Exploring multi-horizon cryptocurrency forecasting: can machine learning drive profitable trading strategies?

This paper explores the integration of machine learning and automated machine learning (AutoML) in the field of cryptocurrency price forecasting, using Solana as the focal case study before extending the analysis to a balanced panel that adds Bitcoin, Ethereum, XRP and Dogecoin. Support vector machines (SVM), random forest (RF) and multilayer perceptrons (MLP) are compared with their MLJAR AutoML counterparts on identical data and splits, each classifying the direction of the next movement at 1, 7, 14 and 30 days. The SVM emerges as the most effective model, reaching over 64% accuracy and balanced accuracy on Solana at the 14-day horizon, whereas RF and MLP overfit the validation split and fail to generalise, mirroring the limitations observed in the AutoML outputs. Across the panel the best model and the best horizon prove asset-conditional. A conceptual trading strategy built on these predictions, net of transaction costs, outperforms both buy-and-hold and a random benchmark on six of the seven asset–horizon–model combinations that are statistically significant and practically accurate. Shapley additive explanations attribute the forecasts principally to momentum and price-level indicators, which take 66% of the attribution, while volume and volatility contribute little; the balance shifts from momentum towards price level as the horizon lengthens, and the leading indicators vary across the panel.