Can Demand Sensing Tell You the Trend Has Changed, Or Is It Just Noise?

Recently, someone asked me whether demand sensing is worth the investment. The pitch they had heard was the usual one: forecasts built from weekly or daily signals are more accurate over the short horizon than forecasts built from monthly history. That is true as far as it goes, but in my experience it is only half the story. The more interesting question is this: when the world changes, how early does one find out? In this blog post, I will explore that question with a worked example, and then discuss the trap that comes bundled with the answer.

Let Us Start With a Worked Example

Say a product sells at a steady rate of 100 units per week, and mid-quarter the true demand steps up to 130 units per week. Perhaps a competitor stumbled, or a new customer came online. The planner does not know this yet; all they see is data.

Consider first the planner working in monthly buckets. A normal month is 4 * 100 = 400 units. The step happens mid-month, so the month comes in at 2 * 100 + 2 * 130 = 200 + 260 = 460 units. Is 460 against an expectation of 400 a trend change? A one-time spike? A customer pulling orders forward? Just noise? In most businesses I have seen, monthly demand routinely bounces around by 10-15%, so 460 sits squarely in the gray zone. The next month comes in at 4 * 130 = 520, which is more suggestive, but one elevated month after an ambiguous one is still not proof (ask any planner who has extrapolated a spike and regretted it). Realistically, it takes 2-3 months of consecutive misses before the change is undeniable and the plan gets rebuilt.

Now consider the same situation viewed through weekly signals: customer orders, point-of-sale data where available, channel inventory movements. After the step, the weekly numbers read something like 128, 133, 127, 131. That is 4 consecutive points well above the historical mean of 100, visible within 3-4 weeks of the change (think control charts; a run of points on one side of the mean is a classic signal). Beyond a more accurate near-term forecast, the weekly view answers the question “has the level shifted?” some 6-8 weeks earlier than the monthly view does.

What are those 6-8 weeks worth? Let us do some rough arithmetic. During that window, the supply plan is wrong by 30 units per week. Over 7 weeks, that is 30 * 7 = 210 units of demand the plan did not provide for. If the margin is, say, $200 per unit, that is 210 * 200 = $42,000 of margin at risk from missed sales, before counting expediting premiums, overtime, and the service-level bruises with the very customer whose business just grew. These numbers are invented for illustration, and your numbers will certainly differ; but the structure of the calculation holds, and in my experience the answer is rarely small.

Now For the Trap

Here is the uncomfortable part: a system that reacts to real signals faster will also react to noise faster. That is the deal one signs up for. I have written before about forecasts that behave like a nervous system, twitching at every data point; a demand sensing process without discipline is exactly that. And a supply chain that chases its own twitchy forecast will manufacture its very own bullwhip, no customer required.

As the old cliché goes, even a broken clock is right twice a day. A planner who reacts to every single elevated data point will occasionally catch a real trend change impressively early, and will point to that occasion for years. What they will conveniently forget are the dozen false alarms in between, each of which moved inventory, production, and blood pressure for nothing.

So a rule of thumb: react to level shifts only when they are confirmed by several consecutive signals in the same direction, say 3-4, and never on a single point. The cutoff of 3-4 is somewhat arbitrary; a business with cheap, flexible capacity might act on 2-3, while one with expensive changeovers might wait for 5-6. Do what works with your data.

An analogy I like is smoke detector placement. The detector in the kitchen fires earliest, and it also goes off every time someone burns the toast. The detector in the hallway hardly ever false-alarms, but by the time it sounds, the fire has had a head start. Neither placement is wrong; where you place the detector is a choice about how many false alarms you will tolerate in exchange for lead time. Demand sensing moves your detector closer to the kitchen. The consecutive-signal rule is what keeps the burnt toast from evacuating the building.

Where Does That Leave Us

Demand sensing is often sold on forecast accuracy, and it does help there. But I would argue the larger prize is detection lead time: learning 6-8 weeks earlier that the level has shifted, while using confirmation rules to avoid chasing noise. The right confirmation threshold is an empirical question, so backtest candidate rules on your own history before trusting any of them (including mine). Which raises a question worth sitting with: how late is your current process at catching trend changes? There is a way to measure that from data you already have, but that is for the next post. In the meantime, have you seen a trend change that your monthly process caught embarrassingly late, or a false alarm that sent everyone scrambling? Please share via comments; I am interested in hearing from you.