Predicting Cryptocurrency Price Movements Using Machine Learning And Technical Indicators

Abstract

Cryptocurrency markets are highly volatile and difficult to forecast, posing challenges for both investors and researchers. This study investigated the use of machine learning models to predict short-term cryptocurrency price movements, using hourly data from Binance for five cryptocurrencies. A wide range of technical indicators and candlestick features were employed as inputs, and both classification tasks (price direction) and regression tasks (future high and low prices) were performed. Automated machine learning (AutoML) with PyCaret was used to systematically evaluate a broad set of algorithms across different input window sizes and prediction horizons. The results showed that the Ridge Classifier generally outperformed other models in predicting price direction. For regression tasks at the one-hour horizon, Huber, Orthogonal Matching Pursuit, and Ridge regression achieved the lowest error rates, while Extra Trees gave better results for longer horizons. Incorporating multiple previous hours of data improved model accuracy up to a threshold, after which additional lagged features yielded little benefit. Prediction accuracy decreased substantially when extending the forecasting horizon beyond one hour. A key finding was that low error metrics, particularly mean absolute percentage error (MAPE), did not correspond to genuine predictive power. Instead, the models often exhibited reactive behavior, replicating the most recent price movement and effectively lagging behind by one time step. This limitation highlights the difficulty of achieving forward-looking prediction in cryptocurrency markets. The study concludes that while machine learning can provide systematic benchmarks and descriptive insights into short-term price dynamics, its practical utility for live trading remains limited without additional predictive signals. Future work should expand datasets, incorporate alternative data sources such as sentiment or blockchain network activity, explore deep learning architectures designed for sequential data, and evaluate models in simulated environments to better assess their real-world applicability.