Why this is a different problem
Every forecasting method described elsewhere assumes a history to learn from. A new product has none, so the question changes from 'what does this product's pattern suggest' to 'which existing products is this one like, and how is it different'.
This is worth saying plainly because teams often apply a forecasting tool to a new line, get a number, and treat it with the same confidence as an established item. It deserves much less.
Building an analogue forecast
The workable method is to pick comparable products and adjust. Done carefully this is defensible; done casually it becomes a number somebody invented with a spreadsheet around it.
- Choose analogues on attributes that actually drive demand - price band, category, pack size, channel - not on what feels similar.
- Use several analogues rather than one, and look at the spread between them. That spread is your honest uncertainty.
- Take the analogue's launch curve, not its steady state. New products sell differently in their first weeks.
- Adjust explicitly for known differences - a higher price point, fewer stores, no launch campaign - and write down each adjustment and why.
- Record the assumptions somewhere they can be checked later, so the next launch is better informed than this one.
Where machine learning genuinely helps
Selecting analogues by hand does not scale past a few launches a year. Where a business launches regularly, a model trained on past launches can predict the first-season volume from product attributes - and it improves with every launch you record.
This works because you are no longer forecasting a time series. You are predicting an outcome from attributes, which is an ordinary supervised learning problem with one row per historical launch. The constraint is simply how many launches you have.
If you have launched a dozen products, that is not enough rows and structured analogue selection is better. Several hundred launches and a model will usually beat manual selection, particularly at choosing which attributes matter.
Plan the correction, not just the forecast
The most useful thing about a new product forecast is how quickly it gets replaced. Early sales carry a lot of information, and a business that can re-forecast after two weeks of real data will beat one that holds its launch plan for a season.
| Stage | Basis for the number | What to decide |
|---|---|---|
| Pre-launch | Analogues and adjustments | Initial buy, allocation |
| Weeks 1-2 | Early sell-through, heavily uncertain | Whether to accelerate or hold |
| Weeks 3-6 | Own data, blended with analogues | Repeat buy, range extension |
| Week 7 onwards | Own history | Normal replenishment |
Designing that sequence in advance - who reviews, on what day, with what authority to change the order - is usually worth more than any improvement in the pre-launch estimate.
Being honest about the range
New product forecasts should be presented as ranges, and wide ones. A single number invites a precision that does not exist and makes the inevitable miss look like a failure rather than an expected outcome.
We would rather tell a client that a launch will most likely land between 3,000 and 7,000 units in its first quarter, with the reasoning attached, than produce 4,850 and watch it get treated as fact.
For a launch, the plan to correct the forecast is worth more than the forecast.