Demand Planning Accuracy

A distributor plans to sell 1,000 units next month and actually sells 800, and calling the forecast 80% accurate is tempting, but it depends on the chosen error formula, level and timing. Demand planning accuracy measures how close a frozen demand forecast came to a defined actual outcome.

Oracle documents several forecast-accuracy measures and their setup in demand planning, and SAP explains forecast error and common metrics in inventory planning, though neither source makes one score suitable for every SKU. Freeze a forecast by recording the version available when the purchasing or production decision was made, because a forecast edited after sales occur cannot fairly assess the plan.

Set the horizon, as one-week, one-month and six-month-ahead forecasts serve different decisions, and do not compare scores without noting lead time. Choose the grain too, since a product-family forecast can look accurate while individual sizes or locations are wrong, so measure at the level where stock decisions happen.

Choose the actual carefully, since orders, shipped units and unconstrained demand are different, and stockouts can suppress shipped sales and make a low forecast appear right. Check lost sales, backorders, returns and calendars, as net sales after returns can differ from units demanded at order date, and moving holidays or weather can shift the week of demand.

Label any estimates and align the actual series with the planning question. Pick an error formula, because absolute error, percentage error and weighted absolute error emphasise different failures, and state both numerator and denominator.

Handle zero actuals with a method suited to intermittent items, since percentage error based on actual demand can be undefined, and note that MAPE can behave poorly near zero while aggregate percentage accuracy can hide opposite errors across items. Separate bias, because positive and negative errors can cancel in a signed average, and show the distribution of errors.

Segment item types, since stable essentials, new launches and intermittent spare parts need different models and one target may punish the wrong group, and review promotions and overrides by preserving the assumptions about timing, discount, channel and reason. For new items with no useful history, use analogous products and a range, then update when actual orders arrive.

Compare a rolling series rather than a single month, since one good month can be luck. Measure both units and value, as unit-weighted accuracy highlights volume while value-weighted analysis highlights money, and test better scores against stockouts, spoilage and excess stock, since a mathematically lower error may not improve service.

Prioritise by lead time, because a small error on a long-lead product may cost more than a larger error on a quickly replenished item, and use scenarios or ranges for uncertain items. Assign owners, with sales contributing customer intelligence, finance challenging economic assumptions and operations testing supply constraints, and report coverage honestly, because a score can improve when difficult low-volume items are excluded.