How many months of history do you need for a reliable demand forecast?

For most retail categories, 18 to 24 months of history per SKU and store. Under 12 months misses seasonality. From 12 to 18 captures a single cycle, without confirmation that it repeats.

The condition that weighs more than the amount is cleanliness. Two years of sales that do not separate sales from shelf-out are worth less than twelve clean months. Granularity is a condition too: SKU by store by day, not the chain’s monthly average, which erases the pattern you want to predict.

The most common exception is a product with no past. A new item, fashion, a launch, has no history of its own, and no amount of waiting fixes that. There the forecast does not come from time; it comes from similar products, by attribute. Anyone waiting to “accumulate a base” for a launch waits forever. And in perishables the ruler changes: history depth matters less, daily sales quality by day of week matters more.

What you can do this week costs nothing. Take one category and count how many months you actually have, with shelf-out flagged and a clean registry, not how many months sit in the database. That number is usually half what the team thinks. It is the same gap as aggregate versus weighted MAPE: the base looks bigger than it is. Before choosing a vendor or model, that is the problem to solve.

If you are sizing a forecasting project and want a second read, send me the category and the history you think you have in one sentence. I will reply within a business day with a paragraph: yes, no, or “yes, if you clean this first”. If it becomes a project, the honest timeline is the next conversation.