Detecting anomalies in cryptocurrency markets is challenging due to their volatility, non-stationarity, and sentiment-driven behaviour. Traditional statistical methods often miss signals or generate false alarms. We propose a hybrid framework combining a CNN-LSTM-AE with LLM-based sentiment analysis. The CNN-LSTM-AE captures spatial patterns, temporal dependencies, and latent representations from price and volume data, while the sentiment module extracts market signals from news and social media. An ensemble refinement aligns sentiment with trading signals, enabling the detection of regime shifts and subtle irregularities. Experiments on multiple crypto pairs show that the framework outperforms baseline models in anomaly detection and trading performance, highlighting the value of integrating deep learning with LLM-driven sentiment analysis for trading.