Cold-Start Demand Forecasting via Text Embeddings and Cross-Product Demand Transfer

Cold-start demand forecasting is the problem of predicting demand for a product with little or no sales history. Forecasting models like MQCNN and MQT (of the MQ-Forecaster family) struggle in this regime because much of their predictive signal comes from historical demand, forcing them to rely heavily on static product attributes for newly launched products. To address this limitation, we introduce a neural forecasting model built on the MQ-Forecaster of multi-horizon quantile forecasting architectures that incorporates three complementary cold-start signals: (1) text embeddings, encoding catalog metadata (title, description) with a large language model into dense embeddings that capture semantics categorical features miss; (2) age-aligned neighbor demand, retrieving the most similar established products by embedding similarity and fusing their lifecycle-aligned demand curve and (3) cross-product aggregate demand, pooling demand over products that share categorical attributes to give a broad category-level prior. On a year-long backtest over 30 million new product launches, our model improves P90 quantile loss by 12--24\% across all cold-start stages versus a production baseline, without regressing established products.