A comprehensive evaluation of statistical, machine learning and deep learning models for time series prediction

How to choose the appropriate model to predict the time series is one of the most prominent activities of temporal data analysis. Empirical evidence is often adopted to select the most suitable model since there is no unified standard for matching data and models. Data characteristics affect model performance to a certain extent and maybe where the factors that determine the balance between prediction accuracy and model complexity are. In this article, Multi-Criteria Performance Measure method considering Mean of Absolute Value of the Residual Autocorrelation was adopted to address this problem. Case studies apply Time-Series Analysis decomposing datasets into trend, seasonality and residue and summarize the limitations and recommendations from the stochasticity of the residue. The results show that the statistical models perform best for datasets with low stochasticity, deep learning models specialize in forecasting fluctuant and long-term time series data, machine learning models could be candidates for datasets that possess numerical characters between the previous two categories. Conclusions could provide suggestions in selecting appropriate models and guide the research community in focusing the effort on more feasible or promising directions.