What is churn prediction?
Churn prediction is the practice of using customer behavior signals to forecast which customers are likely to stop buying — before they actually leave. Rather than analyzing past departures after the fact, a churn prediction model scores every active customer on their current probability of churning within a defined window (say, the next 30 or 90 days), so a marketing or retention team can act while there is still a relationship to save.
Why it matters#
Acquiring a new customer costs more than retaining an existing one — that's the starting point for most retention programs. But the more actionable version of that argument is about timing. By the time a customer's purchase frequency visibly drops or they unsubscribe from your emails, they have often already mentally moved on. Churn prediction tries to catch the signal a few weeks earlier, when an outreach is more likely to land.
The other reason churn prediction has become more practical is data. Ecommerce stores now collect dense event streams — page views, search queries, cart adds, review reads, support contacts — not just order records. That behavioral richness is what separates a real churn signal from a transaction count.
What the signals look like#
Churn is not usually a single event; it accumulates across several dimensions:
- Declining engagement. A customer who used to open emails and browse weekly and now does neither is showing early-stage disengagement, even if their last order was recent.
- Dropping purchase frequency. A previously high-frequency buyer whose inter-purchase interval grows is a classic at-risk pattern. RFM models catch this as a falling Recency score.
- Browse-without-buy behavior. Customers in the consideration phase for a competitor sometimes continue browsing your store — product views without add-to-cart events, repeated visits to the same product that never converts.
- Support contacts. A dissatisfied customer who contacts support and doesn't reorder within a typical cycle is at elevated risk.
- Session recency collapse. No sessions in the last 30 days from a customer who previously visited weekly is a hard signal regardless of whether they've placed a recent order.
No single signal is reliable on its own. Effective churn models combine these dimensions into a composite score — and weight them differently by customer segment, average order cycle, and product category.
Predictive vs. reactive#
The traditional approach is reactive: export a list of customers who haven't purchased in 90 days and send a win-back email. This is better than nothing, but many of those customers churned 60 days ago. A predictive approach scores customers on an ongoing basis and triggers outreach when the probability crosses a threshold — ideally while there's still active engagement to build on.
| Dimension | Reactive | Predictive |
|---|---|---|
| Timing | After churn is confirmed | Before churn completes |
| Data used | Order history only | Event stream + purchase history |
| Trigger | Fixed inactivity window | Probability threshold |
| Audience quality | Includes already-gone customers | Focused on recoverable customers |
| Refresh cadency | Batch (weekly or monthly) | Continuous or near-real-time |
The action gap#
The hardest part of churn prediction isn't the model — it's closing the gap between a score and an action. A model that flags 5,000 at-risk customers is only useful if that list becomes a segment that flows into an email tool, an ad audience, or a proactive outreach campaign. The three failure modes are: the score lives in a data warehouse but never reaches a marketer; the audience is too broad to treat meaningfully; or the outreach arrives too late because the scoring runs weekly instead of daily.
Practical programs usually define two or three at-risk tiers — for example, early warning (probability 30–60%), high risk (60–80%), and critical (80%+) — with a different intervention per tier. Early-warning customers might receive a personalized recommendation email; critical customers might get a direct outreach with a targeted offer.
What good looks like#
A working churn prediction setup has four parts:
- A continuous event stream with session, engagement, and purchase data — not just order exports.
- A scoring model (rules-based or ML) that refreshes frequently enough to catch rapid behavioral shifts.
- Segments that translate scores into addressable audiences automatically.
- Activations — connections to email, SMS, or ad platforms — so the audience reaches a channel without a manual export step.
The last two are where most programs stall. Scoring a customer's churn risk is a data problem; acting on it in time is an operations problem.
How SegOps fits in#
SegOps Customer Intelligence lets you build churn-focused segments that mix behavioral signals, purchase history, and predictive operators — including propensity and LTV scores — in the same rule. A segment like "propensity-to-churn above 0.65 AND no session in the last 21 days AND lifetime orders ≥ 2" is expressible directly in the rule builder or by describing the audience in plain English to the AI segment builder. Segments sync continuously to connected channels via Activations, so there's no manual export step between a score update and an outreach.