Predicted churn-risk winback

By the time a subscriber clicks cancel, the decision was made weeks earlier. The cancel button is the last step of a disengagement that showed up in the data long before, as sessions that got shorter, features that stopped being opened, emails that stopped being read. Most retention programs miss all of it and fire their save attempt at the cancellation screen, which is the worst possible moment: the customer has already rationalized leaving, and a discount offered there reads as a too-late bribe rather than a reason to stay.

The harder problem underneath is that disengagement is an absence. Stacks are built to capture events that happen, and a customer drifting away is defined by events that stop happening. A churn model that only sees positive activity is blind to exactly the signal that matters, and a winback program that waits for an explicit cancel intent has already lost the window where intervention is cheap and credible.

This recipe surfaces the customer while they are still quietly disengaging and routes them into a winback flow before the cancel decision hardens.

Intervene during the disengagement window rather than at the cancellation screen, and save a meaningful slice of subscribers who would otherwise have churned silently. Realistic effect: a reduction in voluntary churn somewhere in the low-single-digit to high-single-digit percent range of the at-risk cohort, depending heavily on how much of the churn is addressable (price-driven churn responds to different levers than value-driven churn) and on whether the product can actually deliver a reason to stay rather than just a discount. Measure it with a holdout inside the at-risk cohort; without one, a win-back program takes credit for customers who would have stayed anyway. The honest framing is that this recipe moves the addressable portion of churn, and a chunk of churn is not addressable by marketing at all.

The second-order outcome is learning which disengagement signals actually precede churn for this specific business, which sharpens every retention decision downstream.

A modeling environment for the churn score (a warehouse-native model in dbt plus a Python or SQL scoring step, or a dedicated predictive layer in the CDP) and an activation layer that can route customers by risk band into differentiated flows. Composable stacks suit this because the score lives in the warehouse where the engagement history already is, and reverse-ETL pushes risk bands to the messaging tool. The capability that decides success is refresh cadence: a churn score recomputed weekly catches drift that a quarterly score misses entirely, and the right cadence is set by how fast customers in this business actually move from engaged to gone.

Work in this order. Signals against each customer's own baseline first, the score second, routing third, and the loop closed at the end.

1. Capture negative signals against a per-customer baseline: declining session frequency, features going untouched, falling email opens.
2. Build the churn-risk model and select predictive features.
3. Refresh the score at the speed customers leave.
4. Route by risk band, with hysteresis, so a customer at the line is not pulled in and out of the flow.
5. Scale incentive depth to risk and value. A discount for everyone above the threshold teaches profitable, slightly disengaged customers that drifting earns a price cut. Early disengagement usually gets a nudge or a value reminder, not money.
6. Exit the flow on re-engagement.
7. Feed outcomes back into training: saved, churned anyway, never at risk. It is the most skipped step, because it is invisible until the model stops working.
8. Hold out a randomized group for the true save rate. Without it, customers who would have stayed anyway count as wins.

Negative signal capture covers step 1, inferred attribute generation steps 2 and 3, threshold-based routing steps 4 to 6, and outcome feedback into model retraining steps 7 and 8.