Forecast accuracy is the wrong target

Almost every planning organization tracks a forecast accuracy number, and almost none of them can say what a point of improvement is worth. That gap is where a great deal of analytical effort quietly disappears.

Accuracy is a proxy. What actually matters is whether the forecast leads to better decisions about buying, building, positioning, and committing capacity. Those decisions have asymmetric costs, and accuracy metrics are symmetric by construction.

Mean absolute percentage error treats an overforecast and an underforecast of the same magnitude identically. Real businesses almost never do. For a promotional item with a hard shelf life, overforecasting produces markdown; underforecasting produces a lost sale and possibly a lost customer. Those are not the same size of mistake, and no symmetric metric can tell them apart.

The consequence is misallocated effort. Teams optimize a metric that weights all items and both directions equally, when the actual cost concentrates in a small subset of items and one direction of error.

Start from the decision. For each item segment, identify what the forecast is used to decide and what an error in each direction costs. Then evaluate forecast candidates on expected cost rather than on average error.

This usually reorders your priorities immediately. Items with poor accuracy but low decision impact stop consuming planner time. Items with acceptable accuracy but severe asymmetric cost get attention they were not getting.

If the decision has asymmetric costs, a point forecast throws away exactly the information needed to make it well. A quantile forecast lets the inventory calculation consume the uncertainty directly rather than reconstructing it from a separate variability assumption that may not agree.

This is not a purely theoretical improvement. Where the point forecast and the safety stock calculation use different variability assumptions, they are two models disagreeing about the same future, and the disagreement is usually invisible.

The last piece is measuring the process rather than the model. Instrument every step where a human touches the number and compare the result against the statistical baseline it replaced. In most organizations this reveals that a meaningful share of manual overrides make the forecast worse.

The point is not to remove planners. It is to concentrate their judgment on the items where it demonstrably helps, and to stop spending it where the model was already better.