Bitcoin Trend Prediction Using Bitcoin Influence Index Weighted Sentiment and Hybrid Ensemble Learning

Akhlis Munazilin (Universitas Diponegoro)
Mochamad Agung Wibowo (Universitas Diponegoro)
Rizky Parlika (University of Pembangunan Nasional “Veteran” Jawa Timur)

Journal of Applied Data Sciences
Vol 7, No 4: December 2026

Predicting Bitcoin trends is a complex problem because market movements are influenced not only by historical data and technical indicators but also by sentiment and rapidly evolving information disseminated through social media. Previous studies have shown that sentiment information can improve cryptocurrency prediction; however, most approaches aggregate opinions without explicitly distinguishing the characteristics of their information sources. This study develops the Bitcoin Influence Index (BII) as a multidimensional representation of social media account influence based on four dimensions: Reach Influence, Credibility Influence, Activity Influence, and Content Influence. BII is integrated with FinBERT based sentiment analysis to construct BII weighted sentiment, which is subsequently transformed into temporal features through lag, rolling average, momentum, interaction, and shock measures. Next day Bitcoin direction is predicted using a hybrid ensemble that combines XGBoost, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) as base learners, with Logistic Regression as the meta learner. The experiment compares two scenarios, Without BII and with BII, using 96 daily observations obtained from the integration of market and social media data. The results show that The with BII scenario increases accuracy from 40.00% to 75.00%, recall from 14.29% to 78.57%, F1-score from 0.2500 to 0.8148, and area under the curve from 0.3809 to 0.6667, while precision changes from 100.00% to 84.62%. Paired t-test yields p=0.030771, and Wilcoxon signed rank produces p=0.034808, while McNemar produces p=0.065430. These findings suggest that opinion source characteristics can provide additional predictive information when integrated with sentiment, technical indicators, and temporal models. The main contribution of this study lies in extending sentiment aware forecasting toward influence aware Bitcoin trend prediction, an approach that considers not only opinion polarity but also the multidimensional characteristics of the information sources conveying those opinions.