Probabilistic Deep Learning and Transfer Learning for Robust Cryptocurrency Price Prediction

Abstract

Forecasting the price of Bitcoin (BTC) with precision is a complex endeavor, given the market’s inherent uncertainty and volatility, influenced by a diverse range of parameters. This research is driven by the central goal of introducing a specialized deep learning model tailored to predict digital currency prices, with a specific emphasis on BTC. To address this challenge, a pioneering strategy has been established, leveraging probabilistic gated recurrent units (P-GRU). This approach integrates probabilistic attributes into the model, facilitating the generation of probability distributions for projected values. The effectiveness of this method is assessed using one year of BTC price history, sampled at a five-minute interval. In parallel, a comparative analysis is conducted against alternative models, including GRU, long short-term memory (LSTM), and variants thereof (time-distributed, bidirectional, and simple models). In pursuit of optimizing model efficacy, a bespoke callback mechanism is deployed. This callback, driven by R2-score tracking, captures optimal model weights based on validation data. Moreover, a transfer learning paradigm is adopted to broaden the study’s horizons. A pre-trained model on BTC data is harnessed to predict prices for six other prominent cryptocurrencies: Ethereum, Litecoin, Tron, Polkadot, Cardano, and Stellar. Consequently, a distinct model is tailored for each cryptocurrency. The outcomes of this investigation conclusively underscore the superior performance of the proposed methodology. In the midst of a volatile and uncertain market landscape, the proposed approach outshines its counterparts, showcasing an enhanced ability for cryptocurrency price forecasting.