Accurate cryptocurrency price forecasting, particularly for Bitcoin’s volatile market, presents a significant challenge in quantitative finance due to the limitations of traditional models in capturing nonlinear dynamics and sentiment-driven fluctuations. To tackle this, we propose CryptoXNet, a hybrid deep learning architecture that integrates three key innovations: (1) an Extended LSTM (xLSTM) with enhanced memory mechanisms for better long-term dependency modeling; (2) a Squeeze-and-Excitation (SE) network for channel-wise feature recalibration; and (3) the systematic incorporation of behavioral finance metrics via the Crypto Fear & Greed Index (FGI). Trained on historical data, CryptoXNet demonstrated strong generalization on future data (2023–2024), outperforming LSTM, GRU, CNN, and WAMC benchmarks with a 33% reduction in Mean Absolute Error (MAE), a 17.6% decrease in Root Mean Squared Error (RMSE), and a 14.42% increase in R-squared (R \(^{2}\) ). Trading simulations further validated its real-world effectiveness, generating 204.34% total returns compared to 109.42% for a passive strategy, while also reducing maximum drawdown by 29%. These results set new standards for temporal-aware cryptocurrency forecasting through integrated memory optimization, adaptive feature selection, and crowd psychology analysis.