Forecasting

When you know something the sales data does not

A trailing sales rate is a good starting position and a bad verdict. When a promotion, a listing change or a wholesale order is about to make the next sixty days look nothing like the last sixty, tell the forecast what you know, then take it back out. The numbers below are invented so the arithmetic can be checked.

The daily sales rate behind a reorder calculation is a trailing average. It looks backwards over some window, divides by the days, and hands the number forward as the best guess for the next window. For a SKU with nothing scheduled, that is a reasonable guess.

For a SKU with a promotion booked, a listing about to be suppressed, or a wholesale order landing next week, it is a guess that you already know is wrong. The information is in your calendar and your inbox. It is not in the history, so no rate computed from the history can contain it.

The honest answer is not “trust the forecast” and it is not “ignore the forecast.” It is to tell the forecast what you know, in units per day, for as long as it is true, and then to take it out again. Three steps, worked on an invented SKU.

Step 1: find the SKU whose next window will not look like its last

Start by looking at the rate over several windows for the same SKU. When the 7-day, 30-day and 90-day rates agree, the trailing average is telling a consistent story. When they fan out, something already changed and the windows are disagreeing about how much of it to believe.

The invented SKU: 40 units a day over the last 90 days, 41 over 30, 39 over 7. Flat. Nothing in the history suggests the next month differs from the last three. Now add the fact the history cannot see: a ten-day site-wide promotion starts in two weeks, and this SKU is in it.

In SKU Compass the comparison is on the same dialog you will use in step 2. Clicking a SKU’s sales-per-day figure on Supply Analysis opens All-Channel Sales/Day Override, and the top of it shows four reference rates side by side: Computed AP, 7-day, 30-day and 90-day. If those four agree and you know something is coming, the trailing rate is the thing that is about to be wrong. If they already disagree, work out why before you add an override on top of the disagreement.

Step 2: size the lift from your own last promotion, then blend it

This is where rules of thumb creep in and where they should not. The question is not “how much does a promotion lift sales?” It is “how much did this SKU lift the last time you ran a promotion like this one?” You have that number. It is in your order history.

The lift multiple

Find the last comparable promotion. Comparable means broadly similar offer, channel, duration and placement — not merely the last time the SKU was discounted. A 40% sitewide clearance and a 10% email-only offer are not the same event, and using one to size the other imports the wrong lift. Take units sold per day during it, and divide by units sold per day in the weeks before it. Say the invented SKU sold 720 units across an 8-day promotion in the spring, against a baseline of 40 a day at the time: 720 ÷ 8 = 90 a day, and 90 ÷ 40 = 2.25. This SKU roughly doubled and a quarter under a comparable promotion. That is your lift multiple, and it came from your own data.

If you have no comparable promotion, say so and be clear about what you are doing: this is a scenario, not an empirical forecast. Pick a multiple you can defend in a sentence, and if the PO is large enough to hurt, run it low, base and high rather than betting the order on one assumption. Plan to replace it with a measured one after this promotion ends. A written guess is better than an unwritten one because it can be corrected.

Blending across the coverage window

Here is the mistake to avoid. If you override the daily rate to 90 and the order maths is covering 60 days, you have just told it the promotion lasts 60 days. It lasts ten. The override has to be the average rate across the whole window the order is covering, promo days and ordinary days together.

The override should represent the average demand across the window you are actually buying for, not the peak demand during the promotion.

Blended: (10 × 90 + 50 × 40) ÷ 60 = 2,900 ÷ 60 = 48.3 units a day. That is the override. Not 90, which over-orders by 2,500 units across the window, and not 40, which leaves you 500 short in the middle of a promotion you paid to run.

Recompute the blend if the coverage window changes. The same 10-day promotion inside a 30-day window blends to (900 + 800) ÷ 30 = 56.7 a day. The override is a property of the promotion and the window, which is why it is not a number you can set once and forget.

In SKU Compass the override is the field labelled Override (units/day) on that same dialog. The dialog’s own text says what it does: a manual value replaces the computed velocity for this SKU across every channel, and it drives days of stock, days until order, and the order estimate. Type 48.3, read the preview line (values round to one decimal), and save. From that point the order maths for this SKU is running on your number, not the trailing one.

Two limits on that number, both worth knowing before you type it. First, it is an all-channel override, while the promotion in this example runs on one storefront. If the SKU also sells somewhere the promotion does not reach, blend across the demand you actually expect everywhere, not just the promoting channel — otherwise you lift the whole SKU on the strength of one channel’s event. Second, the number you enter was blended for one specific window; it is exact for that window and approximate for every other calculation it feeds.

An override is not a better forecast. It is your assumption, written down where the arithmetic can use it and where someone else can see it.

Step 3: clear it when the event has passed

The dangerous half of an override is not setting it. It is leaving it. An override set for a spring promotion and never cleared is still inflating that SKU’s order quantity in autumn, and it does so silently, because the number looks like any other number on the page.

Two disciplines. First, put the clear-by date in the same place you wrote the override. If your tool has no notes field, a line in a shared sheet with SKU, override, reason and end date is enough. Second, after the promotion, measure what actually happened and update the lift multiple you will use next time. The promotion you just ran is the best data you will ever have about the next one.

In SKU Compass clearing is one click: the same dialog shows Clear override once an override exists, and a blank field on save does the same thing, putting the SKU back on its computed rate. Rows carrying an override are marked with a teal dot next to their order quantity, so a stale one is visible on the estimator rather than hidden in a setting. If several have accumulated, the drawer’s Reset per-SKU overrides button clears all of them at once and shows the count, which is a useful number to glance at: if it is higher than the number of events you can name, something is stale.

What an override is for, and what it is not for

For: a known event with a start and an end

A promotion, a wholesale order with a ship date, a listing suppression you are working to lift, a competitor out of stock for a known period. Each has a size you can estimate and a date after which it stops being true.

Not for: fixing a rate you simply disagree with

If the trailing rate feels low and you cannot say why, the answer is to find out why, not to overwrite it. A stockout inside the window, a returns spike, a price change: each of those has a real explanation, and the explanation tells you whether the low rate is temporary.

Not for: a spike that is already in the history

If last week’s spike is inflating the 7-day rate, an override that pushes the rate back down is treating the symptom. The spike will age out of the 7-day window in a week and out of the 30-day window in a month on its own. Overriding it means remembering to un-override it, and the history was going to fix itself for free.

Also worth expecting: the weeks after

A promotion can borrow demand from the period right after it. Customers who would have bought in week three bought in week one instead, so the 50 “ordinary” days in the blend may not all run at the full 40 a day. This does not change the method — you still blend, and you still avoid ordering to the peak — but if you see a soft fortnight after a promotion, that is usually pull-forward rather than a SKU going bad. Judge the event on promotion-plus-recovery, not on the promotion window alone.

Not for: a permanent step change

If a SKU genuinely moved from 40 a day to 60 a day and is staying there, the trailing rate will get there on its own as the new weeks accumulate. An override bridges the gap during the lag, but it needs a clear-by date like any other, otherwise it becomes the forecast.

How a promotion changes the reorder point too

The override changes the order quantity. It also moves the reorder point, because the reorder point is daily demand times lead time, plus safety stock. Raise the daily demand and the level at which you must order rises with it. For the invented SKU with a 45-day lead time and a 200-unit buffer, the reorder point at 40 a day is 40 × 45 + 200 = 2,000 units on hand.

Here is the trap, and it is the one worth reading twice. The 48.3 blend was built for a 60-day coverage window. The reorder point has its own horizon — the 45-day lead time — and the promotion does not distribute itself evenly across the two. All ten promotional days fall inside that 45-day window, so the demand the reorder point has to survive is:

So the event-aware reorder point is 2,300 + 200 = 2,500 units, not the 2,374 you get by pushing the 60-day blend of 48.3 through a 45-day formula. If the SKU was sitting at 2,100 units, it was above the reorder point before the promotion was booked and below it after, without a single unit moving — and it is further below than the blended rate suggests. Put your own numbers into a reorder point calculator under both rates to see the shift on your catalogue.

The general rule this exposes: a blended rate is only correct for the window it was blended over. One override field drives several calculations, and those calculations do not all cover the same number of days. Blend for the decision you are making. If the same override then feeds a shorter or longer horizon, know that it is approximate there.

That is the practical reason to enter a known event early. The promotion is not what makes you late. Finding out about the promotion after the last day you could have ordered for it is.

What to do this week

1. List the events you already know about for the next coverage window: promotions, wholesale orders, listing changes. For each, name the SKUs and the dates.
2. Size each one from your own history and blend it across the window. Lift multiple from the last comparable event, then (promo days × promo rate + other days × baseline) ÷ window days. Enter that as the override. In SKU Compass it is Override (units/day) on the sales-per-day dialog.
3. Write the clear-by date next to it, and clear it. Clear override on the dialog, or Reset per-SKU overrides for all of them. Then measure what the promotion actually did, and keep that number for next time.

A forecast built only from history is a forecast that finds out about your business last. The override exists because you know things your sales history cannot, and the discipline around it exists because you will forget you told it.