Forecasting Demand Without Pretending to Predict the Future
Every inventory system now offers a forecast. Very few explain how the number was produced, which is the only thing that determines whether you should act on it. Here is the arithmetic behind ours, the settings that change it, and the four situations where it is confidently wrong.
A forecast is a claim about the future dressed as a number, and the number always looks equally confident whether it rests on three years of stable demand or on two sales last Tuesday. That is the problem with forecasting features: the presentation is identical, so the buyer cannot tell the difference between a model worth trusting and an average with a coat of paint.
So rather than describe ours as intelligent, here is exactly what it computes. You can then decide for yourself which of your items it should be trusted on — which, done properly, is not all of them.
It starts with velocity, blended from two windows
For each item we take the quantity that actually went out to customers over the last 90 days and over the last 30 days, and turn both into an average daily figure. The two are then blended, with the weighting controlled by a sensitivity setting you choose.
| Sensitivity | Weight on the recent 30 days | When it suits you |
|---|---|---|
| Conservative | 35% | Stable, slow-moving lines where a short burst should not move your buying |
| Balanced | 50% | The default. Recent behaviour and the quarter get equal say |
| Responsive | 70% | Fast-moving or trend-driven stock where last month is genuinely more informative than last quarter |
Alongside the blend we report momentum — how far the blended velocity sits above or below the plain 90-day average. Momentum is often the more useful number of the two, because it tells you something is changing rather than what it will be. A line at +40% momentum deserves a human look regardless of what the forecast total says.
What counts as demand
Outbound movement to customers — invoiced check-outs and point-of-sale lines. Internal use, transfers between locations, damage and manual adjustments are separate movement sources and are not, by default, read as customer demand. That distinction matters: a business that issues heavily to its own projects will otherwise forecast on consumption that no customer ever asked for.
From velocity to a date and a quantity
Once there is a velocity, three practical outputs follow, and each is deliberately plain arithmetic you can check by hand.
One item, start to finish
Illustrative. The risk grading is the part worth internalising: an item is critical when days of cover are at or below the lead time, high within a week of that, and medium when stock has fallen under the reorder threshold but there is still room. It is a comparison between two facts — how long your stock lasts and how long resupply takes — not a judgement about the item.
The suggested reorder quantity follows the same logic: velocity across the lead time, plus safety stock, minus what you already hold, plus a month of cover so you are not back at the same threshold immediately. Lead time and safety stock come from the item where you have set them, and fall back to organization-wide defaults where you have not. The underlying idea is the standard one — see reorder point and safety stock for the theory in isolation.
A forecast is only as good as the question behind it. "How much will we sell?" is unanswerable. "Will this run out before the next delivery arrives?" is arithmetic — and it is the question that actually costs you money.
Where it is confidently wrong
Any honest description of a forecast has to include the cases where it fails, because those are the cases where an unexamined number does damage. There are four, and all four are structural rather than bugs.
-
New items with no history
Ninety days of nothing produces a velocity of nothing, and an item with no velocity reports infinite cover. A new line is a judgement call by a buyer, not a forecast, until it has traded for a quarter.
-
One-off bulk orders
A single institutional order of 400 units inside a 30-day window lifts the blended velocity for the whole following period, and the system will cheerfully recommend restocking for a customer who is not coming back. A conservative sensitivity dampens this; it does not solve it.
-
Anything you have planned but not told it
Promotions, a new branch opening, a tender you just won, a competitor closing. The forecast reads history. It has no knowledge of the future you are actively creating, which is precisely the demand that hurts when you miss it.
-
Real seasonality
There is a simple calendar uplift applied in March, June and December, which is a blunt approximation of common trading peaks — not a seasonal model learned from your own history. If your business has genuine seasonality on a different calendar, treat the forecast as a baseline and apply your own knowledge on top. We would rather say that plainly than let you discover it in a peak month.
Tune it once, then leave it alone
The settings that shape all of this are yours, not ours, and they are worth ten minutes at setup rather than being rediscovered during a stockout.
- Sensitivity — how much the recent month outweighs the quarter.
- Default lead time — used for every item where you have not set a specific one. This is the setting that most often makes forecasts look wrong, because an optimistic default hides real exposure.
- Default safety stock days — the buffer applied when an item has none of its own.
- Minimum demand points — how much history must exist before a forecast is offered at all, which is your guard against confident nonsense on thin data.
- Outlier cap — a limit on how much a single extreme period can distort the average.
- Demand sources — which movement types count as demand, so internal issues and transfers do not masquerade as customers.
What we do and do not do
What AWRA OpsHub does today
- A blended 30/90-day velocity per item, with the weighting under your control.
- Momentum, expressed as the percentage gap between recent and quarterly behaviour.
- Days of cover and a projected stockout date for every item with demand history.
- A risk grade derived from cover against that item's actual lead time.
- A suggested reorder quantity built from lead-time demand plus safety stock less current stock.
- A plain-language narrative over the whole picture, where an AI provider is configured — a summary of the computed figures, not a second opinion on them.
What it does not do
- No machine-learned seasonality. The uplift is a fixed calendar assumption, not a pattern learned from your history.
- No promotion, price or campaign awareness. Nothing you plan is an input; only what already happened.
- No supplier-side probability. Lead time is treated as the number on the item, not as a distribution with a late tail — even though late delivery is the most common cause of an actual stockout.
- No automatic ordering. Every suggestion is a suggestion; a purchase order is raised by a person who can see the reasoning.
The third gap is the one to hold in mind. Your forecast can be perfect and you still run out, because the resupply arrived nine days late. That is why the risk grade compares cover to lead time rather than to a fixed number of days.
Our take
Use the forecast for the question it answers well — will this line run out before resupply arrives — and ignore the precision of the 30-day demand figure, which is an average with a weighting, not a prediction. Set your real lead times, choose a sensitivity that matches how your stock behaves, and treat every new item and every bulk order as a manual judgement for a quarter.
See predictive insights in AWRA OpsHub
Velocity, momentum, days of cover, projected stockout dates and reorder suggestions per item — with the settings that produce them in your hands.
Explore predictive insightsFrequently asked questions
How much history does the forecast need before it is useful?
A full 90-day window on an item that traded steadily throughout it. Below that you are averaging noise, which is why there is a minimum-demand-points setting to suppress forecasts on thin data — set it deliberately rather than leaving it at whatever it came with. For new lines, a buyer's judgement beats any average until the item has a quarter of trading behind it.
Why does the system recommend reordering something we clearly have plenty of?
Usually the lead time. The reorder threshold is velocity across the lead time plus safety stock, so an item with a 45-day lead time legitimately needs a great deal of cover, and what looks like plenty may be six weeks. The second most common cause is a one-off bulk sale inside the 30-day window inflating the velocity. Check the momentum figure: a large positive momentum on a slow item usually means one order, not a trend.
Does it learn from our data over time?
Not in the sense that word usually implies. It recomputes from your recent history every time it runs, so it reflects your data — but there is no model being trained, no parameters being fitted, and no accumulating knowledge of your business. That is a deliberate trade: the arithmetic is inspectable and reproducible, and you can check any number it shows you by hand. A learned model would be better at seasonality and worse at being explainable.
What does the AI narrative add?
Language, not judgement. Where an AI provider is configured, a short plain-English summary is generated over the figures already computed — which lines are at risk, where the pressure is — and it is cached and fails soft, so an unavailable provider means the numbers still work and the paragraph is simply absent. It is a reading aid for people who will not open the table. It does not re-analyse anything and it cannot disagree with the arithmetic.
Should we let it place orders automatically?
You cannot, and we think that is right. Every output is a suggestion for a human to act on, because the four failure cases — new items, one-off bulk orders, planned events the system cannot know about, and real seasonality — are all cases a buyer spots in seconds and an average never will. Automating around a forecast that structurally cannot see your plans converts a good decision aid into a source of dead stock.
