FedSalesNet: A Federated Learning–Inspired Deep Neural Framework for Decentralized Multi-Store Sales Forecasting

Accurate forecasting of market trends and consumer behavior is essential for decision-making in modern retail systems. However, centralized learning approaches often face data privacy, regulatory, and scalability constraints in multi-store environments. In this paper, we propose FedSalesNet, a federated deep learning framework that enables decentralized, store-specific sales forecasting without sharing raw data. Each store trains a local hybrid deep model combining CNNs, LSTMs, and attention mechanisms, followed by secure aggregation to update a global model. Evaluated on two real-world retail datasets comprising over 50,000 sales records across seven store locations, FedSalesNet achieves superior forecasting accuracy compared to common centralized and non-federated baselines, with an MAE of 3.64, RMSE of 4.88, and MAPE of 14.47%. It outperforms centralized LSTM, ARIMA, XGBoost, and Transformer-based baselines by up to 29% in forecasting accuracy. The model converges in under 50 communication rounds and requires only 3.1 seconds per epoch per node, demonstrating strong scalability and efficiency. These results establish FedSalesNet as a viable solution for collaborative retail forecasting under strict data governance constraints.