Most advice on new product demand forecasting starts from the same premise: there’s no history, so you’re guessing. We think that premise is wrong, and that it explains a lot of bad launch forecasts.
A new product has no sales history. It almost always has evidence: a predecessor that sold, attributes it shares with hundreds of SKUs already in your catalog, a retailer’s first order, a test market. The right method depends on which of those you’re holding. Most teams work the other way round. They reach for the method they know, usually a spreadsheet and the sales director’s number, and go looking for evidence afterwards.
That’s an expensive habit. New products represent 27% of sales across all industries, according to Kellogg Insight, and every one is a buy you commit to before a single unit has sold.
This guide covers five methods, a decision table that matches the method to your evidence, and a worked analog forecasting example that ends in a range instead of one number.
Key highlights:
- New product demand forecasting is two jobs: a cold start forecast before launch and a fast-updating forecast after it.
- The method follows the evidence: a close analog, product attributes, early orders, or only the category.
- Publish a low, base and high forecast, and plan for the product’s own sales to take over by week 8 to 12.
What is new product demand forecasting?
New product demand forecasting is the process of estimating how many units of a product customers will buy before it has a sales history of its own. It replaces the missing history with other evidence, such as similar products, product attributes and early orders, and updates once real sales arrive.
For established SKUs, demand forecasting means reading historical patterns and adjusting for drivers like promotions, price changes and seasonality. A new item gives you none of that. The forecast has to borrow its signal from somewhere else, then hand over to the item’s own sales as soon as those can be trusted.
Most planning stacks aren’t built for that handover. According to McKinsey, close to three-quarters of supply chain functions still rely on spreadsheets for planning. That’s how a one-time estimate from sales or marketing ends up frozen in a cell and adjusted by hand for the life of the launch.
Why new product forecasts fail: the two jobs
Through our work with supply chain leaders, we’ve found that most teams fail because they treat demand forecasting for new products as a single task. The process consists of two distinct jobs.
- The first is how to forecast demand for a new product before its launch, when you have no direct sales history. This is the cold start problem.
- The second is how to forecast demand for a new product after launch, when you finally have some data, but not enough to rely on traditional forecasting methods alone.
These two situations require different inputs, model logic and planner behavior. Before launch, you’re asking what the evidence suggests, and a fair answer is a range. After launch the risk flips, because early numbers are loud and often misleading. The first shipments into retail are pipeline fill, and one promotion can make week two look like a trend.
Teams that treat it as one job either freeze the prelaunch number and defend it for months, or throw it out after two weeks of sales and start chasing every bump.
5 methods to forecast demand for a new product
1. Analog (lookalike) forecasting
You borrow the sales history of a similar product and scale it to the new one. This works whenever a close predecessor or sibling exists, which covers most line extensions, pack-size changes and seasonal refreshes. It goes wrong when the analog was picked because it flatters the launch plan. The next section covers it in detail.
2. Attribute-based machine learning
Here a model does the analog-picking. It learns from every past launch in the catalog how attributes such as category, price band, pack size, channel and launch season relate to early demand. AWS describes the approach in its guidance on forecasting new product introductions, where forecasts are “generated many weeks prior to product launch.” The academic DemandForest method also predicts quantiles of launch demand, so it returns a range. It needs a few hundred past launches with clean attributes, and it struggles with attribute values it has never seen.
3. Market research and test launches
Concept tests, purchase-intent surveys, pre-orders and regional test markets collect evidence from customers directly. Nothing else on this list measures the product itself. It’s also slow and expensive, stated intent runs ahead of real purchases, and a test in three stores tells you about three stores.
4. Expert judgment and Delphi
This is structured opinion from sales, marketing and category managers. The Delphi method collects it in anonymous rounds so the loudest person in the room doesn’t set the number. Judgment is the only source for context no dataset holds, like a competitor’s exit or a retailer’s shelf reset. Its weakness is optimism, since the people closest to a launch are the most invested in it.
5. Bass diffusion for new categories
When nothing comparable exists, the Bass model estimates adoption over time from three inputs: the size of the market, the rate at which buyers adopt on their own, and the rate at which they follow other buyers. Researchers writing in Marketing Science note that it “has been a standard for analyzing and predicting the market penetration of new products.” It suits multi-year capacity decisions, not SKU-week planning, and its inputs are guesses until sales arrive.
Analog forecasting: how to forecast from a lookalike product
Analog forecasting does most of the work in prelaunch planning, and the step teams rush is choosing the analog.
How to pick an analog. Match on what drives volume: category, price band, channel and launch season. A premium SKU launched in Q4 through e-commerce is a poor analog for a value SKU going into grocery in March, even when they share a brand. Use two or three analogs if you have them and average the result.
Pick them before you look at their sales, and keep the ones that flopped. Forecasting researcher Paul Goodwin warns about survivor bias when forecasting by analogy: only the products that survived have long sales histories, so the failures drop out of the comparison and the forecast inherits the optimism.
How to scale it. Adjust the analog’s sales for the differences you can measure: distribution (stores, doors or accounts at launch), price position against the analog, and marketing weight.
A worked example. Your analog sold 1,000 units in its first 12 weeks across 80 stores. The new item launches in 60 stores, at a price-value position you rate 10% stronger.
Base forecast = 1,000 × (60 ÷ 80) × 1.10 = 825 units
Don’t stop at 825. University of Tennessee research published in Supply Chain Management Review found that planners underestimate how much demand varies across the SKUs of a new product family, and it recommends “multiple estimates (low, medium, high)” over a single point. A band of plus or minus 30% gives you three numbers, each tied to a different decision:
| Scenario | 12-week forecast | The inventory decision it drives |
| Low | About 580 units | Your firm first order. You’d be comfortable owning this much even if the launch disappoints. |
| Base | 825 units | Total planned supply, including a second order you release only if early sales confirm it. |
| High | About 1,070 units | The upside you keep reachable: reserved supplier capacity, components on hand and agreed expedite terms. No finished goods. |
The 30% is a placeholder. For a better band, take your last 20 launches, compare each prelaunch forecast with what sold, and use the spread you find.
Plan the handover. The analog shouldn’t keep its vote forever. Once sales start, shift the weight toward the item’s own data on a schedule you set in advance:
| Period | Weight on analog | Weight on the item’s own sales |
| Before launch | 100% | 0% |
| Weeks 1 to 2 | 85% | 15% |
| Weeks 3 to 4 | 65% | 35% |
| Weeks 5 to 8 | 40% | 60% |
| Weeks 9 to 12 | 20% | 80% |
| After week 12 | 0% to 10% | 90% to 100% |
Treat this as a starting schedule and tune it by category. Fast sellers earn trust sooner because every week brings more data. Where you can, use sell-through (POS) for the item’s own sales. Sell-in includes the pipeline fill.
Which method should you use? A decision table
Start with what you have, then read across.
| What you have | Method | Best horizon | Main risk |
| A close predecessor or lookalike SKU | Analog forecasting with scaling | Prelaunch through the first 8 to 12 weeks | Cherry-picked analogs; cannibalization of the predecessor ignored |
| Attributes only (category, price band, pack size, channel) | Attribute-based ML trained across the catalog | Prelaunch, months ahead of the first buy | Thin or inconsistent attribute data; attribute values the model hasn’t seen |
| Pre-orders, retailer commitments or test-market sales | Early signal blending with an analog or model baseline | The weeks just before launch and the first weeks after | Pipeline fill mistaken for consumer demand; an unrepresentative test |
| A truly new category with no analog | Bass diffusion curve plus expert judgment | Multi-year adoption and capacity planning | Market size and adoption rates are assumptions; small changes move the curve a lot |
The rows aren’t exclusive. Most launches hold two kinds of evidence, and the best forecast blends them. A line extension with retailer commitments can take its sales curve from the analog and its volume from the commitments. If the only row that fits is the last one, widen your range and tell the business why.
How to forecast demand for a new product in 4 steps with AI
A planner can run these methods by hand for one launch. At a few hundred launches a year, the same logic needs a demand forecasting solution that applies it across the catalog. This is where machine learning earns its keep: a model trained on every past launch can find analogs a planner wouldn’t think of. One deployment in our 2026 forecast accuracy report covers more than 100,000 SKUs with frequent product introductions, well past the point where anyone picks analogs by hand.
1. Establish a foundation with first-party data
External factors such as weather can help, but in our experience about 80% of the predictive signal comes from first-party data: shipment invoices, orders and receipts, point-of-sale (POS) information, and a product master with consistent attributes. For new product demand forecasting, this foundation lets the AI identify which existing items most closely resemble the launch.
2. Generate a predictive baseline forecast
Before launch, the baseline should combine analogous or proxy products, launch and order plans, customer commitments, distribution assumptions, known promotional timing and scenarios for upside and downside risk.
After launch, the baseline should change. The product now has early signals, but short-term distortions and one-off events make them hard to read. The model should update more reactively, using early sales, shipments, consumption patterns and context from similar products, the same short-horizon inputs used in demand sensing.
3. Segment your model into no-touch and touch workflows
Not every new product forecast needs manual intervention. Some are strong enough to adopt as they are. Others need a planner review because the model has low confidence, unusual inputs or incomplete information.
- A no-touch forecast is one that your team can accept without additional manual changes.
- A touch forecast is where a planner should step in because local knowledge or recent business context can improve the outcome.
If planners review every item equally, they waste time on forecasts that were already good enough and may introduce bias where no change was needed.
4. Measure success with forecast value added (FVA)
Forecast value added (FVA) measures whether the final forecast performed better or worse than the baseline. Here the baseline is the model’s prediction before planners adjust it, and the final forecast is the version that moves forward after manual changes or consensus inputs.
FVA is especially useful for launches, because new items attract attention and planners feel pressure to adjust them often. It answers three practical questions:
- Is the baseline strong enough to trust more often?
- Are planners improving the forecast where they intervene?
- Which parts of the portfolio benefit most from manual input?
Where forecasting fits in the new product introduction (NPI) process
The new product introduction process moves a product through stage gates. In NPI supply chain planning, each gate needs a different forecast, and trouble starts when a concept-stage number gets used for a launch-stage decision.
| Gate | The decision | The forecast it needs |
| Concept | Is the opportunity worth pursuing? | A range only, from category analogs or a diffusion curve |
| Development | Capacity, tooling and long-lead materials | An analog or attribute-based forecast with low, base and high |
| Launch | The first committed buy and initial allocation | A committed number inside the range, blended with pre-orders |
| Post-launch | Replenishment and the second buy | A weekly reforecast that shifts weight to the item’s own sales |
Between gates, the launch forecast belongs in your monthly cycle like any other number. NPI plans feed the demand review step of the S&OP process, where the range and its assumptions should be visible to the people committing supply against it.
Use cases by industry
| Industry | Before launch, the forecast draws on | After launch, it updates on |
| Consumer packaged goods (CPG) | Similar SKUs, retail commitments, promo timing and distribution plans | Sell-through data, store-level adoption and regional demand patterns |
| Manufacturing | Comparable product families, customer programs and order expectations | Order cadence, lead times and short-term demand shifts |
| Food and beverage | Lookalike items, weather sensitivity, customer rollout plans and promotional calendars | Separating pipeline fill from repeat consumer demand |
| Oil and gas | Prior product families, contract expectations and account plans | Project timing, regional changes and customer buying patterns |
| E-commerce | Campaign plans, traffic expectations, price point and similar product performance | Clicks, conversion rates and channel-level demand shifts |
| Direct-to-consumer (DTC) | Customer history, subscriber behavior, preorder signals and product affinity across past launches | Onsite conversion, repeat purchase patterns, returns and retention |
For a closer look at how planners and models share this work, read our guide to human and AI collaboration in supply chain.
When the launch does not take off: from new product to long tail
Not every launch finds its volume. Some settle into a few units one week and nothing for the next three. From then on you’re forecasting sparse demand, and a standard model fed a history of mostly zeros will lag every sale and overreact after it.
Set the trigger in advance. If sales at the week 12 review sit below your low scenario and more than half the weeks are zeros, move the SKU to a method built for that pattern. Our guide to intermittent demand forecasting covers Croston’s method and its variants, and how to tell when a SKU should stop being forecast at all.
FAQ
How do you forecast demand for a new product with no sales history?
What is analog forecasting?
Which model is best for new product demand forecasting?
What is NPI in supply chain?
How accurate are new product forecasts?
Forecast your next launch with more than a guess
We built DemandForecast.ai for the part of this that doesn’t scale by hand: finding analogs across a whole catalog, producing a baseline before launch, and telling planners which launch forecasts need them. If you have launches coming and a spreadsheet standing in for a forecast, we’d like to show you what your own data says.
