A churn model can only be as useful as the event it predicts. For a monthly subscription, cancellation may be explicit. For a marketplace or service app, silence is harder to interpret: a customer might simply have a long natural gap between purchases.
Start with the decision
Ask what your team can do differently if a customer is identified as at risk. The useful prediction window must leave enough time for that action. A support call may need days; a product journey change may need weeks.
Use behaviour and business context
Compare candidate inactivity windows against historical return patterns, customer value and seasonality. Malaysian businesses should also consider local festive periods and pay cycles where they materially affect use. Document exclusions, grace periods and reactivation rules.
Validate with people
Review examples with product, CRM, support and finance teams. If the label repeatedly classifies healthy customers as lost, revise it before training a sophisticated model. A plain baseline built on a trusted outcome will outperform an impressive system pointed at the wrong target.
This article provides general information, not legal, regulatory or professional advice for a specific organisation.