How aiku predicts what you run out of

The two numbers

Wherever you are buying — a partner's Browse cards, the Shopping list, a supplier or agent dashboard, an Auto-fill proposal — the same pair follows the item.

Estimated: Would run out in ~N days is what you have on hand divided by how fast aiku thinks it is leaving. It goes red at two weeks or less, amber up to a month. "We run out now" means the shelf is already empty.

Suggested is the quantity that would carry you to the next order and a bit beyond: enough for the supplier's lead time, plus the gap until you would normally order again, plus a cushion sized to how erratic the item is — then minus what is on the shelf and what is already on its way. It is rounded to whole shipping units, because that is what you can actually buy.

Both numbers refresh themselves whenever the stock moves, so they are current when you look at them: the rate an item sells at is worked out each night, and the stock it is divided into is today's. On a purchase order the same forecast shows under Stock as Lasts, for example Lasts: 6 weeks (out around 15 Nov) · could be 4 weeks.

The idea that makes it work: empty days don't count

The obvious way to measure how fast something sells is to average its sales over the last three months. That method quietly ruins a stockroom.

Take an item that sold out in week one and sat empty for the rest of the quarter. Averaged over ninety days it looks like it barely moves — so it is never reordered, so it stays empty, so next quarter it looks even worse. The better it sells, the faster it disappears, the more invisible it becomes. Most stockrooms have a handful of these, and they are usually items people are asking for.

So for any item that has been out of stock a lot lately, aiku doesn't average over the calendar. It reconstructs, day by day, whether the item was actually available, and measures the sales rate only over the days you had it to sell. Days when the shelf was empty are treated as days with no information — not as days with no demand.

That single rule is why a bestseller sitting at zero shows a large suggested order rather than a small one. It is not a bug and it is not the system panicking. It is the system finally seeing the demand that the empty weeks were hiding.

Where the number comes from, and how much to trust it

Not every item has the same quality of evidence behind it, and it helps to know which case you are looking at.

• A forecast of its own sales. The normal case, for items that were on the shelf at least seven days in ten over the last three months. Each night a forecasting model reads up to three years of the item's weekly sales, its seasons included, and predicts the next weeks. The forecasts of every organisation are then checked against what it really dispatched over the last six weeks and scaled to match, so when demand picks up or slows down across the board the numbers follow within days. Tested on our own sales, it was about a third closer to what then sold than the method below.
• Its own recent in-stock days. For items that were out of stock more than three days in ten lately, and when the nightly forecast has not run. This is the empty-days rule above. Steady sellers get a trend-following estimate; slow, lumpy items — the ones that go out three at a time every few weeks — are measured differently, on how big the occasional order is and how long the quiet gaps run, which is the honest way to describe them.
• The same item in a sister organisation. Nothing sold here in the last three months; somewhere else in the group has sold it. aiku borrows their rate and halves it, because a different market is a hint, not a measurement. Treat it as a starting point.
• The family it belongs to. The weakest case: usually a brand new line with no recent sales anywhere, estimated from its neighbours and heavily discounted. This is a placeholder for your judgement, not a substitute for it.

If none of these has anything to go on, there is no estimate: no run-out day and no suggestion. Outside the nightly forecast, an item's older history is not used — an item that sold well last year but not in the last three months is not assumed to pick up where it left off. When that was tried, it asked for too much.

Seasons only where the item's own history shows them. The nightly forecast sees up to three years, so an item that has peaked every Christmas is forecast to peak again. Items on the in-stock-days rule, new items and items too small for a pattern to show get no seasonal lift: a Christmas item in August is then forecast on its August rate. So ahead of a peak you know is coming — the run-up to Q4, a summer line, a trade show — check the suggestion and raise it by hand if it has not risen.

Why a number can look wrong (and often is)

The forecast reads history. Anything that happened outside the history, it cannot know.

• A one-off bulk order. One customer clearing you out looks exactly like sudden popularity. Overrule it.
• A line you are discontinuing. History says it sells; your plan says stop. The system does not know your plan.
• A promotion, a catalogue photo, a marketplace listing going live. Demand about to change for a reason that has not happened yet.
• A known peak, such as Q4. Only items whose own past shows the peak get it; for the rest the suggestion is sized on the quiet months before it. Order up by hand, early enough for the lead time.
• A brand new product. See the family case above — that number is a guess wearing a confident face.
• Something that hasn't moved at all but is worth money. It lands in Dead stock on the dashboard, and it wants a decision from a person, not a reorder.

The rule of thumb: the forecast is better than you at the boring middle of the catalogue — hundreds of ordinary items nobody has time to think about — and worse than you at anything with a story attached. Let it handle the volume, and spend your attention on the exceptions.

Reading it in an Auto-fill proposal

Auto-fill ranks candidates by how soon you run out and tops up the most urgent first until the budget is gone. Every proposed line carries its reason in plain words — "Our sales/quarter ~48 · our stock 0 · we run out now" — which is the forecast showing its working. Read the reasons before you confirm; that is where a wrong number is easiest to catch, and unticking a line takes one click. Nothing is ordered until you press Add items to shopping list.

Looking ahead: what runs out if nothing more is ordered

On the Procurement dashboard the stock outs chart carries on past yesterday with two dotted lines: how many SKOs will be out of stock each day of the next eight weeks, and the sales they would lose a day. Each SKO's stock is played forward with its forecast, and every purchase order and delivery already open lands on its expected day. Orders already past their expected day are left out until they get a new date, so chasing a late order and dating it again brings its stock back into the line. Nothing you have not ordered yet is counted, so the lines show what happens if nobody orders anything more. Order today and tomorrow's line comes down. SKOs marked On Demand are never counted, here or in the history: they are made or bought when ordered, so an empty shelf is normal for them.

The Stock levels list behind each level on the dashboard has a column Lost if not ordered: the sales a SKO would lose over its lead time and the month after, if nothing more is ordered. Sort by it and the gaps that cost the most come first.