Cryptocurrency Price Prediction and Investment Decision Support Using a Transfer Learning-Based Hybrid CNN-BiLSTM-Attention Deep Learning Framework

Abstract

The complexity and volatility of cryptocurrency markets make accurate price estimation a major challenge. Nonlinear, nonstationary, and high-frequency fluctuations in digital asset prices often exceed the modeling capacity of conventional statistical approaches. To address this, we propose a transfer learning-based hybrid CNN-BiLSTM-Attention framework for cryptocurrency price prediction and investment decision support. The framework leverages pretrained LSTM models developed on relevant financial datasets to enhance the learning process. While convolutional neural networks (CNNs) effectively capture short-term trading patterns, bidirectional long short-term memory (BiLSTM) networks identify long-term temporal dependencies in both directions. The integrated attention mechanism further strengthens the model by dynamically selecting the most relevant time intervals, thereby focusing on critical patterns that drive price fluctuations. A historical dataset of major cryptocurrencies—including market capitalization, trading volume, and daily open, high, low, and close prices—was used to train and evaluate the model. Experimental results demonstrate that the proposed hybrid framework outperforms standalone CNN, LSTM, and RNN architectures, achieving an F1 score of 95.5%, recall of 95.2%, precision of 95.5%, and an overall accuracy of 96%. By delivering robust and scalable predictive performance, this study provides investors, portfolio managers, and financial professionals with a reliable tool to support informed decision-making in the dynamic cryptocurrency market.