چکیده:
In this paper، we deal with several time series of share prices and daily returns of different companies which are members of Tehran Stock Exchange. Three prediction methods are used for time series forecasting. The first method is based on the linear models (ARIMA) for short-term and long-term forecasting. The second method is based on the nonlinear neural networks model and the third method is a neural networks model with a special structure. It has been shown that the time series generator process of these companies are complex nonlinear mappings and the methods based on the various linear modelling strategies are unable to identify these dynamics. Also، it has been shown by using the conventional structure of the nonlinear neural networks that one can not obtain a satisfactory result for long term forecasting. Finally، it is shown that the proposed structrure provides accurate next step and the long term share prices and daily returns forecasting.
خلاصه ماشینی:
The results of these studies, published in the form of numerous seminars and symposia in countries with advanced capital markets, questioned the validity of the theories of the 1960s to 1980s; because these theories are based on linear systems and the assumption of a normal distribution, and therefore, the single-variable asset pricing model cannot adequately estimate stock returns or the risk associated with them.
2-3-1-Prediction methods based on nonlinear models (nonlinear neural networks) The use of neural networks to perform mapping and function approximation and ultimately model dynamic processes is, in fact, a generalization of regression analysis and classical statistics.
k_2 (1)-next price It takes; while if the process generating the time series has been identified, the behavior of the process is completely specified with inputs and initial conditions, and in the prediction stage, there will be no need for actual information; in fact, if the obtained models are used for prediction with a horizon of more than one step, for k+t it is necessary to use only estimated output information in the relevant models; that is, during the forward prediction process, the outputs of the linear model (instead of actual values) are used.
A large value (Refer to the image on the page) indicates the ability to extract the structure of the price generation process in the neural network model, which is actually contrary to the efficient market hypothesis.
(refer to the page image) Figures 22 and 23 show the results of the performance of the obtained stock return price model Shahd-Iran with the proposed structure and a prediction horizon of 30 days.
(
, 
, 
)
