaid churn projects are:
- Output file for likely churners in forthcoming 3 months
- Valuable segments to target for proactive retention based on customer value
- Algorithm left behind, will run for example every month; it would be upgraded yearly
- It is possible to run the churn model weekly
- We can integrate derived variables on to data warehouse tables
- Output of project can be base of sophisticated customer value, campaign management, segmentation project
DOs and DON’Ts
- Define very well the churn definition objectively in the beginning
- Collect all possible variables based on various data sources
- Evaluate cost of retention actions to define best ROI after predicting churn
- Do not agree on churn definition last minute
- Do not expect to catch 50% of churners on targeting top 3% suspected churners
Why we are different
- Experience with multi-million prepaid customer base
- International projects in multiple countries
- Successfully completed rotational churn projects
- Planned and implemented full suite of retention data mining roadmaps