What happened
In "Your Churn Threshold Is a Pricing Decision," published on Towards Data Science, the author reviewed 36 publicly available analyses of the IBM Telco churn dataset across Kaggle, GitHub, and data-science blogs.
The findings were stark: roughly nine in ten reported only accuracy or F1 score, just over one in seven plotted a profit curve, and none used survival analysis to compute customer lifetime value before setting a threshold.
Every default-threshold model leaves money on the floor — about $86 per customer in avoidable burn on the standard dataset.
The cost of that default is roughly $86 per customer, because a 0.5 cutoff silently assumes the cost of a wasted retention offer equals the cost of losing a customer — which on the standard dataset is wrong by a factor of about 13.
Why it matters
The deeper insight is that churn is an economics problem, not just a prediction problem. Where you set the threshold decides who gets a retention offer, what it costs, and how much revenue you save or burn. Treating that as a default rather than a decision is leaving the pricing of your own retention up to a textbook.
It is a smart reframing. But the analysis targets voluntary churn — the customers weighing whether your product is still worth the money. That is the hard, expensive kind to win back.
What this means for SaaS founders
Before you tune a single threshold, look at the churn you can recover without any modeling at all: involuntary churn. These are customers who did not decide to leave — their card expired, hit a limit, or got declined at renewal. No probability score is needed to know they wanted to stay. They already paid you.
For most subscription businesses, this is a meaningful share of total churn, and it is the cheapest possible MRR to save. There is no offer to discount, no win-back sequence to design, no model to deploy. You just have to catch the failed payment, retry intelligently, and prompt the customer to update their card before the subscription lapses.
Sequence your recovery work
Tune your churn threshold, absolutely — the $86-per-customer point is real. But recover your failed payments first. It is the one slice of churn where the optimal retention offer costs you nothing.
The bottom line
The Towards Data Science piece is a good nudge to stop treating churn as a pure data-science exercise and start treating it as a revenue decision. The fastest version of that decision: plug the involuntary leak before you spend a dollar modeling the voluntary one.
Sources