Abstract: Financial market series are significantly influenced by noise, non-stationarity, and unexpected events, posing challenges to stock trend prediction in terms of feature representation and generalization stability. Although some existing methods have incorporated multi-indicator inputs, the systematic feature selection and fusion strategies remain insufficient, limiting the effective characterization of market volatility patterns. To address these issues, this paper proposes an two-level Blending attention-based temporal trend forecasting model named Blending-ATT-Ridge. First, feature engineering is applied to construct and select multiple financial technical indicators. Then, attention-based BiLSTM and GRU are employed to model temporal dependencies separately and generate prediction components. Finally, regularization-based weighted fusion is adopted to integrate these components and mitigate overfitting. Experimental results on real stock datasets from leading Hong Kong-listed enterprises demonstrate that our method significantly outperforms other models across metrics such as mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE), and the coefficient of determination (R²). Furthermore, it maintains stable generalization capability under different market conditions, validating the effectiveness and robustness of multi-stage feature engineering and ensemble strategies in stock trend prediction.