Cold-Start Sales Forecasting in Fashion via Feature-Based Ensembles

The fashion retail sector is characterized by high volatility due to rapidly shifting consumer preferences, short product life cycles, and unstable market conditions. These dynamics significantly complicate demand forecasting, particularly for newly designed products lacking historical sales data. This challenge, known as the cold-start problem, poses both operational and strategic risks, often leading to excess inventory or missed revenue opportunities. In response, this study introduces a scalable and interpretable machine learning framework that predicts sales potential using only structured product design attributes - such as fabric type, fit, neckline, and sleeve style - without relying on visual, behavioral, or time-series data. Developed in collaboration with a leading fashion retailer, the proposed approach supports early-stage product evaluation, well before manufacturing decisions are made. The modeling strategy is centered around a stacking ensemble architecture, incorporating XGBoost, CatBoost, and Random Forest as base learners. These ensemble configurations demonstrate robust predictive performance across segmented demand classes (low, medium, and high), which are derived through histogram-based normalization of per-store sales ratios. To address the limitations of observed data, a Bayesian network-based synthetic data generation module is employed, enabling the simulation of realistic, logic-driven product combinations. Candidate designs predicted to have high sales potential are filtered through a confidence thresholding mechanism and subsequently validated by domain experts to ensure operational feasibility. Model explainability is enhanced through interpretable feature importance analysis, providing data-driven guidance for design teams and fostering transparent decision-making.