Beyond the Hype: A Multi-Layer Machine Learning Framework for Cryptocurrency Return Forecasting

Abstract

This study develops and tests a theoretical framework linking liquidity, market integration, and return predictability in cryptocurrency markets. Analyzing high-frequency daily data for five major cryptocurrencies (Bitcoin, Ethereum, Binance Coin, Solana, and Ripple), we employ Random Forest models augmented with rigorous time-series diagnostics and economic significance tests. Our findings establish a fundamental dichotomy: Bitcoin exhibits predictability driven by macroeconomic fundamentals (global risk-free rates, economic policy uncertainty), consistent with its emergence as a macro-asset; altcoins, in contrast, are dominated by internal microstructure (realized volatility, illiquidity) and speculative sentiment. Formal hypothesis tests confirm that (i) lower liquidity predicts higher future returns, (ii) machine learning models systematically underpredict during positively skewed regimes, and (iii) macro integration attenuates microstructure-driven predictability. Out-of-sample R 2 values reach 6.8% for Bitcoin and 9.6% for Ethereum, with Diebold-Mariano statistics rejecting equal predictive accuracy against a random walk at the 1% level. A mean-variance investor would earn a certainty equivalent gain of 2.4% annually by exploiting these forecasts. The results challenge the efficient market hypothesis for 1 digital assets, establish a new taxonomy of cryptocurrency predictability, and provide critical implications for asset pricing, portfolio allocation, and risk management in decentralized finance.