Enhanced Transformer-LSTM Daily Sales Forecasting Model Based on Attention Mechanism

Accurate sales forecasting remains a key challenge for retailers. Existing methods often struggle to integrate local and global contextual dependencies, limiting their predictive accuracy in complex scenarios. In response to these challenges in sales forecasting, we propose an attention-enhanced Transformer-LSTM model. Firstly, we combine the LSTM’s capability to capture time-related sequential data with the Transformer’s global context awareness; secondly, we prioritize the most influential features dynamically through an enhanced attention mechanism for prediction; thirdly, by adopting a hybrid architecture, we effectively address both short-term and long-term dependencies within sales data. Experimental results show that our model not only significantly outperforms traditional forecasting methods but also improves upon recent deep learning approaches in terms of Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R2) metrics.