Auto-TSF: Towards Proxy-Model-Based Meta-Learning for Automatic Time Series Forecasting Algorithm Selection

Time series forecasting (TSF) is a prominent chal-lenge in data analytics, relevant to both scientific research and real-world industrial applications. The rapid increase in high-dimensional time series data has led researchers to develop numerous models capable of handling complex forecasting tasks across diverse scenarios. Nevertheless, selecting an appropriate model and optimizing its parameters-an issue known as the Combined Algorithm Selection and Hyperparameter optimization (CASH) problem-remains a significant challenge. It is worth investigating how to satisfy both accuracy and efficiency in selecting an optimal algorithm and its hyperparameter configu-ration for a given time series with minimal human intervention. Unfortunately, there is no such work in the field of TSF, which has been developed for more than a decade. Existing methods suffer low selection rate of optimal algorithms. Meanwhile, the TSF task is extremely algorithm-sensitive, and the prediction performance of different algorithms and hyperparameter settings on the same data varies greatly. In this paper, we propose a Proxy-Model-based meta-learning TSF-CASH approach named Auto- Tsf. In the offline training phase, Auto- Tsfextracts the historical experience based on the proxy models, which is used to guide the automatic algorithm selection in the online working phase. The historical experience extracted in the offline phase not only significantly reduces the time consumption for algorithm selection, but also the introduction of the proxy model enhances the optimal algorithm selection rate. Moreover, we propose an asynchronous parallel HPO method in the most time-consuming HPO stage, which further improves the efficiency of the whole TSF -CASH. The experimental results demonstrate that Auto- Tsfachieves SOTA in terms of performance and efficiency compared to existing CASH methods.