Academy of Mathematics and Systems Science, CAS
Colloquia & Seminars

Speaker:	Cheng Weitao, University of Science and Technology of China
Inviter:
Title:	Exploring Time Series Predictability using Permutation Entropy
Language:	Chinese
Time & Venue:	2023.09.05 14:00-15:00 腾讯会议：962 4818 2651
Abstract:	Abstract In time series forecasting, model-based prediction error metrics cannot assist researchers in determining whether poor prediction performance is due to an inappropriate model choice or if the data the data inherently lacks predictive information. Intrinsic predictability characterizes "the upper limit of prediction accuracy"for the data, which can help researchers assess the compatibility of the current modeland data. In this paper, we briefly review the concepts of predictability and provide a detailed introduction to the studies of time series predictability based on permutation entropy. Based on this, we propose permutation entropy with covariates to characterize the complexity of target time series when covariates are available and demonstrate its effectiveness through experiments with real glass bubble data. Additionally, we further present a strategy for model selection based on predictability, aiming to choose simpler models and reduce the time cost of modeling and forecasting while maintaining reasonable accuracy. Numerical experiments on economic data validate the efficacy of this strategy.