Magento AI Demand Forecast: plan reorders, stock risk and price decisions

Six units sold in a month can describe a quiet product or a product available for only six useful selling days. Ordering from the sales total alone can repeat the stockout that made the number look small.

AI Demand Forecast brings sales, availability, supplier timing and stock into one planning workflow. It estimates demand, highlights stock risks, suggests purchase quantities and prepares price recommendations for review. Numerical forecasts run locally in PHP. Optional AI writes summaries and explanations from aggregated product figures.

Start with the data, then load enough history

The module collects daily sales, returns, stockout observations, pricing and promotion information. It tracks supported configurable options and, with MSI, the website-linked stocks and their salable quantity separately.

Historical order backfill can load an earlier period, two years by default. It cannot reconstruct stock availability or price observations that were never recorded before installation. Distinguish that imported history from the newer observed stockout data.

Set supplier, lead time, minimum order quantity, case pack, cost and price boundaries where applicable. Product lead time takes precedence over supplier timing and the default. Better lead-time inputs can improve a purchasing decision more than a more elaborate explanation of an incorrect input.

See the planning risks before opening a spreadsheet

The dashboard combines recent units and forecast with stockout risks, suggested purchase value, overstock, slow movers and pending price suggestions. Filter by category, supplier and supported stock to inspect a useful slice of the catalogue.

The displayed forecast-accuracy measure comes from held-out recent history, not a promised future success rate. Ask which products and period contribute before using it to compare models or make a buying commitment.

Open a product and inspect the forecast behind the recommendation

A product forecast shows daily sales, predicted demand, its likely range and observed out-of-stock days. Those unavailable days are not treated as known zero demand.

The module compares candidate methods including seasonal Holt–Winters, damped trend, Croston/SBA for intermittent demand and moving averages. It backtests recent held-out weeks to select a suitable candidate and shows the comparison. A low-volume spare part and a seasonal jacket need not use the same method.

Measured effects can include promotion uplift, holidays, optional temperature and price sensitivity. They are applied only when supported by the history. Holiday and weather services are separate configured inputs, not an AI guess about tomorrow's demand.

Recalculate a single product when useful, and review the interval and error rather than treating the forecast line as a certain order volume.

Convert the lead-time problem into a purchase list

Purchase Suggestions lists products at or below their reorder point, with urgency, days of cover and suggested quantities. The calculation covers configured lead time plus the order cycle, adds safety stock and subtracts salable quantity before applying minimum order and pack constraints.

For example, four units per day over a 14-day lead time plus a seven-day ordering cycle imply 84 units of expected cover before safety stock and stock deductions. These inputs are an illustration; the operational buy still needs supplier and commitment checks.

Export a filtered CSV for your purchasing process. The module does not automatically subtract open purchase orders, so reconcile confirmed inbound stock before placing another order. Suggested purchasing is not an automatic supplier order.

Review markdowns and price pressure with a reason attached

Price Suggestions can highlight overstock markdowns, slow movers, availability pressure or a gap to configured competitor offers. Floors, ceilings and the minimum-margin calculation constrain recommendations using your product inputs.

Nothing changes until an authorised user approves. Depending on the recommendation, apply a special price for the configured period or a new regular price; increases use regular price. Changes are logged and can be reverted, with checks to avoid overwriting a later manual edit. Products with their own store-view prices and unsupported product types have documented restrictions.

A projected demand change is still an estimate. Test campaign economics, tax display, indexing and interaction with existing promotions before approving a broad selection.

Monitor competitors without treating every scraped number as a price

Add product URLs directly or by CSV. The module reads configured pages on its schedule, extracts schema.org price information and reports missing or stale data. It respects robots restrictions and skips blocked access rather than working around it.

Verify the comparable item, variant, currency and offer context. A competitor's displayed number can represent a different pack or size. Old values are excluded by the documented freshness rule, but a fresh value is not automatically a like-for-like comparison.

Use AI where plain-language explanation helps

Optional summaries describe the current risks and opportunities. A scheduled weekly summary is available on Mondays, and authorised staff can request a summary or an explanation of a suggestion. These requests use your configured AI Core provider and budget.

Only aggregated product figures go into these forecasting explanations, rather than customer or order records. The numerical forecasting does not depend on the external language model, so disabling AI explanations leaves the local analysis available.

Check coverage and limitations before trusting the queue

Grouped products are not forecast; bundles are handled as bundles rather than independently buying their components. Configurable options have supported tracking, but configurable price suggestions are excluded. Forecast quality is weaker when history is sparse, availability was unrecorded or supplier assumptions changed.

Public holidays and optional weather effects depend on setup and measured history. Competitor information depends on accessible structured data. Purchase quantities need your inbound-order reconciliation. Each is a reason to review the suggestion, not to dismiss the whole analysis.

Judge repeated decisions, not one impressive forecast

Pilot a representative range. Compare backtest error, stockouts, excess stock, recommendation usability and provider cost for optional summaries over repeated cycles. Record promotions and lead-time changes that alter the conditions.

Visit the AI Demand Forecast product page for the dashboard, guide and admin demo. Pair it with stock-alert demand signals and AI Merchandising when replenishment and catalogue exposure need to be considered together.