Sentiment Trumps Fundamentals: Machine-Learning Evidence on the Drivers of Cryptocurrency Returns
Main article
Abstract
Whether the price of a cryptocurrency reflects its economic fundamentals or merely the mood of the crowd remains a central and unresolved question in digital-asset research. This study confronts the two explanations directly by treating the drivers of daily cryptocurrency returns as a supervised learning problem and letting the data adjudicate. Using a panel of five major coins—Bitcoin, Ethereum, Binance Coin, XRP, and Solana—over the period 2019–2024, we assemble two competing feature blocks: a sentiment block built from social-media tone, search attention, a fear-and-greed barometer, and news polarity, and a fundamentals block built from on-chain network activity and conventional macro-financial variables. Four learners (elastic net, random forest, gradient boosting, and a long short-term memory network) are trained under an expanding-window, strictly out-of-sample protocol, and a horse-race design isolates the marginal predictive contribution of each block. Across every model and coin, the sentiment-only specification delivers out-of-sample explanatory power and directional accuracy that are several times larger than the fundamentals-only specification, while adding the fundamentals block to a sentiment model produces negligible improvement. A model-agnostic Shapley attribution confirms that the ten most influential predictors are overwhelmingly sentiment-based. A simple sentiment-informed trading rule generates a markedly superior risk-adjusted profile relative to a fundamentals-informed rule and to passive holding. The findings provide systematic, methodologically transparent evidence that cryptocurrency prices are driven more by investor sentiment than by economic fundamentals, with direct implications for risk management, market monitoring, and the design of analytics pipelines for digital-asset portfolios.
