Retailers are turning to machine learning to predict demand before a new product has a sales history. A practical case study shows how models can move beyond weak baselines, using product similarity, marketing signals, and transfer learning to cut weighted absolute percentage error from above 70% to 36%.
Why New Product Demand Is So Difficult
Predicting a familiar product is relatively comfortable territory for retailers. Last week's sales, seasonal patterns, regional demand, and promotional history all offer useful clues. A brand-new item has none of that, leaving the forecasting model to work with a blank page.
That missing history is the heart of the cold-start problem. A new product may sell slowly, surge after a campaign, or outperform its established alternative almost immediately. Industry research on new-product forecasting also points to the importance of combining product information with historical demand patterns rather than treating the launch as an isolated event.
The practical consequence is significant. Forecast too low, and shelves empty during the launch window. Forecast too high, and the business is left with excess stock, markdowns, and a rather awkward post-launch cleanup.
Baselines Offer a Starting Point, Not a Final Answer
A straightforward approach estimates units sold per customer, then combines that figure with a forecast of customer traffic. It sounds sensible: one input reflects product popularity, while the other captures demand rhythms such as weekday or regional variation.
In testing, though, that baseline delivered a WAPE above 70%. The weakness isn't surprising. A manually estimated popularity coefficient can vary across locations and time, while multiplying two separate forecasts compounds their errors.
A machine-learning version improved the result to roughly 45% to 55% WAPE by estimating weekly product performance from category, product attributes, calendar variables, and a new-item flag. That's a useful step, but fast-growing launch sales can still slip through the net because the method compresses a changing first week into a single average-like measure.
Product Similarity Becomes the Missing Sales History
When lag features aren't available, the model needs substitutes. Product similarity is one of the most useful. A new flavour, package size, or variation may share enough characteristics with an established item to provide a reasonable demand starting point.
Marketing and popularity signals add another layer. Promotions, launch timing, store coverage, seasonality, and customer activity can explain why an unfamiliar product might suddenly move far faster than its nearest equivalent. This is where the forecast becomes less like traditional time-series extrapolation and more like structured comparison.
Research into multimodal new-product forecasting follows a similar direction, combining sales data with richer product information. In plain English, the model shouldn't just ask, "What did this item sell last week?" It should also ask, "What is this item, and what other products behave like it?"
Transfer Learning Gives Retail Models a Head Start
The strongest result came from transfer learning. First, a gradient-boosting model learned broad retail patterns from the much larger pool of established-product sales. It absorbed clues about seasonality, regional differences, product groups, and typical sales distributions.
The model was then fine-tuned on the smaller set of new-product launches. That second stage allowed it to adjust its expectations for the very different behaviour of recently introduced items. The reported WAPE fell from about 52% after the first stage to 36% after fine-tuning.
This approach is appealing because it balances accuracy with practicality. It doesn't require a giant synthetic dataset or an entirely separate deep-learning system. Tools such as CatBoost can also handle many categorical product attributes, which is handy when retail data is full of departments, brands, formats, and specifications.
How to Choose a Cold-Start Forecasting Strategy
For a small catalogue, a carefully built similarity-based baseline may be enough. Start with comparable products, launch timing, promotion plans, store or region information, and customer traffic. Keep the method transparent so planners can challenge the assumptions rather than treating the forecast as magic.
For larger businesses, transfer learning is a sensible next step. Train on established products first, then fine-tune with historical launches. Crucially, the two datasets need compatible features, and evaluation should mimic the real launch process by retraining before each product release.
More advanced methods, including temporal fusion transformer models, can be useful when the retailer has abundant data and wants to combine static product details, time-varying signals, and external drivers. But complexity isn't automatically better. A smaller gradient-boosting model that's quick to retrain may be more valuable than an elaborate system no one can maintain.
The Forecast Should Improve After Launch
The first forecast is only the opening move. Once a product begins selling, real demand signals should feed back into the system, allowing the business to update inventory, replenishment, and marketing decisions quickly.
That matters especially when launches are promotional or seasonal. A product that starts quietly may need a different plan from one that sells out in its first few days. The best forecasting setup therefore treats cold-start prediction as a temporary bridge between product launch and dependable sales history.
For retailers, the broader lesson is refreshingly practical: don't force new products into a model designed for mature ones. Give them their own logic, use similar products intelligently, and let transfer learning provide the head start.
It's a small modelling shift that can make a major difference to launch decisions.
Disclaimer: This article may have been created with AI assistance and reviewed by our editorial team. It is provided for general informational purposes only. Readers should verify information independently before relying on this content.
