Demand sensing vs forecasting — One extrapolates history. The other reads signals that history has not absorbed yet. Confusing them produces a fast forecast that is wrong in the same way as the slow one.
Demand sensing vs forecasting: bu ne anlama geliyor
“Demand sensing” is often implemented as the existing forecast, run more often. That is not what the term means, and running a bad forecast weekly instead of monthly mostly produces more frequent errors.
A statistical forecast extrapolates a pattern from history. It is good at seasonality, trend and the effect of events that resemble past events.
Demand sensing uses signals that have not yet reached the sales history: point-of-sale data from yesterday, a competitor’s stockout, weather, a promotion that has just started, web traffic on a category page.
The distinction that matters operationally
Ask what happens when something genuinely new occurs. A forecast reverts to the pattern. A sensing model reacts to the signal.
Most organisations need both, on different horizons: the forecast for capacity and purchasing months out, the sensing layer for allocation and replenishment inside the lead time.
Running the first and calling it the second is the common mistake.
What a sensing signal has to survive
Most candidate signals fail one of three tests, and it is cheaper to apply them before building anything.
Does it lead? Web traffic to a category page leads sales by days. A competitor’s price change leads by hours. Last week’s sales lead nothing; they are the thing being predicted.
Is it available in time? A signal that arrives in a monthly data feed cannot inform a weekly decision, however predictive it is. This eliminates more candidates than predictiveness does.
Does it move? A weather signal in a market with stable weather is a constant, and a constant carries no information regardless of how strong the underlying relationship is.
The horizon determines everything
Inside the supplier lead time, purchasing decisions are already committed. What remains is allocation — which location gets the stock that is already coming. This is where sensing pays, because the decision is still open and the signal is fresh.
Outside the lead time, sensing has nothing to add that the forecast does not already carry, and running it there produces volatility in a plan that needs stability.
Organisations that get no value from sensing have usually applied it at the wrong horizon rather than chosen the wrong signals.
The failure of over-reaction
A sensing layer that adjusts on every signal produces a plan that oscillates. Suppliers see erratic orders, warehouses see churn, and planners stop trusting the numbers.
The fix is a threshold rather than a smoother: react when the signal exceeds what normal variation would produce, and hold otherwise. A model that adjusts by three percent every week is not sensing demand; it is fitting noise.
What to build first
Not the model. The measurement of how much the forecast is already wrong inside the lead time, by category.
If the error inside the lead time is small, sensing has little to recover and the effort belongs elsewhere. If it is large and concentrated — promotions, new lines, weather-sensitive categories — that concentration tells you which signals to pursue and which decisions to point them at.
That analysis takes a fortnight and it has prevented several projects that would have taken a year.
The organisational catch
Sensing outputs arrive faster than most planning processes can absorb. A signal that says reallocate today reaches a team whose allocation run is Thursday.
The model is then blamed for being noisy, when what is happening is that its output is being averaged across a cycle it was built to operate inside. Shortening that cycle is usually a larger gain than any modelling work, and it is a scheduling change rather than a technology project.