Demand Sensing: Short-Term Forecasts From Live Signals

The monthly plan was right, and the shelves were still empty

A demand plan built once a month can be accurate over the month and still wrong on the day. A warm weekend sells out the barbecue range in two regions. A competitor's stock problem sends their customers to you on a Tuesday. A school holiday shifts sandwich sales from city centre to seaside. The monthly number does not care; your warehouse and your staffing rota do.

Demand sensing is the practice of correcting near-term forecasts using signals that arrive daily or hourly. The long-range plan still sets purchasing and capacity. Sensing adjusts the next few days to weeks, where decisions are still possible but the plan is already stale.

How demand sensing differs from demand planning

If you do not yet have a reasonable longer-range forecast, start there. Our post on time series forecasting for demand planning covers that foundation. Sensing without a baseline is just chasing yesterday's sales.

Live signals that actually move short-term demand

• Recent sales and order intake. The last few days are the strongest signal of all, as long as the model separates a trend from a blip.
• Weather forecasts. Hugely important for food, drink, garden, clothing and some DIY categories, largely irrelevant for others.
• Web and app activity. Product page views and basket additions often lead store and online sales by a few days.
• Open orders and quotes. In B2B, what customers have booked or requested pricing on.
• Local events and calendars. Football fixtures, concerts, school holidays, Ramadan timing, pay days.
• Stock and availability. Your own and, where visible, competitors', since a rival's stock-out is your spike.

Not every signal suits every business. Part of the work is testing which signals improve accuracy for your products, and discarding the rest, because each one is a data feed someone must keep running.

An illustrative example

Take a regional bakery chain with a central kitchen supplying 30 shops. Production is planned the afternoon before, using a standard order per shop and day of the week adjusted by shop managers. Waste is high on wet days and shops sell out early on sunny Saturdays.

A sensing model producing next-day quantities per shop and product might use the last two weeks of sales, tomorrow's weather forecast, local events and school term dates. Even a modest reduction in both waste and sell-outs pays for the work quickly in fresh food, because the product has no second chance. The same approach applied to a hardware distributor with four-week lead times would achieve far less, since the decision it improves is already locked in.

Where it goes wrong

1. Overreacting: one strong day triggers a large order that becomes waste when demand returns to normal
2. Signal feeds breaking silently, so the model forecasts on stale weather or missing sales data
3. Forecasting demand when the data only shows sales, so stock-outs are learned as low demand
4. Nobody able to act on a daily forecast because ordering or staffing is fixed weekly
5. Too many signals, each adding complexity for tiny gains

The fourth is the most common reason we advise against sensing. A daily forecast is only useful if some decision can change daily. Check the decision cycle before building anything.

When demand sensing is worth the effort

The strongest cases share a few traits: perishable or short-life products, short lead times, demand driven by weather or events, and a meaningful cost to both over- and under-supply. Fresh food, bakery, florists, hospitality, events retail and on-demand staffing businesses fit well. So do eCommerce businesses where warehouse labour is planned daily.

The weakest cases are long lead-time supply chains, stable demand and businesses with few, large customers. For those, improving the weekly plan and safety stock rules will return more.

How we would approach it

At SpiderHunts we start with a back-test: take the existing plan, add recent sales alone, and measure improvement on next-day or next-week accuracy. Then add one external signal at a time, keeping only those that help measurably. Most of the eventual system is data engineering, reliable feeds, monitoring and a way to get numbers to the people ordering or rostering, which is squarely the kind of work our data science and automation teams do together.

Expect a first version in six to ten weeks for a single decision, such as daily production quantities. Resist the urge to sense everything at once.

Frequently asked questions