How to Build a Customer Churn Prediction Model

Since January 2024, we've observed a growing interest among customers and prospects in solutions addressing customer churn. As organizations become more familiar with the benefits and applications of ML and AI, Customer Success leaders are increasingly seeking ways to incorporate AI into their existing processes.

In the current economic climate, retaining existing customers has become increasingly critical, particularly in the SaaS industry. According to Pavilion's “Benchmark It” report in 2023 Sales  the median cost for a B2B company to acquire $1 in ARR was a whopping $1.76. 

High churn rates can negate this substantial investment if customers discontinue service before the company recoups these costs. Research by Bain & Company demonstrates that a 5% increase in customer retention can lead to a 25% to 95% increase in profits. Churn remains a key factor for company profitability and a common denominator in the customer lifetime value equation.

To combat churn, companies must adopt a proactive approach, identifying signals and indicators of customer satisfaction and potential churn. This blog post will guide you through the steps necessary to design and build a Customer Churn Predictive Model, whether or not you're using Forwrd.ai.

‍

Why Focus on Churn?

Customer churn represents the percentage of customers who stop using your product or services over a specified period. Many organizations rely heavily on Customer Success Manager (CSM) sentiment, which can be biased and inaccurate. They may also use basic product usage metrics, such as weekly logins, which provide an incomplete picture.

‍

‍

By leveraging ML and AI, you can utilize all your CRM data and advanced product metrics to build a more accurate Customer Churn Predictive Model. The benefits of accurate churn prediction include:

• Increased retention and revenue protection
• Focused attention on the right customers
• A better understanding of what success looks like

‍

Where to Start While Building a Customer Churn Model?

1. Identify Accounts for Model Training

Begin by identifying the accounts that the model will be trained on. These should include examples of successful accounts, such as those that are currently active and have renewed at least once. Additionally, provide examples of churned accounts to train the model on negative outcomes.

2. Leverage All Your Data

Your organization likely tracks multiple data points for each account, which you should use to train your model:

• Geographic and Firmographic Data: Include basic information such as industry, location, and current ARR. Also, incorporate contact-level data, like the titles of your champions and which departments use your services.
• CS Engagement Data: Include data on your interactions with customers, such as the date of the last email exchange or meeting, and the number of support tickets.
• Product Usage Data: Go beyond simple metrics like user logins. Your product team likely tracks most user actions, which can be leveraged in your model.

3. Clean and Normalize Your Data

After identifying and accessing your data, it should be preprocessed. Work with your data science team to:

• Remove irrelevant data and errors: For example, handle empty values by replacing them with default values based on business logic.
• Normalize data: Instead of using the actual Account ARR, create ARR ranges and categorize your accounts accordingly. Grouping similar accounts can increase the impact of this data.
• Create cohorts for activity-based data: This helps the model recognize changes in activity over time, especially for CS engagement data and product usage. For example, track email exchanges over the last 90, 60, and 30 days.

4. Analyze Data Correlation

With your data prepared, it's time to identify which data points strongly correlate with customer churn and could be used in the model. Employ various statistical models and techniques during this process.

5. Choose Model Factors

While data correlation is important, it must be balanced with business context. Review the correlated data points and select the factors that should be used. Involving champions from your Customer Success team can provide valuable insights and help gain buy-in from the team.

6. Build Your Model

Your data science team can use various techniques to build predictive models, such as regression, decision trees, and neural networks. Collaborate with them to select the best method for your use case. A good practice is to train the model on historical data (e.g., from 2023 or earlier) and test it against your 2024 data.

7. Evaluate Your Model

Review the model's accuracy and confusion matrix to ensure it performs as expected. If not, you may need to revise the model's factors or choose a different modeling method. Be cautious of overly high accuracy levels, which could indicate model overfitting — reassess your factors if necessary.

‍

8. Validate the Model Offline

Use the model to predict churn likelihood for your current customer base and share the results with the CSM team. This can reveal any biases in the model or provide new insights for the CSM team. Gather feedback and determine if further improvements are needed.

9. Operationalize the Model

For the CSM team to act on the model's output, it's crucial to integrate the results into their system of record, whether it's a CRM or a dedicated CSM platform. Most platforms have APIs you can use to push data from your internal tools. Additionally, ensure the model continuously reviews and updates the results by automating data fetching and model execution with the help of your technical teams. This eliminates manual work and keeps the CSM team informed with up-to-date results.

10. Continuously Monitor and Improve

Finally, consistently monitor the model's performance by comparing its predictions with actual outcomes (e.g., whether the customer churned or renewed). Use this data to retrain and optimize the model as needed.

‍

Conclusion

Using ML and AI to build a customer churn predictive model can significantly impact your business by identifying at-risk accounts and enabling a more proactive customer approach. However, this approach is a large-scale endeavor that may require significant resources from multiple departments.

Leveraging solutions that automate Data Science (ADS) processes, such as Forwrd, can greatly reduce the effort involved and enable you to build a new predictive model in just a few days.

‍