Demand estimation and forecasting: what differs

Demand estimation and forecasting answer different operating questions: how much demand belongs to a period with incomplete observations, and how much demand to expect in a future period. A forecast is itself an estimate, so the useful distinction is the target period and the information available when the figure was produced. Record those before comparing either number with actual sales or orders.

What each one measures

In this operational comparison, a demand estimate reconstructs current or completed demand. It might combine shipped units with confirmed requests that remain unfilled, subject to checks for cancellations and duplication. A forecast projects units for a future period using history and documented assumptions.

Terminology varies. Estimation can also mean estimating a demand relationship from price, income or other factors. The table below defines the terms for a business review; it does not imply that every economist uses them this way.

The general demand forecasting guide covers the broader planning question. This comparison focuses on keeping the two figures distinct in the same review.

Start with a demand definition

Choose units, geography, product and date basis before choosing a method. A request placed this month for delivery next month has different timing under order-date and requested-delivery-date measures. Neither measure becomes clearer because it is called demand.

Shipments and orders also measure different events. The Census Bureau's manufacturing definitions distinguish shipments from new orders and define new orders net of cancellations. That distinction is useful inside a company even when its operational review uses units rather than the survey's dollar measures.

For the following example, demand means units requested for delivery during the target month, after cancellations. The shipment export and unresolved-request export share a request identifier. Each request can contribute to the reconstruction only once.

Keep the uploaded fields simple: SKU, site, target month, shipment units, remaining requested units, request status and extraction date. Add an estimate version or forecast issue date to the comparison file. The demand forecasting data checks explain input completeness without turning this comparison into a data cleanup exercise.

The same SKU rows, worked both ways

At the August close, a US distributor reviews demand before all order statuses are finalized. Its provisional reconstruction is:

Estimated period demand = Shipped units + Confirmed unfilled units assigned to that period

This is a stated reconstruction rule for the example, not a universal demand formula. Unconfirmed inquiries are excluded. Unfilled units already counted as shipped are excluded too.

Before September demand is observed, the planner supplies a September scenario with a uniform 10% uplift from this baseline. The uplift is an assumption, not a measured historical relationship. NIST notes that time series analysis considers trend, seasonality and dependence over time; a real planning method should examine those patterns rather than assume a uniform uplift is suitable.

September scenario forecast = August demand estimate × (1 + Supplied uplift)

The same SKU rows now carry a reconstructed August estimate and a September forecast. They do not describe the same period. The check on the scenario is that its SKU total equals the uplift applied to the total baseline: 297 units either way.

Compare each with the correct actuals

August order cleanup changes which requests remain valid. September actuals arrive later. Keep the original figures frozen and compare them separately, using signed error as the figure being assessed minus actual demand. Positive error means the figure was high.

August's net error is zero, but its absolute row errors total 4 + 4 + 0 = 8 units. September's absolute errors total 2 + 6 + 1 = 9 units. The figures describe different work: reconstruction quality for August and future-period accuracy for September.

Using weighted absolute percentage error, August is 8 / 270 = 3.0% and September is 9 / 300 = 3.0%, rounded to one decimal place. Their identical rounded scores do not make the tasks equivalent. The August estimate used evidence from the target period; the September forecast did not.

Forecast evaluation should use outcomes unavailable when the forecast was made, as explained in the authors' guidance on evaluating forecast accuracy. Replacing an issued forecast with a later revision changes the question being tested. Persistent directional errors belong in the separate SKU and site forecast bias review.

When to use which

Use estimation when the immediate question is whether recorded sales understate the demand that actually arrived. Operations owns the supporting availability and request records; sales confirms ambiguous customer requests. Finance checks that the comparison does not quietly switch between orders and recognized revenue.

Use forecasting when a decision must be made before demand is observed. The demand planner owns the issue date and assumptions, while procurement and operations specify the horizon their commitments require. A near-term purchase decision and a capacity discussion can need different forecast versions.

Agree on the handoff in the demand forecasting process: finalized estimates may become historical inputs, but they should not overwrite the original forecast used to judge the planning decision.

Where it goes wrong

• Treating shipments as unconstrained demand. Stock shortages can leave genuine requests unfulfilled. Preserve the evidence behind an adjustment instead of adding an unexplained uplift.
• Double counting the same request. An order partly shipped must contribute its shipped quantity and only its remaining unfilled quantity.
• Scoring different periods together. August reconstruction error and September forecast error need separate actuals and separate denominators.
• Changing the information cutoff. A revised estimate or forecast may be better informed. Preserve both versions so the earlier decision remains reviewable.
• Reading a correct total as correct rows. August's zero net error hid offsets across the products. Check absolute error at the level where people act.

Compare your own estimates and forecasts in Covirage

Bring a file with period demand estimates, issued forecasts, actuals and their SKU and site identifiers. Covirage maps the columns, and its tools compute and check the comparisons; the external AI model explains the results with every figure traced to its rows. Create a free Covirage account in about a minute, then set up your business and bring a sample of your own data.

See forecast analysis for the supporting approach.