Enhancing univariate time series forecasting: a general tree-based model selection approach

The selection of an appropriate forecasting method continues to be a pivotal topic among both theorists and practitioners, as more complex models do not always yield better results. In response, our study advances a pragmatic, data-driven approach for selecting the most suitable forecasting model, tailored to the unique attributes of specific time series data. This approach assesses 22 time series characteristics that span distribution, temporal statistics, and both linear and non-linear autocorrelations, as well as fluctuation analysis. Employing a decision-tree based framework that incorporates seven widely recognized forecasting methods—both parametric and non-parametric—our model facilitates an effective decision-making process among alternative forecasting methods. These methods include statistical and machine learning models, ensuring comprehensive and robust forecasting capabilities. We compare these models using Critical Difference Diagrams and find that some models have no statistically significant performance differences between them. Utilizing a comprehensive subset of the M3-Competition of approximately 2800 time series, which are from socio-economic domains and have a relatively low number of observations, our findings indicate that the decision-tree model reasonably selects an adequate forecasting model.