A May 2026 preprint review by Carlos Baquero of the University of Porto found that no Bitcoin price forecasting model has demonstrated durable superiority over naive benchmarks at one- to six-month horizons despite hundreds of published papers. Complex models including ARIMA, Prophet, random forests, XGBoost, LSTM networks, and N-BEATS consistently failed to beat naive forecasts across five major cryptocurrencies at one-, seven-, and 30-day horizons in a separate study by Francesco Puoti, Fabrizio Pittorino, and Manuel Roveri. The core problem lies in non-stationarity, as Bitcoin’s market structure evolves across cycles (retail-driven 2017, derivatives market of 2021, spot ETF introduction in 2024), rendering relationships learned from historical data obsolete. Popular valuation models like stock-to-flow, Metcalfe’s Law, and power-law frameworks, though intuitively plausible, showed limited or zero out-of-sample predictive ability according to research by Alexander Shelton and Savva Shanaev. The review calls for stronger evaluation standards including non-overlapping holdout windows covering different market regimes, naive benchmark comparisons, and transparent disclosure of how many model variations were attempted.
Source: Read the original article
