This study examines and compares the performance of four-time series forecasting models, including ARIMA, ARIMAX, LSTM, and GRU, in forecasting the stock price of Iran Export Bank over 16 years (2009-2025). The data were prepared for modeling after performing preprocessing steps such as normalization. In the modeling section, the classical ARIMA models and the improved version of ARIMAX, which also consider exogenous variables (such as trading volume, moving average, and volatility), were used along with deep learning-based Recurrent Neural Networks (RNNs), namely LSTM and GRU. The results showed that the deep learning models LSTM and GRU performed significantly better than the classical models. Among the classical models, ARIMAX performed significantly better in forecasting than ARIMA, which had very poor performance. The LSTM model provided the most accurate forecasts and was able to model more than 98.67 percent of the data changes. The GRU model also performed close to LSTM, approximately 98.61, but its accuracy was slightly lower than LSTM. The results show that deep learning models, especially LSTM, perform better than classical models in simulating nonlinear patterns and long-term dependencies in financial data. Overall, deep learning-based models, especially LSTM, are powerful tools for predicting complex time series and can play an important role in investment decisions and analyzing stock market trends.