What Drives Multi-Chain Crypto Forecasting: Model Choice, Feature Selection, and Transferability

Increasingly shaped by heterogeneous on-chain activity rather than a single shared market process, this study investigates 7-day-ahead forecasting using 147 market and on-chain indicators across eight major blockchain ecosystems from October 2023 to April 2025. We benchmark statistical, deep-learning, and foundation-model baselines under multiple feature-selection pipelines using both error metrics and Diebold–Mariano tests. TiRex achieves the best average MAPE ($0.0428$) in a univariate setting without additional optimized covariates, while TFT remains slightly weaker even under its best feature-input configuration (MAPE: $0.0435$; $p=0.9359$ versus TiRex), suggesting a persistent practical advantage for TiRex. Importantly, TiRex’s zero-shot nature confers a substantial efficiency edge: by bypassing feature selection, it delivers comparable accuracy at a fraction of the computational cost. At the same time, feature selection materially affects many model families, with Boruta chosen in roughly 71.7% of best configurations. Taken together, the evidence supports a selective-feature principle: robust forecasting depends on validated, chain-specific features rather than larger feature sets. Feature-importance and overlap analyses further indicate a mixed structure of transferability, where broad market proxies provide baseline context while chain-specific variables drive marginal gains. Overall, this study highlights that effective multi-chain forecasting is primarily a feature selection problem under statistical uncertainty, while also showing that zero-shot designs like TiRex can achieve state-of-the-art accuracy with unmatched efficiency, offering practical implications for building leaner, more robust trading systems.