مطالب

Accurate stock market forecasting is a challenging and complex problem for the market analysts and decision makers. During the past decade’s accuracy of different methods are examined yet there is no consensus on optimum forecasting method. In this regard, the main objective of present study is to investigate eligibility of nonlinear time series, such as exponential smoothing and regime-switching models beside Box-Jenkins scheme in forecasting of stock return time series. Data set consist of daily observations of Apple and Microsoft corporations as of 2024 to 2025. The Terasvirta-Lin-Granger procedure chaotic behavior of data generating process of the selected samples being examined. The Self-Exciting Threshold Autoregressive procedure combined with GARCH component (SETARMA-GARCH) and ARMA model combined with EGARCH component (ARMA-EGARCH) in order to capture the heterogeneous variance of financial time series, which yield dynamic hybrid models. Moreover, due to the overwhelming application of Artificial Intelligence methods in computation, besides the Exponential Smoothing (ES) approach as a non-parametric method, a recently developed Multilayer Perceptron Network (MLP) based on Feed-Forward-Back Propagation (FF-BP) algorithm being developed either. Both of the in-sample and out-sample forecasting are carried out and performance of models is evaluated using standard error criteria. Finally, the Diebold-Mariano test is employed in order to determine the significance of forecasting differences among the models. Findings indicated that the behavior of the return series for the both of the corporations are chaotic and nonlinear methods are appropriate in modeling. The exponential smoothing method outperformed the developed SETARMA-GARCH and ARMA-EGARCH procedures in terms of the majority of error criteria in the both of in-sample and out-sample forecasting. However, the MLP has outweighed the ES model based on every calculated error criteria. The estimated S-statistic of Diebold-Mariano test confirmed results of the forecasting in favor of the MLP method. This finding suggests application of the dynamic nonparametric methods in modeling and forecasting of the selected time series. Implication of such finding recommends use of dynamic nonlinear and nonparametric methods in financial series prediction.