Predicting prices of volatile assets such as Bitcoin is a challenging task because of the nonlinear, nonstationary, and noise-heavy nature of such time series. Classical deep learning approaches, such as Long Short-Term Memory (LSTM) networks, allow a strong sequential modeling, but tend to lose performance when processing raw noisy data. This paper investigates a hybrid modeling strategy that combines Empirical Mode Decomposition (EMD) with LSTM networks to enhance forecasting accuracy. The EMD-LSTM model not only enhances the robustness and interpretability of the model, but also the capture of the temporal structure through decomposing the raw price signals into IMFs for prediction. Using daily Bitcoin data and relevant macro-financial indicators, we evaluate the comparative performance of standalone LSTM versus the EMD-LSTM architecture. The empirical results show that the hybrid model significantly reduces prediction errors, particularly in capturing abrupt price movements. This research highlights the importance of signal preprocessing in enhancing the forecasting capabilities of deep learning models in highly turbulent financial markets.