Hybrid LSTM-CNN with Attention Mechanism for Accurate and Scalable Grocery Sales Forecasting

This paper presents a Hybrid CNN-LSTM with an Attention Mechanism model for accurate and scalable sales forecasting in grocery retail, utilizing the Favorita Grocery Sales Forecasting dataset. By integrating Convolutional Neural Networks (CNN) for capturing short-term patterns, Long Short-Term Memory (LSTM) networks for modeling long-term dependencies, and an attention mechanism to dynamically prioritize influential time steps, the proposed model addresses the complex dependencies inherent in retail time-series data. Experimental results show that the Hybrid LSTM-CNN with Attention outperforms conventional models, achieving a Mean Absolute Error (MAE) of 0.195, Root Mean Squared Error (RMSE) of 0.295, and Weighted Mean Absolute Percentage Error (wMAPE) of 6.50%, yielding an 11.36% improvement in MAE over a baseline CNN-LSTM model and surpassing state-of-the-art methods like the Temporal Fusion Transformer by 7.45% in accuracy. Further analysis across different prediction horizons and feature subsets validates the model’s robustness in handling various data interactions, such as promotions, oil prices, and holiday effects. This work demonstrates that hybrid deep learning models with attention mechanisms can enhance forecasting accuracy and scalability, providing significant benefits for inventory and demand planning in dynamic retail environments.