Skip to content

See a demo

30 minutes with Sasha Zhang · video link on confirmation

Loading scheduler…

CPG glossary

Demand sensing in CPG, explained

What demand sensing is

Demand sensing is short-horizon forecasting: using last week's actual sell-through to correct the next one to four weeks, instead of waiting for the monthly planning cycle to re-plan. On a promoted natural snack SKU at 140 Sprouts stores, that is the difference between shipping against a number locked three weeks ago and shipping against Monday's Retail Link pull.

Sold as short-term demand forecasting, it is the same idea under a plainer name. The statistical forecast sets the shape of the quarter. Sensing adjusts the front of it with signal that arrived after the plan was locked. It does not replace demand forecasting, and a brand that cannot forecast a baseline will not be rescued by sensing a bad one faster.

Demand sensing vs demand forecasting

Demand forecastingDemand sensing
Horizon1 to 18 months1 to 4 weeks
RebuiltMonthly, on the S&OP cycleWeekly, when POS lands
InputBaseline, seasonality, planned promoLast week's actual sell-through
AnswersHow big is the quarterIs the front of it landing where we said
Fails asError against actuals over the horizonA correction nobody has time left to act on

The volatility that decides whether it pays

Sensing earns its keep on variation, not on level. Seasonality is the part that repeats and is already in the model. Volatility is the week-to-week variation left over once you remove it, and its size decides how much a short-horizon correction can buy you.

SKU profileBaseline units/store/wkWeek-to-week variationSensing worth it?
Staple, no promo9±10%Marginal
Promoted natural snack9 to 25±60%Yes
New item, first 12 weeks3 to 14±80%Yes
Seasonal, in season12 to 30±45%Yes
New item, first 12 weeks±80%Promoted natural snack±60%Seasonal, in season±45%Staple, no promo±10%
Sensing corrects the volatile weeks. The staple in gray does not have volatile weeks to correct

A staple at ±10% does not need sensing. It needs to be left alone. The volatile weeks are where the miss lives, so they are the only weeks a weekly correction pays for.

What a weekly correction actually recovers

Here is a different four-week window on the same 10 oz roasted almond SKU, units per store per week, with a 25%-off TPR in W3. These are not the four weeks demand forecasting scores on that page, and the actuals differ accordingly. The locked plan was set at the start of the month. The sensed forecast applies one rule: carry last week's actual-to-plan ratio forward, damped 50%, so a single soft week does not whipsaw the next one.

Forecasts are shown to two decimals because the error column is computed from the unrounded value.

WeekLocked planActualLocked errorSensed forecastSensed error
W112933.3%12.0033.3%
W2121020.0%10.505.0%
W3252119.0%22.929.1%
W412119.1%11.040.4%

The W2 correction: W1 came in at 9 against a plan of 12, a ratio of 0.75, damped to a 0.875 factor, so 12 x 0.875 = 10.50 against an actual of 10. Carry the same rule through W3 (10/12 damped gives 0.9167, so 25 x 0.9167 = 22.92) and W4 (21/25 damped gives 0.92, so 12 x 0.92 = 11.04).

MAPE on the locked plan is (33.3 + 20.0 + 19.0 + 9.1) / 4 = 20.4%. On the sensed forecast it is (33.3 + 5.0 + 9.1 + 0.4) / 4 = 12.0%. The correction removed 8.4 points of error, about 41% of it.

I never got a sensing process approved on the strength of a vendor's average. What got it approved was running the rule over our own last two quarters and showing which specific weeks it would have moved, which is a much shorter list than the pitch implies.

Note which week did not improve. W1 has no prior signal to carry, so sensing cannot touch it. That is the honest shape of the benefit: sensing buys you weeks 2 through 4 of a window, never the first one, and the vendor deck that shows an even improvement across all four is showing you a backtest, not a forecast.

Lead time is what caps it

A correction you cannot act on is a more current number to be wrong with. Before buying sensing, write down what decisions are still open inside a one-week signal.

DecisionTypical lead timeCan Monday's signal move it?
Allocate on-hand DC stock by banner1 to 3 daysYes
Expedite a partial truck3 to 7 daysYes
Cut in or cancel a retailer promo2 to 4 weeksSometimes
Co-packer production run4 to 8 weeksNo
Ingredient and packaging buy8 to 16 weeksNo

If the co-packer needs six weeks and the distributor holds four weeks of cover, a signal that arrives on Monday cannot change what ships this month. It can still change allocation between retailers, and on a promoted SKU that is often the decision worth the most. This is the same constraint that separates a forecasting problem from a planning one on the replenishment planning side of the loop.

Where Scout fits

Scout is a demand-side analytics layer. It connects your SPINS, Circana, or retailer POS exports so last week's actual sell-through lands as a series you can compare against the plan you locked, by retailer and SKU, and so you can score whether a weekly correction would have helped before you build a process around it. It is not an S&OP system or a supply-planning engine. It does not place orders, hold your lead times, or generate the replenishment plan. It measures the signal and the size of the error a correction removes.

The short version

  • Demand sensing is short-horizon forecasting: correct the next one to four weeks with last week's actual sell-through.
  • It pays on volatility, not on level. A staple at ±10% week-to-week variation does not need it.
  • On the worked example a damped one-week correction took MAPE from 20.4% to 12.0%, removing about 41% of the error.
  • It cannot improve the first week of a window, because there is no prior signal to carry.
  • Lead time caps the whole idea. If nothing downstream can move inside a week, sensing buys a more current number and no decision.
See your CPG data answer questions in plain English — book a Scout demo

Want the rest of the CPG analyst's glossary?

Drop your email and we'll send the full set of CPG and retail-data definitions as one reference sheet.