Comparison of deep learning methods with traditional financial time series models for forecasting the TEPIX index in the Tehran Stock Exchange

Stock price forecasting is complex, influenced by economic conditions, company performance, market trends, and news. Predictive methods include stock chart analysis, technical analysis, fundamental analysis, mathematical models, and artificial intelligence algorithms. Accurate stock price predictions can directly impact capital investment. This research employs auxiliary algorithms, specifically LSTM, RNN, and CNN, to predict the total stock index of the Tehran Stock Exchange, comparing them to traditional financial models like ARIMA. The LSTM-3D model among the models presented in this research (LSTM-1D, LSTM-2D, LSTM-3D, RNN-1D, RNN-2D, RNN-3D, CNN-1D, CNN-2D, CNN-3D, and ARIMA(1,1,2)) has been introduced as the best option for forecasting according to RMSE, MAE, MSE criteria. This model provides better results than other deep learning methods by using explanatory variables and an appropriate time step (5). • This article compares deep learning models (LSTM, RNN, and CNN) with traditional models, utilizing one-dimensional, two-dimensional, and three-dimensional architectures. • It incorporates suitable time steps and explanatory variables to enhance forecasting accuracy, demonstrating that 3D deep learning models outperform the others in stock market predictions. • Building on this foundation, the study delves into the specific implementations of LSTM, RNN, and CNN architectures, highlighting their unique advantages and limitations in the context of financial forecasting. • The LSTM model, characterized by its ability to retain long-term dependencies, proves exceptionally effective in predicting stock price movements over time. By carefully selecting time steps that correspond to significant market events, the model enhances its predictive capabilities, capturing the underlying patterns that influence pricing.