Forecasting the Prices of Selected Cryptocurrencies Using Various Machine Learning Algorithms

Abstract

The cryptocurrency market, due to its extreme volatility, non-normal behavior, and high sensitivity to trading fluctuations, poses a serious challenge for forecasting algorithms, and sole reliance on classical time-series models in such a noisy environment offers limited reliability. This study aims to identify the most accurate and stable machine learning and deep learning algorithms for short-term forecasting of selected cryptocurrency prices, as well as to examine the role of technical, trading, and intermarket variables in explaining price movements. The research approach is based on the analysis of daily data from five major cryptocurrencies and the application of a wide range of models, including linear regression, support vector machines (linear and RBF), random forest, XGBoost, LSTM networks, and ARIMAX and Lag-1 models. Model performance was evaluated using indicators such as MAE, RMSE, MAPE, R², and directional accuracy, and differences in performance were assessed using the Friedman and Diebold–Mariano tests. The results indicate that the linear-kernel support vector machine is the most accurate and stable algorithm at the one-day forecasting horizon, recording the lowest errors and the highest directional accuracy for most assets. Linear regression also demonstrated performance close to that of the best model and yielded particularly strong results for Ethereum. Tree-based models such as random forest showed higher efficiency for certain assets; however, their cross-asset stability was lower. Among advanced models, XGBoost and LSTM, although capable of effectively extracting nonlinear patterns, exhibited limited point-forecast accuracy and performed better in identifying the direction of market movements. Feature-importance analysis revealed that price–trading variables and technical indicators account for the largest share in explaining price changes, while intermarket variables play a complementary role. Overall, the findings suggest that simple and stable linear models are more reliable for short-term forecasting, whereas more advanced models primarily generate directional value added.