Algorithm vs. Architecture: Testing Machine Learning Methods Against Traditional Econometric Models in Forecasting the Price of Crypto Assets

Cryptocurrency prices are highly volatile and difficult to forecast in the current economy, and the capital now exposed to them makes reliable forecasting an important question.Existing studies that compare machine learning with econometric models usually rely on a single benchmark and a single asset, or compare machine learning models only with one another, so claims that flexible models forecast better rest on thin evidence.This paper tests whether a machine learning model can outperform traditional econometric models in forecasting Bitcoin's daily closing price, and what the results reveal about the market.It compares ARIMA, linear regression and random forest on daily Bitcoin data from 2015 to 2026, using a 80/20 split into training and testing sets.The forecasts of the models were evaluated one day ahead and over the full test period, using RMSE and MAPE and a naive benchmark that predicts tomorrow's price as today's.Linear regression performed best, with ARIMA a close contender, while the random forest was slightly worse and did not benefit from its added flexibility.