In any contest, it sure is helpful to know what your opponent is going to do next. In the perpetual tussle against customer churn, organizations would do anything for the power to predict which customers are about to leave.

Customers are not the enemy of course, but when they exercise their freedom to go elsewhere it can do real harm to businesses. The trick is being able to identify and anticipate this before it happens, thereby aiding customer retention.

And here lies the challenge and the opportunity of predicting customer churn: if every customer gave you a chance to change their mind before they left, customer retention would be a whole lot simpler. But they don’t, and so you’re left to attempt two things on two different timescales:

  • Immediate to short-term
    • Finding and responding to customers where your intervention could make all the difference between them leaving right now or staying for longer
  • Medium to long term
    • Working out how to predict customers who might be on the road to churning, long before they take any such decisions

Short-term: spotting customers that are angry at you

Perhaps the easiest group of potential churners to identify are those that are very dissatisfied. There could be any number of triggers for this – what’s important is that they are so upset by something that their leaving (and never coming back) is a clear and present danger. However, this doesn’t necessarily mean organizations get to hear about these reasons.

Sometimes – regardless of how angry customers get – organizations are oblivious to the threat. This is because only a very small percentage (about 4%) of people with good cause to complain actually go ahead and make a complaint. That’s an awful lot of people who aren’t telling the organizations who serve them how they really feel.

Customer feedback mechanisms can really help here, particularly when they’re used frequently and in relation to key events in the customer journey. Providing an easy way for customers to express satisfaction or dissatisfaction reduces the effort and makes it more likely to happen. 

Rescuing negative customers from churning

With feedback systems in place, organizations can immediately respond to negative feedback to understand what’s gone wrong and try to fix it. These are out-and-out rescue missions, where the primary objective is to stop the customer from churning. It is the equivalent to putting out a fire and usually attracts disproportionate time and resources. Apologies will be offered, promises made, rearrangements and fixes carried out, and new priorities given.  

As prediction models go, this isn’t very sophisticated but can be highly effective. It’s true that you can confidently predict, with a high degree of probability, that very unhappy customers will soon churn. If you get to them in time, and respond appropriately, you can rescue them. You can even restore their level of satisfaction and loyalty to a higher equilibrium than if they never become upset in the first place. This is called the service recovery paradox. 

However, as accurate as it is, this model is very high risk. The customer is already upset and may – quite understandably – have closed their minds off to the idea of offering redemption. Rescuing customers in this way can feel like fighting a losing battle. And it’s also a bad look in terms of public reputation. Do you want to be the kind of organization that waits until customers are very upset before taking action to mitigate their chances of leaving?

Is “just being great at everything” a strategy?

The glaring question when confronted by angry customers is not so much “how do we make these problems go away” as “how do we stop these problems happening in the first place”. Being great at everything in terms of customer service and the quality of the products and services you provide can certainly help. But things will still go wrong. 

Organizations with low churn work really hard to achieve that. Unquestionably, they do less of the things that upset customers and more of the things that meet or exceed their expectations. What’s really interesting is how they manage to know what to focus on to stop customers from churning.

Again it’s customer feedback that plays a key role here, by feeding customer intel about likes and dislikes into a change program within the organization. This is important with regards to churn prediction because it introduces the concept of pattern recognition.

Pattern recognition in customer feedback

Pattern recognition is used widely in machine learning and advanced automation, but at its simplest level is just a case of spotting commonalities in sets of data. We’ll cover how this is an important building block for prediction later in this post, but for now just consider pattern recognition in the case of customer feedback. Let’s use the example of a hotel chain and doing pattern analysis on the feedback of lots of different customers over time:

  • Customers are more dissatisfied after they’ve dealt with customer agent A than on average
  • Guests at the Miami Beach property are consistently more dissatisfied than average with the check-in/out experience
  • Female guests tend to be more dissatisfied with the cleanliness of the restroom facilities 
  • Customers booking via the mobile app spend significantly more time and effort navigating the ‘room selector’ tool than any other aspect of the digital customer journey

In each of these examples, it’s possible for the hotel chain to make changes that increase the probability of happier, more satisfied customers. This in turn reduces customer churn.

The relationship between churn probability and the impact of your actions

What we learn from the basic ‘ rescue’ approach is that a high predicted probability of churn correlates with a high risk of the customer churning. In other words, the more convinced you are that a customer is close to leaving, the less chance you have of making a difference to them. But is the opposite true – that the lower the probability of churn, the higher the chance of making a difference to customer loyalty?

To illustrate this, consider the following example:

Malik has insured his car and home with the same insurance broker for the last 2 years. The latest renewal was 8 weeks ago and Malik thought the premium was OK; a little less than the year before. He gets on well with the broker and seems to get good service. Six months ago, he referred a friend to the broker and received a $25 Amazon voucher for his trouble. By any assessment, Malik is a satisfied customer. Not super-enthused but, still, a stable bet for not churning unless some unforeseen factor were to arise.

In this example, the probability of Malik leaving any time soon is fairly low. So what if the broker called Malik up one day and offered to double the coverage on his home contents for free for the rest of the year? It’s a gesture that might cost the broker a few dollars a month, but surely it would be massively influential upon Malik in terms of his future loyalty. 

Long term: predicting those on the road to churning

The basic ‘rescue’ model described earlier relies on data about individual customers. If customer A is upset, someone needs to make them happy again so they don’t leave. As we’ve already covered, this game of ‘Whack-a-Mole’ can be effective but risky. Every business needs to engage in this kind of cycle or they wouldn’t be very good at fixing problems. However, a better prediction model would be highly beneficial alongside this.

In the quest for a proper customer churn prediction model, two aspects are going to need to be radically different from the basic ‘rescue’ model:

  1. All customers will need to be included, not individuals on a case-by-case basis
  2. Far richer behavior data will need to be interrogated, not simply whether the customer has given a clear signal they are about to churn

Identifying data types to include in the churn prediction model

Companies that continually collect data about customers are at a distinct advantage when creating any predictive analytics model. Without it, the result is literally guesswork.

The essential core of the model is recognizing the behavior patterns of potential churners and correlating this with:

  • Decreasing customer satisfaction
  • Customers leaving/churning

Applying this to various data sets can throw up some interesting results. If you were to examine the behavior of the last 10 customers who churned, the following data might be instructive:

  • Number of calls to customer support (high/increasing)
  • Average time to resolution of the customer’s tickets (high)
  • Customer satisfaction score (low/falling)
  • NPS score (low/falling)
  • Service usage (low/falling)

With some fairly rudimentary mathematics, it would be possible to chart the correlation between data readings and churn probability. 

To collect this sort of data, organizations will need to operate and source the following:

  • Customer feedback platforms
  • CRM systems
  • Service desk/help desk/PSA
  • Digital analytics 

Defining thresholds and responses

With these patterns established, testing will bring you closer to determining their effectiveness as prediction tools. Ultimately you will want to set thresholds at which you can reliably wait until intervening with an appropriate response. If, for example, NPS scores are flatlining at a mediocre level, this wouldn’t normally warrant initiating a ‘rescue’ response. But in the prediction model, it could be an early sign that the customer is uninspired and unmotivated to think positively about you. And so an early response would be warranted, though the content of that response would need to be carefully considered. 

Predicting churn timing is not the same as predicting why customers may be about to churn. In other words, you aren’t really any closer to knowing how to prevent churn from happening, even if churn has been accurately predicted. This introduces a degree of jeopardy into how you respond. Again, testing is vital here so that you can optimize a procedural response. All your possible responses are going to be variations on something positive and attentive to the customer’s individual needs, so it’s likely to be all upside – just a question of how much. 

The more data you have, the more sophisticated your prediction model can become. But beware the challenges of complexity, especially when trying to infer causal relationships between lots of different factors. Having combinations of thresholds and triggers to triangulate potential churners could increase prediction accuracy if you get it right. 

Coming to terms with anomalies and exceptions

You also need to bear in mind that prediction models aren’t going to work all the time, or possibly even most of the time. There are plenty of examples that confound analysis because customers have chosen to churn for unpredictable reasons. Customers who exhibit the purchasing behavior and feedback of a satisfied and loyal customer could suddenly leave in direct contradiction to your model. Also, when customers don’t respond to customer feedback requests it can leave you with a data analysis deficit. Often, in these cases, they aren’t just failing to provide you with reliable data., they’re actually signaling a disengagement with you that is itself a potential churn red flag.  


Predicting anything accurately is very difficult to get right. To build a model for customer churn prediction you need lots of customer behavior and feedback data that’s been repeatedly observed. Then you need lots more real-time data to check against threshold triggers when tracking individual customers. If you can find ways of getting to these customers early enough with the right action, you stand a better chance of making a difference. Don’t forget that customer feedback can support what you do next with individual customers you suspect might churn, as well as feeding valuable intel into an improvement program that can make all customers more loyal.

Finally, if you can build predictive analytics for customer churn then you’re well on your way to predicting other aspects of customer behavior too. The same underlying systems and sources, combined with similar probability calculations and data science, can help you identify who’s likely to upgrade, downgrade, refer or recommend.

Get instant feedback to help stop customer churn

Start getting engaging feedback from your customers in minutes. Our free trial should give you plenty of opportunities to experiment with fast, effective feedback surveys. We integrate with practically every platform and generate some of the best response rates in the business. Simply fill out the form below to get more feedback from your customers.