| Abstract [eng] |
Cryptocurrency markets are difficult to analyse because they are highly volatile, continuously traded, sensitive to news and affected by changing market conditions. Although machine learning models are increasingly used for cryptocurrency prediction, their value is limited when model decisions cannot be interpreted. Therefore, this thesis focuses not only on predicting market direction, but also on explaining which variables influence model decisions and whether predictions have economic meaning. The object of the thesis is the market dynamics of Bitcoin, Ethereum and Solana. The aim is to evaluate machine learning models for analysing cryptocurrency market dynamics using explainable artificial intelligence methods. The literature review examines cryptocurrency market behaviour, price drivers, volatility, regime changes, news sentiment, machine learning-based prediction, explainability methods and economic evaluation of trading signals. The methodology develops an empirical framework based on daily OHLCV market data, derived technical indicators, crypto-specific and macroeconomic news variables, macro-financial indicators and HMM-based market-state features. Random Forest classifiers are used to predict UP/DOWN market direction, while feature importance, SHAP analysis, PDP/ICE plots and backtesting are applied to interpret the models from technical and economic perspectives. The results show that the Random Forest models achieved above-random directional prediction performance for all three cryptocurrencies. Bitcoin produced a weaker but meaningful result, with accuracy and macro F1 of about 0.63, while Ethereum and Solana achieved stronger results of about 0.75-0.76. Explainability analysis showed that recent asset-specific return variables were the most important predictors. Positive recent returns generally increased the predicted probability of an UP movement, while negative returns reduced it, indicating short-term momentum-like behaviour. External variables, including news sentiment, macroeconomic indicators and HMM-based states, had weaker direct influence on model predictions than return-based variables, although they remained useful for market-context interpretation. Backtesting showed that model probabilities could be transformed into economically interpretable trading signals. The strategies did not consistently outperform buy-and-hold, but sometimes helped reduce exposure during weaker market periods. Therefore, the practical value of the thesis lies in applying explainable models to support interpretation of price-change signals, risk conditions and position-management decisions rather than in creating a universally profitable trading system. |
