Our goal when working with churn and with clients in general is to extend their life cycle and increase LTV. A Starbucks customer who has been visiting this chain for 20 years will bring in $ 14,099 for the entire period. That’s a hefty sum of $ 3.5 cups of coffee. The question is – how to make a person visit a cafe for 20 years in a row?
In this article, we will focus on two areas: retail (retail) and telecom operators.
Let’s look at the main types of customer churn and the reasons.
- change of residence / temporary use,
- changing needs.
2. Reasoned refusal:
- dissatisfaction with the level of services or product quality,
- a company was found where the cost of similar goods or services is lower with the same quality,
- a company was found where the quality of similar goods or services is higher at the same cost,
- Brand fatigue,
- the client has completely abandoned the product or service (for example, quit smoking).
Read more about: Top Data Analytics Companies
3.Hidden – a gradual decrease in the volume and regularity of consumption of the company’s products due to:
- parallel use of competitors’ services,
- the use of substitutes.
Outflow is always present in the database, and it is important for a business to understand what share is a natural outflow, and with whom it is possible and necessary to work.
There are market benchmarks and competitor benchmarks to look up to. But be sure to do your own research – talk to customers, conduct surveys to find out why they abandoned services or products.
At the same time, it is important to consider the reasons for refusal in the context of client characteristics – at least using segmentation on the best, with potential, uninteresting ones.
But before calling past customers, churn needs to be defined internally. This is not always so obvious, especially for retail or, for example, a network of clinics. Usually they choose between half a year and a year without visits / purchases.
But you don’t always need to look so straightforward. Pay attention to the proportion of those who return after this period, and if this proportion is significant, try to move the border further. Thus, your first metric will be the churn rate over a period of time.
At the second stage, we determine the frequency of control – how often we will measure churn and what values we expect to receive.
Here it will be relevant to conduct a study and determine for yourself the level of natural outflow and, possibly, the factors influencing it. For example, the current economic environment is helping to increase this indicator.
Also, the churn rate can be considered not “in the hospital as a whole,” but for specific segments – the same best and clients with potential.
One step before the outflow
But measuring the amount of outflow is a statement of a fait accompli. Our task is to prevent it. To do this, we need to define:
- forecast horizon;
- parameters indicating that the client is at risk of outflow;
- experimental design – confirmation of the effectiveness of the hypotheses put forward.
The forecast horizon is the period we need to take customer retention actions. And here we need to ask questions:
- how quickly can we contact the client, what communication channels are available?
- what methods of retention do we use?
- how long does the hold take?
Experience confirms the fact that the earlier the communication starts, the higher the probability of retention. If the client has already made the decision to go to a competitor, it will be difficult to reverse it. Therefore, it is so important to choose the right forecast horizon.
Determining the parameters that signal that the client is at risk is a critical and creative task at the same time.
In one of our projects, we found out that the low average duration of incoming and outgoing calls can be such a signal for a telecom operator. We will talk about this in more detail a little later.
Our work ends with testing hypotheses in practice – an experiment. This is the only way to understand whether we have guessed or not.
Retail churn prediction
To identify patterns of churn, you can start by analyzing the following parameters:
- frequency of purchases;
- check size / number of purchased goods;
- prescription of purchases;
- participation in the loyalty program (there is a greater risk of churn for those who do not participate);
- contacting the support service;
- participation in communities;
- Purchase returns;
- other parameters (social, demographic, psychographic, geographic).
When we say that a person is at risk of outflow, we most likely mean that he began to visit us less often. Accordingly, if the frequency of purchases has changed, has the value of the receipt changed? What about the prescription of purchases? These 3 criteria are parameters of classical RFM analysis that we can evaluate regularly.
When conducting RFM analysis, we divide clients into groups by age, purchase frequency and check amount.
We get the relevant segments: the best customers with potential, new customers, good customers at risk , bad customers at risk, former best customers, former customers.
For each group, you can develop your own strategy and work on changing the current situation.
RFM can be done in many programs, ranging from Excel to advanced tools from IBM or SAS – it all depends on the volume of data, the range of analyst / user capabilities, and budget.
The analytics platform is great for performing RFM analysis, especially on data volumes over 5 million rows. An algorithm without programming is created within 10 minutes and allows further analysis to be carried out every day and to see how the segments change, whether the strategy of working with them is effective.
Another option for quick one-button analysis is clustering, which allows you to evaluate the change in one or more parameters.
We usually start worrying about churn when a person lowers their frequency – purchases, service usage, etc. Considering each client in this case is very long, so you can cluster the audience and identify patterns of behavior that are characteristic of departed and departing clients.
In , clustering is started in a few clicks, also without programming. It is important to quickly select clients whose frequency has changed, and start interacting with them, working to retain them. And this is true not only for retail and retail, but also for other areas.
Outflow on the example of telecom operators
Let’s highlight the parameters that need to be analyzed for clients at risk:
- participation in the loyalty program;
- duration of calls;
- internet traffic;
- visiting competitors’ sites;
- social dem criteria;
- tariff plans;
- contacting support.
In practice, it is not always possible to get access to all the necessary data. This can be due to technical problems or bureaucratic obstacles, such as who owns the data and their willingness to share it.
As part of a consulting project, we did outflow analytics for a regional telecom operator.
Our task was to identify the main signal – a characteristic that indicates that the client is at risk and additional parameters that must be taken into account when forming a pattern that identifies an outflow.
As a result of the project, the marketing department of the telecom operator received a methodology for identifying subscribers at risk.
The methodology is based on the following parameters: average duration of incoming and outgoing calls, age, participation in the loyalty program, tariff plan.
For six months, various hypotheses were tested and, as a result, a 16% reduction in churn was recorded, which made it possible to extend the life cycle for the most interesting subscribers for the operator.
How to reduce churn by 15% or more?
The most important thing is the attitude towards the subscriber. Until now, not all operators are interested in feedback – what the client liked, what did not like. Think how many times you were allowed to sell something, and how many times to find out how you are doing and whether you are happy with the service.
We suggest considering the following specific measures:
- Selecting subscribers visiting competitors’ sites, monitoring user experience.
- A system of proactive notification of service quality degradation.
- Segmentation and profiling of subscribers.
- Welcome Calls.
- Formation of recommendations for the preparation of outbound communication scripts for customer retention based on the cluster / segment to which it belongs.
- Formation of a matrix of proposals for retention, taking into account the task of preserving ARPU.
- Surveys and NPS Assessment.
Does it make sense to include machine learning?
If the diagnostics are completed, there are basic parameters and hypotheses, cause-and-effect relationships are determined, machine learning can be started.
Binary classification is most often used – the algorithm identifies patterns of behavior based on two classes: outflow and non-outflow. The model must be probabilistic, that is, as a result, we must understand with what probability the client is inclined to churn.
What you need to know to run machine learning:
- training should take place on the same data on the composition as used;
- a control group is required to track performance;
- it is desirable to clarify the impact of temporal trends on learning;
- it is important to ensure that the model does not “rot”, since services and business processes change, history also changes, and therefore you will have to retrain the model on a regular basis.
As a result, we have to come to the conclusion that we have an automatic customer scoring that is recalculated every day.
For example, a client contacted technical support – the scoring lowers the score, accepted the offer to change the tariff plan – the score increased. Thus, we can not formally calculate the customer retention rate, but “personally” see the mood of people. Such tracking of the entire database is possible only when artificial intelligence is connected.
Procedure for working with churn:
- We determine what the churn is and whether it is necessary to work with it. We highlight different retention strategies for different segments.
- We visualize the data and select the parameters that signal the outflow. We form hypotheses.
- We test hypotheses and track their effectiveness.
- Improving the model, adding new parameters, removing unnecessary ones.
The analytical platform allows you to visualize data and conduct diagnostics: identify causal relationships, come up with hypotheses and track their viability. We use machine learning when you need to automate processes.