Demand Sensing

A monthly forecast may not react quickly to an unexpected surge in one item, so demand sensing looks at recent evidence and adjusts the short-term view. Planners can then decide whether to move stock or change production.

Kinaxis describes high-frequency data such as point-of-sale transactions, open orders and promotions, and treats sensing as an addition to a consensus demand plan; a vendor's performance claims are not guaranteed outcomes for every business. A fictional grocer expects 100 units of an item next week, but recent local sales and an upcoming event suggest 130.

The sensed forecast moves to 130 only after the signals are checked for stockouts and promotions. Sales are not always the same as demand, because if the shelf was empty, recorded sales understate how many customers wanted the item, so include availability data before training a model on a stockout.

Each signal needs judgement. A large one-off order may be a real near-term need but not a recurring trend, so flag it separately or later weeks may be overforecast.

Weather can matter for drinks, heating equipment or outdoor goods, and online search interest may lead purchases, but both relationships differ by region and product, and a viral mention does not guarantee orders; test whether each signal improves accuracy rather than assuming it does. The useful forecast horizon depends on lead time, so a daily signal may be too late for a product imported over months.

It can still help allocate existing inventory among locations. A fictional retailer moves surplus umbrellas from one branch to another after a short-range weather forecast, checking transport time and stock needs at both sites, because moving too late would add cost without availability benefit.

Data quality is central, since late POS feeds, duplicate orders and wrong product codes can create false trends, so show when each signal was last refreshed. Represent promotions explicitly with their start and end dates, since a price cut can raise sales temporarily and treating the spike as organic demand can lead to excess stock afterwards.

A model should be measured against a suitable baseline by product and horizon, because if its near-term forecast is no better than a simple recent-sales average, the added complexity may not help. Forecast error can be expressed as absolute units or a percentage, but percentage error becomes awkward when actual demand is zero, so use measures that fit intermittent products, and record the key signals and a confidence range so planners can see why a forecast changed.

Avoid overreacting to noise, because one busy day should not trigger a costly replenishment if demand normally varies widely, and set review thresholds to the product and risk, since running out of medicine may matter more than holding extra stationery while perishable overstock can be expensive. Sensing can also help fulfilment teams prepare labour and help a fictional manufacturer that cannot accelerate supply to communicate lead times and prioritise allocation, provided it uses only permitted data, since a better forecast does not justify unrelated surveillance.