Deep learning for cryptocurrency price forecasting: A taxonomy, benchmark, and gap analysis

Abstract

Relevance. Cryptocurrency returns are volatile, non-stationary, and heavy-tailed. Purpose. This study develops a unified taxonomy and benchmark with reproducible scoring. Tasks. Representative architectures are compared and measured limitations are separated from proposed remedies. Methods. Ten statistical, linear, recurrent, attention-based, selective state-space, and pretrained models are evaluated on daily Bitcoin/Tether and Ethereum/Tether data from two thousand eighteen to two thousand twenty-six. Three expanding-window folds use horizons of five, ten, and twenty days. A target-history-only control shares post-warm-up calendar rows and normalization. Evaluation uses absolute error, symmetric percentage error, normalized-return directional agreement, and paired tests corrected for temporal dependence and multiple comparisons. Scientific novelty. The benchmark combines architectural coverage, aligned information-set control, and a zero forecast reference. Practical significance. It informs accuracy and resource trade-offs under weak predictive signal. Results. Best absolute-error skill over zero is approximately zero. Bitcoin autoregressive and recurrent model differences are not supported after correction. The recurrent model leads learned-model Ethereum mean absolute error, with many, but not all, pairwise advantages supported. Directional agreement remains near zero point five. TimesFM is the strongest native-input foundation reference; larger Chronos scale gives no consistent gain. Lag-Llama produces over-dispersed forecasts, and its low symmetric percentage error does not establish superior accuracy. Conclusions. This low-signal benchmark supports protocol-specific comparisons rather than exploitable return forecasts. Coverage is limited to two assets, three folds, and zero-shot foundation evaluation.