Moving Windows Quadratic Autoregressive Model for Predicting Chaotic Time Series
Aiguo Li · Systems Engineering - Theory & Practice · 2004
A novel model for predicting nonlinear time series is proposed in this paper, namely moving windows quadratic autoregressive (MWDAR) model. The model is constructed by using historical data and the quadratic items of data, and the parameters of the model are estimated by linear least square algorithms. It is necessary to specify the size of the windows, and the orders of the model before prediction process. In every crisp time point, the parameters of the model are estimated according to the data in current window, then, the future value is predicted as a result of the model parameters and the current input vector. The MWDAR model not only works very well on small data sets, but also has high computing efficiency on large data sets. Single and multi-step predictions experiments of comparing the MWDAR model with well-known local linear model are done on synthetic (Henon map) and real data (stock price movements) respectively. The results are excellent: MWDAR model not only has higher precision, but also has higher computing efficiency than local linear model.