Application of artificial intelligence in cryptocurrency trading

Abstract

This work deals with the possibility of using recurrent neural networks of the LSTM and GRU type for predicting daily logarithmic returns of selected cryptocurrencies (BTC, ETH, LTC, BNB). Based on daily OHLCV data from the period 2017–2024, three sets of input variables are constructed: a basic set (transformed price and volume variables and logarithmic returns), a set of technical indicators, and a set of technical indicators together with macroeconomic data. The LSTM and GRU prediction models are calibrated for different memory lengths (10, 20, 50 days), numbers of neurons, and all three sets, and their performance is evaluated using RMSE, MAE, and directional accuracy. The results show that daily returns of cryptocurrencies are difficult to predict from the point of view of a one-time prediction: random walk remains a very strong benchmark in all cases, and the best neural networks only come close to it. Subsequently, simple Long/Short strategies are constructed based on the predictions and compared with the passive Buy&hold strategy. For all four cryptocurrencies, configurations are found whose strategies achieve higher annualized returns and Sharpe ratios than Buy&hold in the test period, especially for more volatile altcoins. However, this outperformance is conditioned by ex post selection of the “best” models and neglect of transaction costs, and therefore it must be interpreted with caution as an illustration of the potential and limits of deep neural networks in short-term prediction of cryptocurrencies.