Demand Sensing

A classical time series model treats every period as one observation of equal weight, so a level shift needs several periods before it moves the parameters. Demand sensing shortens that lag: it works on a horizon of days to a few weeks, weights recent observations more heavily and can absorb signals from further down the chain – point‑of‑sale data, orders already placed, weather, promotional activity. The output is not a second forecast but a correction of the existing one, usually at item‑location‑day level.

• Horizon – days to a few weeks, not months or a whole season.
• Input – recent demand plus downstream signals, not only the history of the item itself.
• Output – a correction of an existing forecast, not a replacement for the planning model.

Demand sensing in practice

The correction is worth only as much as the decision it can still change. Where the supplier lead time and the order period together cover several weeks, the quantity for that period is already on order by the time sensing registers the shift; what stays open is allocation between locations and redistribution of stock that is already in the network. Sensing pays off first in short cycles – fresh goods delivered daily, store replenishment from a central warehouse, e‑commerce – and in the allocation decision rather than in the purchase order.

The second condition is the input. Recent weeks have to describe demand, not sales: an item that was out of stock shows depressed sales, and a sensing layer reading raw sales cuts the forecast, orders less and deepens the stockout. Promotions behave the same way – where the promotional calendar is dense, a recent deviation is more often the promotional peak than a level shift, so the correction belongs on a cleaned baseline. And with intermittent demand, a few weeks of observations cannot separate a level shift from randomness at all.

Veritico STOCK uses demand sensing as one layer of the forecast: it compares the last several weeks against long‑term history and adjusts the forecast accordingly. It calculates on history cleaned of extremes, stockouts and promotional effects, and it predicts actual demand rather than censored sales – which is what stops a short‑horizon correction from amplifying an availability problem. The corrected forecast then feeds replenishment and allocation.

Related terms

• Demand forecasting
• Demand planning
• Baseline demand
• Time series forecasting
• New Product Forecasting