Due to the extensive developments in the field of artificial intelligence and deep learning, financial market forecasting has become one of the key topics in modern econometrics. However, research conducted in the Iran capital market often faces two fundamental methodological challenges: first, data leakage caused by the use of price-dependent variables such as transaction value, which leads to the creation of spurious correlations; and second, the failure to observe the principles of validation in time series data, which makes it impossible to generalize models to the future. In this study, with the aim of eliminating these shortcomings, the forecast of the logarithmic returns of two large companies in the Iranian capital market, namely Bank Mellat Iran and Mobarakeh Steel Company of Isfahan, has been carried out using artificial neural networks, linear regression, group method, majority vote method, random forest method and gradient boosting method. The main innovation of this research is in two axes: the strategic elimination of leakage variables to ensure the realistic nature of the results, and the use of evaluation on out-of-sample data to preserve the temporal structure of the data and prevent bias in model evaluation. The research findings show that although complex models such as artificial neural networks have high potential in learning nonlinear patterns, the proximity of the coefficient of determination (R^2) to zero in most models indicates high market efficiency and the quasi-random nature of returns in the time period under study. These findings provide a realistic and scientific view of the potential and limitations of artificial intelligence in the Iranian capital market.