One action-list slot in three was a phantom. That is the number that stopped me when I read this paper. In a live B2B marketplace, the churn early-warning system was flagging accounts for intervention, and a third of those flags dissolved the moment the label was lined up properly. Your retention team was chasing accounts that were never really declining.

The bug is subtle. A standard way to build a churn label compares an account’s next k months of activity against its trailing k months. Sounds reasonable, except those two windows cover different calendar months. For a seasonal business, the trailing window might hit high season while the outcome window lands in low season. The model reads “decay” when the account is just seasonally lower. Seasonality gets confused with decline.

Retail, tourism, fitness, education, any B2B with an annual cycle: this applies to you. And the fix is surprisingly simple, once you see it.

Why This Paper Matters

Every flagged account consumes intervention capacity. Account managers have limited hours, and each alert pulls them off whatever else they were doing. If a third of those alerts are noise, you are quietly burning your most expensive resource on accounts that were never at risk.

The authors measured this in a deployed production system, not a lab. Between 28% and 50% of the decay events flagged in production had no counterpart under a seasonally aligned definition. On three public panels the number ran even higher, from 37% to 69%.

Here is the part that should make you uncomfortable: your own churn model might have this exact bug and you would not know. The accuracy looks fine, the alerts look plausible, and the seasonality confound stays invisible until someone lines the labels up the right way. This is a data-quality issue at the foundation of how your warning system is taught, and it is precisely the kind of thing a Chief AI Officer should be auditing for.

Methodology, Explained Simply

Imagine a ski resort. In February you compare its next three months of bookings against its trailing three months. The trailing window sits in high season, the outcome window slides into spring. Bookings drop, the ratio fires, and the model says decline. But the resort is not declining. It is just being compared to its own best month.

That is the adjacent-window label in plain English. Those two windows used to build the target cover different parts of the calendar, so for anything seasonal the threshold-ratio construction confounds seasonality with decline. The event rate literally depends on which anchor calendar month the label happens to start from.

The proposed fix is a year-over-year comparison. Instead of comparing an account to its trailing k months, you align the baseline to the same k calendar months one year prior. Same season, same calendar position. A purely seasonal entity suddenly reads parity instead of decay, and the curve flattens at the source.

The authors formalized the mechanism, measured it on three public panels and the production panel, and costed the trade. A fix like this is not free. It needs extra history, it trades in a blind spot for decline-then-stabilization, and it applies a stricter cut under sustained growth. But those costs are small next to a third of your action list being fake.

Results and Practical Insights

With the warning model held fixed, production holdout ROC-AUC rose from 0.767 to 0.864 for decline. Thats a large jump you can expect to feel, and it happened without retraining the classifier at all. The gain came purely from fixing the labels, which is the cheapest kind of improvement available to you.

The served action list tells the operational story. It shrank from 119 accounts to 79. Your account managers either stop chasing seasonal dips or, better, point that freed capacity at a shorter list of genuinely at-risk accounts. The same people, the same hours, aimed at accounts that actually need saving.

This is a reminder that for any executive carrying an AI roadmap, the most useful work is often not a fancier model. It is checking the ground truth and the labels that teach the model what to look for. A Chief AI Officer that audits label construction can deliver gains that a thousand more training runs never will.

Key Takeaways for Revenue and Operations Leaders

Thanks to All Authors