Sales & Inventory Forecasting with Machine Learning

Aitium forecasts two things: how much you will sell, and how much stock you need. Both are built from your own trading history using managed time-series machine learning on AWS.

Scope

Sales demand by SKU over a horizon you choose, and that demand set against current FBA and merchant-fulfilled stock. See sales forecasting and inventory forecasting.

That is the full scope. Aitium does not forecast staffing, website traffic or company cash flow.

Model selection

No algorithm wins on every product. AutoML can evaluate candidates against your data and pick the best performer, or you can choose yourself.

• CNN-QR — large catalogues; learns across related SKUs.
• DeepAR+ — many related series with yearly seasonality.
• Prophet — strong repeating seasonality and holiday effects.
• ARIMA / ETS — simpler catalogues and shorter histories.
• NPTS — sparse or intermittent demand, common in B2B.

History required

Your horizon cannot exceed one third of your training history, so a 90-day forecast needs about 270 days of data. Aitium backfills up to two years, which satisfies every model above. Where a SKU is too new to forecast reliably, Aitium says so rather than returning a confident-looking number.