Bitcoin Price Movement Prediction: A Machine Learning Comparison

In this work, a machine learning benchmark is set for the prediction of the evolution of the price of Bitcoins from OHLCV data available for 2,713 trading days from 2014 to 2023. The three most basic algorithms (logistic regression, random forest and XGBoost) are tested under time-based validation aimed at giving some financial plausibility, resulting in the corresponding accuracies of 54%, 48% and 49%.Overall these outcomes identify a difficult yet exciting landscape for applying machine learning to cryptocurrency markets, especially when the class imbalance influences prediction accuracy in a negative way for times that are beyond rising bull markets. There seems to be a “hidden”, regular signal in the price data of BTC as even these linear models perform well better than the nonlinear on the raw price data. The overall insight these outcomes raises is an enhanced focus on time validation of financial ML, in order to avoid any data leakage or overfitting. Above findings present important baselines to the studies in the future and motivate us to look into better methods in volatile asset predicting. There are more to study for the feature engineering and imbalance correction in future research as well.