Ghadeer Jasim Mohammed Mahdi, Zahraa Ibrahim Al-Share

ARIMA-NN Model for Drugs Sales Forecasting in the United States

Introduction

Arima-nn model for drugs sales forecasting in the united states. Discover the ARIMA-NN model, combining ARIMA with Artificial Neural Networks for enhanced drug sales forecasting in the United States. Evaluated using U.S. Census Bureau data.

Abstract

This study proposes a new version of the Autoregressive Integrated Moving Average (ARIMA) model using Artificial Neural Networks (ANNs) denoted by ARIMA-NN. The new model incorporates a multi-layer perceptron with matrix multiplication within a feed-forward network. The logistic, hyperbolic tangent (tanh), and sigmoid activation functions are used for weight updates in ARIMA-NN. A new forecasting algorithm is proposed, and one-step and multiple-steps forecasting procedures are rigorously analyzed. The proposed model was evaluated against existing forecasting model using performance metrics such as the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) to assess its effectiveness. The U.S. Census Bureau (www.census.gov) provides a dataset of monthly drug sales spanning ten years (2014-2024), which is utilized in the study. The ARIMA-NN model is applied to generate forecasts for drug sales in the U.S. for the next four years to demonstrate the models' utility and efficacy. All the computations and visualizations are performed using various R packages in version 4.3.2.

Review

This study introduces a promising approach to drug sales forecasting in the United States through a novel hybrid model termed ARIMA-NN, combining the Autoregressive Integrated Moving Average (ARIMA) with Artificial Neural Networks (ANNs). The importance of accurate drug sales predictions for healthcare planning, logistics, and resource allocation cannot be overstated, making the proposed research highly relevant and potentially impactful. The abstract outlines an ambitious endeavor to enhance existing forecasting capabilities by integrating a sophisticated neural network architecture within a traditional time-series framework. Methodologically, the paper proposes a specific implementation of the ARIMA-NN model, incorporating a multi-layer perceptron (MLP) with matrix multiplication in a feed-forward network. The authors specify the use of logistic, hyperbolic tangent (tanh), and sigmoid activation functions for weight updates, suggesting a deliberate design choice for the neural network component. A "new forecasting algorithm" is highlighted, along with a rigorous analysis of both one-step and multiple-steps forecasting procedures, which is critical for practical applications. The model's evaluation is based on a substantial ten-year dataset (2014-2024) of monthly drug sales from the U.S. Census Bureau, a credible source. The performance assessment will utilize metrics like the Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC), and the model's utility will be demonstrated by forecasting drug sales for the subsequent four years, with all computations performed using R packages. While the abstract presents an intriguing framework, the full paper would benefit from clarifying the precise novelty of this "new version" of ARIMA-NN and the "new forecasting algorithm," as the described ANN components (MLP, specific activation functions) are relatively standard. For evaluating forecasting effectiveness and comparing against "existing forecasting models," the inclusion of more direct out-of-sample performance metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), or Mean Absolute Percentage Error (MAPE) alongside AIC and BIC would significantly strengthen the empirical analysis and provide a clearer understanding of predictive accuracy. Detailing the specific existing forecasting models used for comparison would also enhance the study's contribution. Addressing these points will undoubtedly bolster the clarity and impact of the research findings.

Full Text

Comments

You need to be logged in to post a comment.