Enhanced cryptocurrency price prediction using integrated sentiment and price features

In the era of advanced technologies, digital currency such as cryptocurrency has attracted immense interests among traders and investors. The volatility and drastic shifts in cryptocurrency market make price predictions difficult. Thus, experiments were conducted to find the most optimised results. Majority findings have found that sentiment data improves cryptocurrency price prediction performance. But there exists a gap in research that discovers sentiment data for cryptocurrency price prediction for some languages (non-English) are not catered for. Therefore, the research aims to enhance the cryptocurrency price prediction performance and to extract sentiment features from online news and Twitter in English and Malay language, then builds a sentiment deep learning model integrating self-attention mechanism to investigate the effect of sentiment features on cryptocurrency price prediction. The methodology consists of three main phases, which are, data collection, feature extraction, and price prediction. Two corpora were constructed which are price and sentiment corpora in the first phase. During feature extraction phase, price features (open, high, low, close, and trading volume) were used directly, while text pre-processing was performed on the eight sentiment datasets. Sentiment data were then being annotated manually and automatically to be fed into Generalized Autoregressive Pre-training for Language Understanding (XLNet) model. Finally, the sentiment features are aggregated with the cryptocurrency price features and fed into the multi-layered bidirectional deep learning model with attention to predict the cryptocurrency closing price of the next day. The experimental results show that integrating sentiment features into deep learning models enhances cryptocurrency price prediction accuracy, with the Bi-GCL-A model outperforming traditional methods by leveraging self-attention mechanisms with an adjusted r2 of between 0.907 and 0.918, with 2.4% improvement. Moreover, there is an improvement of 0.130 (Bitcoin) and 0.127 (Ethereum) in average adjusted r2 for price prediction with price and sentiment features in comparison to price prediction with price features only. The research contributes to improving cryptocurrency price prediction performance while addressing the underrepresentation of the Malay language in Natural Language Processing (NLP) to make informed decision-making in optimising cryptocurrency trading profits.