Comparative analysis of time series forecasting: Traditional Statistical Approaches vs. Machine Learning Methods

In the evaluation of time series predictors, the model selection processes for traditional forecasting methods and machine learning (ML) techniques differ significantly. This study aims to assess their performance across multiple forecasting horizons using an extensive set of univariate time series. By comparing the out-of-sample accuracy of popular ML methods with traditional statistical approaches, the results indicate that traditional methods consistently outperform ML methods across all accuracy measures and forecasting horizons considered. However, ML techniques can be improved through preprocessing methods such as Box-Cox and log transformation. The results show that log transformation enhances the performance of ML techniques compared to traditional statistical forecasting methods. ML techniques must improve in accuracy, efficiency, and interpretability to be competitive. The primary contribution of this research is demonstrating the superior accuracy of traditional statistical methods over ML methods and highlighting the urgent need to identify the underlying causes and develop strategies to improve ML performance.