Inventory forecasting is treated as a prediction problem, but the more damaging property is that its errors are self-reinforcing. Each mistake corrupts the data used to make the next forecast.
Stockouts erase the demand they cannot serve
Sales data records what was sold, not what was wanted. When an item is unavailable, the demand simply does not appear in the record.
The next forecast, built from sales history, therefore reads the shortage as low demand and orders less, producing another shortage.
Escaping the loop requires tracking lost sales — out-of-stock page views, backorder requests, substitutions — which are separate signals most systems do not capture by default.
Overstock inflates the record it later distorts
Excess inventory is usually cleared through discounting, and discounted units sell at volumes the full price would never have produced.
Those units enter the sales history without a marker for the price that generated them, so the following forecast reads temporary discount demand as baseline demand.
The result is another over-order, another markdown, and a product whose apparent demand curve has been shaped mainly by past pricing mistakes.
Lead time variability multiplies the effect
Forecast error and supply timing interact: a longer lead time means the forecast has to reach further into the future, where it is less accurate.
Variability in that lead time is worse than length alone, because safety stock has to cover the late deliveries rather than the average one.
Suppliers who are slow but consistent are therefore easier to plan around than suppliers who are usually fast and occasionally very late.
Aggregation hides the errors that matter
Category-level forecasts often look accurate because individual overestimates and underestimates offset each other in the total.
Purchasing happens at the item level, so the offsetting is arithmetic rather than operational. The cash is still tied up in the wrong items.
Measuring accuracy at the level at which orders are actually placed usually reveals a much worse picture than the summary report suggests.
Simple forecasts with clean inputs beat sophisticated ones without
Because the dominant errors come from contaminated history rather than from weak mathematics, better modelling on unadjusted data delivers limited improvement.
Cleaning the record — flagging stockout periods, marking promotional sales, separating one-off bulk orders — improves any method applied afterwards.
It is unglamorous work that shows up as fewer emergency reorders and less end-of-season clearance, both of which are easier to notice in cash than in a forecast accuracy metric.
The cleaning also has to be continuous rather than a one-time correction, because every subsequent stockout and every promotion introduces the same contamination into the record again.