Customer churn prevention

All models are fake. But some can be useful

Start with behaviour - not data - when implementing predictive models for customer churn prevention

When using predictive modelling and analytics to prevent customers from churning, all too often, organisations are caught in a state of apophenia followed by analysis paralysis. The trick to overcome this is to let go of the “we need more and better data” excuse and instead start by asking yourself “what can and will we do if, in fact, we already knew who is about to churn?”

All models are fake. But some can be useful

Data-driven customer engagement

“Blondes have more fun, but brunettes sell more lingerie”. At first sight, this statement from a CNN article might seem both sexist and rather primitive, when in fact it is the key learning of a highly data-driven approach to customer engagement. In less than five years after the lingerie start-up Adore Me launched their business, they have been able to achieve a three-year revenue growth exceeding 5,500%, hence becoming one of the fastest growing companies in New York.

The same product shot on the same model in a different posture can nudge sales a few percentage points in either direction.

Source: fast company

Their approach is rather comprehensive, specific and methodical, yet still extremely simple: For every product line and item, Adore Me shoots multiple versions of photos to use in their web shop. The photos are only slightly distinct; same model, same outfit and same props, but a different posture will have a measurable impact on how well specific product lines sell compared to the exact same setup only with a different model or a change in hand position. And so on. And they test all photos this way – every month.

No matter if you are a retailer wanting to increase revenue by giving your customers the best web experience through relevant and personal recommendations by utilising the leading graph database technologies like Walmart and eBay do, or if you are in any other line of business, the ultimate goal of becoming a data-driven organisation must be to grab any opportunity to lead, so that every touchpoint in every channel on the customer journey will result in both increased energy in terms of customer engagement, satisfaction and loyalty as well as direction in terms of customer lifetime value.

Establish an infinite feedback loop

Essentially, what Adore Me is doing, is A/B testing their different offerings to create maximum impact on customer engagement and revenue. Their secret ingredient is how they have managed to build an organisation and operating model around an infinite feedback loop, turning data into meaning into change in behaviour and, finally, into real market impact.

all models are fake

Figure 1: Engaging the market by establishing an infinite (near) real-time feedback loop based on data.

Any attempt to utilise quantitative methodologies and techniques must take an outset in supporting this infinite feedback loop.

This is done by creating a culture and an organisation, focusing – not on data and technology alone – but more so on how to support and drive behavioural change by acknowledging the importance of the human factor, when it comes to closing the gap between data and impact.

Close the gap by managing the change in the interaction between technology and people

Traditionally, the “excuse” for not being able to proactively engage and delight customers in a multi- or omnichannel setup would be something like “we need more and better customer behavioural data, before we can …” and subsequently “and then we need a new tool to help us analyse the data, before we can …”.

However, the problem is no longer poor data or lack of technology – it is more like “we have too much data, and technology makes us see patterns everywhere …”.

The biggest hurdle is being able to tell if a data set or pattern is both valid and relevant. In order to do that, you need to add a wee bit of human to the recipe – you need to manage the change in the interaction between technology and people. Getting to that point requires overcoming three stages of organisational inertia.

Stage 1: Too much data

First of all, there is already too much data and standing on the brink of what has been coined “the third big wave of the Internet” or Internet-of-Things, it is clear, that the ocean of data is only going to get bigger, deeper and more fluctuous by the hour.

Beacons, context-aware mobile devices and other data-producing and consuming items will begin to take an even more prominent place in the digital market arena. And no matter how hard we try, we will never be able to neither drink nor boil that ocean of data.

Therefore, do not rely on “more and better data” to be your saviour, when it comes to spotting who is about to terminate the customer relationship.

Stage 2: Technology has become a commodity

Secondly, the type of technology needed for gathering, storing and analysing these types of large-scale customer behavioural data has more or less become a commodity. Being an organisation wanting to explore or exploit quantitative computing, you no longer have to rely on a team of NASA scientists building it for you from scratch.

All you need now is a credit card, and you can access anything-as-a-service – from machine learning platforms like Microsoft Azure ML, Amazon ML or IBM Watson Analytics and crowd-sourcing your data prediction with the help of to open source technologies for distributed big data processing like Hadoop and fast network and cluster analysis, using the Neo4j Graph Database or similar.

Therefore, do not rely on a “lack of tools” to be a valid excuse for not delighting your customers in due time.

Stage 3: Patterns, patterns, everywhere! 

Thirdly, people are prone to see patterns where there are none, and as organisations get their hands on even more detailed and diverse data as well as faster and cheaper technology to process these data, the risk of falling into a state of organisational apophenia followed by analysis paralysis is increased dramatically.

Correlation and causation is two very distinct attributes of data sets and variables – there is absolutely no causation between “US spending on science, space and technology” and “suicide by hanging, strangulation and suffocation”, although there is a very clear correlation.

All models are fake

Figure 2: Correlation is not causation


So, to succeed, organisations must focus on what matters most: Answering how the organisation can and will respond to early warnings of churn, before they start focusing on their ability to collect and process more data and, essentially, spot more relevant patterns of these early warnings of churn.

A structured approach for succeeding with churn detection

Basically, managing churn can be approached in three ways:

  • Untargeted. The churn is prevented by product superiority and other means of increasing brand loyalty.
  • Proactively targeted. The churn is prevented by proactively identifying customers, who are likely to churn at a later time, and present them with special service offerings or other incentives that will increase their likelihood to stay.
  • Reactively targeted. The company does not act until the very last minute, when the customer tries to cancel the relationship.

Following a targeted proactive approach is, at first hand, likely to have a cost advantage as the cost of the “bribe” that the company has to pay is expected to be lower than the cost of persuading a customer to stay, once they have one foot in the door.

On the other hand, there is also a risk of the opposite being the fact, when a proactive targeted approach turns out to be more expensive than a reactive targeted approach. This would be due to a waste of incentives on customers who would have stayed anyway – which is likely to happen when the churn prediction model produces inaccurate results, hence before an organisation engages itself in an attempt to proactively reduce churn by modelling, simulation and predictive analytics, it is important to observe the three most prominent pitfalls and causes for failure to create impact:

  • Not starting off by ensuring the link between the model behaviour and the desired organisational behaviour, i.e. not answering the question “what can and should we do in terms of churn prevention, if, in fact, we already had the knowledge required to locate and proactively engage the individual customer in risk of churning – perhaps even before they showed signs of increased churn risk”.
  • Assuming that “one size fits all forever” when it comes to decide the technique for churn detection (logistic regression, decision trees, n-tuple neural networks, discriminant analysis, cluster analysis, Bayesian logic or an ensemble of multiple classifiers, using e.g. Bagging or AdaBoost) as well as deciding on the type and weight of input variables. Each customer is different, and the relationship evolves over time, hence, a churn detection model will have a certain life expectancy
  • Failing to kill model complexity and keeping a narrow focus when deciding on “how much of reality” is to be modelled and which input variables are to be included in order to find an optimal trade-off between minimising the number of variables and maximising the discriminatory power, as this inevitably will lead to organisational apophenia followed by analysis paralysis.

All models are fake

Figure 3: Four-phased approach to customer churn prevention

To reduce the risk of falling into any of the pitfalls, an organisation planning a churn detection and prevention initiative should follow the following structure; the four-phased approach:

Scope and approach

  • Engage stakeholders and key opinion leaders to agree on the “framing question” – and what would be different in the way that the customers/segments are engaged – if we already had the answer – in order to set a specific and narrow scope.
  • Determine the right qualification models, variables, quantitative approach and technologies by acknowledging that one size does not fit all. No two individual customers, segments or campaigns are defined by the same set of variables – and no two questions have the same implications on customer engagement.

Build and validate

  • Prepare the data needed in a structured approach for importing, processing, exploring and statistical analysis. Make sure to distinguish between training, validation and testing data.
  • Critically analyse the behaviour and sensitivity of the churn detection model and scrutinise the results to avoid wrong conclusions, i.e. evaluate the relative importance and magnitude of false positives and false negatives on the overall validity, quality and correctness of the model, e.g. by using “confusion matrices”.

Deploy and engage

  • Deploy the model and the required analytical tools on a scalable platform in order to be able to handle a sudden increase in data volume or formats without suffering from reduced speed or accuracy.
  • Engage and prepare the organisation for the upcoming A/B test of churn mitigating initiatives to make sure that all affected touchpoints are handled properly and does not become an isolated initiative – but rather an integrated initiative across all channels, departments and competences.

Learn and adapt

  • Follow up on the required link between ambition, business impact and suppor - ting behavioural changes – i.e. make sure that “someone” actually does have the means and will to respond to a certain trigger, reaching a certain threshold.
  • Adapt the operating model to make sure it reflects how churn is best mitigated based on the newly acquired customer segment insights.
  • Take means to ensure that the model is governed and adjusted over time according to its life expectancy.


For a list of references and sources, please refer to the downloadable PDF.