Retail Sales Forecasting using Advanced Deep Learning Models: A Comparative Study

In real-world decision-making processes, anticipating customer demand allows businesses to minimize stock shortages, reduce excess inventory, and improve overall operational efficiency, ultimately leading to better customer satisfaction and increased profitability. In our study, we explore the potential of advanced machine learning and deep learning techniques to enhance time series sales forecasting. Using a rich dataset of several years of historical sales records, we develop and compare four models: a LightGBM, a bidirectional LSTM network, a Transformer-based attention model, and the Informer model, specifically designed for long-sequence forecasting. While traditional models such as LightGBM rely on carefully engineered features such as time-lag patterns and rolling statistics, deep learning models automatically learn complex, time-dependent relationships. For example, LSTM networks encode a given time window into a hidden representation, whereas Transformer and Informer architectures use attention mechanisms to dynamically focus on the most relevant past information. Using Bayesian optimization with Optuna for hyperparameter tuning, we achieve robust performance improvements. Our findings demonstrate that integrating external influences with deep learning not only increases accuracy but also provides a scalable framework for smarter resource allocation and strategic decision-making in retail.