PREDICTING CLOSING STOCK RETURNS USING SEQUENTIAL ARTIFICIAL NEURAL NETWORK

Abstract

Financial time series exhibit complex characteristics, including volatility, nonlinearity, and temporal dependencies, which make accurate forecasting challenging. This study investigated the use of a dynamic Long Short-Term Memory (LSTM) neural network to forecast stock prices and market direction. Augmented Dickey-Fuller tests confirmed stationarity of the return series, while model performance was evaluated using RMSE, MAPE, Directional Accuracy, and Theil's U-statistic against a naive forecasting benchmark. The model achieved low forecasting errors, with Apple and Microsoft recording RMSE values of 4.56 and 5.34 and MAPE values of 2.34% and 2.87%, respectively. Nvidia and Tesla recorded MAPE values of 2.98% and 4.12%. Directional Accuracy ranged from 56.7% to 62.1%, while Theil's U-statistic ranged from 0.67 to 0.85, indicating improved performance relative to the naive benchmark. Diebold-Mariano tests also indicated superior predictive performance compared with selected traditional and static neural-network models. Residual diagnostics showed fat-tailed errors but no significant residual serial correlation. The findings indicate that dynamic LSTM networks can effectively capture nonlinear temporal dependencies in financial time series and have potential applications in financial forecasting, algorithmic trading and portfolio risk management.