Predicting Bitcoin Price Using Random Forest Regression: A Non-Linear Machine Learning Approach to Cryptocurrency Forecasting

Abstract

This paper examines the predictability of Bitcoin price changes through a Random Forest Regression model that is used to predict the future price using past daily OHLCV (Open, High, Low, Close, Volume) data between 2011 and 2025. Since the cryptocurrency markets are non-linear and heteroskedastic, as well-documented, the customary linear models are not sufficient to understand the intricate dynamics of Bitcoin prices. The approach of the present paper consists of an ensemble machine learning strategy, with engineered features, including lagged prices, moving averages, log returns, and volatility measures, to increase the predictive accuracy. The non-parametric model is justified as linearity diagnostics such as QQ plots, OLS residual analysis, rolling variance, and Pearson-Spearman correlation comparisons demonstrated that data have a non-linear form. On the test set, the Random Forest model has reached an R 2 value of 0.9878 and a mean absolute percentage error (MAPE) of 3.29% which is an outstanding predictive performance. The results are relevant in the emerging body of research on machine learning applications in financial forecasting and emphasize the need to use non-linear modelling in volatile asset markets. Feature importance analysis further identified lagged prices and moving averages as the dominant predictors, confirming that short-term price momentum is the primary driver of Bitcoin price behaviour.