How to Forecast Demand for a New Product: A 2026 Guide for Consumer Brands

A planner is sitting in a launch-prep meeting. Marketing is excited about a new shade range dropping in six weeks. Operations need a buy quantity. Finance wants to see an opening order. The planner looks at her tools — a sales-history file, a baseline forecast, an exponential-smoothing model — and realizes none of them have anything to say about a product that doesn't exist yet.

This is the cold-start problem in demand planning. Forecasting demand for a new product is structurally different from forecasting demand for an existing one — and the brands that handle it well treat it as two distinct jobs rather than a single task.

This guide is for demand planners and supply chain leaders at consumer brands — CPG, D2C, fashion, beauty, electronics — who launch new products at any meaningful cadence and need a practical framework for forecasting them. We cover what's structurally different about new product forecasting, the five methods that work, when to use each one, and the six-step process that ties them together.

Why Forecasting a New Product Is Structurally Different in Retail?

Traditional demand forecasting starts from history. A statistical model looks at how much you sold last year, last quarter, and last week and extrapolates a future. For an established SKU with 12+ months of clean sales data, this works well — accuracy at 10–15% WMAPE is achievable for stable items.

For a new product, none of that applies. By definition, there is no sales history. Most planning teams handle this with a manual judgment override on top of a baseline that doesn't apply — and the launch ramp ends up 40–60% off, locking up working capital in slow-moving stock or under-supplying a bestseller. McKinsey's research on consumer brand product launches has consistently flagged this as one of the largest hidden sources of margin erosion in the industry.

“As a fashion brand, we continually introduce new SKUs in various styles to cater to our customer base. While we can manage demand planning for our established items using techniques such as moving averages on Excel, we struggle when it comes to estimating demand for newly launched products,” expressed the Founder of a fashion brand.

Consumer brands feel this acutely. Retailers and DTC brands typically launch 15–30% of annual revenue from products that didn't exist a year ago. In categories like fashion, beauty, and snacks, that share is even higher. A planning function that can't forecast new products is, mathematically, getting a meaningful slice of the portfolio wrong every cycle.

The good news: the problem has known solutions. The five methods below have been tested across thousands of consumer brand launches, and they all share one principle — they replace historical sales data of the new product itself with structured information about similar past launches and product attributes.

Two Distinct Jobs: Pre-Launch and Post-Launch for a New Product

Most demand planning content treats new product forecasting as one task. In practice, it's two — and they need different methods.

Pre-launch forecasting is the cold-start problem. You have no direct sales data on this SKU, but you have the launch date, product attributes, marketing plan, channel mix, and an analog set of past launches. The job here is to produce a defensible baseline forecast that drives the opening order — typically the buy quantity you commit to with the factory.

Post-launch forecasting is the calibration problem. From the moment the product hits the channel, you start gathering real data — but you have too little of it to forecast with traditional statistical methods. Three weeks of POS don't yield a reliable model on its own. The job here is to blend the pre-launch baseline with emerging actuals, weighted toward whichever signal is currently more reliable. By roughly week 12–16 post-launch, you typically have enough history to retire the pre-launch model and move the SKU to standard forecasting.

Both jobs require different methods and different feedback loops. Confusing them is the most common source of new-product forecasting error — teams either keep using the pre-launch model long after they have real data, or abandon it the moment first-week sales come in and overreact to a partial signal.

The 5 Methods That Work for New Product Forecasting

1. Attribute-Based Forecasting

This is the method TrueGradient was built around, and it's the strongest single approach for consumer brands with rich product attribute data.

The model extracts product attributes from existing products and cross-learns them with all products in the same category — as well as other categories possessing matching attributes — to predict demand for the new product. Let's illustrate with an example from a fashion apparel brand.

Suppose they launch a new shirt. The SKUs for this new shirt, based on size and colour, may look something like this:

Now, the demand planning of the new shirt will happen at the most granular level — i.e., SKU ID. For SKU1, the model takes into account the existing assortment of shirts and algorithmically cross-learns the demand for red shirts, shirts with size XS, collared shirts, plain shirts, and shirts sold on the website to predict the demand for SKU1, which in this case would be a red, solid, collared shirt in XS size to be sold on the website. Naturally, these algorithms are more intricate than they might appear, as numerous factors are considered on the backend to ensure accuracy.

The same logic generalizes across categories. For a beauty brand launching a new lipstick shade, attributes include shade family, undertone, finish (matte/satin/gloss), price tier, and channel — and the model forecasts the new shade by cross-learning from how similar past shades performed. For a CPG brand launching a new flavor extension, attributes include flavor family, pack size, price tier, and channel. For electronics, attributes include screen size, processor tier, price band, and form factor.

The compounding accuracy gain across launches is what makes attribute-based forecasting one of the largest single levers in modern consumer brand planning — see ToolsGroup's published Aston Martin case, which reported a 10% WMAPE reduction and a 30% new launch accuracy lift after adopting attribute-based clustering.

A common misconception worth addressing here: don't substitute competitor data for your own attribute model. Every business is unique in terms of growth stage, distribution channels, and brand positioning. A product that performed well for a competitor at the same price tier may behave entirely differently in your portfolio. Use your own analog set.

2. Analog / Lookalike Forecasting

Where attribute-based forecasting cross-learns mathematically across the full assortment, analog forecasting picks a small set of behaviorally similar past products — typically 3–5 launches — and uses their first 12–16 weeks of sales as the template.

The discipline here is in the analog selection. A new lipstick shade launching at $24 with a TikTok-led campaign behaves nothing like a $40 lipstick shade launching to wholesale channels — even if the attribute overlap looks high. Strong analog forecasting weighs not just product attributes but also marketing intensity, channel mix, price tier relative to the line, and lifecycle stage of the category.

Modern AI planning systems automate analog selection by ranking historical launches against the new SKU across all these dimensions, removing the bias that creeps in when planners pick analogs manually.

3. Causal and Judgmental Methods

For entirely new categories — markets where you have no analog set because you've never sold anything like this — pure quantitative methods fall short. Three qualitative approaches earn their place here:

• Delphi method. Expert panels (internal R&D, category leaders, retail partners) iteratively converge on a forecast through structured rounds of input. Most useful for category-creating launches.
• Sales force composite. Aggregated estimates from the sales team and key account managers. Strong for B2B-heavy consumer brands with relationship-driven distribution.
• Market research and customer panels. Surveys, pre-order signals, and customer-panel testing produce structured demand-intent data. Most useful for category extensions where consumer comprehension is uncertain.

These methods work best as inputs into a quantitative model rather than as standalone forecasts. The pattern that works: use judgment to set a credible base, then layer attribute-based or analog modeling around it as a sanity check.

4. Bayesian Ramp Models

A Bayesian ramp model starts with a prior — typically an attribute-based or analog forecast — and updates the forecast as actual sales data comes in. Each week of real sell-through tightens the model's confidence and shifts the forecast toward the actual ramp curve.

This is the bridge between pre-launch and post-launch. By Week 4 post-launch, the Bayesian model is typically weighting actuals at 30–40% and the analog prior at 60–70%. By Week 12, that flips. By Week 16, the prior is retired and the SKU moves to standard time-series forecasting.

For the technical foundation here, see our piece on probabilistic modelling using prediction intervals.

5. Diffusion / Bass Models

Originally developed for durable goods (where adoption follows a clear S-curve — early adopters, early majority, late majority, laggards), diffusion models like the Bass model have re-emerged in consumer brand planning for products with strong adoption-curve dynamics: electronics, premium beauty, novel CPG categories.

The Bass model fits two parameters — coefficient of innovation (early adopter pull) and coefficient of imitation (network/social spread) — and produces a full adoption curve from launch to maturity. For high-growth DTC brands with weekly product drops, simplified Bass-style models inform the buy quantity decision by predicting whether a launch will trend toward viral, steady, or niche adoption.

Methods Comparison: When to Use Each Forecasting Method for a New Product

The strongest planning teams don't pick one method — they layer them. Attribute-based modeling generates the baseline; judgmental inputs provide a sanity check; a Bayesian framework blends pre-launch and post-launch signals as actuals arrive. The platform you use should handle this layering automatically; if a planner has to manually choose between methods per SKU, the system is too low-level for production use.

What is the Process for Demand Forecasting a New Product That Works?

Step 1 — Define what you're forecasting

This sounds trivial. It isn't. "Forecast the new shirt" can mean total units across all sizes and colors, or unit demand per SKU per channel per week. The granularity you commit to here drives every downstream decision. For consumer brands with high SKU complexity, forecast at the SKU-channel-week level — aggregate up when you need higher levels, never the reverse.

Step 2 — Build the analog set and attribute map

For each new SKU, identify the 3–5 most similar past launches (analog set) and the product attribute fingerprint (attribute map). The analog set drives lookalike methods; the attribute map drives attribute-based methods. Most platforms do this automatically once the product master is set up properly — covered in more depth in the top 3 data readiness concerns of a mid-market CPG and retail player.

Step 3 — Generate the pre-launch baseline

Run the attribute-based and analog models against the new SKU. Pull in judgmental inputs from marketing and category teams as structured adjustments with reason codes (not free-form overrides). The output is a pre-launch baseline forecast with a confidence range — typically expressed as P10 / P50 / P90 (10%, 50%, 90% probability scenarios).

Step 4 — Decide on stage-gate buy quantity

The buy decision isn't simply "the forecast." It's a structured decision based on the forecast, the service-level target, the working capital constraint, and the cancellation/return cost. Most mature brands operate this as a stage-gate: a conservative opening buy with rapid reorder cycles if the launch performs above expectations.

Step 5 — In-flight signal calibration (Weeks 1–4)

The first four weeks post-launch are signal-rich and noise-prone. POS data starts flowing, marketing spends activate, retail partners place reorders or hold off, and search/social signals indicate consumer interest. The discipline here is to read the signals as Bayesian updates to the pre-launch prior, not as ground truth on their own. A Week 2 sales spike could mean a hit — or it could mean a campaign pulse that won't recur.

Step 6 — Lock the post-launch forecast and retire the analog model

By Weeks 12–16 post-launch, the SKU typically has enough history to forecast with standard statistical or ML methods. The pre-launch model retires; the SKU moves into the regular planning cycle. The post-launch performance feeds back into the analog set for the next launch — the system learns.

Demand Forecasting Examples for Newly Launched Products Across Different Industries

Fashion — the original red-shirt example (above)

Attribute-based forecasting at the SKU level — color × size × style × channel — using the existing assortment as the analog universe.

Beauty — new lipstick shade

The same model applies. Attributes: shade family (red/pink/nude/berry), undertone (cool/warm/neutral), finish (matte/satin/gloss/lacquer), price tier, packaging size, channel. A new "burgundy matte at $24 launching to DTC + Sephora" cross-learns from how past burgundy-family, matte-finish, $24-tier products performed in the same channel mix. The shade name and the marketing concept don't enter the model; the attribute pattern does.

CPG — new flavor extension

Attributes: flavor family (sweet/savory/spicy), pack size (single-serve/family/club), price tier, channel (Amazon Vendor Central/retail / DTC), launch season. A new "spicy single-serve at premium tier launching to Whole Foods" cross-learns from past spicy single-serves at premium tiers in natural-grocery channels. Promotional support and shelf placement enter as multipliers.

Electronics — new model SKU

Attributes: screen size, processor tier, RAM/storage, price band, form factor, launch timing relative to category cycle. A new mid-range laptop launching pre-holiday cross-leverages from past mid-range laptops launched at the same time of year.

The method is the same. Only the attribute set changes by category. This generality is what makes attribute-based forecasting a single capability that scales across a consumer brand's full portfolio — from fashion to beauty to CPG to electronics — rather than requiring a different model per category.

What Changes with AI-Driven NPI Forecasting?

Three things change when you move from manual to AI-driven new product forecasting:

The analog set selection becomes automated and unbiased. Manual analog selection is shaped by planner intuition — strong planners pick well, but the bias is unmeasured. AI-driven analog selection ranks every past launch against the new SKU on a multi-dimensional similarity score, removing the unmeasured bias.

Attribute weighting becomes adaptive. A manual model treats every product attribute as equally important. An AutoML-based system learns from past launch outcomes, which attributes matter most for which categories, and the weights update as new data arrives.

Outputs become probabilistic and explainable. Instead of a single point forecast, modern systems output a range with confidence intervals (P10 / P50 / P90) and an explanation of which attributes and analogs drove the number. Explainability is what makes planners trust the forecast enough to use it — covered in factor contribution in demand forecasting.

For two angle-specific deep dives, see how high-growth brands solve demand planning for new products (D2C/DTC with weekly drops) and how CPG brands approach new product demand planning (enterprise CPG with structured stage-gate launches).

Common Mistakes Brands Make While Forecasting for New Products

1. Substituting competitor data for your own analog set. Every business is unique in terms of growth stage, distribution channels, and brand positioning. What works for one business may not necessarily apply to others.

2. Using one method for every launch. A category-creating launch needs Delphi + market research; a line extension needs attribute-based forecasting; a relaunch needs analog modeling on the original. Forcing one method across all launches forces every assumption to compromise.

3. No stage-gate buy quantity. Treating the buy as a single decision — rather than a conservative opening with rapid reorder cycles — locks all the risk into the initial commitment and removes the brand's ability to respond to early signals.

4. Overreacting to first-week sales. A Week 1 spike or shortfall is a signal plus a lot of noise — campaign pulses, retailer ordering quirks, channel-mix anomalies. Strong planning teams treat Weeks 1–4 as Bayesian updates, not as ground truth.

5. Not segmenting launches by type. Category-creating launches, line extensions, color/flavor expansions, and replacement SKUs all behave differently. A planning function that treats them uniformly forces the same accuracy target on launches with radically different forecasting difficulty.

FAQs on Demand Forecast of New Product Launch

How do you forecast demand for a new product with no sales history? Use attribute-based or analog forecasting. Attribute-based methods cross-learn demand patterns from the full assortment using the new product's attributes (color, size, flavor, price tier, channel). Analog methods pick 3–5 similar past launches and use their early sell-through curves as the template. Both approaches replace the missing sales history of the new product itself with structured information about similar past products.

What is attribute-based forecasting? Attribute-based forecasting predicts demand for a new product by extracting its attributes (color, fabric, flavor, price tier, channel, etc.) and cross-learning from how products with similar attributes performed in the past. It's the strongest single method for consumer brands with rich product master data and is particularly effective for line extensions, color/size variants, and category extensions.

How accurate can new product demand forecasting be? Accuracy depends heavily on the method, the analog quality, and the launch type. For attribute-based methods on line extensions in established categories, 15–25% WMAPE at launch is achievable, tightening to standard accuracy by Week 12 post-launch. For entirely new categories with no analog set, accuracy expectations are lower (30–50% WMAPE) and judgmental methods carry more weight. ToolsGroup's published case work on Aston Martin reported a 30% improvement in new launch forecast accuracy after adopting attribute-based clustering.

What's the difference between pre-launch and post-launch forecasting? Pre-launch forecasting is the cold-start problem — you have no direct sales data, only product attributes, an analog set, and the launch plan. Post-launch forecasting is the calibration problem — you have actuals but not enough to forecast statistically. Most brands handle both with a Bayesian framework: the pre-launch model produces a prior, and each week of actuals updates the forecast. By Weeks 12–16, the SKU moves to standard time-series forecasting.

Can AI forecast demand for a new product? Yes — and in fact, AI methods are structurally better suited to new product forecasting than to forecasting established SKUs. AI handles the multi-dimensional similarity matching that drives attribute-based and analog methods, automates the analog set selection, learns attribute weighting from past launch outcomes, and produces probabilistic outputs with driver explanations. Manual methods cannot match this at scale.

How long after launch can you stop using new product forecasting methods? Typically 12–16 weeks. By that point, the SKU has enough actual sales history to forecast with standard statistical or ML methods. The pre-launch model retires, the SKU moves into the regular planning cycle, and the post-launch performance feeds back into the analog set for the next launch.

What data do you need to forecast a new product? Three categories of data: (a) product master with attributes (color, size, flavor, ingredient, price tier, lifecycle stage, channel), (b) historical sales data on the analog set — past launches in similar categories, ideally at SKU-channel-week granularity, (c) launch context — launch date, marketing plan, promotional calendar, channel allocations. Without rich product attributes, attribute-based methods can't work; without an analog set, lookalike methods can't work.

Should I use competitor data to forecast my new product? No, as a primary input. Every business is different — your distribution, your brand positioning, your pricing strategy, your channel mix. A competitor product that performed well at your target price tier may behave entirely differently in your portfolio. Competitor data is useful as one input among many (category size, growth direction, channel trends), not as a substitute for your own analog set.

Plan Your Next Launch With Confidence

TrueGradient is an AI-native planning OS for consumer brands. We help CPG, D2C, fashion, beauty, and electronics teams forecast new products from day one — through attribute-based modeling, automated analog selection, Bayesian post-launch calibration, and explainable forecasts that planners can trust. The platform connects new product forecasting natively to assortment planning, inventory optimization, merchandise financial planning, and demand planning — so a new launch flows from forecast to buy quantity to allocation in one connected workflow.

If you're launching products at a meaningful cadence and want to see what attribute-based forecasting looks like applied to your portfolio, write to us at info@truegradient.ai or book a demo · talk to us.

Related reading:

• How high-growth brands solve demand planning for new products
• How CPG brands approach demand planning for new products
• Why traditional forecasting fails in fast fashion
• Probabilistic modelling using prediction intervals
• How to reduce dead stock in apparel
• Pre-season and in-season planning in fashion retail