New product forecasting without sales history

By Elie Dufeu, CTO & Co-Founder, Metreecs. Published 8 October 2026.

New product forecasting without sales history works in four steps: borrow demand from products that have already sold, shape it with a launch curve, size the first buy as a range, and correct the forecast every week once real sales arrive. Each step replaces a guess with something you can check against data.

Priya buys for a home goods retailer with 40 stores. In March she approved 1,200 units of a new linen cushion cover, forecast as "about the same as last year’s cushions" and split evenly across stores. Six weeks in, 12 stores had sold out and 15 stores had sold fewer than a third of their allocation. The total was close to plan, yet the plan had said nothing about which stores would sell the cushion. (Priya is an illustrative example with round numbers, but most buyers have lived a version of it.)

The difficulty is plain: a statistical model needs history, and a new product has none. This guide covers the four methods that work, how to combine them into one repeatable process, and where new product forecasts tend to break. If you want to see this process on your own launches, you can test launch forecasts on your next collection.

Key Takeaways

• The workable sequence has four steps: build an analog set from product attributes, apply a launch curve, size the first buy as a range, and re-forecast weekly from early sales.
• A 2025 Winter Simulation Conference paper by Ruihan Zhou reports that combining product attributes, reference trajectories, and early sales cut pre-launch mean absolute error by more than 15% and early post-launch error by more than 20%, against the strongest benchmark tested.
• Skenderi and colleagues built a public dataset of 5,577 new fashion products and found that adding Google Trends signals improved WAPE by 1.5%, so external signals refine an analog-based forecast and do not replace it.
• In the worked example below, an 8-week forecast of 240 units becomes about 360 units after two weeks of sales running at 1.5 times plan.
• With €18 of margin lost per missed sale and €10 lost per leftover unit, the buy should sit at roughly the 64th percentile of your demand range, not at the median.

What is new product forecasting without sales history?

New product forecasting without sales history is the practice of estimating demand for a product or variant that has never sold, using indirect evidence: the sales of similar products, the product’s own attributes, its launch plan, and any early sales that arrive after launch. It replaces a time-series model, which needs history, with comparison and correction.

The same problem appears in several forms. A brand-new product, a new variant such as a color or size, a replacement for a discontinued item, and a product launched in a new location all start with little or no direct history. The first buy for each is the highest-risk decision in the cycle, because the order is placed before the market has said anything.

Our new product introduction guide goes deeper on the launch process. This article focuses on the forecast itself.

Which new product forecasting method fits which launch?

No single method covers every launch, and most teams end up combining two or three. The table compares the four that matter in retail.

Research points the same way. Ruihan Zhou’s paper on cold-start forecasting of product life cycles, presented at the 2025 Winter Simulation Conference, combines product attributes, reference trajectories from similar products, and early sales in one model. On a microprocessor dataset, it reduced pre-launch error by more than 15% against the strongest benchmark. Microprocessors are not retail, but the structure of the problem is the same.

AI-powered demand planning applies this combination at product and location level, so a launch gets a forecast for each variant in each store rather than one number for the line.

Build the analog set and the launch curve

Start with three to five analogs, meaning products that have already sold and resemble the new one. Rank candidates by shared attributes: category, price band, style or material, season, channel, and launch support such as in-store placement or marketing. Weight the closest match most heavily.

Then make the histories comparable. Convert each analog to units per store per week over its first eight weeks of availability, remove weeks with stockouts or heavy promotion, and keep the shape of the curve separate from its level.

Here is a worked example for Priya’s cushion cover, launching in 40 stores. Three analogs sold 0.6, 0.75, and 0.9 units per store per week, and the middle one is the closest match, so the weights are 25%, 50%, and 25%. That gives a weighted rate of 0.75 units per store per week, or 240 units over eight weeks (40 stores × 8 weeks × 0.75).

The curve tells you when those 240 units sell. Averaging the analogs gives the weekly share below.

Across Metreecs’ work with home décor, fashion, and jewelry retailers, the choice of analogs moves the first buy more than the choice of statistical model does. One mistake we repeatedly see is picking analogs by category and price alone. Style, finish, and placement usually drive the launch shape, and they sit in attributes that were never cleaned up in the product catalog.

Lena, a planner at a footwear retailer, learned this the hard way. She forecast a chunky leather boot from a canvas sneaker in the same price band, and the boot sold 40% below plan, leaving her with excess stock by mid-season. When she rebuilt the analog set around style and material, her next boot launch landed within 12% of actual sales. (Lena is an illustrative example with round numbers.)

Size the first buy as a range, not a point

A range shows how uncertain a new product forecast is, which a single number hides. The spread among your analogs is a floor on that uncertainty, because it ignores any mismatch between the analogs and the new product.

In the cushion example, the analogs imply 192 to 288 units over eight weeks. Widening that for mismatch gives a low case of 170, a median of 240, and a high case of 340. The widening is a judgment call, so write down why you chose it.

The buy should reflect what each kind of error costs. This is the classic newsvendor logic. If a missed sale costs €18 of margin and a leftover unit costs €10 in markdown and holding cost, the ratio is 18 / (18 + 10), or about 64%. You buy at the 64th percentile of demand, not the 50th.

If you treat the high case as roughly the 90th percentile and interpolate between it and the median, that points to about 275 units. Where the supplier allows a reorder inside the selling window, buy closer to the median and keep the option open.

How you allocate matters as much as how much you buy. A common pattern is to send part of the buy to stores in proportion to each store’s analog sales rate and hold the rest centrally for a second wave. Our article on how to allocate stock to maximize full-price sales covers that step, and inventory optimization software applies the cost logic across a whole assortment at once.

Correct the forecast from early sales

The first sales are the best new information you will get, so build a weekly early-read routine. Compare actual sales to the launch-curve forecast for each week, calculate the ratio, and apply it to the remaining weeks.

Back to the cushion cover. Week 1 sold 72 units against a forecast of 48, and week 2 sold 65 against 43. Together that is 137 units against 91, a ratio of about 1.5. Applying it to the full eight weeks gives roughly 360 units, which leaves about 223 units still to sell in the final six weeks, against 149 in the original plan.

Tom plans for a beauty brand with 25 stores and an online shop, and he ran this routine on a new serum launch. His week-2 read showed sales at about 1.5 times plan, so he pulled a reorder forward by three weeks and avoided roughly ten days of stockout on the best-selling shade. (Tom is an illustrative example with round numbers.)

Apply the ratio with some care. Early buyers can be unrepresentative, so many teams blend the early-read ratio with the original forecast and shift the weight toward actual sales as the weeks pass. Reading at product level also matters, because one hot variant can hide three slow ones in a line-level total.

Skenderi and colleagues tested what outside data adds. Their dataset, VISUELLE, holds 5,577 new products sold by an Italian fast-fashion company, with images, attributes, and Google Trends series. Adding the trend signals improved WAPE by 1.5%, which is useful but small next to the effect of choosing good analogs and reading early sales.

Where new product forecasts break

New product forecasts miss for structural reasons, and the causes are rarely about effort. Four show up most often.

• Thin attributes. If the catalog records only category and price, the analog search cannot tell a statement piece from a staple.
• Biased analog history. Analogs that sold out early understate true demand unless you correct for stockout weeks. Analogs that were heavily promoted overstate it.
• Launch conditions that differ. The planners we work with often underestimate how much a launch curve depends on store count, placement, and marketing support. An analog launched with a front-window display will not predict a launch placed on a back wall.
• Slow reaction. A forecast set at launch and reviewed at season end cannot use the best signal available, which is the first two weeks of sales.

The cost of missing is real on both sides. Too little stock loses sales on the winners, and too much creates markdowns on the misses. Our piece on the cost of inaccurate sales forecasts breaks down how that cost adds up, and the same logic applies to launches.

In our experience deploying product-level forecasting, teams get the most from a simple discipline: record which analogs you chose and why, then review them after eight weeks. That review is what makes the next launch forecast better than the last one.

FAQ

How do you forecast demand for a product with no sales history?

Use the sales of similar products as a proxy. Pick three to five analogs based on shared attributes, convert their first weeks of sales to units per store per week, and apply that rate and curve to the new product. Then replace the forecast with actual sales as they arrive.

What is the most accurate method for new product forecasting?

No single method wins in every case. Research on cold-start forecasting finds the best results when product attributes, reference trajectories from similar products, and early sales are combined in one model. For most retailers, that means analogs for the starting point and a weekly early-read correction after launch.

How do you choose an analog product?

Match on the attributes that drive demand, not only on category and price. Style, material or finish, season, channel, and launch support all matter. Remove stockout and promotion weeks from the analog’s history so it reflects normal demand, and keep notes on why you picked it.

How much stock should you buy for a new product?

Buy to a range and weigh the costs of error. If a missed sale costs more than a leftover unit, buy above the median demand estimate. If the supplier allows reorders, keep the first buy near the median and hold part of it centrally for a second wave.

How quickly can you trust early sales?

Two weeks is enough to detect a large gap, such as sales at 1.5 times plan. Treat the first reads cautiously, because early buyers can be unrepresentative. Blend the early ratio with the original forecast and shift the weight toward actuals each week.

Conclusion

New product forecasting without sales history comes down to a short, repeatable process: choose analogs from attributes, apply a launch curve, buy to a range, and re-forecast weekly. To see how this works on your own launches, book a demo with our team.

Sources

• Ruihan Zhou, "Cold-Start Forecasting of New Product Life-Cycles via Conditional Diffusion Models," Winter Simulation Conference 2025: paper. Supports the pre-launch and early post-launch error reductions.
• Skenderi, Joppi, Denitto, and Cristani, "Well Googled is Half Done: Multimodal Forecasting of New Fashion Product Sales with Image-based Google Trends": arXiv. Supports the VISUELLE dataset size and the 1.5% WAPE improvement.