churn prediction models Customer churn is a realistic part of doing business – but much of it is preventable when you have the right insights. While B2B attrition rates may vary across different SaaS and technology companies, one thing remains the same:  predicting why and when customers might leave empowers your business to take action to preserve those relationships.

That’s where churn prediction models can be a critical asset.

B2B customers often make a substantial investment in tech solutions and subscription-based services, and they need to see ongoing business value. Various pain points could trigger them to exit, and at various points in the customer journey. Is onboarding too challenging? Performance too low or support too slow to respond? Maybe they’re frustrated with poor communication or complicated functionality. Predictive modeling for churn prediction can help you identify the risks of attrition for segments or individual customers.

So how does it work, and is it right for your business? Let’s unpack it with a quick primer on important things to know that can help guide your decisions.

3 Key things to know about churn prediction models

1 – How can you reduce churn with predictive modeling?

Looking at churn after it happens may point to customer experience issues that need fixing – but you want to spot those indicators before you lose customers. Churn prediction models use machine learning (ML) to uncover patterns in your customer data that can provide early warning signals. From there, the system can predict future behavior and the likelihood of attrition.

It’s all about unlocking value from your customer data such as demographics, historical activity, service interactions, and transactions. For predicting churn, data scientists often start with a baseline ML algorithm like logistic regression or decision trees, and “use that model’s performance as a metric to compare the prediction accuracy of more complex algorithms.”1

Beyond predicting the who or when of customer churn, predictive models like decision trees also help the business understand why people might leave. And depending on where people are in the customer lifecycle, different models and datasets may generate better actionable insights.

What does the workflow look like? Here’s a quick snapshot… First, data scientists define what data to include. Then, selected data is prepared, preprocessed, and transformed in a form suitable for building machine learning models. Next, they identify the best methods to train and fine-tune the models. Once they determine which models make predictions with the highest accuracy, analysts can put them into production.2

2 – What are the business benefits?

Reducing attrition is the most obvious win from using churn prediction models. Yet gaining deeper insight into customer behavior can deliver other advantages as well. For example:

  • Optimize engagement strategies. Predictive modeling can help customer success and sales teams better understand which customers to engage and when. You can then more effectively tailor communications to address needs and concerns based on where they are in the customer journey, relationship lifecycle, etc. For example, is it better to “not poke the bear” or reach out to a customer proactively?
  • Improve customer experience and risk management. Identifying why customers are unhappy can inform you about areas for improvement in your products, pricing, or processes. The more you know, the more you can minimize risk.
  • Increase cross-sell/upsell effectiveness. Behavioral insights enable you to better understand customer preferences and expectations, so you can customize offers and pricing strategies to better align with what people want.
  • Improve forecasting. Predicting churn can be valuable for budget planning, as it helps you factor in potential revenue losses due to attrition.

3 – What are the challenges and how do you address them?

Not all churn prediction models are created equal. The level of accuracy and effectiveness depends on many factors related to your data and your process. To help you minimize challenges, here are some important considerations:

  • Develop a robust modeling dataset. For many companies, data is distributed across silos and it can be difficult to rein it all together. Ideally, your predictive models should include data that reflects transactions, billing, service quality, product/service usage activity, and more. These days, it is less about the volume of data. You have to compile and stage the right data to enable an exploration of the potential hypotheses.
  • Define clear goals for modeling. Churn modeling can focus in different directions to answer different questions. For example, models can calculate how many customers are churning, forecast future churn rate, or predict the risk of attrition for specific customers. Clearly defining your goal will help ensure you use the right models and data to get the answers you need.
  • Optimize more than one model. A single model is often not enough to provide deep context about why specific customers might leave, which at-risk customers are more valuable, or how much time before potential churners might leave. Multiple models give you more nuanced insights to help you prioritize and strategize engagement for the greatest retention impact.
  • Keep models fresh. Older predictive models may not reflect recent changes in market trends, new competitors, and how customer expectations are evolving over time.
  • Access the right expertise. To increase the accuracy of churn prediction, it’s valuable to work with data scientists who have a strong blend of domain knowledge, ML modeling skills, and business strategy expertise to build creative, sophisticated models that drive better results.

Often people want to jump right into making models. However, the stages of data curation and data preparation are key success factors in the process. As Beyond the Arc data scientist Bruce Johnson notes, “the process of creating the first attrition models usually reveals certain pain points for the business. All of the idiosyncrasies in how people enter data into their systems come to light. This is why up to 80% of the effort in machine learning goes toward selecting and preparing the data.”

One final thought… There’s a misconception that since attrition is a common problem, a generic attrition model can be used.  But that’s not how it works. Each company serves a different mix of customers and has unique business practices. To predict the behavior of your customers, you need models that are tailored to your business.

How Beyond the Arc can help —Machine Learning churn prediction models

Our data scientists are passionate about using machine learning to help companies solve problems and take action.

Interested in learning how to deploy predictive modeling to reduce attrition? Let’s start a conversation >



1,2   Customer Churn Prediction Using Machine Learning, KDNuggets, May 2019

Johnson Controls is using AI to reduce churn and identify over $100M a year of protectable revenue


Johnson Controls had no usable data sets, no data science team or data engineers.

How could they rapidly build a global data team with new AI/ML capabilities to improve business outcomes on a major scale?

Is your company struggling with how to implement predictive analytics?