Your demand history is lying about your best items.

There is a defect in almost every demand history I have ever looked at, it points in a consistent direction, and it makes your planning system most confident about exactly the items it should be least confident about. It's called censored demand, and correcting it costs almost nothing compared to what it recovers.

The loop

A customer wants an item. It's out of stock. What happens next depends on your order entry practice, and there are only two possibilities. Either the line never gets entered — the rep says "we're out," the customer moves on, nothing is recorded. Or the line gets entered and immediately zeroed out, leaving no trace of the requested quantity.

Either way, your demand history for that item, that week, records a number lower than the truth. Often it records zero.

Now run that history through any forecasting method — statistical, judgmental, machine-learned, it does not matter. The method sees weak demand and recommends carrying less. You carry less. You stock out again, faster. You record another suppressed number. The recommendation drops again.

This is a self-reinforcing loop, and the direction it runs in is the cruel part: it starves the items customers wanted most, in favor of items that were always available and therefore always fully recorded. The item that sat on the shelf all year has a complete, honest demand history. The item that flew out the door and stocked out in week three has a history that says nobody wanted it after week three.

It corrupts your service metric too, in the same direction

This is the part that tends to land in an executive room. If the shortfall never reaches an order line, your fill rate never sees it. The demand that you failed to serve isn't in the denominator.

Which means a company can drive its reported service level up by being reliably out of stock on the same items for long enough that customers stop asking. The metric improves precisely because you failed, consistently, over a long enough period to change customer behavior. I have watched teams celebrate that improvement.

Any service metric built on order lines has this hole. It's not a reason to abandon the metric — it's a reason to fix the capture, because the same fix repairs both the forecast and the measurement.

The fix is a process fix

There is no clever statistical repair for data you never captured. There are academic methods for estimating censored demand — you can model the truncation, fit a distribution, impute what the order would have been. They're real, and I'd use them on history that already exists. But they estimate. The fix for going forward is boring and much better:

Enter the customer's full intended order. Then cancel the unfillable lines with a reason code.

That's it. The demand now exists in the history at the quantity the customer actually wanted. The shortfall is attributable to a cause. Both the forecast and the fill rate see what really happened. You've converted an invisible failure into a measured one, which is the only kind you can act on.

The two objections, and what to do with them

"That's extra work for customer service." True, and small — it's keystrokes on a line that's already open. It's also the objection that always surfaces first and should be answered with a pilot rather than an argument. One willing account, one willing planner, a few weeks. Process changes that touch order entry either prove themselves on a small surface or die in a committee.

"Cancelled lines will confuse the forecasting system." This one used to be a legitimate architectural objection, in the era when forecasting platforms consumed net shipments. Many now consume order-line history through extracts, in which case the objection has quietly expired — and I've seen it survive years past the constraint that produced it. Before you accept it, go check what your platform actually ingests today. The answer has changed at more companies than have noticed.

What to do about the history you already have

Three things, in order of effort:

1. Flag the periods. If you have stock status history, you know when each item was unavailable. Mark those periods as unreliable rather than treating them as observed demand. Even excluding them from the fit is better than believing them.
2. Estimate the gap where it matters. For A-class items with meaningful stockout history, impute what demand likely was during the censored periods. You don't need this for the whole catalog — you need it for the items where being wrong is expensive.
3. Start capturing properly now. Everything above is a patch on data you can't fully recover. Twelve months from now the only thing that matters is whether you started capturing intended demand today.

Why I look for this first

In a diagnostic, this is near the top of my list, for three reasons. It's universal — any business that ever runs out of anything has it. It's cheap to fix relative to what it recovers, because the fix is process rather than technology. And it's diagnostic of something larger: a company that has never noticed this is usually a company where nobody has looked hard at what the planning data actually represents, which tells me where else to look.

The tell is easy to check. Pull your ten worst-serviced A-class items and look at what the forecast recommends for them. If the recommendation is "carry less," you're in the loop.