A Comparative Study of Forecasting Models Using the Temporal Fusion Transformer in Pharmacy Chain Sales Systems

Abstract

This study investigates the application of the Temporal Fusion Transformer (TFT) model for sales forecasting in pharmacy chain systems. Accurate forecasting is essential for optimizing inventory management in retail automation. The TFT architecture uniquely integrates Long Short-Term Memory (LSTM) networks with interpretable multi-head attention mechanisms, enabling effective processing of heterogeneous data, including static metadata and time-varying inputs. Using a real-world dataset comprising 113,946 transaction records, this research rigorously compares TFT against Linear Regression, Random Forest, XGBoost, and standard LSTM models. Performance is evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination ( ). Experimental results demonstrate that TFT achieves superior overall performance (MAE: 0.960, : 0.948), outperforming traditional models in capturing complex temporal dependencies and sales volatility. The findings validate TFT as a robust solution for enhancing decision-making in retail supply chains.