Bitcoin price forecasting has accumulated numerous competing methodologies, from scarcity-based models built around the halving schedule to machine-learning systems processing market and macroeconomic data. Yet research suggests these sophisticated approaches often fail to outperform a much simpler baseline: naive forecasts that rely solely on current market information.
A May 2026 preprint by Carlos Baquero of the University of Porto reviewed Bitcoin prediction research and reached a significant conclusion: across peer-reviewed literature, no model demonstrated durable superiority over the appropriate naive benchmark at horizons of one to six months across several market regimes. Baquero examined 23 papers selected based on their methods, influence, and use of genuine out-of-sample evaluation from hundreds available in the literature.
Why Simple Forecasts Win
Naive forecasting works because financial prices are persistent. A model predicting today's price tomorrow produces tiny percentage errors even with limited information about direction or returns. As forecasting horizons expand beyond a day, however, Bitcoin's volatility should theoretically create room for models to add value. Instead, the relationships learned by complex models tend to decay as markets evolve.
The problem stems from non-stationarity—when relationships between variables do not remain stable enough for past observations to describe the future. Bitcoin's user base, liquidity, regulation, and access methods have all changed substantially. A model calibrated to the 2017 retail-led cycle encountered a different derivatives structure in 2021, and spot ETFs created new capital routes in 2024. Each era supplies historical data from a version of the market that no longer exists.
A separate study comparing statistical, machine-learning, and deep-learning forecasts applied 12 approaches to five major cryptocurrencies at one-day, seven-day, and 30-day horizons. Simple naive models consistently outperformed ARIMA, Prophet, random forests, XGBoost, LSTM networks, and N-BEATS models.
The Backtest Overfitting Problem
Many Bitcoin models appear strongest after researchers have reviewed the entire historical period used to build them. By trying different variables, lookback windows, and architectures before publishing results, researchers can inadvertently win a large lottery conducted on the same price history—a phenomenon known as backtest overfitting.
A single chronological split offers minimal protection against this problem. A researcher can train a model through 2020 and evaluate it in 2021, producing an apparently out-of-sample result that owes much of its performance to a single bull market. Walk-forward evaluation is stronger, as is evaluating the same method across multiple non-overlapping holdout windows covering different regimes.
Baquero's review found that none of the peer-reviewed papers examined evaluated the same approach across several non-overlapping holdout windows covering different regimes. Information leakage—when a feature calculated with future data gives a model information about the answer—can also produce false confidence, particularly when complex architectures place multiple transformations between raw data and reported forecasts.
Popular Valuation Models Under Scrutiny
Bitcoin's best-known valuation frameworks appeal because they translate a complicated asset into intuitive explanations. Stock-to-flow attributes value to scarcity and the halving schedule. Metcalfe-style models tie value to network expansion. Power-law models propose that Bitcoin's history follows a stable mathematical relationship between price and time.
Alexander Shelton's 2024 peer-reviewed examination found that stock-to-flow and Metcalfe variables helped explain returns using historical data but offered limited or zero predictive ability on new data. When time effects entered the stock-to-flow regression, its statistical force disappeared. Bitcoin's supply ratio increases on a predetermined schedule, and its price also climbed for much of its history, making two time-linked series appear economically connected despite no causal relationship.
Metcalfe's Law faces a related identification problem because network activity and price can climb together through multiple mechanisms: adoption raising value, higher prices attracting users, or both variables following a common time path. One study using instrumental variables across six proof-of-work assets found that once researchers addressed autocorrelation and the two-way relationship between activity and price, the positive effects previously attributed to hashrate and transaction count disappeared.
Power-law models remain more complicated because their corridors have captured much of Bitcoin's historical path. However, the academic issue lies in the strength of inference. A high R-squared on a log-log chart establishes that a line fits observed data, but formal support for a power law requires evidence about residual distribution and comparisons with other time functions. Baquero's review found that current Bitcoin power-law literature had not yet completed this work, including sensitivity analysis to starting dates and performance on future observations.
Standards for Better Forecasting
An honest forecasting standard would publish naive benchmarks alongside models, report results separately for each market regime, and include trading costs. Public code and data should allow reproduction, while disclosures on attempted variations would reveal how surprising any winning backtest actually is. Valuation narratives should be separated from point forecasts, and reported ranges should reflect the asset's uncertainty rather than presenting single targets.
