Optimizing Forecasting Performance via AI Algorithm Modifications: An Analytical Study

Abstract

This survey paper studies the breakthrough improvements of artificial intelligence (AI) algorithms for improved forecast in diverse fields. The paper reviews how changes in traditional AI frameworks have resulted in significant gains with respect to prediction accuracy, computational efficiency and dynamic environmental accommodation. By meta-analysing recent works, we highlight which algorithmic alterations have shown to be most effective, such as hybrid model architectures, attention mechanisms, (multi-)task/trans-fer learning methodologies and interpretability improvements. Our results suggest that these changes have collectively tackled long-standing issues in forecasting, including the management of non-stationary data, the representation of complex temporal patterns, and the measurement of uncertainty. The paper also examines methodological trends across 87 papers between 2018 and 2024, concluding that there is a rising attention to ensemble methods and self-adaptive algorithms. We close by briefly proposing new research trends and unexploited opportunities for enhancing AI forecasting approaches to enable a high adaptation to the specific domain, as well as to make use of causal inference techniques to push the field ahead.