Sign in

Blog · Forecast and pipeline · Supply chain

Demand forecast bias by SKU and site: the plan that is always high in one place

How a supply chain team measures demand forecast accuracy and bias per SKU per site from the forecast snapshots and the actual demand: error and signed bias, the SKU-site pairs that are consistently over-forecast and carry the excess stock, the ones consistently under-forecast and carry the stock-outs, why bias at one site is usually a planner's override, and the adjustment the team's own history supports.

The short answerDemand forecast bias per SKU per site is the mean signed difference between the forecast made at a fixed lead time and actual demand, over a trailing window, from forecast snapshots and the demand history. Error ignores direction; bias keeps it. A SKU-site pair over-forecast for six months running is carrying excess stock that the bias explains; one under-forecast for six months is stocking out for the same reason. Bias concentrated at one site across many SKUs is usually a planner's override, and the report shows the override history beside it.

A demand planner's accuracy figure is 72 percent across the network and stable. Inside it, one distribution centre is over-forecast on every fast-moving SKU by a fifth, and has been for a year, and the excess stock there is the bias made physical. This guide sets out forecast bias per SKU per site from snapshots, the pairs that carry it, and the site-wide pattern that usually means an override.

The measures

Per SKU, per site, per period:

Forecast at lead time L = the snapshot L periods before Signed miss = forecast − actual Bias = mean signed miss ÷ mean actual, trailing window Error = mean absolute miss ÷ mean actual

Per site:

Share of SKUs with bias beyond a stated threshold, in each direction

The rows you need

  • Forecast snapshots: SKU, site, period forecast, forecast made date, quantity.
  • Actual demand: SKU, site, period, quantity.
  • Override log: SKU, site, period, system forecast, planner forecast, where held.
  • Inventory: SKU, site, period, on hand, where held.

SKU and site identifiers only.

The assertion

every period has one forecast per SKU-site at lead time L, or is marked missing

Periods with no snapshot at the lead time are excluded with a count, not filled from a later one.

A worked view

Lead time 8 weeks. Trailing 26 weeks.

SKU Site Bias Error Consecutive periods same sign Stock vs target Reading
K-104 DC-North +21% 24% 24 of 26 +$180,000 Over-forecast; excess
K-104 DC-South +2% 18% mixed on target Fine
K-109 DC-North +19% 22% 23 of 26 +$60,000 Over-forecast; excess
K-227 DC-West −17% 21% 22 of 26 6 stockouts Under-forecast; short

Two SKUs at one site over-forecast by a fifth for almost every week of the window, with the excess stock to show for it. The same SKU at the other site is fine. The site is the pattern.

Per site

Site SKUs Over-biased share Under-biased share Overrides last quarter
DC-North 1,840 41% 6% 620
DC-South 1,790 9% 8% 40
DC-West 1,210 7% 19% 210

DC-North's planner overrides six hundred forecasts a quarter, upward, and two fifths of the site's SKUs carry the bias. That is one conversation, not eighteen hundred.

Where it goes wrong

Accuracy without bias. Seventy-two percent and no direction.

Latest forecast measured. Always accurate; never informative.

Network-level only. The site with the override is inside the average.

Adjustment applied silently. The planner's number changes and nobody knows why. Shown, then discussed.

Every month, bias by pair and by site

Mapped once, the snapshots, the actuals, the override log and the inventory produce bias and error per pair, the site pattern and the suggested adjustments every month. Covirage builds this from the exports as they are. The supply chain page describes the setup, and the lead time variability guide covers the other input the same excess stock usually traces to.

Questions people ask

Which forecast is measured?

The forecast as it stood at a fixed lead time before the period, the same lead time every month, from the snapshot. A forecast revised the day before the period is always accurate and tells nothing about the plan that drove the purchase orders.

How is bias separated from a demand shift?

A demand shift moves actuals away from a forecast that was right until then, and the bias appears and then closes as the forecast catches up. Persistent bias with a stable actual is the forecast itself. The report shows both series so the reader can see which.

What is done about it?

Per SKU-site pair, an adjustment by the trailing bias is computed and shown, not applied. Per site, the override log is reviewed where the bias is site-wide. The planner sees the pairs and the direction, and most bias is corrected once it is visible.