Tree-based vs. deep learning in time series: A prescriptive framework for performance-driven model selection

A primary criticism of modern artificial intelligence models is the significant energy consumption required for their training. This issue is particularly relevant in time series forecasting, where a wide variety of models is available, yet the prevailing trend often favors maximizing predictive accuracy at the expense of substantially higher computational cost and environmental impact. This work poses a fundamental question: to what extent is this trade-off worthwhile? It seeks to quantify the marginal gain in accuracy relative to the increase in model complexity. Specifically, the study investigates whether it is possible to determine the characteristics that make a time series suitable for simpler, tree-based methods to yield sufficiently accurate forecasts, thereby rendering more resource-intensive deep neural networks unnecessary. To address this question, Regression Tree, M5P, Random Forest, XGBoost, Long Short-Term Memory networks, Convolutional Neural Networks, N-BEATS, Transformer and iTransformer are systematically compared across over 50,000 diverse real-world time series, divided into 31 different forecasting problems. The evaluation jointly considers predictive accuracy, computational efficiency, and sustainability-related aspects including energy consumption and CO 2 emissions. For each dataset, a comprehensive set of descriptive characteristics is computed and used as input to a decision-tree meta-learner. The proposed framework not only predicts the forecasting algorithm most likely to minimize error for a given dataset but also generates interpretable rules that clarify why one model is preferable over another. The results identify the specific conditions under which tree-based or deep learning approaches are more effective and demonstrate the potential of the framework to provide actionable, transparent, and generalizable guidance for model selection across diverse forecasting scenarios.