Are predictive churn models a waste of time?

You might think this is an odd statement from someone who has spent much of their career advocating the benefits of predictive modelling. I am not, however, suggesting that you should abandon predictive modelling. Instead, I believe that they can be used much more effectively, in other ways then simple churn or attrition models.

 What other ways, though?

Let me explain. Most people would probably agree that most industries share a core business goal of maximizing shareholder value, or something similar. It does not really matter whether the industry is telecommunications, banking, insurance, retail, or any other B2C business. To a greater or lesser extent, retaining customers, or not losing them (churn), will drive that core business goal. I am not convinced, however, that knowing a customer’s probability of churning—the outcome of a predictive churn model—is actually very valuable.

Knowing that a particular customer has x% chance of churning only really leads to more questions: either ‘why?’, or ‘what should we do about it?’ These questions apply whether x is zero, 100% or any point in between.

Always start with ‘why’.

The ‘why?’ question is interesting. Root cause analysis or customer profiling could lead to action to correct technical problems or operational issues. It could also allow marketers to create an offer to encourage that customer to stay longer, renew a service or upgrade a product. The follow-up question, ‘what should we do about it?’, however, isn’t answered at all by knowing the churn rate. At best, the answer is ‘something’ or ‘nothing’, but not what.

I think it is more productive to spend time and effort answering this follow-up question. This will add much more value to a business, and this is where predictive modelling is far more worthwhile. Models can help to identify the expected customer response and the expected value of that response to each possible offer. This makes it much easier to decide what you should do for a particular customer.

 In reality, most potential offers to a customer have already been created. They often just get more attractive when a customer is more likely to churn, (which is another potential waste of resource). The only decision to make is which offer to use. If you know a customer’s response and expected revenue for each one, then notionally this becomes easy: you just pick the biggest likely return.

Here comes the arbitration

This challenge can be far more analytically interesting. You have to arbitrate effectively across multiple offers, driven by multiple models, and model accuracy becomes a challenge. There is a danger that models for similar offers will be too similar - and it will become too hard to arbitrate, or there will be no uplift between the similar offers.

This challenge can be met, (potentially with more models on segmented customer bases), or perhaps with new techniques, such as attribution, net lift modelling, and optimisation. None of which necessarily require more effort.

Ultimately, and crucially, a different approach to predictive modelling will drive more business value than simply working to get a better churn model, whatever algorithms or techniques are used.

 Learn more at MWC19

If you are going to MWC in Barcelona, and want to discuss this or other Customer Intelligence issues, why not stop by the SAS booth? My colleagues and I will be running a number of analytics clinics. Better still, book an appointment so you are assured a good customer experience!