A May 2026 preprint examining Bitcoin price forecasting models found none consistently outperformed naive benchmarks across one- to six-month prediction horizons and multiple market regimes, according to research by Carlos Baquero of the University of Porto.
Baquero’s review selected 23 peer-reviewed papers based on methodological rigor, influence, or genuine out-of-sample evaluation. The finding challenges the assumption that sophisticated forecasting approaches add predictive value beyond simple models that rely only on current market information, such as today’s price, zero return, or random walk assumptions.
Complexity Does Not Guarantee Performance
A parallel study by Francesco Puoti, Fabrizio Pittorino, and Manuel Roveri tested 12 forecasting approaches across 5 major cryptocurrencies at 1-day, 7-day, and 30-day horizons. Their results showed simple naive models consistently outperformed complex methods including ARIMA, Prophet, random forests, XGBoost, LSTM, and N-BEATS.
Bitcoin’s market structure has shifted multiple times since 2017. Retail adoption drove a cycle in 2017, followed by a bear market in 2018, a liquidity shock in 2020, a derivatives structure change in 2021, and the creation of spot ETFs in 2024. These regime changes complicate the stability of any forecasting relationship.
Historical Fit Does Not Equal Predictive Power
Several classes of models show strong historical alignment with Bitcoin’s price path but face validation challenges. Power-law models capture much of Bitcoin’s historical trajectory, yet high R-squared values on log-log charts only establish that a line fits the observed sample, not that the model predicts future price movements.
Stock-to-flow models, which base projections on Bitcoin’s halving schedule and scarcity mechanics, and Metcalfe’s Law, which links network activity to valuation, both helped explain returns when tested on historical data. However, Alexander Shelton’s 2024 peer-reviewed examination of Bitcoin return prediction found these variables offered limited or zero predictive ability when tested out of sample.
Backtest Overfitting and Evaluation Standards
A core methodological problem undermines many forecasting studies. Backtest overfitting occurs when researchers test multiple model variations and publish only the best result, inflating apparent performance. Walk-forward evaluation and multiple non-overlapping holdout windows provide stronger evidence than single chronological splits.
Bitcoin forecasting methods span scarcity models based on halving schedules, on-chain models using address or transaction activity, power-law charts, and machine-learning systems incorporating market and macroeconomic data. Naive forecasts use only current market information and serve as a baseline against which all other methods should be measured.
Non-stationarity, the condition where relationships between variables do not remain stable over time, poses a structural challenge. Bitcoin’s evolving market structure means that relationships established in one regime may not persist through the next.
Baquero’s preprint remains under review and has not yet been peer-published.
