What data is needed to test a demand forecast?
Start with dated sales or usage at the level where a decision will be made, then add stock availability, returns, cancellations, prices, promotions, locations and product changes where they affect the history. Supplier lead times, order packs, minimum quantities and current replenishment rules are also needed if the forecast will inform purchasing. The useful history length depends on the forecast horizon, seasonality and data quality; this page does not assume a universal minimum.
How should a candidate forecast be evaluated?
Keep later periods out of model fitting and compare the candidate with an agreed baseline on those holdout periods. Score the forecast at the actual decision horizon and by relevant item, location and demand segment. Review absolute error, percentage-based measures where valid, bias and operational exceptions. One average score can hide systematic over-ordering, stockout periods or poor results for important products.
Does a better error score prove better inventory results?
No. A statistical score does not include every purchasing constraint, supplier delay, minimum order, expiry risk, substitution or service target. Test how forecast outputs pass through the proposed replenishment rules, and compare the resulting recommendations with the current process before changing orders. Savings, availability and waste must be measured separately using agreed business data.
How are stockouts and zero-sales periods handled?
Zero recorded sales may mean zero demand, unavailable stock, a closed location or missing data. Mark known stockouts and availability gaps before training or evaluation so sold units are not automatically treated as unconstrained demand. Agree how intermittent items, long runs of zero demand and new items will be segmented, scored and handled when evidence is weak.
Can weather, events or promotions be included?
They can be considered when the data is licensed, consistently available and known at the time each forecast would have been produced. Historical evaluation must prevent future information leaking into earlier predictions. Compare the added signal against a baseline and keep it only when it improves the relevant holdout result without creating unacceptable cost, delay or dependency risk.
How do new products or locations get forecast?
A new item or location has limited direct history, so any estimate depends on explicit assumptions or a reviewed comparison group. Document the chosen analogue, launch plan, range, price, channel and uncertainty. Use conservative ordering rules, staff approval and frequent review rather than presenting an untested estimate as measured demand.
Will the system place purchase orders automatically?
That is a separate scope decision. Start with read-only data and recommendations that staff can review. Define supplier constraints, approval thresholds, duplicate-order protection, rollback, audit records and behaviour when data or a connected system is late or unavailable. Enable write actions only after representative testing and named business owners approve them.
How do we start and what does a consultation cost?
Bring one replenishment decision, representative history, the current baseline method and the constraints staff apply today. The 30-minute consultation is free for businesses with 20+ full-time staff; otherwise it is AUD200 including GST. Data preparation, evaluation, implementation, provider usage and ongoing review are scoped and priced after the systems, data and acceptance criteria are understood.