Price trend prediction in financial markets remains a major challenge in quantitative trading and risk management. This paper proposes a Multi-task Focused Attention Mechanism (MFAM) to address the limitations of traditional Long Short-Term Memory (LSTM) models, including their inability to effectively handle varying time-step feature importance, susceptibility to overfitting, and constraints of single-task modeling. The MFAM model integrates an attention mechanism within a multi-task learning framework. A data preprocessing method combining local mean sliding-window normalization, the Hampel estimator, and rolling interquartile range is employed to reduce forward-looking bias and improve adaptability to non-stationary financial data. Experimental results on EUR/USD hourly exchange rate data show that MFAM significantly outperforms the baseline LSTM + Attention model. Specifically, MFAM improves the coefficient of determination (R2) to 0.9529, reduces Root Mean Squared Error (RMSE) by 5.3%, Mean Squared Error (MSE) by 11.5%, and Mean Absolute Error (MAE) by 3.6%. It also achieves notable improvements in trend classification metrics, including precision, accuracy, recall, and F1-score. This study demonstrates the effectiveness of attention-guided multi-task learning in financial prediction and provides a promising approach for modeling complex financial market environments.