Pattern recognition and classification in time series analysis

摘要

A time series often exhibits a certain pattern and it may form a basis for discriminating between different classes of underlying events. With phenomena in economics or life science being, in particular, dynamic and time dependent, time series analyses have great scope in the extension of classical pattern-recognition techniques to dynamic data. Model-free design of neural networks for nonlinear time series classification is studied. We use the networks to classify the linear ARIMA(1, 0, 0) model between the bilinear BL(1, 0, 1, 1) model. The results of our investigations show that the designed networks have a significant rate of correct classification if the coefficient of the bilinear term is large. In the empirical examples, we trained networks to identify which is the cause of seismic wave, and the rate of correct classification approaches 90%.