An Efficient Self‐Attention Based Hybrid Deep Learning Model for Stock Prediction

ABSTRACT Stock predictions have always been a trending topic and are considered a future development expectation of companies. Providing an accurate prediction is an interesting task because of non‐linearity, inherent dynamics, and complexity. Thus, this study proposed a new hybrid deep learning architecture for an efficient stock prediction process. Initially, the study collected the input samples from a publicly available dataset, and pre‐processing was done to maintain the quality of the inputs through min‐max normalization and Z‐score normalization. From the pre‐processed data, prediction is accomplished by proposing a new hybrid method called Self‐Attention based Dense Capsule Dual Stack Enhanced Auto Encoder (SADC‐DSEAE). The parameters of the proposed model are fine‐tuned by utilizing Reptile Search optimization. The simulation outcomes show that the proposed study has attained reduced prediction error in a 1‐day lead time with 0.38 of MAPE and 3.35 of MAE. The comparative analysis proves the strength of the proposed work over other prevailing methods.