Detecting Bitcoin Sentiment: Leveraging Language Model Applications in Sentiment Analysis for Bitcoin Price Prediction

• Published in: Neural processing letters Vol. 57; no. 5; p. 77
• Main Authors: Jung, Hae Sun, Lee, Haein, Kim, Jang Hyun
• Format: Journal Article
• Language: English
• Published: Dordrecht Springer Nature B.V 01.10.2025
• Subjects: Ablation; Accuracy; Artificial intelligence; Blockchain; Data collection; Deep learning; Dictionaries; Digital currencies; Financial analysis; Institutional investments; International finance; Investment strategy; Investor behavior; Language; Large language models; Machine learning; Modular design; Natural language processing; Neural networks; Performance evaluation; Prices; Securities markets; Sentiment analysis; Social networks; Time series; Trends; Volatility
• ISSN: 1573-773X; 1370-4621; 1573-773X
• DOI: 10.1007/s11063-025-11787-1

As Bitcoin continues to establish itself as a global asset and discussions around relevant regulations become more active, there is an increasing demand for a comprehensive price prediction framework. To address this necessity, this study aims to enhance the accuracy of Bitcoin price predictions by integrating sentiment information with technical indicators, on-chain data, and cryptocurrency price data. Recognizing Bitcoin’s sensitivity to market sentiment, the proposed framework incorporates sentiment features derived from both lexicon-based methods and large language models. As unsupervised sentiment tools can introduce label noise particularly in domain-specific or ambiguous financial contexts, this study combines the outputs of multiple sentiment models at the feature level to construct a more stable representation. This design improves the robustness of downstream regression performance and distinguishes the framework from previous hybrid models that relied on a single sentiment source without component-wise evaluation. Experimental results using a dataset spanning 2700 days showed that the long short-term memory (LSTM) model with a 3-day window achieves the best performance with mean absolute percentage error (MAPE) of 3.93% and R-squared value of 0.99106. Feature importance analysis further demonstrates sentiment index as the most impactful feature, as excluding it resulted in the largest decline in predictive accuracy. Additionally, the model's performance was evaluated under four major volatility periods, revealing MAPE values ranging from 1.49 to 4.03%, highlighting the framework’s practical capability in rapidly adapting to sudden market shifts. In summary, integrating sentiment information attained from multiple language models significantly enhanced prediction accuracy compared to single source approaches. These findings highlight the framework’s practical value for sentiment-informed investment strategies and risk alerts, with a modular design that enables flexible adaptation and potential integration into automated trading systems.