Demand sensing: what it is and when it helps retailers

By Elie Dufeu, CTO & Co-Founder, Metreecs. Published 13 October 2026.

Demand sensing is a short-horizon forecasting method that adjusts the next one to four weeks of demand using the most recent signals, such as daily sales, promotion activity, and stock position, instead of waiting for the next monthly or weekly planning cycle. It helps most where demand moves faster than the plan does, and it helps very little where history is thin or the signals arrive late.

Amara, a planner at a home goods retailer, launched a three-day promotion on a ceramic dinnerware set with a weekly forecast of 400 units. By the end of day two, 190 units had sold against 114 expected for those days. The weekly plan did not change until Monday, and the product stocked out on day five. (Amara is an illustrative example.) Nothing was wrong with the forecast she had made on the previous Monday. It simply could not see what had happened since.

This guide explains how demand sensing differs from demand forecasting, which signals it uses, when it pays off, and where it does not work. If you want to see how short-horizon adjustment would behave on your own assortment, you can test demand sensing on your own sales data.

Key Takeaways

• Demand sensing revises the next 1 to 4 weeks of a forecast using recent signals and sits on top of a longer-horizon baseline forecast, which it does not replace.
• A 2011 benchmark study of nine multinational consumer products companies, reported by Supply Chain Management Review, found demand sensing cut weekly forecast error by 40%. Treat that as evidence of direction, because it covers consumer goods suppliers rather than retailers.
• Traditional ordering processes take 7 to 14 days for a shelf sale to turn into a retail order, according to AMR Research analysis published by SCDigest. Sensing targets that lag.
• In the worked example below, a promoted product that sells 190 units in two days against 114 expected moves from a 286 unit remaining forecast to 340 units, a 19% revision made in time to reorder.
• Sensing adds little for new products, intermittent demand, and any product whose supplier lead time is longer than the sensing horizon.

What is demand sensing?

Demand sensing is the practice of updating a short-term forecast, usually daily, from fresh demand signals so that near-term replenishment and allocation decisions follow what customers are doing now. The horizon is short, typically one to four weeks, because that is the window where recent data still carries information the baseline forecast lacks.

An older description from AMR Research analyst Lora Cecere, published by SCDigest in 2009, frames it as turning downstream data into insight on what is selling and whether promotions are working, with minimal delay. The idea has not changed much since. What changed is the amount of data retailers can read each day and the compute available to re-run the numbers.

Two ideas sit inside the definition. The first is a refresh rate: the forecast is recalculated as often as new data arrives, rather than once a cycle. The second is a signal hierarchy: point-of-sale data, stock levels, and promotion calendars get more weight over a few days than a long history does.

How is demand sensing different from demand forecasting?

Demand forecasting estimates demand over a long horizon, often weeks to months, from history, seasonality, trend, and planned events. Demand sensing corrects the first few weeks of that forecast as actual signals come in. Buying and production follow the first, and replenishment and transfer decisions follow the second.

Layering matters here. A sensing model that has no good baseline underneath it chases noise: one strong afternoon becomes a trend, and the next weak day cancels it. Across Metreecs' work with retailers, one mistake we repeatedly see is switching on short-horizon adjustment before the baseline forecast is stable at the product and location level, which pushes daily noise straight into purchase orders. For a deeper view of what a good baseline needs, see our piece on combining internal and external data for forecasting.

Which signals does demand sensing use?

Demand sensing works from signals that are fresh, close to the customer, and available within a day. The more of them you can connect, the better the short-term read, although the first two usually deliver most of the value.

• Point-of-sale and e-commerce orders, by product, variant, and location, refreshed daily or faster.
• Current stock position, including stock in transit, because a product that sold out looks like a demand drop when it is really a supply gap.
• Promotion and price changes, so the model can tell a promotional spike from a real change in baseline demand.
• External signals such as weather, local events, or web traffic, which matter in some categories far more than in others.

Fresh data also has to be clean. Returns, cancelled orders, and stockout days all distort the recent history, and a sensing model reads them as demand. The planners we work with often underestimate how much stockout days distort recent history, and correcting the data before it is read matters more than the tooling.

Siobhan runs replenishment for a garden and outdoor retailer. Two weeks of unseasonal heat lifted fan and cooler sales across the south of the country, while the northern stores saw nothing. A weekly plan by region treated them the same. A daily signal by location showed the split within three days, and she moved stock between stores before the south sold through. (Siobhan is an illustrative example.)

When does demand sensing help most?

Demand sensing earns its keep when four conditions hold: demand moves quickly, the retailer can act on a short notice, history is reasonably rich, and the cost of being wrong is high. Those conditions are common in promotions, weather-driven categories, fast-moving consumables, and any product with a short replenishment cycle.

The 2011 benchmark study mentioned earlier is the most cited public data point. It covered more than 7 billion cases and over $200 billion in sales across nine multinational consumer products companies, and reported that demand sensing reduced weekly forecast error by 40% across participants, while the same study's weekly demand planning error remained close to 50%. The study is over a decade old and covers suppliers rather than store networks, so use it as evidence that short-horizon signals improve near-term accuracy, not as a benchmark for your own results.

The same body of work points to why the lag matters. AMR Research, in the SCDigest article cited above, reported that traditional processes take 7 to 14 days for shelf sales to turn into a retail order. It also gave a promotion example: a 30% sales lift can produce an order for 30% more product the following week, which arrives after the promotion has ended and leaves a stockout during it and excess after it. Sensing shortens that loop by acting on the first days of the lift.

A worked example

Take a promoted product with a weekly forecast of 400 units, or about 57 units a day. Over the first two days it sells 190 units against 114 expected. The remaining forecast for the week is 400 minus 114, or 286 units. A sensing model that weighs the early lift against the promotion calendar and stock position might raise the remaining forecast to 340 units. That is 54 more units, or a 19% revision, and the week now totals 530 units rather than 400.

Those numbers are illustrative, and the model would not carry the whole early lift forward. The point is the timing: the revision lands on day three, while a replenishment order can still arrive before the shelf is empty. Under the weekly plan, the same information would have been available on Monday, after the stockout. For retailers reviewing how replenishment logic reacts to signals like this, our article on how AI is changing replenishment covers the ordering side.

When does demand sensing not work?

Demand sensing has limits, and they are worth naming before a project starts. It cannot create information that is not in the data, and it cannot move stock faster than the supply chain allows. Four situations come up most often.

Lead time longer than the horizon. If a supplier needs eight weeks, a four-week sensing signal changes nothing about what you can order. It may still help with allocation across stores, which is a faster lever.

New products without history. There is little baseline to correct, and the first days of sales are noisy. In the same 2011 study, new products carried a weekly error of 65% in their first year against 46% for existing products, which shows how hard the first year is whatever the method. New product forecasting needs its own approach, usually based on similar products and attributes.

Intermittent demand. A product that sells a few units a month gives the model almost nothing to sense. A two-unit day is a coin flip, and no model reads a trend into it.

Slow or incomplete data. AMR Research noted that in medical supplies, consumption data from care facilities is hard or impossible to obtain, so manufacturers rely on distributor shipments. Any category where sell-out data reaches you late will see a smaller benefit.

In our experience deploying short-horizon forecasting, the teams that get value from it are rarely the ones with the most advanced models. They are the ones whose planners can act on a daily change, with the authority to move stock or adjust an order without a weekly approval cycle.

How to start with demand sensing

You do not need a large program to test whether sensing pays off. A small pilot on a narrow set of products is enough to judge the value before widening it.

1. Pick products with fast, frequent sales and short lead times. Promoted items and fast sellers in a few locations are the best test.
2. Fix the baseline first. Check that the long-horizon forecast is stable at product and location level, using WMAPE and bias over several weeks.
3. Measure the short horizon separately. Track one to four week error for the baseline and for the sensed version, so the gain is not blended into the monthly number.
4. Decide what action each signal triggers. A revised forecast that nobody acts on has no value. Define the reorder, transfer, or allocation rules in advance.
5. Watch bias, not only accuracy. A sensing model that over-reacts to promotions can improve accuracy on some days and still tilt orders upward. Reviewing bias by week catches that early.

Tomás leads planning for a regional pharmacy group and ran a pilot on twenty fast-moving products across forty stores. He tracked the one-week error before and after, and found the gain concentrated in the five products with active promotions. He kept sensing for those and left the slow sellers on the baseline. (Tomás is an illustrative example.) That narrower scope was cheaper to run and easier to defend internally than a blanket rollout.

If the goal is cutting overstock and stockouts together, short-horizon signals work best next to sound stock targets. Our inventory optimization page shows how safety stock and reorder points use the sensed forecast, and Metreecs' AI demand planning approach builds the baseline from product and location history. The cost of getting this wrong is covered in the hidden cost of inaccurate sales forecasts.

FAQ

What is demand sensing in simple terms?

Demand sensing updates the near-term forecast, usually the next one to four weeks, using the latest sales and stock data. It lets a retailer adjust orders and transfers to what customers are doing now, rather than what a weekly or monthly plan assumed.

What is the difference between demand sensing and demand forecasting?

Forecasting covers a longer horizon from history, seasonality, and planned events, and the first weeks of that baseline can then be revised using fresh signals. Most retailers run both.

How accurate is demand sensing?

It depends on the category, the data, and the baseline. A 2011 benchmark of nine consumer products companies reported a 40% reduction in weekly forecast error, but that study covers suppliers and is more than a decade old. Measure the gain on your own products with a controlled pilot.

Does demand sensing work for new products?

Rarely, in the first weeks. With no history there is little baseline to adjust, and early sales are noisy. New products are better served by methods built on similar products and attributes, with sensing added once a few weeks of data have accumulated.

What data do I need for demand sensing?

At minimum, daily sales by product and location, current stock including in-transit, and a promotion and price calendar. Weather or web traffic can help in some categories. Data quality matters more than the number of sources, since returns and stockout days distort recent history.

How is demand sensing different from real-time forecasting?

Sensing is defined by its short horizon and the decisions it changes, such as reorders and transfers. Real-time forecasting describes how often the model updates. A sensing model can refresh daily without being real time, so the terms overlap without being interchangeable.

Conclusion

Demand sensing is a short-horizon layer on top of a stable baseline forecast. It pays off for promoted, fast-moving, short lead time products, and it adds little where lead times are long, history is thin, or data arrives late. To see how this would work on your assortment, book a demo and bring your own sales history.