New Product Demand Forecasting: Methods With No Data

New product demand forecasting estimates how much a brand-new SKU will sell when you have zero sales history to lean on. You replace missing history with three things: analogs (similar products that already launched), structural models like the Bass diffusion curve, and the product's own attributes mapped against your catalog. Then you forecast a range with explicit assumptions instead of a single number, commit a small first production run, and correct fast as early sell-through comes in.

I sat in that meeting many times running planning at a $250M furniture manufacturer. New collections were where we either tied up a quarter's cash in goods nobody bought, or sold out in three weeks and watched the launch momentum die on backorder. Both are forecasting failures. Here's how to forecast a product with no data and still hand the business a number it can plan against.

Why "no data" is the hardest forecast you'll make

Every other SKU in the plan has a past to smooth. The new SKU has a launch date, a spec sheet, a sales VP who swears it'll be huge, and a marketing budget. Somebody has to commit a production run and a working-capital number against all of that.

The stakes are real, not theoretical. A study in Marketing Letters tracked 83,719 new consumer-packaged-goods SKUs and found roughly 25% had stopped selling within a year and around 40% within two (Victory et al., 2021). Build to an optimistic single number and you fund the inventory that becomes next year's write-down.

That write-down is expensive even before it hits the P&L. Most businesses carry 25-30% of their working capital in inventory, and capital frozen in goods nobody wants can't go to growth or debt (IMCO Software, 2023). New launches are where that trap gets sprung.

The first move: kill the single number

A point forecast for a brand-new SKU is false precision. Nobody knows whether the launch does 4,000 units or 18,000. Pretending you do is how you commit to one production run and get it wrong.

Forecast a range with explicit assumptions instead. Three scenarios — low, base, high — each tied to a stated belief about cannibalization, marketing reach, and conversion.

The planning value isn't the base case. It's that everyone in the S&OP room now argues about the assumptions instead of the number, and you can pre-decide what you'll do at each level: which run size, which deposit, when you reorder. If your team isn't running consensus demand planning yet, a new launch is the right place to start the habit.

Method 1: Analog forecasting (look-alike modeling)

No history for this product doesn't mean no history at all. You've launched products before. The discipline is picking the right ancestors. Forecasting by analogy estimates a new product's demand and life-cycle shape from similar products that already ran their course (Institute of Business Forecasting, 2023).

How to build the analog set

• Find 3-5 prior launches that genuinely resemble the new one — same channel, comparable price tier, similar buyer, similar marketing push.
• Pull their actual demand curves: week-1 sell-through, ramp to peak, the shape of the tail.
• Average them, weighting toward the closest analogs, and overlay your new product's known differences (bigger launch budget, wider distribution, premium price).

The trap

The sales team will point you at the one product that went vertical. Force at least one disappointing analog into the set. Real launch portfolios have flops, and your forecast should price one in. This is also where measuring forecast bias by person pays off — wishful analog selection is just bias wearing a data costume.

Method 2: Bass diffusion model

When you're launching something genuinely new — a category buyers haven't seen — the Bass model is the standard. Frank Bass published it in Management Science in 1969, and it remains one of the most-cited papers in the journal's history (Bass, 1969). It splits adoption into two forces:

• Innovators (p) — buyers who adopt on their own, from advertising and discovery.
• Imitators (q) — buyers pulled in by word of mouth as the installed base grows.

The parameters

Feed it three inputs: market potential (m, total addressable units), coefficient of innovation (p, typically 0.01-0.03), and coefficient of imitation (q, typically 0.3-0.5). Out comes the classic adoption S-curve — slow start, steep middle, saturation.

You borrow p and q from analog product categories where adoption is already known. Bass is the right tool when word-of-mouth drives demand and the early ramp is what you need to plan capacity around. It's weak for fashion-style demand that spikes and dies, so don't force it onto a seasonal color refresh.

Method 3: Attribute-based forecasting

This is the underused method that pays off if you have a real product catalog. Instead of treating the new SKU as a unit, decompose it into attributes — color, size, material, price band, feature set — and forecast from how those attributes have historically performed across your line. The academic version is conjoint analysis, which estimates demand from the value buyers place on each attribute (Green & Srinivasan, 1990).

If walnut finishes outsell oak 1.6:1 across your existing range, and the new piece comes in walnut, the model already knows something about it before a single unit ships. Attribute-based forecasting is how you get a defensible number on a product that's new as an assembly but built from familiar parts.

It scales, too. Launch 40 new SKUs in a season and you're not hand-forecasting 40 times — you're scoring each one against the same attribute table. That's the same logic AI demand forecasting automates at scale once you have the attribute history clean.

Method 4: Structured judgment (done right)

Sales and marketing input is data — if you collect it without letting the loudest voice win.

Run a Delphi-style round

Each stakeholder submits a number independently and in writing, with reasons, before anyone sees anyone else's. Then you reveal, discuss the spread, and re-poll. The Delphi technique improves on unstructured group judgment by stripping out anchoring and the dominant-personality effect (Rowe & Wright, 2001).

Decompose, don't guess whole

Breaking a hard estimate into parts — addressable accounts × attach rate × units per account — beats guessing the total. Judgmental decomposition can cut forecast error materially under high uncertainty, which is exactly the regime a new launch lives in (Armstrong, 2001).

Track bias by person

If your VP of Sales runs 30% high on every launch, that's a known, correctable bias — apply the haircut. Most companies never measure this, so they re-absorb the same optimism every season.

Choosing the method

Situation	Best method
Variation on existing line (new color/size)	Attribute-based forecasting
Similar to past launches	Analog / look-alike modeling
Genuinely new category, word-of-mouth driven	Bass diffusion
High uncertainty, strong stakeholder opinions	Structured judgment + scenario range
Any high-stakes launch	Two methods, then reconcile the gap

The rule: never trust one method on a launch that matters. Run analog and attribute-based, see where they disagree, and the disagreement tells you where your risk is. Reconciling that gap is exactly what the S&OP monthly cycle is built to do.

Plan the launch as a sequence, not a bet

The forecast is wrong on day one — guaranteed with zero history. What separates good launch planning is how fast you correct.

1. Commit the smallest viable first run that hits your launch service level. Don't build to the high scenario. Build to base, and keep the option open.
2. Watch the first signal hard. Week-1 and week-2 sell-through against your scenarios is the most valuable data you'll ever get on this product. By week 3 you usually know which scenario you're in.
3. Pre-negotiate the reorder. Line up the supplier and the lead time before launch so you can pull the trigger on a replenishment run the moment the signal says high. The cost of being caught flat-footed on a winner is a stockout during peak buzz.
4. Pre-decide the markdown trigger. If week-4 sell-through tracks the low scenario, the markdown clock starts then — not at end of season when the inventory is stale and the cash is fully trapped. Obsolete stock ties up capital indefinitely; the longer you wait, the deeper the cut (Eagle Rock CFO, 2024).

That sequence — small first run, fast signal read, pre-staged reorder, pre-set markdown — turns a single high-stakes bet into a controlled series of small decisions. It's the difference between a launch that ties up a quarter of working capital and one that funds the next launch.

Where AI fits (and where it doesn't)

The honest answer: a machine-learning model can't conjure history that doesn't exist. What it can do is run analog matching and attribute scoring faster and across more candidates than a human team, then update the forecast continuously as early sales arrive.

McKinsey's work on forecasting in data-light environments makes the case that structured analog and attribute methods, paired with AI, beat naïve guessing even when the target product has no record of its own (McKinsey, 2022). The value shows up after launch — fast signal reads and automatic reforecasting — more than before it.

Don't let a vendor sell you AI as a no-data magic forecast. The discipline is the same with or without the model: pick good analogs, decompose the assumptions, and correct fast. The model just lets you do it across the whole launch slate instead of the top three SKUs.

What good looks like

A mid-market manufacturer with real new product demand forecasting should have:

• A scenario range with named assumptions for every launch, not a single number.
• At least two forecasting methods on any launch above a dollar threshold, reconciled in S&OP.
• An analog library and attribute history that make each new launch faster to forecast than the last.
• A launch playbook with pre-staged reorders and pre-set markdown triggers tied to early sell-through.

Frequently asked questions

How do you forecast demand for a product with no historical data?

You substitute borrowed history for missing history. Use analog (look-alike) products that already launched, a structural model like Bass diffusion for genuinely new categories, and attribute-based forecasting that scores the new SKU against how its color, size, material, and price band have performed across your catalog. Combine at least two of these, express the result as a low-base-high range, and correct it against the first weeks of real sell-through.

What is the Bass diffusion model used for?

The Bass model, published by Frank Bass in 1969, forecasts the adoption curve of a genuinely new product or category. It splits buyers into innovators who adopt on their own and imitators pulled in by word of mouth, producing the classic S-shaped adoption curve. It's most useful when you have no sales history but can borrow innovation and imitation parameters from a comparable product category.

How accurate is new product forecasting?

Far less accurate than forecasting an established SKU, which is why you forecast a range rather than a point. Expect to be wrong on day one and plan to correct: week-1 and week-2 sell-through against your scenarios usually tells you which case you're in by week three. The goal isn't a precise number, it's a number tight enough to commit a safe first production run against.

What is analog forecasting?

Analog forecasting, also called look-alike modeling, estimates a new product's demand from the actual demand curves of similar products you've already launched. You pick 3-5 genuine analogs matched on channel, price tier, buyer, and marketing push, then average and adjust their curves for the new product's known differences. The discipline is including at least one disappointing analog so the forecast prices in the real chance of a flop.

How much inventory should you build for a new product launch?

Build the smallest viable first run that still hits your target launch service level — base case, not the high scenario. Pre-negotiate the reorder with your supplier so you can replenish fast if early sell-through signals the high case, and pre-set a markdown trigger if it signals the low case. This staged approach caps your downside on a flop while keeping the upside open on a winner.