Comparative Time-Series Forecasting of Amoxicillin Demand Using ARIMA Variants and LSTM

Accurate short‑term medicine forecasts are critical to prevent shortages and plan production. This study compares statistical and machine learning models for predicting Iranian demand for the amoxicillin group, using Iranian Food and Drug Administration data from ۱۴۰۰–۱۴۰۳ aggregated into ۲۴ half‑month periods per year. Models tested were univariate SARIMA, Holt‑Winters, temperature‑augmented SARIMAX, and multivariate LSTM with national average minimum temperatures (°F) from Weather Spark. A ۲۴‑period hold‑out set was assessed with MAPE, RMSE, MAE, and R&sup۲;, with Diebold–Mariano tests. Holt‑Winters (MAPE ۷.۲۶ %) and LSTM (۶.۴۴ %) outperformed the Seasonal Naïve benchmark; SARIMAX underperformed, while LSTM with temperature achieved the best accuracy, indicating nonlinear models exploit climatic signals effectively. Results guide model choice by balancing accuracy, interpretability, and operational constraints in pharmaceutical supply chains.

Keywords:

Amoxicillin demand forecasting , Seasonal ARIMA , SARIMAX , Holt‑Winters , Long Short‑Term Memory (LSTM) , exogenous temperature , pharmaceutical supply chain , Iran , time‑series analysis.