<![CDATA[Sales forecasting plays a crucial role in retail decision-making, particularly in multi-product environments with heterogeneous demand characteristics. Unlike most previous studies that focus on finding a single best model globally, this study conducts a standardized comparative evaluation between statistical (TES, SARIMA) and deep learning (LSTM, CNN–LSTM) models using six years of real retail sales data covering eight products with stable seasonal patterns to volatile and nonlinear demand, thus representing the complexity of retail scenarios. To ensure the validity and reliability of the evaluation, data leakage was prevented by splitting the data seasonally (80% training, 20% testing), applying all transformations (scaling, log transform, sequence generation) estimated only on the training data, and using a consistent multi-step autoregressive scheme across all models. Model performance was evaluated using MAPE, MAE, MSE, and RMSE. The adaptive ensemble strategy was implemented by combining SARIMA and CNN–LSTM through optimization of product-specific weights (α ∈ [0,1], interval 0.01) using grid search to minimize MAPE over the testing horizon. The results show that the deep learning model outperforms products with fluctuating and nonlinear demand, while the statistical model remains competitive on products with stable seasonal patterns. Aggregated across products, the adaptive ensemble yielded the lowest average MAPE (0.2184), lower than all individual models, indicating better stability in the face of demand heterogeneity. This study confirms the product-dependent nature of forecasting performance and offers a replicable adaptive framework for multi-product forecasting. By enhancing data-driven supply chain efficiency and supporting skill development in advanced forecasting methodologies, this research aligns with Sustainable Development Goals (SDG) 8 and SDG 4.]]>
