Skip to content
Predict.ai

Data & drivers

Forecasting new products and locations (the cold start)

A cold start is a forecast with almost no history of its own. You borrow shape from similar items or places, use leading signals, keep the horizon short, and replace the borrowed prior as soon as real sales show up.

Updated Aug 7, 2026·8 min read

The hole in the series

A cold start is a forecast with almost no history of its own. New flavor. New store. New region. The usual teacher — the series itself — has not shown up yet. You need a prior, a short horizon, and a plan to hand off.

Borrowed shape, then a hand-off

Cold start versus analog seriesanalognew SKU
Gray is a similar SKU's season. Lime is the new item as it starts to sell. Until the lime line has a spine of its own, the analog (or a zero-shot model) is the prior — not a verdict.

Analogs and hierarchies

Cluster on what ops already uses: price band, category, channel, trade area. Start from that seasonal shape. Scale by distribution and price. In a hierarchy, shrink the child toward the parent until the child has a spine. Top-down only is unfair to the SKU. Bottom-up only is noise. Reconciliation exists for a reason.

New stores undershoot then overshoot more often than the slide deck says. Copying last year’s hero location produces a confident ramp into a different parking lot. Keep the interval wide.

Leading signals

Traffic, waitlists, appointments, search, preorders — if they exist, they are how you nowcast a series that has not learned to speak yet. Treat them with the same known-ahead rule as any covariate. A waitlist you will only export on Friday cannot help Monday’s inference.

Handing off to its own history

The cold start is over when a model trained only on this series beats the analog on a fair backtest. That can be weeks for a fast web SKU and a couple of seasons for a slow spare part. Until then, do not pretend the borrowed path is a measurement. And do not keep the analog in the driver’s seat out of habit.

FAQ

How do you forecast a product with no sales history?
Cluster it with similar SKUs (price band, category, channel), start from their seasonal shape, scale by distribution and price, and tighten the horizon. Update weekly until you have a season of its own.
What about a new store?
Use nearby stores, trade-area data, and the chain's typical ramp. New stores often undershoot then overshoot. A wide interval is better than a confident ramp copied from last year's hero location.
When is the cold start over?
When a model trained only on this series beats the analog on a backtest. That can be weeks for a fast web SKU and a couple of seasons for a slow spare part.

Keep going

Try it on your data.

Connect an outcome and its history. Predict.ai finds the drivers, races the models, and keeps the forecast live — in the workspace, over the API, and through agents.