Accurate demand forecasting for new product introductions in Fast Moving Consumer Goods (FMCG) is constrained by the cold-start problem, where products with limited sales history behave differently from established Stock Keeping Units (SKUs). Traditional methods struggle with sparse, volatile demand patterns, while deep learning models theoretically benefit from transfer learning across related products. This thesis evaluates whether global neural forecasting models, specifically N-HiTS, can outperform simple baselines by leveraging patterns learned from mature (>50 weeks of sales history) products to forecast sparse, short-history SKUs (12 to 50 weeks of sales history). Using a decade of NielsenIQ retail data from Coca-Cola Europacific Partners (CCEP), covering beverage products across two major Dutch retailers, Albert Heijn and Jumbo, N-HiTS was compared against a Naïve persistence baseline and a Croston intermittent demand baseline using Mean Absolute Scaled Error (MASE) across a rolling origin evaluation protocol. Two N-HiTS variants were tested: one without covariates and one incorporating static product attributes such as brand and caloric content. Results show that N-HiTS performs consistently relative to both baselines across retailers, yet fails to achieve statistically significant improvements over the Naïve baseline at either retailer. The addition of static product covariates did not yield statistically significant improvements in forecast accuracy. The difference of distribution shift between training and test volume regimes emerges as the primary explanatory factor for performance divergence between retailers. These findings suggest that global neural models are a viable but conditional alternative in cold-start FMCG settings, where model adoption should be contingent on the volatility profile of the specific retail environment rather than applied as a universal forecasting solution.