Abstarct: The capital market, as one of the fundamental pillars of any country's financial system, plays a crucial role in the efficient mobilization and allocation of resources. Within this context, the Tehran Stock Exchange (TSE), as the largest trading institution in the country, has consistently attracted the attention of researchers and investors in the field of stock price analysis and forecasting. Given the dynamic, nonlinear nature of this market and its susceptibility to various economic, political, and behavioral factors, traditional statistical methods face limitations in stock price prediction. In recent years, significant advancements in machine learning, and particularly in deep learning, have opened new horizons for financial data modeling. The primary objective of this research is to propose and evaluate a deep learning-baxsed model for stock price prediction on the Tehran Stock Exchange with appropriate generalizability. To this end, using historical price data of 50 index constituent symbols from the beginning of the Iranian year 1393 (March 2014) to the beginning of 1404 (March 2025), several deep learning architectures, including Multilxayer Perceptron (MLP), Recurrent Neural Network (RNN), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM), were trained. The data were normalized using min-max scaling and split into training and evaluation sets using the K-Fold method. Model performance was assessed using the Mean Squared Error (MSE) criterion. To obtain the best model for each architecture, hyperparameter tuning experiments were conducted on the number of laxyers, hidden dimensions, and input window length. Furthermore, to evaluate the generalizability of the obtained models, 10 symbols which were not included in the initial training and evaluation process (i.e., not among the 50 symbols) were used as a test dataset. The experimental results showed that using deep learning-baxsed models, especially MLP, along with a large number of symbols from diverse industry groups, achieves low error rates comparable to other methods and high generalizability for unseen symbols. This indicates the capability of deep learning-baxsed models to extract nonlinear patterns and temporal dependencies from Iranian stock market data. The findings suggest that employing these models can serve as an effective tool for investors and portfolio managers in decision-making and can pave the way for the development of intelligent decision-support systems in Iran's capital market.