Forecasting a brand-new product is the cold-start problem: there is no sales history to extrapolate from. The usual fallbacks, a planner’s gut, or “copy the closest existing product,” are either arbitrary or brittle. A hierarchical Bayesian approach does the sensible thing more rigorously: it borrows strength from similar products through their attributes, and updates fast once real sales arrive.
The idea: borrow strength through attributes
A new SKU has no history, but it has attributes: category, price band, size, brand, season, channel. Existing products with similar attributes do have history. A hierarchical model uses that structure: products inherit a prior from the groups they belong to (their category, their price band), so the new item starts with the expected demand of its kind rather than a blank slate.
- No history? Lean almost entirely on the attribute-derived prior.
- A few weeks of sales? The model blends prior and observed data, weighting the data more as it accumulates.
- Plenty of history? The prior fades and the item’s own data dominates.
That graceful shift from “products like this” to “this product” is exactly what a launch forecast needs.
Where embeddings come in
Hand-defined groups (category, price band) are coarse. An embedding of the product’s attributes and description places each item in a continuous space where similarity is richer than shared category labels, so a new “stainless travel mug, 16oz, premium” sits near the right neighbours even across category boundaries. The prior is then drawn from its embedding neighbourhood, not just its assigned category. (This is the same semantic-similarity tool used for SKU deduplication, pointed at a different problem.)
# illustrative shape
prior_mean(new_sku) = weighted_avg(demand[n] for n in embedding_neighbours(new_sku))
posterior = bayesian_update(prior, observed_sales_so_far)
forecast = posterior_mean # prior-dominated at launch, data-dominated laterWhy this beats the usual fallbacks
- Better than a single analog. “Copy product X” bets everything on one guess; the hierarchical prior averages over many similar items and quantifies uncertainty.
- Self-correcting. It does not stay wrong; each week of real sales pulls the forecast toward reality automatically.
- Honest about uncertainty. It produces a distribution, so you can size launch safety stock for a new item deliberately instead of guessing.
The takeaway
Cold-start forecasting is not hopeless, it is a borrowing problem. Build a prior from similar products (ideally via attribute embeddings, not just category), update it as real sales land, and you get a defensible launch forecast that improves itself. It is the principled middle ground between a planner’s gut and waiting for history that, by definition, a new product does not have. For the broader toolkit see demand forecasting.
Implementing this at your scale?
The walkthrough above comes from production work. AvanSaber’s inventory practice has implemented variations of this pattern across multiple customer engagements.
If you are building this and want expert review of your design, or would rather have the team that built this build yours, book a discovery conversation or describe your situation at [email protected].