Bitcoin Trend Prediction Using Bitcoin Influence Index Weighted Sentiment and Hybrid Ensemble Learning (Journal of Applied Data Sciences)

Summary

• This study proposes the Bitcoin Influence Index (BII) to weight social media sentiment using four influence dimensions: Reach, Credibility, Activity, and Content Influence.p.1
• The experiment evaluated next-day Bitcoin direction using 96 daily observations combining market and social media data.p.1
• Under the scenario without BII, the prediction model achieved an accuracy of 40.00%, recall of 14.29%, F1-score of 0.2500, and ROC-AUC of 0.3809.p.1
• Integrating BII features increased accuracy to 75.00%, recall to 78.57%, F1-score to 0.8148, and ROC-AUC to 0.6667.p.1
• Precision changed from 100.00% in the baseline model to 84.62% in the BII-enhanced model.p.1
• Statistical evaluation showed significant performance differences based on a paired t-test of p = 0.030771 and a Wilcoxon signed-rank test of p = 0.034808.p.1

Key insights

1. Adding multidimensional BII features improved next-day Bitcoin trend prediction accuracy from 40.00% to 75.00%.p.7
2. The model's recall registered a gain from 14.29% without BII to 78.57% with BII.p.7
3. Paired correctness tests showed significant statistical differences at alpha = 0.05 for both the paired t-test (p = 0.030771) and Wilcoxon test (p = 0.034808).p.8
4. McNemar's exact test resulted in p = 0.065430, which did not meet conventional statistical significance at the 5% level.p.8
5. Technical indicators such as SMA-14 and SMA-7 remained the most dominant features, complemented by lag-1 BII-weighted sentiment.p.8