The canonical paper on forecast accuracy measurement found that percentage-based error metrics become unstable or undefined as actual demand approaches zero, and proposed a scale-free alternative for exactly this reason. Demand forecasting in supply chain management runs into that instability constantly at SKU-location level, where a large share of actual-demand values sit at or near zero on any given day.
In demand forecasting in supply chain management, getting the metric wrong at this grain doesn’t just misreport accuracy. It feeds a distorted error signal into every downstream buffer calculation. These are the four places that measurement usually goes wrong.
The zero-denominator problem
MAPE divides the error by the actual, and at SKU-location level, the actual is frequently zero or close to it, especially for slower-moving items at smaller locations. A single zero-demand day can make an otherwise reasonable forecast register as infinite percentage error. A handful of these blowups then drags the whole portfolio-wide MAPE average away from what typical performance actually looks like. Scale-free measures build from the average absolute change in the series itself. That reference point avoids MAPE’s zero-denominator problem, and it holds up at the SKU-location grain where MAPE breaks down.
The cancellation problem
Averaging error across many SKU-location combinations lets an overforecast at one location cancel an underforecast at another, producing a portfolio number that looks stable while individual locations run the opposite biases underneath. This is a different failure from the aggregation problem of measuring at category or network level. It happens even when you are measuring at SKU-location grain, as long as the errors are then averaged together with sign intact. Track bias alongside absolute error, at each SKU-location separately, and that catches it. A portfolio bias near zero can still hide every location running persistently high or low.
The cold-start problem
A new SKU-location combination, an existing item added to a new depot, or a new item added to an existing one, has no history to measure error against yet. Portfolio-wide accuracy reporting mostly either excludes these combinations silently or blends them into the same average as mature SKU-locations, and both choices misstate what’s actually being measured. New combinations need their own tracking window and their own realistic accuracy expectation until enough periods accumulate to measure them the same way as everything else.
Why this feeds a multi echelon inventory optimization model
Error measured wrong at SKU-location level doesn’t stay a reporting problem. A multi echelon inventory optimization model sizes safety stock and positioning decisions directly against the error rate reported at each node, so a MAPE distorted by zero-demand blowups or a bias hidden by cancellation feeds a buffer calculation built on a number that was never real. Getting the four measurement problems above right gives a multi echelon inventory optimization model something honest to size against. Oritiq’s multi-echelon inventory optimization explainer covers how that positioning decision uses the error signal once it’s measured correctly. For the other three sources of forecast error beyond measurement itself, see this guide to reducing forecast error.
Frequently asked questions
Why not use MAPE everywhere for simplicity?
Because MAPE’s instability near zero shows up constantly at SKU-location level. It affects a meaningful share of the portfolio on any given day. Simplicity that produces an unreliable number isn’t actually simpler once someone has to explain why the metric spiked.
How much history does a new SKU-location need before it counts in reporting?
Enough periods to establish a baseline pattern, typically 8 to 13 weeks for weekly-reviewed items, though this varies by demand frequency. Before that point, track it separately. Folding it into a mature-item average this early just hides the gap.
Does this level of measurement matter for slow-moving items too?
If anything, it matters more for slow movers. Slow-moving SKU-locations are exactly where zero-denominator and cancellation problems concentrate, and where a distorted error signal does the most damage to a safety stock calculation sized against it.
Closing
Demand forecasting in supply chain management is only as trustworthy as the error measurement underneath it, and at SKU-location level, that measurement breaks in specific, repeatable ways. Fix the metric, the averaging, the cold-start handling, and the window, and an error number turns into something a buffer calculation can actually use.
Talk to our team to see how your current error measurement holds up at SKU-location level.
Contact us to review your forecast error measurement together.