A Transfer Learning–Enhanced Hybrid Deep Learning Framework for Bitcoin Price Forecasting Using Market Sentiment and Time Series Data

Abstract

The extreme volatility of Bitcoin markets makes accurate price prediction notably difficult. This paper proposes a new hybrid deep learning model that incorporates a Gated Recurrent Unit (GRU), a Bidirectional Long Short-Term Memory (Bi LSTM) model, and a Multi Head Attention mechanism to permit the model to utilize both historical price data and sentiment information from Twitter. We constructed the model utilizing a two-stage transfer learning approach: we first pretrained the model on data from 2017−2019 to learn lower-level fluctuation behaviors, then we fine-tuned the model on data from 2021−2023 in order to be sensitive to recent market behaviors. The model performed exceptionally well against multiple state-of-the-art baselines using root mean square error (RMSE) and mean absolute error (MAE) metrics, reporting RMSE values of 679.61 and MAE of 452.95, achieving considerable improvement over the baseline models. Our experimental results show that leveraging Twitter sentiment greatly improved trend prediction. In addition, our benchmarks showed that our method performed better than the existing methods. Furthermore, our ablation studies illustrated how each particular feature performed. Overall, our results demonstrate that multi-scale temporal modeling combined with social media sentiment integration produces a scalable and resilient solution to combat the challenges of volatility to forecast cryptocurrency prices accurately and efficiently.