Rethinking Churn Prediction: From Traditional Machine Learning to Tabular Foundation Models

Many companies treat churn like a customer-support metric, something Customer Success owns and reports next to Net Promoter Score and ticket volume in a quarterly review. But they are missing the cost of churn and how it shows up in revenue forecasts.

Customers rarely announce that they're about to leave. By the time it shows up as a number in a board deck, the value is already gone.

What is churn? Churn is the rate at which customers stop doing business with you in a given period. But "churn rate" can mean three different things: how many accounts left (customer churn), how much revenue left with them (revenue churn) and whether expansion from existing customers offset the losses (net vs. gross revenue churn).

For a CFO, the problem looks different. Churn shows up in the growth numbers. Every customer acquired comes with an expected lifetime. If that customer leaves earlier than expected, part of that revenue is gone.

Churn Is a Financial Forecasting Problem

Leaders track customer churn because it's the easiest metric to measure. A churn rate alone doesn't say which customers left or how much revenue went with them. McKinsey's 2016 analysis of nearly 200 growth-stage SaaS companies found that the strongest growth performers were simply the ones losing less revenue to customer attrition. A decade later, ChartMogul found the same pattern.

Churns belongs on a CFO's desk. It quietly degrades the accuracy of everything built on top of a revenue forecast, cash planning and hiring plans included.

It's just as much a CDAO's problem. The data that would explain a slipping forecast (billing history, usage logs, support tickets) is already sitting somewhere in the business. Some teams have a model running to help predict and manage it; others have the infrastructure to build one. Their job is to catch churn as it happens, not months later.

Not All Churn Is the Same

Some churn has nothing to do with performance. A meaningful share of SMB churn simply comes from small businesses going out of business. No retention program fixes that. Other times, customers just leave.

To manage churn effectively, you have to break it down:

• Unavoidable churn is the business closing or the need genuinely disappearing. No retention offer will change that.

• Involuntary churn is a payment failing, a card expires, a charge bounces. It's the easiest to fix, since it just requires catching the failure before it becomes a cancellation.

• Voluntary churn is a customer deciding to leave (dissatisfaction, a competitor, a price increase, quietly declining engagement).

A good model finds the churners you can still save.

What Losing a Customer Actually Costs

HubSpot saw monthly MRR churn fall from 3.5% to 1.5% over five quarters, back in 2011-2012. Expected customer lifetime rose from about 29 months to 67 months. CAC didn't move. Churn decided how much of that spend got paid back.

Net revenue churn above 2% a month is a warning sign. At that rate, a company loses roughly 22% of its revenue base over a year. At a million customers, 3% monthly churn means losing 30,000 customers every month.

Convincing numbers. But they don't explain why so many business teams still catch churn late.

Finance Teams Catch It Too Late

Knowing churn is expensive doesn't make it easier to prevent. The problem is timing. By the time finance sees churn in the numbers, the window to act is usually gone. The issue isn't a lack of retention programs but reaching the customer early enough.

Treating every customer as equally at risk gets expensive fast, a discount handed out indiscriminately across the whole base. Waiting for the actual cancellation is worse, the moment's already passed by then.

What's left is a short window, between the first warning signs and the cancellation, while the customer is still persuadable.

What Actually Moves the Number

A churn score sitting on an unread dashboard has zero business value. What works instead: identify at-risk accounts, intervene, and measure the impact.

McKinsey documented a telecom operator that used targeted analytics to issue retention offers, cutting churn by 10-15% over 18 months.

A peer-reviewed study tested the same approach across 7.5 million telecom customers. Predicted churners were split into two groups: one received a retention deal, and one was left alone. About 47% of those given an offer stayed, cutting overall churn by roughly 1.5 percentage points.

Prediction tells you who. The offer and the measurement move the needle.

Where Prediction Gets Hard

Churn is hard to predict. Operational data is scattered across billing, usage, support, and sales, meaning different teams view the same customer through completely different lenses.

On top of that, actual churners are rare. The model needs to rank a small group of likely churners correctly. By the time a cancellation is logged, the window to act has usually closed, because the behavior that predicted it happened weeks earlier.

The signals change as the product, pricing and competitive market change. Every retraining cycle takes time. Meanwhile, the old model quickly goes stale.

The two mistakes also have different costs: missing a churner costs a full customer's remaining value, while retaining a loyal one costs a discount. Getting that trade-off right is a business decision as much as a modeling one.

Not an Algorithms Problem

Most enterprises with a data team already have some kind of scoring model in place, whether they built it themselves or use a standard machine learning approach. Building the first model usually isn't the hard part. Keeping it delivering is.

Research on production machine learning shows that teams spend far more time preparing data, maintaining pipelines and adapting models to changing conditions than they do building new predictive workflows.

A new pricing tier, a product launch, a shift in customer mix or a data schema update can all trigger another round of work. The work might mean retraining, rebuilding feature pipelines, revalidating and redeploying, then monitoring the result. That work multiplies fast once a business is running more than one of these models.

Tabular foundation models are designed to reduce that work.

Enter the Tabular Foundation Model

This is where tabular foundation models come in.

With traditional ML, teams usually build and tune a model for each prediction problem. A tabular foundation model is pretrained across many datasets. Instead of training a new model from scratch, it uses the new dataset as context and makes predictions directly (a process called in-context learning).

For churn, this matters because the data changes as the business changes, and the model needs to keep up. A foundation model gets from new data to a useful prediction faster, without losing accuracy.

In practice, most teams aren't choosing between a foundation model and a perfectly tuned one. They're choosing between having a model now or waiting months to get one into production (or sometime never getting one built at all).

TFMs change the modeling cycle. Teams no longer need to rebuild a model every time the data changes. That's the difference between catching churn in time and finding out too late.

The Bottom Line

Churn is one use case. The overhead behind it is the general point.

Every prediction problem in an enterprise carries the same overhead: prepare the data, build the model, validate it, deploy it, then maintain it as conditions change. That overhead is why most prediction problems never get a model. The ones that do are the ones big enough to justify the work.

Tabular foundation models lower that cost. Fraud, credit, pricing, demand, inventory all run on the same tabular data, sitting in the same systems.

When engineering capacity is no longer the bottleneck, the question left is which decisions to accelerate first.
‍

Learn more about Seldon, Neuralk's tabular foundation model.
If you're exploring how TFMs can fit into your AI strategy, get in touch.

‍‍

References

• Allen Miller, Ben Vonwiller and Peter Weed, "Grow fast or die slow: Focusing on customer success to drive growth" McKinsey & Company, October 26, 2016.

• David Skok, "SaaS Metrics 2.0 – A Guide to Measuring and Improving What Matters"

• A customer's expected lifetime, in months, is roughly one divided by the monthly churn rate.

• Pallav Jain and Kushan Surana, "Reducing churn in telecom through advanced analytics" McKinsey & Company, December 12, 2017.

• Abdelrahim Kasem Ahmad, Assef Jafar and Kadan Aljoumaa, "Customer churn prediction in telecom using machine learning in big data platform" Journal of Big Data.

• Kyle Poyar (Analyst-in-Residence, ChartMogul), "The SaaS Retention Report: The AI Churn Wave" ChartMogul, 2025-2026.

• "A Study on the Importance of and Time Spent on Different Modeling Steps"

• "Hidden Technical Debt in Machine Learning Systems" Advances in Neural Information Processing Systems 28 (NeurIPS 2015).

• "Accurate predictions on small data with a tabular foundation model", Nature‍

• TabBench V2: The new version of the open evaluation suite for tabular classification, Neuralk