Time-series forecasting is an essential component of data analysis and predictive modelling. The primary objective of time-series forecasting techniques is to minimise the percentage prediction error associated with future observations as much as possible. To achieve this goal, numerous statistical approaches and machine learning models have been proposed to capture temporal patterns and generate reliable forecasts.However,traditional statistical forecasting models often suffer from significant limitations in capturing long-time dependencies, addressing abrupt fluctuations, and dealing with non-linear interactions.To overcome these shortcomings, machine learning approaches have emerged as a powerful tool that has fundamentally revolutionised data analysis in recent years.Nevertheless, their forecasting performance remains sensitive to hyperparameter configuration, whereas manual tuning across high-dimensional hyperparameters can often be time-consuming, inefficient and suboptimal. Consequently, automatic hyperparameter optimisation (HPO) methods have been increasingly adopted. Among these, metaheuristic optimisation algorithms have emerged as innovative and powerful optimization techniquesand are currently among the most widely employed methods, with primary attention given to hyperparameters associated with machine learning architectures. By optimising the hyperparameters of machine learning models, these algorithms can substantially enhance forecasting accuracy and generalisability. This study reviews the performance of different metaheuristic optimisation algorithms and machine learning forecasting approaches, and further proposes a classification of metaheuristic optimisation algorithms and provides a clearer direction for the future design and implementation of hyperparameter optimisation strategies for deep learning models in time-series forecasting.