Fundamentals
How much history do you need to forecast?
Enough to see the pattern you are betting on. A weekly seasonal business wants more than one year; an hourly series can learn a week of shape faster. Quality of the target beats another dusty decade of the wrong grain.
Updated Aug 8, 2026·7 min read
A usable rule of thumb
You need enough history to see the pattern you are betting on. If the bet is “weekends look different,” a few months of daily data can already show it. If the bet is “December is another planet,” you want more than one trip around the sun — or a cousin series to steal the shape from.
How much history, roughly
Hourly, weekly season
4–8 weeks
Daily, weekly season
3–6 months
Weekly, yearly season
2+ years
New SKU, no cousins
cold start
Seasonality sets the floor
Weekly season on hourly data shows up fast. Yearly season on weekly data does not. People bring three months of a brand-new product and ask for a Christmas forecast. You can give them a prior. You cannot give them last Christmas.
Foundation models change the floor a little. Zero-shot can draw a plausible path from a short window because it has seen other windows. It still has not seen your promo calendar. Test it. Do not skip the analog SKU.
When you only have months
Shrink toward related series. Use leading indicators (traffic, waitlists, appointments). Keep the horizon short. Report a wide interval so nobody confuses a prior with a measurement. That is a cold start, and it is a first-class job — not a failed forecasting project.
When more history hurts
Five years is not automatically kinder than two. A pandemic year, a rebrand, a warehouse move, a new POS — those years teach a world you do not live in. Prefer recent, comparable history. If you keep the old years, down-weight them or mark the break, or the model will average two companies.
FAQ
- How much data do I need to start forecasting?
- For a weekly series with yearly seasonality, two years is a comfortable floor and one year is a start. For daily series with weekly seasonality, a few months can already show the shape. Less than that, lean on related series or a foundation model.
- Is five years always better than two?
- Not if the business changed. A pandemic year, a rebrand, a warehouse move — those years can teach the model a world you no longer live in. Prefer recent, comparable history over a long, mixed bag.
- What if we just launched the product?
- That is a cold start. Borrow shape from similar SKUs, use leading indicators (traffic, waitlists), and keep the horizon short until you have a season of your own.
Keep going
Guide
Forecasting new products and locations
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.
Guide
Choosing a forecast horizon
The horizon is how far ahead the forecast looks; the grain is how finely it ticks. Pick both from the decision they serve — lead time, planning cycle, and how fast the world can change — not from what the model can print.
Guide
Zero-shot forecasting
Zero-shot means you get a forecast without training on your series first. A foundation model uses what it learned from many other series, then reads yours. Fine-tuning is the later step, once you have enough of your own data to make it worth the spend.
Use case
Retail demand forecasting
Anticipate demand by product, store, channel, and region before buying or allocation decisions are locked.
Use case
Promotion performance forecasting
Estimate promotion lift, product substitution, margin impact, and the demand that remains after a campaign ends.