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 multi-product daily sales forecasting in pharmacy chain sales systems. Using real transactional data from a pharmacy chain, the study constructs a forecasting framework based on the top 20 best-selling products. Historical sales records from 2025, comprising 327,726 transactions, are divided into training and validation sets, while sales data from January 2026, comprising 25,284 transactions, are reserved as an unseen test set for multi-horizon evaluation. To ensure a fair comparison, TFT is benchmarked against Linear Regression, Random Forest, XGBoost, and standard LSTM models. The experimental design incorporates daily aggregation, temporal feature derivation, rolling-mean smoothing, and sequence construction to model heterogeneous retail information, including static attributes and time-varying inputs. Forecasting performance is evaluated using Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and the Coefficient of Determination (R2). Experimental results show that TFT achieves the best overall performance, with MAE = 104.165, RMSE = 149.717, and R2 = 0.986, outperforming all baseline models. These findings indicate that TFT provides a robust and interpretable forecasting framework for pharmacy retail automation, inventory planning, and replenishment decision support.