Financial forecasting improvement with LSTM-ARFIMA hybrid models and non-Gaussian distributions

This article introduces a groundbreaking method for accurately forecasting financial stock market returns. The approach utilizes a hybrid neuro-autoregressive model, combined with a multi-objective decision-making phase, to determine the optimal distribution, offering significant relevance in modern finance. The proposal harnesses the impressive capabilities of the long short-term memory (LSTM) recurrent neural network, synergistically coupled with the autoregressive fractionally integrated moving-average (ARFIMA) model across various distribution options. This synergy enables precise management of a wide range of both linear and nonlinear time series data. Utilized on two prominent American stock market indices (Dow Jones Industrial Average (DJIA) and Dow Jones Islamic Market International Titans 100 (IMXL) between 1/2/2015 and 12/10/2020), the experimental findings unequivocally illustrate the hybrid model's supremacy over baseline models in accuracy and computational efficiency. Notably, the forecasting experiments conducted in both tranquil and turbulent periods underscore the stability and robustness of this approach. The model's adaptability and resilience make it a promising tool for precise financial stock market return forecasts, particularly crucial in informing decision-making within the financial industry. Furthermore, this proposed approach contributes to the expanding research on decision support systems for financial forecasting, potentially influencing policy and strategic financial management, particularly in addressing both stable and volatile market conditions.