Demand Sensing vs. Demand Forecasting: How to Run Both Without Making Your Plan Worse

Steve Morlidge studied more than 300,000 forecasts and found that 52% of them were worse than a naive random-walk forecast, according to Gilliland, Tashman, and Sglavo’s Business Forecasting: Practical Problems and Solutions. More than half the time, doing nothing beat the forecasting process. That is the number to keep in mind before adding demand sensing on top of any planning process. Our demand sensing guide covers what each approach is. This article covers the harder question: how to measure them, where one hands off to the other, which to fix first, and what breaks when a company runs both.

Do you need demand sensing, demand forecasting, or both?

Most companies that need demand sensing need both, because the two do different jobs. Forecasting produces the baseline plan for months and quarters ahead. Sensing adjusts the nearest weeks of that plan as fresh signals arrive. Sensing without a decent baseline has nothing sound to adjust, and a baseline without sensing stays frozen until the next planning cycle, whatever the market does in between.

The useful question is not “which one” but “which one is the bottleneck.” A company whose baseline forecast is poor gains little from a fast layer on top of it. A company whose baseline is solid but whose short-term plan misses on promotions, weather, and sudden shifts is exactly where sensing pays back.

How do you measure demand sensing and demand forecasting differently?

Forecasting and sensing should be measured on the same error metric but at different horizons, and both should be judged against a simple benchmark. Judging sensing on a monthly accuracy number, or forecasting on a weekly one, hides what each is actually being paid to do.

Question	Demand forecasting	Demand sensing
What horizon do you score?	Months out, at the lag the plan was actually frozen	The nearest one to four weeks, at the lag where decisions get made
What is the benchmark?	A naive forecast, such as last period’s actuals	The existing baseline forecast for the same week
Which errors matter most?	Systematic bias, over- or under-forecasting month after month	Large misses on promotions, weather events, and demand spikes
What proves it works?	Beats the naive benchmark consistently	Beats the baseline it adjusts, week after week

The discipline that ties these together is forecast value added, or FVA: the change in an accuracy metric attributable to one step in the forecasting process, a concept Michael Gilliland introduced. Every layer, including sensing, should have to prove it beats the layer beneath it. A sensing model that does not beat the baseline forecast it adjusts is adding cost and complexity for nothing, and the Morlidge finding above shows how common it is for forecasting steps to fail that test.

Where does demand sensing hand off to demand forecasting?

Demand sensing takes over inside the short window where fresh signals still change what you can do, typically the first one to four weeks of the planning horizon, while the consensus forecast owns everything beyond it. The handoff point should be explicit, a specific number of weeks written into the planning process, not something planners decide case by case.

Beyond the handoff, the sensed signal has to be reconciled with the plan it adjusts. Practitioners use caps, dampening, and exception thresholds so that a noisy signal cannot swing replenishment wildly from one day to the next. Without those guardrails, a sensing model that reacts to every fluctuation creates its own volatility, which is the opposite of what it was bought to do.

Two design decisions matter most at the handoff:

• Where the sensed number lands. It should adjust the operational plan for the near weeks, replenishment and allocation, without silently rewriting the consensus forecast that finance and sales signed off on. Overwriting the baseline destroys the ability to measure whether sensing helped.
• What triggers a change. A move in the sensed signal should only change the plan when it crosses a defined threshold. Small movements are noise, and acting on them creates work without value.

Which should you fix first, your forecast or your sensing?

Fix the baseline forecast first. Sensing corrects short-term deviations from a plan; it cannot rescue a plan that is systematically biased or built on unreliable data. Layering a fast model onto a weak baseline makes the errors move faster, not smaller.

Use four checks to decide whether the baseline is ready for a sensing layer:

1. Does the baseline beat a naive forecast? If it does not, the problem is upstream of sensing. The Morlidge result suggests this is worth checking before spending on anything faster.
2. Is bias under control? A forecast that consistently runs high or low needs its process fixed, not a short-term correction bolted on.
3. Do current signals exist for the categories in question? Point-of-sale visibility, promotion calendars, and web traffic have to actually reach a model on a near-continuous basis. If they arrive in weekly batches, the speed advantage of sensing disappears.
4. Are the products fast-moving enough to benefit? Slow-moving items and long-cycle B2B orders gain little from a short-horizon layer, however good the model is.

If the first two fail, invest in the forecasting process. If the last two fail, sensing is the wrong tool for that part of the portfolio. Only when all four pass does a sensing layer earn its cost. Data quality sits underneath all four: a baseline built on inconsistent or unvalidated inputs will not pass the first two checks, which is why the advanced analytics work behind either approach starts with the data, not the model.

What goes wrong when companies run both?

The most common failures are organizational, not technical: two competing versions of the truth, manual overrides that quietly destroy accuracy, and no one accountable for the sensed number. The models rarely cause the damage. The process wrapped around them does.

• Two sources of truth. Supply planning works from the sensed number while finance and sales work from the consensus forecast, and nobody reconciles them. Every planning meeting then starts with an argument about which number is real.
• Overrides that make the forecast worse. Manual adjustments feel like judgment and often reduce accuracy. FVA analysis exists precisely to find the overrides that destroy value, and it applies to sensed numbers as much as statistical ones. Track every override against the number it replaced.
• Sensing measured against the wrong benchmark. Comparing a sensing model to last month’s forecast flatters it. It has to beat the baseline for the same week, at the lag decisions were made.
• No owner for the sensed number. If nobody is accountable for acting on it, or for challenging it when it looks wrong, it becomes a dashboard that planners glance at and ignore.
• Treating the model as finished. Signals that predicted demand last year may not next year. Both layers need periodic retesting against fresh outcomes.

These are process problems, and process problems respond to measurement. The practical minimum is one accuracy report per layer, scored at the right horizon against the right benchmark, reviewed on a fixed cadence by a named owner. For strategy teams that also need a longer view, our guide to market forecasting methods and models covers the medium-to-long-range side that sensing never touches.

Infomineo transforms complex data into predictive insights and actionable intelligence through end-to-end analytics services, including data quality assurance and validation and data science and predictive modeling.

Talk to our data analytics team →

What does a combined planning rhythm look like in practice?

A combined rhythm runs on two clocks: a slow one for the forecast and a fast one for sensing, with a review point where the two meet. The slow clock is the monthly or quarterly cycle where sales, finance, and supply agree the consensus forecast. The fast clock is a daily or weekly refresh of the sensed signal for the nearest weeks. What ties them together is a short, regular review that compares both against actuals.

1. Monthly: lock the consensus forecast and record it as the benchmark that sensing has to beat.
2. Weekly: refresh the sensed signal, apply the thresholds, and release only the changes that cross them to replenishment and allocation.
3. Weekly review: score the sensed number and the baseline against last week’s actuals, and log every manual override with the reason for it.
4. Quarterly: retest which signals still predict demand and drop the ones that stopped, then review overrides that reduced accuracy.

The cadence is deliberately dull. Companies that get value from running both layers tend to be the ones that turn the comparison into a routine, not the ones with the most sophisticated model.

Frequently Asked Questions

Can you use demand sensing without a demand forecasting system?

Technically yes, but it rarely works well. Sensing adjusts a baseline plan, so without a reliable baseline it has nothing sound to correct. Companies that skip the forecasting foundation usually find the sensing layer amplifying noise instead of removing it.

What is forecast value added, and why does it matter here?

Forecast value added measures the change in an accuracy metric that a single step in the forecasting process contributes. It matters because it tests whether each layer, including demand sensing, actually beats the layer beneath it, rather than assuming that more sophisticated means more accurate.

How far ahead does demand sensing look?

Practitioner frameworks typically place demand sensing’s highest value in the nearest one to four weeks, with some extending to a few months. Beyond that window, demand forecasting built on historical patterns and market intelligence does the work.

Who should own the sensed demand number?

One named owner in supply or demand planning, with a defined process for challenging it. Shared ownership usually means no ownership, and an unowned sensed number ends up ignored or overridden without anyone tracking whether the override helped.

Is demand sensing worth it if my forecast accuracy is already high?

Only if the remaining errors are concentrated in the short term, on promotions, weather, or sudden shifts. If your misses are long-range or come from slow-moving items, a sensing layer will not fix them, and the money is better spent elsewhere in the planning process.