The problem statement
At the heart of what we as corporations/businesses do is a cycle of acquiring, retain and grow customers. In simple terms we are interested to predict the following:
- Which of my customers are most likely to be active in the future
- the levels of transactions we could expect in the future periods from my individual customers and groups of customers
- the expected/predicted LifeTime Value (LTV) from my customer
An important point here is that we are looking into the crystal-ball and asking these questions of our data. It is not a posterior or a hindsight view but a forward-looking perspective.
To support this analysis we need to pay attention to churn. At what rate do we expect our customers to engage with our product:
- how long will they remain active or alive ?
- will they abandon us altogether?
- will they engage in a new transaction?
- Will they renew?
And of course, the behavior of our customers is influenced by the business setting. Fader et al have done a great job of classifying the business into the following quadrants:
types of relationships that businesses have with customers
Continuous refers to customer transactions that have a high frequency associated with them. Things like buying coffee, grocery shopping, credit cards, utilities. Discrete have a lower frequency of interaction between the customer and the brand. Think of these as one-time events such as hosting a trade-show.
These transactions can be categorized as either contractual or non-contractual. These are self-explanatory but to recap:
- contractual refers to an explicit contract between the customers and the company. the contract life could be monthly, annual, etc.
- The non-contractual relationships refer to the ones where the users can come in and out of the "relationship"
With that out of the way, our approach will need to take into account the business context and apply it to the churn analysis.
Churn analysis should also be at the heart of retention management programs . It is recommended that retention management programs have the following steps:
- Who is at risk of churning?
- Why? What are the underlying causes?
- Whom to target? Not all churners are good targets for retention programs.
- When to target? (pro-active/reactive)
- What are the interventions?
- Are my retention programs delivering value?
This post focuses on the predictors of churn. I will be writing follow-up posts that cover other aspects of retention management.
Predictors of Churn
So how can we predict which of our customers are going to churn? The good news is that to date that there has been a rich body of work to analyze whether a customer will churn.
Some of the predictors of churn are:
- transactions (recency, frequency, the time between transactions  and monetary value)
- demographics (who)
- customer satisfaction (NPS scores, emotions, attitudes)
- psychographics (preferences and lifestyle behaviors)
- product and brand engagement (in product usage, time spent on the site, product journeys)
- outreach responses (marketing, offers, promotions)
- barriers to switching (how easy or difficult is to switch, auto-deduction, etc)
- social connectivity etc
Also note that an important difference between contractual and non-contractual models is that unlike contractual relationships, non-contractual relationships do not have an explicit termination of relationship signal but more of a slow diffusion. Researchers refer to this as latent attrition.
As you read these predictors, I wanted to stress that prediction models need care and analysis. They cannot be a set it and forget it. This is true for the following reasons:
- predicting churn is an evolution and requires constant analysis. It can start with the transactions and you can layer on additional attributes.
- prediction models assume that the world around them has not changed. So doing a reality check becomes an important part of the exercise.
Last but not least, the addition of attributes should take the business case in mind. So start by asking the question: how would this work of adding attributes contribute to my expected LTV?
Before we get into prediction models, it is important to get a few things out of the way:
- regardless of the jargon and terminology, the building blocks of models are numbers
- we have the luxury of using multiple models, the art is picking the ones that work for our situation
- Not all models need Machine Learning or Deep Neural Networks or fancy techniques. Prediction models can be done in Excel. We don't always need Neural Networks.
- Repeat: it boils down to numbers. As a general rule of thumb, you cannot model what you cannot translate to a number.
Now that is out of the way, that prediction models can be built using multiple approaches:
- Deterministic (where we do not assume randomness)
- Probabilistic or stochastic
- Machine learning
- Deep Neural Network Models (DNN)
Examples of such models are:
- Probabilistic: Beta Geometric models (BG), Boosting and Bagging (BnB), Negative Binomial Distribution (NBD), Pareto -NBD, Beta-Binomial
- Machine learning: Simple Vector Machines (SVM), Random Forests, Decision Trees, Neural Networks, etc..
What does that mean for you, a non-data scientist?
I am not sharing these models to showcase my model regurgitation skills but to highlight a few points that have helped me. Look at the rich choice of algorithms as an opportunity, not something to be intimidated by.
- this rich choice gives us an array of tools that can be applied at different stages of the analytics journey. For example, we can start with the basic attributes such as RFM (Recency, Frequency and Monetary Value) and perhaps use Excel, R or a Python library to get started.
- It gives us an opportunity to evaluate the missing pieces and create an actionable roadmap. A roadmap can include sharpening the models in production or introducing new layers.
- Last but not least, It also gives us a cost/benefit analysis decision-making framework. The benefits of building, deploying and maintaining models should exceed costs.
The next blog post will cover the causes of churn. Not just predictors of churn.