How Do You Forecast Demand for New Products?

New product launches face a forecasting paradox. Inventory decisions get made before any sales data exists, but those decisions determine whether the launch succeeds or fails. Order too little and stockouts kill the launch momentum. Order too much and markdowns destroy the margin. Both happen because nobody knows what demand will actually look like.

The technical name for this is the cold-start problem. It's one of the harder challenges in retail forecasting, and it's where modern AI techniques deliver their largest gains over traditional methods. Here's how the problem actually works, the four approaches that solve it, and the lifecycle phases that determine which approach to use when.

The Two Components Every Cold-Start Forecast Estimates

Most articles on new product forecasting jump straight to techniques. The conceptual foundation matters more than the specific method. Every cold-start forecast separates into two distinct estimates that need different inputs.

Sales pattern: the shape of the curve

The pattern captures how demand evolves over time, including ramp-up, peak, plateau, and decline. It also captures seasonality, day-of-week effects, and response to promotions. Patterns are reasonably transferable between similar products, which is why borrowing patterns from comparable SKUs is the standard cold-start technique.

Scale: the magnitude of demand

The scale capt[3] ures how much the product actually sells. A new lipstick shade and an existing one might share the same shape (slow start, accelerate over four weeks, peak around week 8), but the new one might sell at a fraction of the volume because of lower brand awareness, smaller marketing budget, or limited store distribution.

Scale is the harder estimate. Industry research from Impact Analytics highlights that scale depends on buy quantity, price point, store penetration, marketing investment, and customer adoption rate. Bass diffusion models, originally developed in marketing science to predict adoption of new technologies,originally developed by Frank Bass in 1969 to predict how new products spread through a market, are particularly useful for estimating customer adoption curves.

Key takeaway: Pattern and scale need different techniques. Pattern can be borrowed from comparables. Scale needs additional signals about distribution, marketing, and customer adoption. Treating them as a single estimate is one of the most common cold-start mistakes.

The Four Cold-Start Modelling Approaches

Modern AI forecasting platforms use four distinct approaches to the cold-start problem. The right choice depends on what kind of new product is launching, how much information is available pre-launch, and what category dynamics apply.

1. Supersession modelling

When a new product directly replaces an existing one (V2 of a phone, a refreshed version of a packaged good, a new model year of a car), the historical sales of the predecessor become the forecast foundation. This is the highest-accuracy approach when it applies. The model adjusts for known differences (price changes, distribution changes, feature improvements) but inherits the bulk of the demand pattern from the predecessor.

2. Attribute-based clustering

This is the workhorse method for most cold-start situations. The platform groups existing SKUs by attributes (category, brand, price tier, size, color, material) and identifies the cluster the new SKU belongs to. The forecast borrows the demand pattern of comparable SKUs in that cluster, weighted by similarity. This is what we've been calling attribute-based modelling throughout the cluster.

3. Visual similarity (especially for fashion)

For products where appearance drives purchase (apparel, beauty, home goods, design), structured attributes alone miss important signal. Recent research published in the International Journal of Data Science and Analytics (2026) demonstrates how deep learning architectures like ResNet-50 and InceptionV3 extract visual features from product images, then match new products to historically similar ones based on what they actually look like. This catches signals that structured metadata can't capture, like style, vibe, and visual appeal.

4. Meta-learning

The cutting-edge approach for high-volatility categories. Instead of training a model on historical data and applying it to new products, meta-learning trains a model on how to learn from minimal data. The model gets exposed to many product launches and learns the meta-pattern of how forecasts should adapt as early signals arrive.

Key takeaway: No single approach handles every cold-start situation. Modern platforms apply the right technique for each launch type, often combining methods (attribute-based plus visual similarity for fashion, supersession plus meta-learning for tech upgrades).

The Four Lifecycle Phases of a New Product Forecast

Cold-start forecasting isn't a single problem solved at launch. It's a sequence of four phases, each requiring a different forecasting approach as more real data becomes available.

The transitions between phases are smooth, not sharp. A good platform blends attribute-based forecasts with actual early sales data progressively, reducing the weight on comparables as real data accumulates and confidence improves. Forcing a hard switch at any point ("now we have four weeks of data, throw away the comparables") tends to produce volatile forecasts. Smooth blending produces stable ones.

Key takeaway: New product forecasts evolve through phases. Each phase needs its own approach, and the transitions between them should be gradual rather than abrupt. A forecast that looks the same in week 2 and week 12 is doing something wrong.

The Censored Demand Problem for New Products

This is the cold-start version of the shortage-censored data problem that affects established products. Because new product demand is inherently uncertain, retailers often under-order on the first run, which causes early stockouts. The recorded sales reflect what was available, not what was actually demanded.

If the model trains on the censored data without correction, it learns that demand is lower than reality. The next order will be too small. The next launch in the same cluster will be under-forecast. The bias compounds across launches.

The technical fix is censored demand correction using algorithms like Expectation-Maximization (EM). Recent research in fashion demand forecastingAdel et al. (2025), published in the International Journal of Data Science and Analytics, demonstrates EM algorithms reconstructing what would have sold without inventory constraints, then feeding the corrected demand into the forecasting model rather than the constrained sales data. This is increasingly standard in modern AI platforms but remains absent from most legacy forecasting tools.

Key takeaway: New product stockouts compound forecast errors across future launches. Censored demand correction matters more for new products than for any other category, and it's one of the easiest wins for any retailer with frequent product introductions.

Probabilistic Forecasting for New Products

New product demand is inherently uncertain. The conventional response is to produce a single point forecast and hope it's close. The better response is to produce a probability distribution and make decisions accordingly.

A probabilistic forecast might say something like "there's a 50% chance this product sells between 800 and 1,200 units in the first month, with a long tail upside that could reach 2,500 if it goes viral." The distribution itself is the actionable output, not just the median value. Inventory decisions can then balance the cost of over-ordering against the cost of stocking out, with both probabilities explicit.

This is particularly valuable for products with skewed adoption distributions, where most launches are modest but a small fraction become breakout hits. A point forecast averages these out and gets every launch wrong. A probabilistic forecast captures the bimodal distribution honestly and lets planners hedge.

Key takeaway: New product forecasts should be distributions, not point estimates. The uncertainty is the most important information for inventory decisions, and a forecast that hides it is misleading even when its central value is correct.

How OnePint.ai Handles New Product Forecasting

New product forecasting is one of the highest-stakes capabilities in any AI planning platform, particularly for retailers with frequent product launches. OnePint.ai handles the cold-start problem through multiple integrated approaches. Pint Planning applies attribute-based clustering across the existing SKU catalogue, generates supersession forecasts when products directly replace predecessors, and produces probabilistic distributions rather than point forecasts. The platform handles censored demand correction natively, so early stockouts don't compound errors across future launches. As real sales data arrives, the forecast transitions smoothly through the four lifecycle phases without abrupt shifts.

Customers using the platform see 20 to 30% better forecast accuracy, up to 85% fewer stockouts, and 10 to 20% lower fulfilment costs. The accuracy gains are typically largest on new product launches because the legacy tools most retailers replace handle cold-start either poorly or not at all. OnePint.ai was also recognised as a 2025 Gartner Cool Vendor in Supply Chain Planning Technology.

Frequently Asked Questions

What is the cold-start problem in demand forecasting?

The cold-start problem is the challenge of forecasting demand for a new product with no historical sales data. Traditional time-series models can't be applied because they need historical observations to train. The standard solution is attribute-based modelling, which infers the new product's demand pattern from comparable existing products grouped by features like category, price tier, and size.

How accurate are new product demand forecasts?

Significantly less accurate than mature product forecasts, and the goal isn't perfect prediction. Industry benchmarks for new product forecasting target 40 to 55% accuracy (45 to 60% WAPE) compared to 75 to 85% for established categories. The job of cold-start forecasting is confident decision-making under uncertainty, with proper probability distributions and confidence intervals rather than misleading point estimates.

How long does it take for a new product forecast to stabilise?

Typically 8 to 12 weeks of post-launch sales data is enough for forecasts to transition from attribute-based to standard time-series methods. Some platforms can adapt with as little as seven days of data using meta-learning approaches. The exact timeline depends on category volatility, sales velocity, and how well the new product fits its attribute cluster.

What's the difference between supersession and attribute-based forecasting?

Supersession applies when a new product directly replaces an existing one. The forecast inherits most of the predecessor's demand pattern. Attribute-based forecasting groups the new product with similar SKUs by features and borrows patterns from the cluster. Supersession is more accurate when applicable but only works for direct replacements. Attribute-based is more general but slightly less precise.

Why do new products often experience stockouts at launch?

Because demand uncertainty makes retailers conservative on the first order. The result is censored demand data when stockouts occur, which then under-trains the forecasting model for the next launch in the same category. Modern platforms correct for this using algorithms like Expectation-Maximization that reconstruct true demand from constrained sales, breaking the compounding error pattern.