Forecasting with sparse sales data
In many catalogs, a large share of products sell only a few units per week, with many weeks of zero sales. Classic time series methods assume smoother demand and tend to produce forecasts that are biased or unstable for these items.
What works better
For intermittent demand we usually separate two questions: how often does the item sell, and how much when it does? Methods built on that idea, such as Croston-style approaches and their modern variants, are simple to explain and hold up well in backtests.
Pool information
Slow-moving items can borrow strength from similar products, stores or regions. Hierarchical models let a forecast for one item lean on its category when its own history is thin.
Measure the right thing
Percentage errors behave badly when actual sales are zero. We prefer scaled error measures and, more importantly, the business outcome: stockouts and excess inventory.