How to Predict Customer Churn Without Coding
What Data You Need
The foundation of a churn model is a dataset of past customers where you know the outcome: did they cancel or did they stay? Each row represents one customer, and you need columns that describe their behavior before the outcome was known.
Useful features for churn prediction include:
- Account age: Days or months since the customer signed up
- Recent activity: Days since last login, last purchase, or last interaction
- Usage metrics: Logins per week, features used, pages viewed, actions taken
- Spending patterns: Monthly spend, spend trend (increasing, flat, declining), average order value
- Support interactions: Number of support tickets, unresolved issues, complaint count
- Engagement: Email open rate, response to campaigns, feature adoption rate
- Plan or tier: Which pricing plan they are on, number of upgrades or downgrades
- Target column: A "churned" column with values like "yes" or "no" (or 1/0)
Export this data from your CRM, billing system, or application database. Combine multiple sources into one spreadsheet if needed. See How to Prepare Your Data for Machine Learning for detailed preparation steps.
How to Build the Model
Pull at least 500 past customers (both churned and retained) with the features listed above. More is better, but 500 well-prepared rows is enough to get started. Make sure both categories are represented, aim for at least 100 churned customers in the dataset.
Save your data as a CSV file and upload it through the Data Aggregator app. The platform reads your columns and lets you select which one is the target (the "churned" column) and which are the input features.
Random forest is the best starting point for churn prediction because it handles mixed data types, captures non-linear relationships, and resists overfitting. Start training and the platform handles the rest. Training time depends on data size but typically takes seconds to minutes.
After training, review the accuracy metrics. For churn prediction, pay special attention to recall (what percentage of actual churners did the model catch) and precision (when the model says someone will churn, how often is it right). A model that catches 70% of churners with 60% precision is useful, you catch most at-risk customers while accepting some false positives. See How to Test Model Accuracy.
Export your current active customers with the same columns (minus the "churned" column) and send them through the model. Each customer gets a churn probability score. Sort by score to see who is most at risk. This prediction step costs zero credits per customer.
What to Do With Churn Predictions
A churn prediction on its own does nothing. The value comes from the action you take based on it:
- Proactive outreach: Have your support team reach out to high-risk customers before they decide to leave. A simple "Is there anything we can help with?" email sent at the right time can prevent cancellations.
- Targeted offers: Send discount codes, extended trials, or upgrade offers to customers flagged as at-risk. The cost of a discount is usually far less than the cost of acquiring a new customer.
- Product improvements: Look at what features the highest-risk customers are not using. This reveals gaps in your onboarding or product that cause people to leave.
- Resource allocation: Assign your best account managers to the highest-value at-risk accounts where saving the customer has the biggest revenue impact.
You can automate these actions using workflow automation. Set up a chain command that runs churn predictions nightly, filters for high-risk scores, and triggers appropriate follow-up actions automatically.
Example: SaaS Subscription Churn
A SaaS company exports 3,000 customer records from the past two years. Features include monthly active users per account, feature usage breadth (how many distinct features they use), support ticket count, billing plan, account age, and whether the account cancelled. They train a random forest classifier and achieve 75% recall with 65% precision.
They score their 800 active accounts. The model flags 120 as high churn risk. The customer success team contacts the top 30 highest-value flagged accounts. Within a month, they save 12 accounts that were actively considering cancellation, recovering an estimated $144,000 in annual recurring revenue. The entire ML pipeline cost less than $5 in training credits.
Predict which customers are about to leave and take action before they cancel. Train a churn model on your own data today.
Get Started Free