machine learning framework for predicting cryptocurrency return trends: a comparative study

Cryptocurrency markets are highly volatile, making return forecasting both valuable and difficult. this study develops a machine learning framework for predicting cryptocurrency return trends, formulated as a three-class problem (negative, near-zero, and positive next-day returns) rather than the conventional binary one. xgboost, random forest, and support vector machine (svm) were evaluated on bitcoin, ethereum, and solana using daily data from december 2024 to july 2025. the features combined technical indicators with macroeconomic variables, and hyperparameters were tuned via bayesian optimization using the optuna framework. performance was assessed by classification accuracy and per-class precision, recall, and f1-score. average accuracies generally exceeded 60%. svm achieved the highest average accuracy, while xgboost was the most stable across assets and return classes. positive returns were predicted more reliably than near-zero and negative returns, though strong positive-class recall was not always matched by precision. the results suggest the potential of combining machine learning, technical indicators, and macroeconomic variables for return-trend prediction. as they derive from a single test window and three assets, they should be read as indicative rather than definitive.