By Eric Wilson & Somi Agarwal CEO / Co-founder at Synthefy
New products are expected to drive a significant share of your company’s growth. But how do you forecast demand for your product with little or no sales history to work from?
New product forecasting is a challenge for most companies since there is little to no historical demand data. IBF research shows that new-product forecasting continues to be a significant challenge, even for line extensions where the brand, market, and customer base are already known.
Most systems use methods designed for established products that do not always translate easily. Yet these forecasts influence expensive decisions: production capacity, inventory, supplier commitments, marketing investment, launch timing, and revenue expectations. Despite the importance and investment, many organizations continue to accept high forecast error as simply the cost of launching something new.
Part of the issue is that we are still planning new products the way we have always done it. The most common approach remains subjective judgment. About 43% of companies rely primarily on educated guesses when forecasting items with little or no history. That judgment may come from sales, marketing, product development, or an executive with strong confidence in the launch. The result is often less of a demand forecast and more of a sales target, business case, or wishcast.
Another 17% attempt to apply statistical methods through supersession or like-item forecasting. The new item is matched to a single similar product, and its history becomes the starting point. This can be useful, but it assumes you selected the right product, it will be sold at the same price, and that the market will respond in roughly the same way. Those are significant assumptions.
Why is Forecasting a New Product with no sales history so difficult?
New product forecasting creates particular problems for demand planners. For most, their work is built around using univariate models and sales history to identify patterns and predict what comes next. With a new product, that foundation disappears. The question is no longer simply which model to use. It is which data, if any, should be trusted.
Planners often start by searching for an analogous product. They compare attributes, price points, channels, customers, and launch timing, then borrow the demand curve from the item that appears most similar. But no two launches are identical. Other factors and data may influence the launch such as promotional support, competition, or market conditions. The selection itself can become another subjective judgment hidden inside what looks like an analytical forecast.
The challenge with new-product forecasting isn’t only finding that the product has no history. Traditional machine learning approaches like XGBoost also require you to train and tune a model on the data you do have, adding additional data-science work.
Considerable time is spent debating the analog, adjusting the curve, gathering opinions, training models, and reconciling competing assumptions. Ironically, many of these products eventually become relatively small contributors to the overall portfolio.
A planner can spend a disproportionate amount of time forecasting an item that may never generate enough demand to justify the effort.
When the forecast is too high, the company carries excess inventory, consumes capacity, and eventually discounts or writes off the product. When it is too low, the launch experiences shortages, missed sales, frustrated customers, and pressure to expedite supply. Either way, the cost extends beyond forecast accuracy.
The same uncertainty follows the forecast into the S&OP meeting. Sales defend the market opportunity. Marketing points to launch investment. Supply highlights inventory exposure. Finance questions the expected return. The planner is asked to explain the number but may have little more than an analog curve and a collection of assumptions.
That lack of explainability matters. If decision-makers cannot understand what drives the forecast, they are less likely to trust it. The meeting then shifts from evaluating risk and making decisions to negotiating whose opinion should carry the most weight.
How Can You Forecast a New Prodict with No Sales History?
Structured data Foundation Models offer another approach. These models are pretrained in advance to make predictions from structured, tabular data. Instead of training and tuning a new model for each prediction problem, they can use examples from a customer’s existing data as context for a new prediction.
Nori is Synthefy’s structured-data foundation model. For new-product forecasting, a new SKU may have no sales history of its own, but the rest of the product portfolio does. Nori can use examples from across that existing data as context when making a prediction, without training and tuning a new customer-specific mode
Synthefy tested this using the public Online Retail II dataset. Both Nori and XGBoost were given the same underlying data and features. In this test, Nori reduced forecast error by roughly 10% relative to XGBoost on brand-new products with zero history, using the same underlying data and features. The potential shift is bigger than accuracy alone. If every new prediction problem doesn’t require starting another bespoke machine-learning project, more forecasting problems may become practical to tackle. That is still something we need to prove. But it raises a different question: instead of asking “Is this forecast worth building a model for?” teams can begin asking “What predictions could we make if building the model were no longer the bottleneck?”
What to do Monday
Technology does not remove the need for a strong collaborative process. Sales, marketing, product development, finance, supply, and demand planning must still align on launch timing, distribution, assumptions, risks, and execution. Better models will not rescue poor inputs or a launch plan that changes without anyone telling the planner. But those decisions should be supported by timely insights rather than an analog selected because it was the least-wrong option available.
So, what should you do before your next new product launch. Begin by examining how the forecast is created today. Who selects the like item? What relevant data is being ignored? How much time is spent manually adjusting curves? Can anyone clearly explain why the final number changed? The goal should be to create a process that uses more of the information already available while reducing complexity, manual effort, and unsupported judgment.
- Still keep it simple, but not simplistic. A process that is too basic may overlook valuable signals, but greater complexity does not automatically create greater value. Spending days selecting attributes, preparing data, or training models may simply shift more work to the planner. Ideally, the model should do more of the analytical work and make new product forecasting easier.
- Move beyond the single-analog approach. A new product may not have its own sales history, but that does not mean the business lacks relevant data. Product records, previous launches, pricing, promotions, customers, channels, and market conditions may all contain useful signals. The objective should be to learn from patterns across all the information you have rather than force the new product to match one manually selected item.
- The final forecast is still your responsibility through collaboration. A model is a decision-support tool, not a replacement for collaboration or judgment. The objective remains an unbiased, unconstrained projection of demand, with the uncertainty clearly understood. If meetings are spent debating which number, the team may miss the more valuable discussion: what is driving the forecast, what could change, and which strategies should be developed in response.
This is where structured foundation models such as Synthefy’s Nori are changing the approach. They can evaluate patterns across a much broader set of tabular business data without requiring months to build and tune a model for each problem. They can also show which signals such as price, channel, customer segment, timing, and promotional support contributed most to the forecast. This gives planners something they can explain and decision-makers something they can challenge.
It is time to stop treating a manually selected analog and an adjusted demand curve as the only or best evidence available. It’s time your system and models work for you and provide insights and save time.
The objective is not to remove the planner from new product forecasting. It is to remove more of the guesswork, give the planner better evidence, and return time for what planners should be doing: evaluating risk, improving assumptions, and helping the business make better decisions.
Thank you Somi Agarwal CEO / Co-founder at Synthefy for collaboration and insights on this article.