Retail’s demand forecasts are still built around an apex the calendar has stopped producing. Glossy ran through the moves this week: Prime Day stretched into Prime Week, Black Friday into Black Friday Month, back-to-school opening in June, holiday shopping opening in October. Forecasting systems learn the shape of a season from the seasons before it, and each of those shapes is now obsolete before the goods it sized have landed.

Cyber Week has shrunk into a minority event inside the season it names. Adobe measured the six days at $44.2 billion against a season that finished at $257.8 billion, which puts the most heavily planned week of the retail year at roughly one dollar in six. Vivek Pandya, Adobe’s lead analyst for digital insights, said the persistent deals running through the week “pushed consumers to shop earlier, creating an environment where Black Friday now challenges the dominance of Cyber Monday.” A curve whose second-highest point starts contesting its highest is a curve losing its apex.

October is now growing faster than the season it feeds. US online spending that month rose 8.2% year over year in 2025, a figure that means nothing until you set it beside what the full season delivered. The season finished at 6.8% growth year over year. The shoulder is still outrunning the peak, if by a narrower margin than the calendar’s loudest month would suggest.

The stretching is a competitive race with no natural stopping point. Kelly Pedersen, PwC’s global retail lead, told the Glossy Podcast that retailers “always want to capture consumers during these moments, and they want to be first,” and that the prize is acquiring the customer rather than banking the window’s revenue. First is a relative position, so any retailer who opens earlier than last year’s field obliges the field to open earlier still. PwC’s back-to-school research, which Pedersen cited on the podcast, puts average household spend at around $920 across a window that has already migrated into June. The NRF found 42% of shoppers planning to start their holiday buying before November. Pedersen’s own account has consumers being trained into that habit rather than driving it.

A retailer with ten years of daily sales history has ten years of data and exactly one observation of the calendar it is about to trade through.

Seasonality at annual frequency is estimated from years, and years are the one input nobody can buy more of. Fildes, Kolassa and Ma, updating their review of retail forecasting for the International Journal of Forecasting, examined how the pandemic broke established sales patterns and pushed forecasters toward leaning more heavily on recent data to keep models current. That approach shrinks the usable sample at the exact moment an analyst most needs a large one. Huber and Stuckenschmidt, writing in the same journal about daily demand in bakery retail, call forecasting for special days “a key challenge” because demand on those days departs so far from the ordinary weekly pattern. Stretch a special day across a month and the feature that encodes it stops marking anything.

The strongest case against all of this is that models relearn, and faster than the pessimists allow. Marco Zanotti, testing ten machine-learning and deep-learning models across two large retail demand datasets, found that less frequent retraining held forecast accuracy while cutting computational cost, which suggests these systems tolerate staleness better than their operators fear. Read that forward and the elongation is a transition cost: two or three seasons of the new shape and the models settle onto it. The argument here fails if that condition holds, and the condition is that the calendar settles first. It has not, because settling would require every retailer to stop trying to be first, and first is defined against whatever the others did last year. Tolerance for stale data is a virtue when the target is stationary; against a target still in motion it is a slower route to the same error.

Apparel pays for this in a currency the fast-moving categories do not use. A seasonal buy is committed on lead times measured in months, so the forecast that sizes it is made long before anyone knows what shape the promotional calendar will hold when the goods land. A model expecting an apex over-buys the apex week and under-buys the weeks on either side, and both errors deposit units in the wrong week. In apparel a unit in the wrong week is a markdown, and Adobe expected apparel discounts to reach 25% off list. If the calendar keeps stretching at anything like the current rate, that error stops being noise that averages out over a season and becomes a bias with a sign, and the sign reads: too much, too late.

Peak season used to be weather, and forecasting weather is a respectable problem. It is now a schedule, and the schedule is written by the same merchants who then ask a model to predict it. Treating a variable you set as one you observe is a misspecification, and no vendor and no quantity of compute repairs a misspecification. The retailers who gain ground over the next three years will be the ones who stop asking when the peak arrives and start declaring when they intend to make it, then forecasting against their own published plan instead of against the ghost of a curve that last held its shape in 2019. That is a merchandising decision in a data-science costume, which is why it keeps getting routed to the wrong department.