Why Churn Models Keep Mistaking Seasonality for Decline

On August 17, a new paper — "Seasonal false alarms in customer churn and decline early-warning systems" by Md Rezwanul Islam and Wael Mohammed — landed on a blunt finding: a large share of accounts flagged as "at risk" by typical churn models were never actually churning. The model was simply confusing seasonality with decline.

The mechanism

Most churn and decline detectors work the same way: compare an account's trailing k months of activity against its next k months. For any seasonal business, that's a trap — a normal post-holiday slowdown reads identically to real decline.

Testing across four datasets (including M5 and an online retail set) plus a live production system, the authors found that 28–69% of flagged "decay events" were seasonal noise, not weakening accounts. In their production system, a third of the flagged accounts (119 down to 79) disappeared once seasonality was accounted for.

Old way vs. new way

Old way: rolling trailing-vs-next-window comparisons, which treat any calendar-adjacent dip as a decline signal regardless of season.

New way: align the baseline to the same calendar months one year prior, so the comparison is year-over-year instead of window-over-window.

The payoff, and the cost

With the classifier held constant, that single change took production ROC-AUC from 0.767 to 0.864. It's not free — it needs more historical data per account, and it's less sensitive to genuine decline-then-stabilization patterns — but the trade looks worth making for any business with real seasonality.

The uncomfortable implication

Teams may have been sizing retention budgets and CS outreach around false churn signals for years without knowing it. Open question worth arguing about: how many other early-warning systems (fraud, demand spikes, inventory shortfalls) have the same calendar-blindness baked in — and would anyone have caught it without going back and checking the labels?

Our take

We've hit this exact failure mode building churn models for marketplace and subscription clients, which is why calendar-aware baselines are a default step before any decay score ships.