This study investigates the effectiveness of traditional econometric models and deep learning architectures in forecasting Bitcoin's 7-day realized volatility over the period from 2021 to 2025. Four models, GARCH, ARIMA, LSTM, and CNN-LSTM, are evaluated using a combination of error-based metrics, information criteria, statistical tests, and visual diagnostics. The analysis highlights key trade-offs between model interpretability, complexity, and adaptability to market conditions. Rather than identifying a universally superior approach, the study emphasizes that model selection should be guided by the specific forecasting objectives, data characteristics, and application context.