Efficient Model Selection for Time Series Forecasting via LLMs
Wang Wei, Tiankai Yang, Hongjie Chen, Ryan A. Rossi, Yue Zhao, Franck Dernoncourt, Hoda Eldardiry
TL;DR
The paper tackles the costly problem of selecting forecasting models for new time-series tasks by introducing an LLM-driven zero-shot model selection approach that does not require a preconstructed performance matrix. It formulates model selection as a dataset-prompt to model mapping, leveraging four prompt designs (with/without meta-features and with/without chain-of-thought) to guide the LLM in choosing a suitable forecasting configuration. Compared against traditional meta-learning and heuristic baselines, the method demonstrates competitive accuracy on hit@$k$ and MSE while offering substantial reductions in inference time and avoiding expensive training on historical performance data. The study includes ablations on meta-features, CoT prompting, and data representations, and highlights practical gains in efficiency and scalability, with limitations including evaluation limited to univariate data and ongoing questions about the underlying mechanisms of LLM-driven selection.
Abstract
Model selection is a critical step in time series forecasting, traditionally requiring extensive performance evaluations across various datasets. Meta-learning approaches aim to automate this process, but they typically depend on pre-constructed performance matrices, which are costly to build. In this work, we propose to leverage Large Language Models (LLMs) as a lightweight alternative for model selection. Our method eliminates the need for explicit performance matrices by utilizing the inherent knowledge and reasoning capabilities of LLMs. Through extensive experiments with LLaMA, GPT and Gemini, we demonstrate that our approach outperforms traditional meta-learning techniques and heuristic baselines, while significantly reducing computational overhead. These findings underscore the potential of LLMs in efficient model selection for time series forecasting.