Complex Bitcoin Price Models Often Fail to Outperform Simple Benchmarks

Researchers have examined dozens of Bitcoin price‑prediction approaches, ranging from basic scarcity models that tie the halving schedule to price, to sophisticated machine‑learning pipelines that ingest market and macroeconomic data. Despite this variety, a systematic review highlights a persistent shortfall: most models do not reliably exceed simple naive forecasts.

Naive benchmarks dominate

Naive forecasts—such as assuming tomorrow’s price will be today’s price, or using a random walk for direction—serve as a low‑bar baseline. Because Bitcoin’s price series is highly persistent, even a trivial forecast can achieve a small percentage error, making it difficult for more complex models to demonstrate added value.

Academic review findings

A May 2026 preprint by Carlos Baquero (University of Porto) evaluated 23 peer‑reviewed Bitcoin forecasting studies that used genuine out‑of‑sample testing. The review concluded that none of the models showed durable superiority over the naive benchmark across one‑ to six‑month horizons and multiple market regimes.

Machine‑learning models underperform

In a separate study, Francesco Puoti, Fabrizio Pittorino, and Manuel Roveri compared 12 statistical, machine‑learning, and deep‑learning methods on five major cryptocurrencies. Simple naive models consistently outperformed ARIMA, Prophet, random forests, XGBoost, LSTM networks, and N‑BEATS at daily, weekly, and monthly horizons.

Challenges of backtesting

Many studies suffer from backtest overfitting: researchers test numerous variations on the same historical data, increasing the chance of finding a spurious “winner.” Robust evaluation requires walk‑forward or multiple non‑overlapping holdout windows to expose models to diverse market conditions (bull runs, crashes, sideways periods).

Valuation formulas vs. forecasting

Well‑known valuation frameworks—stock‑to‑flow, Metcalfe’s law, and power‑law price corridors—provide intuitive narratives but have limited out‑of‑sample predictive power. Recent peer‑reviewed work shows that stock‑to‑flow and Metcalfe variables explain returns in‑sample but add little predictive value when tested on new data.

Recommendations for future research

• Publish naive benchmark results alongside model performance.
• Report results separately for distinct market regimes.
• Include trading‑cost adjustments and disclose the number of model variations tested.
• Make code and data publicly available for reproducibility.