Predicting Bitcoin prices is inherently difficult due to the market’s highly volatile nature, nonlinear behavior, and sensitivity to global financial conditions. Traditional statistical models often fail to capture complex patterns, while standalone deep learning approaches may overlook underlying linear structures. To address these limitations, this paper presents a hybrid forecasting framework that combines ARIMA, Long Short-Term Memory (LSTM), and an attention mechanism in a unified, residual-driven architecture. Initially, the ARIMA model is applied to identify and model the linear components of the time-series data after ensuring stationarity. The residuals obtained from this stage, which represent nonlinear and unexplained variations, are then used as input to an LSTM network to learn long-term dependencies. Furthermore, an attention layer is incorporated to assign adaptive importance to different time steps, allowing the model to focus on the most influential historical patterns. This integrated approach enhances both prediction accuracy and interpretability, particularly under unstable market conditions. Experimental evaluation using standard performance metrics such as RMSE, MAE, and directional accuracy demonstrates that the proposed model consistently outperforms individual and baseline methods. The results highlight the effectiveness of combining statistical and deep learning techniques for robust cryptocurrency price forecasting.