Leveraging Multiple Models to Enhance Sales Demand Forecasting in Brazilian E-commerce

Abstract

In e-commerce, customer preferences and market trends change rapidly, making sales forecasting crucial for business performance. Accurate forecasting involves maintaining optimal inventory levels, optimizing processes, and enhancing customer satisfaction. We leverage machine learning, particularly deep learning techniques, in addressing complex patterns found in voluminous datasets, like the Olist dataset analyzed in our work. This capability is especially valuable in dynamic markets like Brazil, where adaptability is essential. While existing research often neglects the behavioral information incorporated in transactional data, our study proposes a novel approach that integrates advanced feature engineering techniques. We extract customer behavior insights through recency, frequency, and monetary (RFM) analysis, and account for the nuanced impact of seasonality at the product category level. We also consider the impact of national holidays, which often lead to noticeable shifts in consumer behavior. Our research evaluates a range of deep learning models, including sequence-based architectures (LSTM, GRU, TCN), hybrid models (CNN-LSTM, CNN-GRU), and attention-enhanced networks (LSTM-Attention), to assess which approaches demonstrate the most significant predictive accuracy. Our evaluation identifies the LSTM-attention architecture as the top performer, achieving an MSE of 0.0045 and RMSE of 0.0067.