| Dataset route | What you can honestly build | What remains unproven |
|---|---|---|
| IBM Telco sample | A classification pipeline and leakage audit; inspect which version and fields you downloaded | A single customer snapshot is not a historical scoring ledger. Supplied churn scores, reasons and post-outcome statuses must not become predictors. It cannot establish offer effects |
| UCI Iranian Churn | A small tabular comparison with documented target and feature provenance | Without reconstructable decision timestamps and follow-up, it does not prove future-cohort validity. Subscription length is not a recorded churn time; do not invent survival durations |
| UCI Online Retail | Monthly transaction-based snapshots, RFM features, a stated inactivity proxy and a time backtest | Inactivity is not confirmed termination. Clean cancellation invoices and verify usable customer identifiers. The finite record limits matured cohorts and horizons; no campaign assignment is recorded |
| Your permissioned subscription and event history | A realistic risk/survival backtest, CRM integration and delayed-outcome monitoring | You still need an intervention study, or explicitly defensible causal assumptions, to claim retention benefit |
| Clearly marked synthetic campaigns | Exercising assignment logs, delayed-label logic, policy allocation and rollback | Synthetic effects come from your generator; they are not evidence of real model uplift or profitability |
IBM documents a fictional-company sample and separates provided churn labels from supplied churn scores and churn reasons, which makes it useful for a leakage exercise. Dataset variants have different fields, so check the dictionary instead of copying a public notebook. UCI documents Iranian Churn as first-nine-month feature aggregates with churn status at month twelve, a three-month planning gap. That supports its stated customer-table task, without per-customer event dates.
Online Retail provides invoice generation times, with no ingestion or version history. A snapshot lab that assumes immediate availability must state that assumption and test simulated late arrivals. This release cannot verify historical production availability. Limit matured inactivity labels to the documented observation window and the customers you can identify. These are practice datasets with different limits. Keep the dataset licence, attribution, checksum and provenance in the project README.