A unified GRU model for cryptocurrency price prediction and harsh price movement detection using enhanced sentiment analysis

Abstract

Predicting cryptocurrency price movements using social media sentiment remains challenging due to the noisy, heterogeneous, and rapidly evolving nature of online signals. While prior studies commonly combine sentiment analysis with deep learning models, less attention has been given to how sentiment signals are constructed, aggregated, and aligned with price dynamics. This study investigates the impact of sentiment representation and price change labeling on short-term Bitcoin price movement classification. Over 1.1 million Bitcoin-related tweets spanning April to August 2021 are analyzed using a RoBERTa-based sentiment model, incorporating both sentiment probabilities and user-level activity metrics. These features are consolidated via Principal Component Analysis (PCA) and aggregated over time using a decay-weighted scheme to emphasize recent information. Price movements are categorized into discrete regimes using a data-driven K-means clustering approach, with controlled Gaussian noise applied to improve boundary robustness. Multiple predictive models, including a Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), LightGBM, and multinomial logistic regression, are evaluated. Although the GRU achieves the highest overall performance, an extensive ablation study demonstrates that the primary performance gains arise from the proposed sentiment construction and labeling framework rather than the forecasting architecture alone. Removing PCA-based aggregation, adaptive clustering, or noise injection leads to substantial degradation, particularly for extreme price movement classes. The findings highlight the importance of sentiment feature design and class definition in cryptocurrency prediction and provide empirical guidance for constructing robust sentiment- driven financial models.