In the fashion industry, demand forecasting for new season products is challenging due to data sparsity, lost sales, stock constraints, and in particular, the lack of historical records for new items. This paper presents a data-centric approach that systematically improves data quality and enriches training datasets through domain-specific label correction, similarity-based augmentation, and statistic-based feature extraction. We address these issues through a sell-through-based adjustment mechanism, which corrects observed sales during stock-out periods. The method emphasizes the early weeks of a product’s lifecycle when demand signals are most informative. We validated our proposed method through a case study conducted with a fast fashion company in Turkiye. Experimental results on real-world retail data from a fashion retailer demonstrate that these techniques significantly reduce forecasting errors. This work emphasizes data-centric approaches can outperform model-centric baselines in complex, sparse environments such as retail fashion forecasting. The experiments show that our proposed model reduced mean absolute error and root mean squared error compared to initial expert-based distributions. Mean bias deviation is decreased by over 19% for high-demand items, offering more reliable forecasts under stock-constrained scenarios.