Особливості комбінованого моделювання лінійних та нелінійних компонент фондових індексів

Abstract

This work is devoted to assessing the efficiency of combined modeling for stock indices, namely, hybrid architectures based on ARIMA and LSTM. The basic principles of time series decomposition into linear and nonlinear components to improve medium-term forecasting accuracy were considered. Practical experiments on the S&P 500 and NASDAQ index data were conducted to evaluate the models using standard RMSE and MAE metrics. The research confirms that statistical models effectively filter linear trends, while deep learning networks successfully capture complex nonlinear patterns. The conclusion highlights the practical utility of this integrated approach for risk management tasks, such as Value-at-Risk calculation, while outlining the limitations related to computational costs.