ENHANCING TIME SERIES FORECASTING ACCURACY WITH DEEP LEARNING MODELS: A COMPARATIVE STUDY

Abstract

This study offers a detailed comparison of both traditional and ad-vanced deep learning models in the context of time series forecasting, with a specific focus on ARIMA, Random Forest, Recurrent Neural Networks (RNNs), Long Short-Term Memory networks (LSTMs), and Gated Recurrent Units (GRUs). In line with open science principles, it utilizes publicly accessible datasets to guarantee the reproducibility of its findings and broaden their relevance. The research meticulously ap-proaches preprocessing and thoroughly investigates model architectures and hyperparameters to establish solid benchmarks for performance evaluation. It uniquely employs the Root Mean Square Error (RMSE) as the primary metric to assess forecasting accuracy across different datasets. This singular focus on RMSE enables a precise understanding of model performance, highlighting the exact conditions under which each model excels or falls short, considering dataset characteristics such as size and complexity. Additionally, the study explores the interpretability of these models to provide insights into the decision-making processes underlying deep learning predictions. The results of this analysis yield essential recommendations for selecting optimal modeling techniques for time series forecasting, significantly advancing theoretical knowledge and practical applications in the field. By narrowing the gap between advanced machine learning techniques and their effective deployment in forecasting tasks, this study guides practitioners and researchers toward informed model selection based on RMSE performance.