Problem
A demand forecast says how many units of each item will sell. Planning tools often need more: orders, each with a region, a number of lines and an item mix. Store labor planning needs orders to simulate which store ships what. Fulfillment-center planners need to know which items will travel together in multi-item orders, because consolidating them lowers ship expense.
Target's data science team built Demand Profiler (DP), which generates synthetic digital orders for a future date.
Example from the article: DP forecasts the share of single-line and multi-line orders, and planners reserve stock for items likely to appear in multi-item orders.
Solution in brief
DP learns the distribution of each order feature (region, items per order, item set, quantity) from historical orders and samples new orders from those distributions. A restriction step caps the sampled units per item at the forecasted demand, so the orders add up to the existing forecast instead of becoming a second one.
Questions this case answers
- How to expand an aggregate per-item demand forecast into a realistic stream of future transactions without producing a second, conflicting forecast
- How to turn an item-level demand forecast into realistic future orders that downstream planning tools can consume?
- How to plan store labor for ship-from-store demand before the orders arrive?
- How to find items likely to ship together so their inventory can be reserved at one fulfillment center?
Tools and methods
- Conditional probability models — estimate the distribution of each order feature from historical transactions.
- Restricted random sampling — generate orders, then align item units to the forecast.
- Order allocation simulator — consumes DP orders to size store labor.
- Apache Spark (PySpark) — implementation stack.
Google or a one-time email code. No password.