A different question from reordering

Inventory forecasting asks a short question: what should I order this week?

Demand prediction asks a longer one: what is going to happen over the next few months, and what should we do about it?

The horizon changes what the answer is for. A reorder proposal feeds a purchase order. A demand forecast feeds capacity planning, cash flow, hiring, supplier negotiation and what you decide to promote.

It also changes how confident you can be. Accuracy falls sharply the further out you look, and any forecast presented without that caveat is being oversold.

What it is used for

Five decisions that need a view beyond next week.

Production capacity. Do we need another shift, another machine, more subcontracting?

Cash flow. Big stock purchases have to be paid for. Knowing three months ahead changes how you plan.

Supplier commitments. Volume pricing needs a volume commitment. Better forecasts mean better negotiating positions.

Staffing. Seasonal businesses hire ahead of demand, not during it.

What to promote. Pushing a product you cannot supply is worse than not promoting it.

What it uses

Beyond raw sales history:

• Seasonality — repeating patterns across the year
• Trend — steady growth or decline underneath the noise
• Cycles — quarter-end effects, month-end effects
• Product relationships — items that sell together or replace each other
• Promotion history — what happened last time you ran one
• Customer behaviour — regular buyers versus one-off orders

The best forecasts also take in things from outside your data: planned promotions, known customer commitments, a product being discontinued. Odoo can hold this context; it cannot know it unless somebody enters it.

Reading the number honestly

Three points that determine whether a forecast helps or misleads.

Accuracy decays with distance

Next month is a reasonable estimate. Six months out is a direction. Two years out is a conversation, not a number.

Use forecasts at the horizon they support. Treating a twelve-month figure with the same confidence as a one-month figure is how planning goes wrong.

A range is more honest than a point

“We expect 400 units” implies a precision that does not exist. “Between 340 and 470” is more useful, because it tells you how much room to leave.

If your system only gives point estimates, mentally add a range — and plan for the lower end of it when the cost of over-committing is high.

It predicts the past continuing

Every forecast assumes tomorrow resembles yesterday. That is usually true and occasionally very wrong.

A new competitor, a price change, a lost major customer, a regulatory shift — none of these are in your history, and all of them break the forecast.

Where it works and where it does not

Works well: established products with years of history, recognisable seasonality, a stable market, and enough volume that patterns are visible above the noise.

Struggles: new products, a few large unpredictable orders, project-based businesses, markets that have just changed structurally.

A business selling ten thousand units a month across a stable range gets genuinely useful forecasts.

A business winning four large contracts a year does not. The pattern it would need to learn does not exist, and a forecast will be produced anyway.

Knowing which you are is the most important thing in this article.

The promotion trap

Worth its own mention, because it catches people repeatedly.

If you ran a promotion last March, your history shows a spike in March. The forecast may repeat that spike next March — even if you have no promotion planned.

Equally, if you plan a promotion this June, the forecast will not know unless you tell it.

Tag promotional periods in your data. Otherwise the model learns that certain months are naturally busy when really they were artificially busy, and it will keep expecting a lift you no longer create.

Using it in practice

Forecast at the level you decide at. If you plan capacity by product family, forecast by product family. Individual product forecasts are noisier and you probably do not act on them at that level anyway.

Review monthly, not quarterly. Compare last month’s forecast against what happened. That comparison, done regularly, is how the forecast — and your judgement about it — improves.

Keep human overrides visible. When someone adjusts a forecast, record why. In six months you will want to know whether the override or the model was right.

Track your error. How far off were you, on average? A forecast you have never measured is a number you should not be planning on.

What it will not do

Predict something unprecedented. A pandemic, a competitor collapsing, a regulation change. By definition these are not in the history.

Replace judgement. Your sales team knows a major customer is about to change supplier. The model does not.

Be accurate on new products. No history, no pattern. It will produce a number and the number is decoration.

Work on bad data. Sales history with wrong dates, missing records or unrecorded returns produces a confident wrong forecast.

The short version

Demand prediction is for decisions that need weeks or months of notice — capacity, cash, commitments, hiring.

It works where you have years of history, stable patterns and enough volume for the signal to show. It does not work on new products, lumpy demand, or a market that just changed.

Read it as a range, not a number. Match the horizon to the decision. Measure your error every month. And keep the people who know what is coming in the conversation, because the forecast only knows what already happened.