This paper presents a deep learning approach for predicting asset returns, leveraging a hybrid architecture that combines LSTM layers with an attention mechanism. The model processes historical price and volume data, along with engineered features such as moving averages and volatility indices, to forecast future returns. The attention mechanism dynamically weights different time steps, allowing the model to focus on the most informative periods. The methodology is validated using walk-forward analysis on a diverse set of stocks and ETFs over a 10-year period, ensuring robustness against overfitting and market regime changes. Key results demonstrate that the deep learning model significantly outperforms traditional machine learning baselines, achieving a Sharpe ratio of 1.8 compared to 1.2 for LSTM and 0.9 for XGBoost. The model also reduces prediction error by 15% and maintains profitability during turbulent market conditions. The attention weights provide interpretability, highlighting the importance of recent price movements and trading volume. This work advances the application of deep learning in quantitative finance, offering a practical tool for asset return prediction.