Churn prediction for a foodtech subscription platform

The client is a UK meal-subscription platform in a crowded, promotion-heavy foodtech market where retention economics make or break the business. Acquiring a subscriber is expensive, so keeping them is where the margin lives.

They had rich behavioural data (ordering, delivery, and support signals) but were not yet using it to intervene intelligently, so retention spend was applied bluntly and much of it was wasted.

Retention offers went to every subscriber or none, spraying discounts at people who were never going to leave while missing those quietly on their way out. The blunt approach eroded margin without meaningfully moving retention.

The business could not tell who was actually about to cancel, so marketing spend and margin both leaked. Without a reliable read on individual risk, every retention decision was a guess dressed up as strategy.

Leadership had seen enough dashboards and wanted a rigorous, measurable read on retention, something that would demonstrably change behaviour and pay for itself, not just look busy in a report.

We built a churn model on ordering, delivery, and support signals, retrained regularly and scored daily so the picture stayed current as customers' behaviour shifted. A stale model is worse than none, so freshness was a design goal.

Rather than stop at a risk score, we surfaced a next-best action in-app for at-risk subscribers, turning prediction into intervention at the moment it mattered. The model's drivers were made legible, so the retention team could understand and trust why a subscriber was flagged.

Crucially, we measured the whole programme against a held-out control group, so the reported lift reflected genuine behaviour change rather than seasonality or wishful attribution. That discipline is what let leadership trust the number.