Why this exists
A supplier takes a week to deliver. You only review stock once a week. Tomorrow’s sales forecast alone cannot tell you what to buy: your decision must cover both periods, account for stock already available, and allow for forecast errors.
DemandCast makes that arithmetic visible. The original project estimated safety stock from variation between future point forecasts. A flat prediction produced zero uncertainty, even when actual sales were volatile. This workbench measures historical errors instead.
How it works
-
Read complete daily sales.
Import one item’s history. Missing dates and negative quantities block the calculation; they can hide closures, returns, or stockouts.
-
Use a transparent baseline.
A trailing 28-day mean supplies the forecast. A four-week weekday alternative had identical weekly totals in the benchmark, so the extra selection step was removed.
-
Measure the errors that matter.
Replay fully observed protection periods. Take the requested quantile of actual period sales minus the forecast. Add only a nonnegative buffer.
-
Account for your stock.
Target = expected period sales + error buffer. Position = on hand + incoming − backorders. Order = max(0, target − position), rounded up to the supplier pack.
Where this can fail
Sales are not demand when you were out of stock. Promotions, new products, supplier delays, bulk purchases, and sudden changes can break historical patterns. The error windows overlap; the quantile has no guaranteed future coverage.
This is one-item decision support for a fixed lead time. It does not optimize a purchasing budget, simulate purchase orders, infer lost sales, or account for expiry dates. All incoming stock is assumed to arrive within the protection period; check exact arrival dates separately.
What would prove the idea wrong?
If the buffer cannot improve held-out cycle coverage without doubling target stock, the current mechanism is too expensive. The evidence lab reports that threshold directly. A better next step would be stock-availability labels and delivery-time history, tested on multiple retailers.