A Hierarchical Platform-Aware Multimodal Framework for Sentiment-Driven Cryptocurrency Trend Prediction under Extreme Market Volatility

Abstract

Cryptocurrency trend forecasting is a major problem due to the high volatility of the market and the non-linearity of the decentralized social proof mechanisms. Contemporary sentiment-aware models are largely based on data taken from Twitter (X); however, they tend to neglect the complex, community-itself that less well-studied platforms are inherently known to involve. Accordingly, this manuscript presents a novel multimodal sentiment data pipeline along with Hierarchical Cross-Modal Platform-Aware Multi-Attention Network (HCMP-MA-Net). The proposed framework assimilates sentiment signals extracted from comments on Reddit, telegram, YouTube in connection with conventional news feeds and Twitter streams. To reduce the effect of noise in the data, a Relational Graph Convolutional Network (RGCN) is used for the topological bot activity detection, while a Platform-Aware Gating (PAG) mechanism is employed to dynamically re-weight social inputs for periods of increased market turbulence. Empirical assessments using three different multisource data sets show that the HCMP- MA.Net achieves 97.60% direction forecasting accuracy and Mean Absolute Error (MAE) of 0.0065, as compared to both univariate LSTMs and news-only models, significantly outperforming both. Comparative analysis further shows that expert commentary on Telegram is a key high frequency catalyst for intraday price movements, while consensus coming from Reddit is a medium term leading indicator. Back - testing simulations support the practical value of the framework, with a Sharpe ratio of 4.21 and minimum maximum drawdown of 1.05% which represents a new front-line standard for sentiment - informed cryptocurrency analytics.