Neural Product Embeddings for (New) Product Sales Prediction

Marketers extensively use neural product embeddings to capture product attributes. Our research shows the predictive power of product embeddings in forecasting product sales success beyond the captured attributes. Across a unique dataset comprising over 100 online and offline retailers with more than 6 million products and 1 billion baskets, such sales predictions are accurate for various product categories and retail types. A deeper analysis of sales prediction at a brick-and-mortar grocery retailer shows that product embeddings enhance predictive accuracy for product sales by 32%, surpassing the information captured by traditional product attributes. To predict sales of \emph{new} products before their launch, we first show how we can derive product embeddings before a product’s introduction, and then use the embedding to predict future sales of the new products. We assess under which conditions product embeddings yield reliable sales predictions and establish the boundary conditions for the proposed approach. Intuitively, the quality of product embeddings is crucial for sales predictions and is furthermore related to industry, basket size and assortment structure.

Working Paper