NewStart your own marketing agency on askotter.See how it works
Predictive forecasting

Ecommerce revenue forecasting model: when daily beats weekly, and when it does not

An ecommerce revenue forecasting model can run on daily or weekly numbers. Daily models catch weekday swings and short promos, but they need more orders and they are noisier. Weekly models are steadier, but they blur the days that ad budgets depend on. Pick the one that fits the decision the forecast feeds.

Nov 19, 2026Brian Chiou8 min read
the short answer
Use daily
For ad budgets and staffing over the next couple of weeks, when the store has enough orders each day.
Use weekly
For inventory and cash plans further out, or when daily totals swing on a handful of orders.
Before you trust either
Flag promos and stockouts, read the range, and backtest against a simple baseline.

This guide is for ecommerce and DTC operators who want to know if their forecast can carry inventory and ad-budget decisions. Demand forecasting granularity is the first real choice you make, and most tools make it for you without saying so.

How much history each granularity actually needs

A model can only learn a pattern it has seen repeat. The weekly rhythm repeats every seven days, so a few months of daily data show it many times. The yearly pattern repeats once a year, so you need at least a full year of history to see it at any granularity. Two years is better, because then the model sees each holiday twice.

GranularityWhat it can learnWhat it needsWatch for
DailyThe weekly rhythm, short promos, the days around a holiday.Enough orders each day that one or two sales do not swing the total. A full year to see the yearly pattern.Noisy for small stores and for single products.
WeeklyThe trend and the yearly pattern.Fewer orders per period. Still a full year, better two, because a year is only 52 points.Hides which day sold. Blurs promos shorter than a week.

Granularity is about products and places too

A daily forecast for the whole store may work well. The same forecast for one product in one location may be mostly zeros. When counts get that small, forecast at a level with enough sales, such as a product category by week. Then split the total down by each item's recent share of sales.

Weekday and payday seasonality in DTC revenue

Most stores sell more on some days of the week than others. Which days, and by how much, differs from store to store, so measure your own. A weekly model hides this rhythm. That is fine for a reorder you place once a month. It is not fine for an ad budget you set each morning.

Payday and the month

If many of your customers are paid on set dates, orders may rise around those days. Look at revenue by day of the month, not only by day of the week. If a pattern shows up in every month you check, add it to the model. If it does not, leave it out.

A quick test before you choose

  • Take the last twelve weeks, and leave out any week with a promo.
  • Average revenue by weekday.
  • If the gap between the best and worst day is small next to the normal week-to-week swing, a weekly model loses little. If it is large, go daily.

Count each sale on the right day

A daily model is only as good as the dates it is fed. Shopify writes each order the moment it is placed. Many POS systems settle and export sales in end-of-day batches, so check which day a late sale lands on. Stripe charge dates and payout dates differ by days. Forecast on the charge date, never the payout.

Handling promos, launches, and stockouts as regressors

A regressor is an extra input that tells the model something unusual was going on. Without one, a model treats a promo spike as normal. It may expect the same spike at the same time next year, or raise its baseline for every week after.

  • Promos. Flag each promo day. Flag a few days after it too, because customers who bought early do not buy again next week. A weekly model can carry a promo flag, but it cannot tell which day the promo ran.
  • Launches. Flag the launch and the weeks after it. A new product has no history of its own, so do not let its first weeks set the trend for the store.
  • Stockouts. A day with no sales because nothing was in stock is not a day with no demand. Mark those days so the model does not learn them as weak days. Shopify data includes inventory levels, so you can find them.

The model only knows the future you tell it

A promo flag helps the forecast only if next month's promos are on a calendar the model can read. The same goes for ad spend. If you feed planned spend in, the forecast now assumes that plan, so say so when you share it. Be careful with recent ad results too. Google Ads credits a conversion to the date of the click, so the last few days can keep rising for a while.

Reading confidence intervals before committing spend

A forecast is a range, not one number. An 80% range means the real number should land inside it about eight times in ten, if the model is honest. Daily ranges look wide. That is the truth about a single day, not a broken model.

  • Stock you cannot return. Buy closer to the low end, unless running out costs you more than leftover stock.
  • Stock you can reorder fast. Plan for the middle, and keep a reorder ready for the high end.
  • Ad budgets. When the range is wide, commit in steps. Add budget as the actual numbers come in near or above the middle.

Be wary of a range that looks too tight. A narrow range that the real numbers keep missing is worse than a wide one that holds. The backtest below shows which one you have.

Matching forecast horizon to the decision it feeds

A forecast should be no finer than the decision it feeds. It should also reach at least as far ahead as the time it takes to act, such as a supplier's lead time.

DecisionHow far aheadGranularity
Daily ad budget pacingThe next one to two weeksDaily
Packing and fulfillment shiftsThe next few weeksDaily, by weekday
Inventory reordersSupplier lead time plus a bufferWeekly
Quarterly budget and cash planThe next three to twelve monthsWeekly or monthly

So the daily vs weekly sales forecast question often has two answers in one store. Run daily for the next couple of weeks and weekly beyond that. Then check that the daily forecast, added up by week, roughly matches the weekly one. If they disagree, find out why before either goes into a plan.

Backtesting: the accuracy check most teams skip

A backtest asks one question: had you run this model in the past, how close would it have been? Pick a date a few months back. Forecast using only the data up to that date, and compare the forecast with what really happened. Then move the date forward and repeat, so one lucky month does not decide the result.

  • Beat a simple baseline. Compare the model with the same week last year, or the average of the last four weeks. If the model does not beat that, use the baseline.
  • Score at the level of the decision. If inventory is planned by week, add the daily forecast up to weeks before you score it. That is the fair way to compare daily and weekly models.
  • Check the range. Count how often the real number fell inside the 80% range. Far less than eight in ten means the range is too tight.
  • Score promo weeks on their own. A model can be close on normal weeks and far off on the weeks that matter most.
  • Measure error in dollars. Percent error breaks on days with almost no sales, so add up the error in dollars over the period.

Run the backtest again when something big changes, such as a new sales channel or a price change. A model that was right last year may not be right now.

askotter forecasts with live numbers from every connected account. Its forecasts update as your numbers change, with a likely range attached. Shopify, Stripe, Google Ads and Meta Ads connect today. WooCommerce and POS connections are on the integration roadmap. See Shopify + POS forecasting for online and in-store sales in one forecast, Shopify + Meta Ads and Shopify + Google Ads for spend beside orders, and Stripe + WooCommerce for WooCommerce stores.