Predicting Customer Churn Using Data Science

Odds are that you have some sort of monthly subscription or membership, whether it's with Netflix, the gym, or something you signed up for 10 years ago and have completely forgotten about. You've probably also done something called churning at one point or another, meaning that you canceled your subscription.

When thinking about customer relationship management, the last thing firms want is for a customer to churn. As my professor Daniel M. Ringel stated in last week's lecture, acquiring a new customer is seven times more expensive than retaining an existing customer, which explains why predicting and preventing churn is so important. To predict churn in my data science class, we applied our knowledge of supervised machine learning, which is when we teach a machine to predict outcomes by using a set of "labeled" data that acts as an example to imitate. Within the realm of supervised machine learning, firms today largely use models known as classification models, which help reach common yes/no conclusions (e.g., whether a customer will churn or not). Sure, we can visualize our data to observe trends and reach conclusions about the past, but a classification model is what enables us to predict future customer churn based on data that the machine hasn't yet seen.

Beyond that, I learned that data processing pipelines are essential to efficiently compare various classification models and, ultimately, predict customer churn as accurately as possible. Thanks to pipelines, we can rapidly load and clean data, construct new variables (a process called feature engineering), fit our models, and compare accuracies--and best of all, we only have to call the code once. With a few simple clicks, we easily compared how well a number of classification models--including ones called Logistic Regression, Support Vector Machine, and Random Forest--could predict the bank's customer churn. Finding the most accurate model enables the bank to identify its on-the-verge customers and then take the necessary steps to convince them to stay. In the end, the bank retains more clients and the customers get extra incentives. Sounds like a win-win to me!

I'm excited for potential insights that machine learning pipelines will offer beyond this class in my future role as a Digital Marketing Analyst. Classification models like these not only predict churn, but also analyze other yes/no questions I'll encounter on the job, such whether a customer will respond to an ad campaign or click on a Google result. It's been so cool to learn about these interactions between data science and marketing so far this semester, and I can't wait to continue learning more!