Which forecasting approach fits which demand?
There is no single best way to forecast demand. Rolling and weighted averages, statistical models, machine learning, and foundational models each win on different items. Here's what each one does, when it fits — and how to choose without becoming a forecasting expert.
No single approach wins for every item
A fast-moving, steady product and a slow item that sells a few times a year are not the same forecasting problem — and the method that nails one will mislead you on the other. The real skill isn't picking a favourite technique; it's matching the right approach to each item's demand pattern, and proving it on real data.
Five ways to forecast demand
Rolling Average
The baselineAverages an item's recent months into next month's number.
Weighted Average
Recency-tunedA weighted blend of recent months, where the latest weeks count for more — often tuned to your business.
Statistical Models
Learns each itemFits an item's own history to capture its level, trend, and seasonality — and returns a calibrated confidence band.
Machine Learning
Learns your whole catalogOne model trained across every SKU, using dozens of signals — seasonality, stock, price, region — to forecast each item.
Foundational Models
The frontierVery large models pre-trained on vast volumes of demand data, so they can forecast an item with little or no history — bringing patterns learned across the world to your catalog.
Which approach for which demand pattern
Classify each item by its demand shape first — then the right approach follows. Tap a pattern.
Smooth → Steady, high-volume — machine learning captures the trend and seasonality.
How to choose an approach — or a tool
Six questions cut through the noise. The honest answer to most of them is that you need more than one method — and a system that picks for you.
- 01What demand patterns fill your catalog?
Most catalogs are a mix of smooth, intermittent, lumpy, and erratic items — so no single method covers all of them.
- 02Do you need honest confidence, or just a number?
A forecast with no confidence band is a guess in a suit. You should be able to see which numbers to trust.
- 03Does it measure accuracy — or just claim it?
The only real proof is a walk-forward backtest (MASE) that beats a naive baseline on your own data before you trust it.
- 04Does it forecast the long tail?
Half your catalog is slow-moving. If a tool only handles the fast movers, the tail runs on gut feel.
- 05Does it correct for stockouts?
A month you were out of stock reads as low demand. Without correction, the forecast quietly under-orders the items that were selling.
- 06How fast can it go live on your data?
The best method is worthless if it takes a year to deploy. Look for weeks, on the data you already have.
You shouldn't have to choose
Picking the right approach for every one of tens of thousands of SKUs by hand is impossible. Likwid's ARIA engine runs multiple approaches on every item, classifies its demand pattern, and picks the one that measurably wins on a backtest — adding newer methods like foundational models only once they beat the incumbents on real data. You get the best method per SKU without becoming a forecasting expert.
