Evaluation of Heuristics for Taken’s Theorem Hyper-Parameters Optimization in Time Series Forecasting Tasks

Abstract

This study compares three methods for optimizing the hyper-parameters m (embedding dimension) and τ (time delay) from Taken’s Theorem for time-series forecasting to train a Support Vector Regression system (SVR). Firstly, we use a method which utilizes Mutual Information for optimizing τ and a technique referred to as “Dimension Congruence” to optimize m. Secondly, we employ a grid search and random search, combined with a cross-validation scheme, to optimize m and τ hyper-parameters. Lastly, various real-world time series are used to analyze the three proposed strategies.

Keywords

• Taken’s Theorem
• time-series
• SVR forecasting
• mutual information
• dimension congruence
• random search