Nonlinear time series forecasting: A recurrent neural network approach

This dissertation presents a study of recurrent neural network based nonlinear time series forecasting models. Traditional forecasting methods based on linear models use the past internal patterns in the time series to generate a model that is used to forecast the future value of the series. These models do an adequate job of forecasting if the future values of the series mimic its past values and the series exhibits Gaussian characteristics. However real systems are dynamic, complex, iterative, nonlinear and non-Gaussian, thus linear models often give incorrect predictions.; Recently there has been an increase in the use of nonlinear models for time series forecasting. To overcome the problem of selecting the appropriate nonlinear model for any given time series we propose the use of recurrent neural networks as a nonlinear time series model. Recurrent neural networks are data driven models that learn to identify the underlying dynamics producing a particular time series during the process of training. This obviates the need for an a priori nonlinear model specification. We show that these networks are rich enough to model a wide range of linear and nonlinear time series. The dissertation is divided into two parts.; In the first part, a comprehensive comparative study is done comparing the forecasting performance of recurrent neural networks with that of other linear and nonlinear models that have been proposed in the past. To overcome the limitation of finite memory capacity, a modification to the supervised training algorithm is proposed for the online training of recurrent neural networks.; In the second part of this dissertation an improvement on the forecasting performance of the previously described recurrent neural network models is presented. The approach taken is to break down the original time series into different scales (resolutions) via a wavelet transform. The recurrent neural network trained with the output of the wavelet transform is shown to have a superior forecasting performance.