Meta-Learning Based Forecasting for Non-Stationary Time Series
Xinying Li, Anqi Hong, Qiang Gao, Ru Zhang · 2024
The prediction of non-stationary time series is of great importance in various fields such as finance, meteorology, and network traffic. Analyzing and predicting non-stationary time series can provide decision-making guidance, help forecasters determine future trends, and reduce operational risks across different domains. In this paper, we address long non-stationary time series by dividing them into relatively stationary short-term segments and propose a forecasting model based on meta-learning for non-stationary time series. This meta-learning forecasting model consists of a predictor module and a meta-learner module. Based on this model, We design both a meta-learning based linear AR forecasting model and a meta-learning based nonlinear CNN forecasting model. Experiments on the Dow Jones index data and wind speed data demonstrate that the forecasting model based on meta-learning for non-stationary time series outperforms traditional AR forecasting model and CNN forecasting model in terms of prediction accuracy. Moreover, the meta-learning based nonlinear CNN forecasting model, which extracts the nonlinear features of the time series, performs better than the meta-learning-based linear AR forecasting model.