This paper addresses the challenge of forecasting cryptocurrency returns, complicated by the market’s speculative nature. Previous research has focused on historical data, macroeconomic indicators, and public interest, often overlooking inter-cryptocurrency relationships. We leverage temporal data from core cryptocurrencies like Bitcoin and Ethereum to automatically select features for predicting the future returns of other cryptocurrencies. We also developed a fundamental network representing cryptocurrencies across multiple dimensions–industry, technology, and investor interactions. Using a transformer-based architecture with dynamic gating, we demonstrate that our framework outperforms previous methods, with long-short portfolios yielding higher returns than simple buy-and-hold strategies.