Reti neurali per la previsione di serie storiche dinamiche

Time series forecasting has gained much attention in the last few years due to its practical applications. The dynamic behavior of most of the real time series in real life with its autoregressive and inherent moving average terms issue the challenge to adequately forecast this dynamic. Among the various models proposed in literature, linear and nonlinear, neural networks are considered a powerful technique that has been successfully used for time series forecasting. However, in this study we show that common neural networks are not efficient to recognize the behavior of nonlinear time series or the dynamic the time series with moving average terms. Let me underline that few results are available on forecasting time series using Neural Networks. Very recent papers show that the hybrid methodologies are powerful tools in forecasting. this work is inspired by this last point of view. In particular we propose a model that includes Fuzzy Logic and high-order neural networks with output error feedback, called Ridge Polynomial Neural Network with Error Feedback (RPNN-EF). Applying this model to financial time series regarding stock prices we show how the overall forecasting performance for the network is improved compared with the traditional statistical forecasting models.