In today’s dynamic and uncertain economic environment, accurately forecasting product demand presents significant challenges for retail businesses. Overstocking generates excessive costs that burden operational overheads, while under-stocking compromises profitability and customer satisfaction. Traditional statistical methods provide interpretability for demand forecasting, yet suffer from linearity assumptions, stationarity requirements, and an inability to capture long-term dependencies in complex datasets. Machine learning emerged as the first alternative approach to address these constraints. However, the revolutionary breakthrough came with deep learning technologies that fundamentally eliminated the core limitations of statistical methods. This study investigates and compares four hybrid deep learning models for short-term item demand forecasting, with an emphasis on their practical applications and operational implications in retail environments. These models create a powerful hybrid architecture that leverages complementary strengths to achieve superior forecasting performance. They combine an automatic feature extraction from time series data with sequential memory capabilities that model long-term dependencies, maintain context from previous time steps, and capture the inherent temporal nature of demand data. These hybrid models are comprehensively evaluated and compared using multiple performance criteria to identify the most effective approach for providing accurate and robust demand forecasts.