Predicting local maxima of nonlinear time series with a neural network and edit distance

Abstract

In this article, we introduce an edit distance for event series into the prediction network, which considers both time and value of the events like local maxima. We use the time series data generated from the Rössler system to conduct a prediction experiment. We compared the prediction results of inputting 0,1,2, and 3 edit distances into the neural network. We found that the root mean square error of prediction decreases while the input number of the edit distances increases from 0 to 3. We discuss how the edit distance contributes to improving the prediction accuracy because the edit distances effectively describe the relationships between maxima states from different time.