Complex Bitcoin Models Often Fall Victim to Backtest Overfitting

Bitcoin price forecasting has become a complex and multifaceted field, with numerous approaches attempting to outperform simple naive forecasts. These models range from basic scarcity models that convert the halving schedule into a price prediction to highly contested power-law charts and machine-learning systems fed with market and macroeconomic data.

The academic literature on Bitcoin price forecasting has struggled to beat the naive benchmark, which uses only current market information. A recent preprint review by Carlos Baquero found that no model demonstrated durable superiority over the naive benchmark at horizons of one to six months across several market regimes.

However, some studies have shown that short-horizon order flow and daily return forecasts can produce real predictive value. The problem with longer-horizon price forecasts is that they often memorize noise rather than capturing stable patterns in the data.

A key issue in Bitcoin price forecasting is non-stationarity, where relationships between variables change over time due to changes in market conditions, user base, and regulation. This makes it difficult for models to capture durable relationships and leads to backtest overfitting, where a model performs well on historical data but fails to generalize to new situations.