It is just statistics and computer science hiding behind a very charming name. Not sure about you, but I definitely feel like there is something very special about “machines” somehow “learning”.
While it’s no secret that Machine Learning (ML) has potential to help businesses save money and improve their sales, many perceive ML to be too complicated to be practically applied. This means that majority of businesses never get the chance to take advantage of ML models available.
As such, this is a continuation of the series in which we discuss different business applications or Machine Learning. In this review I will be focusing on customer churn modelling.
Customer churn i.e. “Please don’t leave us”
What is it?
Customer attrition is a tendency of customers to stop being a paying client to a business. The proportion of customers lost during a particular time period is referred to as attrition rate.
There are many reasons to monitor and predict attrition rates, for instance:
- Being able to predict future attrition rate helps businesses estimate their expected revenue.
- It is more difficult to attract a new client than to retain an existing one. Identifying customers as being “high-risk” of leaving a business allows us to target that individual in an attempt to prevent them from leaving.
- Targeting only at-risk customers with retention strategies can help save cost and resources.
Wouldn’t it be great if you could know when a customer is about to leave you? You could then send them an incredible promotion and, suddenly, they remember how great and valuable your business is to them. They realise that they are about to make a mistake. They don’t go anywhere. They are with your business forever. Perfect…
While it is clearly not that simple, Machine Learning certainly makes tracking attrition easier. ML can help us gather customer data and identify behaviour patterns that would identify a customer as being at high-risk of abandoning your business.
How does it work?
Considering the obvious business value of making accurate churn predictions, there is a lot of literature available examining accuracy and predictive power of different statistical techniques that can be used.
As with majority of data science tools, there is a simple trade-off between complexity of the model and its relative predictive power. There is little value in using an over-complicated model, if its predictive power is not better than that of a basic model.
i.e. Is all this extra mathematics really worth it?
Previous research has shown that Logistic Regression and Bayesian Network have a consistently high churn prediction rate. Both of these are relatively straightforward to implement and they demonstrate an excellent trade-off between complexity and predictive power.
- Logistic Regression is always a good place to start, as it is essentially an extension of a linear regression, which majority of us will be familiar with. It uses business input and makes a prediction of whether a person is likely to belong to class 0 (non-churn) or class 1(churn) and provides you with statistical confidence interval, giving you probability of the prediction being incorrect.
- Bayesian Networks are a widely-used class of probabilistic graphical models. The algorithm expresses conditional independencies and dependencies among random variables associated with nodes. Importantly, while Bayesian Networks can be prepared by experts, they can also be learned from Data in an unsupervised manner.
Where Can I see it used?
Businesses across many industries are taking advantage of churn prediction. However, this is more predominant in some sectors:
- Telecom services (Internet/Television/Mobile networks). Because of how highly competitive these products and services are, customer attrition is particularly high. This means that revenue can only be maximized through diligent tracking of attrition and targeted customer retention.
- Subscription-based Businesses. Majority of tech giants, such as Apple, Amazon and Netflix still use Machine Learning to monitor at-risk subscribers, to start proactively engaging them with special discounts and offers. For subscription-based businesses it is important to find customers who are at risk early in advance.
Where to start?
The first step in implementing churn prediction is collecting the data. This data can then be used as a base for your Machine Learning predictive model. Quality of the original dataset will largely limit how good predictions made by Machine Learning model will be.
Data collection, preparation and processing is made a lot easier with Microsoft Cognitive Services. And, if your business doesn’t have resources/trained employees specialising in Data Science, partnering with company such as Coeo, who specialise in predictive analytics, can be a good solution.
If you missed it...
This blog is part of our Machine Learning series so if you would like to read the previous blog post then click here.