Investigation of ensemble approaches and hybrid architectures for financial time series forecasting

Abstract

The problem of short-term forecasting of financial time series characterized by nonlinearity, non-stationarity and sensitivity to external shocks is investigated. The application of ensemble methods of combining models (bagging, boosting, stacking), as well as the construction of hybrid neural architectures combining LSTM, GRU, CNN and Autoencoder are considered. A theoretical review of the conceptual features of these approaches is provided, their advantages and limitations are discussed, and the conditions for effective application are determined. A comparative analysis of the forecasting results using actual AMD data and four architectures is presented. It is shown that the hybrid Autoencoder + LSTM model achieves the highest accuracy, demonstrating the advantages of integrating mechanisms for automatic feature selection and modeling long-term dependencies. The obtained results confirm the feasibility of using hybrid ensemble approaches to improve the efficiency of financial forecasting.