Forecasting Cryptocurrency Market Trends Using Machine Learning on Multidimensional Time-Series Data

Abstract

The cryptocurrency market is characterized by extreme volatility and strong behavioral biases, which often lead to irrational decision-making by both retail and institutional traders. To address this challenge, we present a systematic, data-driven forecasting framework that combines classical machine learning and deep learning models to predict short-term price direction in the Bitcoin market. Our approach leverages a decade of hourly Bitcoin price data and constructs a rich set of engineered features across four key domains: momentum (rate of change), candlestick psychology, volatility patterns, and volume dynamics. These features capture both structural and behavioral signals that are critical in high-frequency trading environments.
We evaluate ten supervised learning models including tree-based ensembles, logistic regression, recurrent and convolutional networks, and hybrid architectures—on their ability to forecast 4-hour price movements. To enhance robustness, we implement multiple ensemble techniques: majority voting, weighted soft voting, and rule-based threshold combinations. Experimental results show that ensemble models significantly outperform individual classifiers. Weighted soft voting achieves a precision of 0.6375, while rule-based ensembles reach over 0.85 precision with ultra-low activity, making both methods suitable for high-confidence entry signals. By aligning predictive accuracy with trading practicality, our framework provides a modular and emotion-agnostic decision-support system for cryptocurrency trading. The results demonstrate the potential of combining structured feature engineering with model diversity to navigate the complexities of a highly volatile, behavior-driven market.