Why New-Product Demand Planning Is Broken for Growing Brands, and What Actually Works

Every growing brand hits the same wall the moment it launches something. The demand planning that works for established SKUs is built on their sales history, and a new product has none. For a brand launching every few weeks, that is not an edge case. It is a growing share of the catalogue, and the part where getting inventory wrong costs the most.

The traditional response, taught in every buying office, is analog forecasting: pick two or three similar products, use their history as a proxy, place a conservative first order, and adjust once real sales arrive. It is a reasonable instinct. But it rests on two assumptions that quietly break for a modern brand: that a person can reliably pick the right analog, and that a single pre-launch number is the output you need. Both are flaws. For the fundamentals of how new-product forecasting works, see our full guide to demand planning for new products in retail. This page is about why the traditional approach breaks and what replaces it.

Why new-product demand planning is broken for modern commerce

The core issue is a well-known one in forecasting: the cold-start problem. As the peer-reviewed operations research on new products puts it, inventory, capacity, and marketing decisions "must often be made with little or no product-specific history" (Hu et al. 2019; Lei et al. 2023). Traditional time-series models cannot help here, because they infer demand from a product's own history, and there isn't any. That leaves the highest-stakes inventory decision a brand makes, the launch buy, with the least data to support it.

The traditional workaround has three structural weaknesses that compound for a fast-launching brand.

It depends on a person picking the right analog. The buying-office standard is to select two or three "similar" products by judgment. But similar on what axis: price, category, customer, season? A buyer's mental model of similarity is subjective, inconsistent between people, and does not scale past a handful of launches. Pick the wrong analog and the whole forecast inherits the wrong demand shape. This is the single biggest point of failure, and it is the part everyone treats as unavoidable.

It produces a single number when it should produce a range. New-product demand is genuinely uncertain, more uncertain than any established SKU, so compressing it to one figure discards the most important information: how wide the range of outcomes is. A launch that could do 5,000 or 25,000 units needs a very different inventory posture from one confidently around 12,000, even at the same expected value.

It treats the forecast as a one-shot guess, not a system that converges. The richest signal, actual early sell-through, arrives right after launch, yet the traditional process makes the pre-launch call and at best revises it manually much later. Treating the launch forecast as finished once the PO is placed wastes the data that would correct it fastest.

None of this means new-product forecasting is hopeless. It means the traditional method is a coarse tool for a problem that now has much better ones.

Why spreadsheets fail for new products

The spreadsheet version of this concentrates every one of those weaknesses. A spreadsheet can hold a chosen analog's history, but it cannot search thousands of candidates to find the right analog, so that stays a manual guess. It naturally produces a point estimate, because a cell holds one number, not a distribution. It has no mechanism to update from early sell-through except a person rebuilding it. And the analog choice and its adjustments live in one planner's private logic, undocumented and unrepeatable.

The result is a launch forecast that is subjective, single-valued, static, and opaque, which is the worst possible combination for the highest-uncertainty decision a brand makes. The broader pattern is why traditional forecasting fails in fast fashion, where launch velocity makes the spreadsheet approach break fastest.

The working-capital cost of a bad launch forecast

Getting a launch forecast wrong is uniquely expensive, because both directions cost real cash and there is no history to soften the error.

You over-forecast and create instant dead stock, inventory bought against demand that never existed, with no established baseline to sell through over time. It goes straight to markdown and the working capital is frozen from day one. Under-forecast and you stock out during the only launch you get, losing not just those sales but the algorithmic and word-of-mouth momentum a launch depends on, which rarely returns.

The industry evidence shows how much is at stake here specifically. One AI planning vendor reports that the accuracy gains from proper cold-start methods "are typically largest on new product launches," precisely because the legacy tools most retailers use handle cold-start poorly or not at all (OnePint.ai). The working-capital math is unforgiving because the error is large (high uncertainty), immediate (no baseline to average it out), and one-shot (you cannot re-run the launch). That asymmetry is exactly why the single-number approach is so costly: it gives you no way to plan the buy against the range of outcomes.

Attribute-based forecasting: similarity learned, not guessed

The first real fix is to stop hand-picking analogs and let the data define similarity.

Attribute-based forecasting predicts a new product's demand from its attributes, category, price band, colour, material, size curve, and style, by learning how products with those attributes have performed across stores and channels. As Grid Dynamics notes, attribute-based forecasting "utilizes product attributes for forecasting, especially useful for new products without historical sales data." The principle underneath it is simple and, in the industry's words, foundational: "similar SKUs tend to behave similarly; this is the foundation on which we build a new SKU pattern estimation".

The important shift is who decides what "similar" means. Instead of a buyer choosing three analogs, an AI clustering approach identifies similarly behaving SKUs "by considering factors such as product attribute data (color, material, etc.), pricing information," and more (Impact Analytics), effectively weighing every past product by attribute similarity rather than relying on one person's pick. This is not a novelty either. Retail-planning patents describe creating a "placeholder" product from a "combination of attributes and attribute values" precisely so a planner can forecast demand for an item that does not yet exist (US 11030574). And it fits how modern retailers actually operate: as Impact Analytics observes, brands "constantly churn out new assortments" and "often create products by modifying prior SKUs," which is exactly the situation attribute similarity is built to exploit.

It depends on the same driver attribution that decomposes any demand signal, applied across products rather than within one. The practical requirement is an attribute taxonomy, so the richer your product-attribute data, the sharper the cold-start forecast. As one 2026 industry playbook puts it plainly, "AI forecasting for new items works by reading attributes, not history" (Impact Analytics).

Probabilistic life-cycle forecasting: a curve, not a number

The second fix is to forecast the launch as a distribution over its life-cycle, not a single figure.

A new product does not have "a demand number." It has a life-cycle: a ramp, a peak of some timing and magnitude, and a decline, all uncertain. The right output is a probabilistic forecast of that whole curve, a low, expected, and high trajectory with likelihoods, because that is what lets you plan the buy against the actual risk. The recent academic work is explicit that the goal is forecasting "the full demand distribution of new products across launch stages" (Guo et al. 2025), not a point estimate, and that attribute and promotion information should feed a life-cycle view rather than a single number (Lei et al. 2023).

With a distribution in hand, you can stock a can't-miss launch to a high band and accept the extra inventory as insurance, or protect cash on a speculative one by planning closer to the expected case. That is a deliberate service-versus-cash decision the single-number method cannot express.

Self-correcting forecasts: updating from early sell-through

The third fix is to treat the launch forecast as a system that converges, not a guess you defend.

Self-correcting forecasts: updating from early sell-through

The third fix is to treat the launch forecast as a system that converges, not a guess you defend.

Pre-launch, the forecast conditions on attributes and analogs. Then, as the first days and weeks of real sales arrive, the forecast should update coherently, blending the prior (attributes, analogs) with the emerging product-specific signal. The industry calls this in-season learning, where "predictions update as early sell-through signals arrive, and they update often," refreshing "on a rolling cadence, sometimes daily," until a product graduates to standard models "once it has a real sales history," usually after a full season (Impact Analytics).

There is a subtle trap that good systems handle explicitly: early stockouts censor demand, and if you learn from raw sales, you teach the model that a sold-out product had low demand. Mature platforms apply "censored demand correction" so that "early stockouts don't compound errors across future launches" (OnePint.ai). This is where agentic AI matters: the forecast re-optimises continuously as the launch unfolds, so the second replenishment is far smarter than the first order, and the third smarter still.

How to forecast a new product, step by step

1. Build the attribute taxonomy first. Define the traits that drive demand in your category, price band, material, style, size curve, because attribute-based forecasting is only as good as the attributes you capture. Start it before you need it.
2. Let the model find analogs by similarity. Instead of hand-picking two or three comparables, use attribute-based forecasting to weigh all relevant history by similarity. See AI demand forecasting.
3. Forecast the life-cycle as a distribution. Produce a probabilistic curve (ramp, peak, decline) with low, expected, and high bands, not a single number.
4. Choose the buy against the range and your risk. Stock a can't-miss launch to a high band; protect cash on a speculative one, feeding inventory optimization and replenishment.
5. Update from early sell-through, with censored-demand correction. Let the forecast converge as real sales land, and make sure early stockouts are corrected for rather than learned from.
6. Tie it to the financial plan. Connect the launch buy to the merchandise financial plan so the assortment lands within the dollar plan. The onboarding arc is in what the first 90 days look like.

How high-growth brands solve new-product demand planning

The brands that launch well share a pattern. They invest in attributes, not analogs, treating the product-attribute taxonomy as infrastructure, because it is what makes every future launch forecastable. They plan launches as ranges, so the buy is a risk decision rather than a single bet. They learn in-season, updating forecasts from early sell-through instead of waiting for the post-mortem. And they treat classification and launch together, so a new SKU's attributes place it in the assortment and its ABC-XYZ segment from day one rather than defaulting it to "unknown."

The payoff is measurable. Vendors reporting on cold-start specifically cite "20 to 30% better forecast accuracy" and "up to 85% fewer stockouts" on new launches once attribute-based, probabilistic, self-correcting methods replace the legacy approach (OnePint.ai). Re-verify those figures for your own context, but the direction is consistent across the industry: the biggest accuracy gains available to a growing brand are on exactly the launches the traditional method handles worst.

Common mistakes in new-product demand planning

• Treating cold-start like normal forecasting. Running a time-series model on a product with no history produces confident nonsense; attribute-based methods exist precisely because the standard approach cannot work here.
• Relying on a single hand-picked analog. One comparable carries one product's idiosyncrasies into your forecast; similarity should be learned across many products, weighted by attribute.
• Forecasting a number instead of a range. The most costly mistake, because it hides the uncertainty that should drive the buy.
• Learning from censored demand. If early sell-through includes stockouts, raw sales understate true demand; without correction, one sold-out launch poisons the next.
• Treating the pre-launch forecast as final. The best data arrives right after launch; a forecast that cannot update from early sell-through wastes it.
• Skipping the attribute taxonomy. Without captured attributes, attribute-based forecasting has nothing to connect to. Start the taxonomy early.

New-product demand planning FAQs

Why is demand planning for new products so hard? Because a new product has no sales history, and traditional time-series methods need roughly 8 to 12 weeks of data to work. The peer-reviewed research frames this as the cold-start problem: decisions "must often be made with little or no product-specific history." That leaves the highest-stakes inventory decision, the launch buy, with the least data. The traditional workaround, picking a few analog products by hand, is subjective and produces a single pre-launch number, which is a coarse tool for a genuinely high-uncertainty problem.

What is analog forecasting, and what is wrong with it? Analog forecasting uses one or more similar existing products as a proxy for a new item's demand, the standard approach in retail buying offices. The instinct is sound, since "similar SKUs tend to behave similarly," but it has two weaknesses: the analog is hand-picked by human judgment, which is subjective and does not scale, and it typically produces a single-point forecast. Attribute-based forecasting improves on it by learning similarity from data across all products rather than relying on a buyer's choice of two or three.

What is attribute-based forecasting for new products? It is forecasting a new product's demand from its attributes, category, price band, colour, material, and size curve, by learning how products with similar attributes have performed across stores and channels. In the industry's words, "AI forecasting for new items works by reading attributes, not history." Instead of a planner hand-picking analogs, the model weighs all relevant history by attribute similarity, which is especially strong in fashion and specialty retail where attributes are rich.

How does AI forecast a product with no sales history? By transferring demand patterns from similar products through their shared attributes rather than needing the new product's own history. AI clustering identifies similarly behaving SKUs using product attribute data such as colour, material, and pricing, then produces a store-level forecast from day zero. It updates that forecast as early sell-through arrives, correcting for early stockouts so censored demand does not compound errors, and converges on the product's real demand over its first weeks.

Should a new-product forecast be a single number? No. It should be a probability distribution over the product's life-cycle, because new-product demand is genuinely uncertain. Recent academic work targets "the full demand distribution of new products across launch stages" rather than a point estimate. A distribution lets you plan the buy against the real range of outcomes and choose how much launch risk to cover, stocking a can't-miss launch to a high band or protecting cash on a speculative one.

How do you improve a new-product forecast after launch? Let it update from early sell-through. Industry practice is in-season learning, where predictions "update as early sell-through signals arrive, and they update often," sometimes daily, until the product graduates to standard models after a full season of history. The important nuance is correcting for censored demand, so that early stockouts are not mistaken for low demand and do not poison future launches.

Why do spreadsheets fail for new-product forecasting? Because they concentrate every weakness of the traditional approach: they cannot search thousands of candidates to find the right analog, so it stays a manual guess; they naturally produce a single number rather than a distribution; they cannot update from early sell-through without someone rebuilding them; and the analog choice lives in one planner's undocumented logic. That is the worst combination: subjective, single-valued, static, and opaque, for the highest-uncertainty forecast a brand makes.

Where TrueGradient fits

TrueGradient forecasts new products the way the problem actually demands: attribute-based, probabilistic, and self-correcting. AI demand forecasting predicts a new SKU's demand from its attributes rather than a hand-picked analog, produces a probabilistic life-cycle forecast rather than a single number, and updates continuously as early sell-through arrives, so the launch buy is a deliberate risk decision and each replenishment is sharper than the last. It flows into inventory optimization, replenishment, and the merchandise financial plan on one substrate.

Typical time to first measurable outcome is 8 to 12 weeks. See what the first 90 days look like.

Book a demo · Talk to us

Related reading: Demand planning for new products in retail · Why traditional forecasting fails in fast fashion · Factor contribution in demand forecasting · Probabilistic modelling using prediction intervals · How the apparel industry can reduce dead stock