Crypto foretell: a novel hybrid attention-correlation based forecasting approach for cryptocurrency

Cryptocurrencies function as a digital exchange medium operating on network-based technology, where records are secured using cryptographic algorithms such as Secure Hash Algorithm 2 (SHA-2) and Message Digest 5 (MD5). These cryptocurrencies utilize blockchain technology to provide transparent, reliable, and immutable transactions. Consequently, cryptocurrencies have gained significant traction across multiple sectors, particularly finance. However, their value is still prone to considerable fluctuations, which raises concerns about the risks associated with investments. The emerging discipline of cryptocurrency forecasting has gained popularity worldwide, and academics are employing a variety of deep learning (DL) and machine learning (ML) techniques to investigate the elements that influence cryptocurrency values. Among the various DL methods, LSTM has demonstrated noteworthy efficiency. Nevertheless, there are intrinsic downsides to LSTM, notably due to its sequential nature, which hinders parallelization and complicates the modeling of both short- and long-term dependencies. To address these shortcomings, the Transformer architecture has emerged as a potent solution. The Transformer is widely used in DL for its exceptional parallelization capabilities and its capacity to extract broad, distant data dependencies. Recent studies have explored Transformer-based approaches for cryptocurrency price forecasting, particularly for modeling long-term dependencies. However, these models often exhibit limitations in capturing high-frequency, short-term fluctuations, making them less suitable for short-term prediction tasks. Our proposed methodology introduces a novel Transformer-based hybrid framework designed to enhance forecasting accuracy across various time scales. We evaluate the forecasting accuracy for 10 cryptocurrencies at hourly, daily, and yearly frequencies. The findings show that, in comparison to other DL techniques such as LSTM, RNN, and baseline Autoformer, our model achieves superior accuracy. Furthermore, we benchmark our method against prominent Transformer variants such as Informer and FEDformer, and observe improved performance in both short- and long-term forecasting scenarios. These results indicate that our proposed model consistently outperforms existing state-of-the-art Transformer-based approaches in cryptocurrency price prediction.