Traditional demand forecasting methods struggle when applied to assortments dominated by new products. For apparel and footwear retailers, where newness can account for 40–60% of seasonal assortments, the lack of historical sales data introduces significant uncertainty into planning decisions.
As consumer expectations rise and planning cycles compress, retailers need more advanced ways to understand demand for products that have never been sold before. Embeddings — an AI technique that captures complex product relationships — are increasingly becoming a foundational capability for addressing this challenge.
The unique challenge of forecasting new products
Newness forecasting requires planners to estimate demand without relying on prior sales history. In some cases, teams may be forecasting a product that has sold in one location but not another. In others, the product is entirely new to the assortment.
Relying on intuition or broad analogs often leads to inconsistent outcomes. Overestimating demand results in excess inventory, markdown pressure, and margin erosion. Underestimating demand leads to missed sales opportunities and dissatisfied customers. As assortments become more dynamic and trend cycles accelerate, these risks grow more pronounced.
To manage newness effectively, retailers need forecasting approaches that move beyond simple historical comparisons and capture the more granular attributes that influence consumer behavior.
Understanding embeddings: Capturing product similarity at scale
Embeddings are mathematical representations that encode a product’s characteristics into numerical vectors. These vectors allow machine learning models to understand similarity and relationships between products in a way that traditional attribute matching cannot.
Text embeddings analyze product descriptions, specifications, and marketing copy to capture semantic meaning. For example, they recognize that two dresses may appeal to similar customers even if their descriptions use different terminology.
Image embeddings analyze product imagery to capture visual attributes such as silhouette, pattern, and design details. This allows models to identify similarities that may not be easily expressed in structured data, including style or aesthetic cues.
By combining text and image embeddings, retailers can build a richer, more nuanced understanding of how products relate to one another across an assortment.
How embeddings improve demand forecasting for newness
Embedding-based models allow forecasting systems to learn from relationships across products, rather than relying solely on past sales of the same item. When a new product is introduced, the model can infer expected demand by referencing patterns observed in similar products from previous seasons or other parts of the assortment.
These models can also distinguish between products that substitute for one another and those that complement each other — an important capability for understanding cannibalization and halo effects.
As new sales signals emerge in-season, embedding-driven forecasts can adjust quickly, propagating changes across related items. This enables planners to respond proactively to emerging trends rather than reacting after performance has already diverged from plan.
Merchandising use cases enabled by embeddings
Embedding-based forecasting supports several critical merchandising workflows:
Assortment strategy
By revealing relationships between products, embeddings help teams understand assortment balance, identify gaps, and reduce redundancy. This improves category mix decisions and supports margin-focused assortment planning.
Pre-season planning
For products with no sales history, embeddings help identify truly comparable items from prior seasons, improving initial buy quantities and allocation decisions.
In-season management
As trends shift or early performance data becomes available, forecasts can be updated across related products, supporting faster and more confident inventory adjustments.
Considerations for operationalizing embeddings
While embeddings offer clear advantages, implementing them effectively requires careful consideration. Not all forecasting models are designed to work with embedding data, and advanced architectures are often needed to capture their full value.
Organizations must also address data quality, computational requirements, and model governance to ensure outputs are reliable and interpretable. Embeddings are most effective when integrated into a broader planning framework that connects demand signals with financial, supply, and inventory decisions.
Looking ahead: Embeddings as part of connected planning
As retail planning continues to evolve, embeddings are becoming a key component of more intelligent, demand-driven forecasting approaches. By capturing the subtle relationships that influence consumer behavior, they help organizations move beyond guesswork and toward more resilient planning for newness.
For retailers seeking to improve forecast accuracy, reduce risk, and better align merchandising decisions with financial outcomes, embedding-based approaches represent an important step forward.
