Integrating Statistical, Machine Learning, and Fuzzy Time Series Models for Robust Bitcoin Price Forecasting

Abstract Bitcoin price forecasting is inherently challenging due to extreme volatility, nonlinear dynamics, and frequent regime shifts, which limit the effectiveness of traditional statistical approaches. This study provides a comprehensive, leakage-free comparison of statistical (ARIMA, SARIMA, Exponential Smoothing), machine learning (Random Forest, Gradient Boosting, Support Vector Regression, Multi-Layer Perceptron), and fuzzy time series models (first-order Chen FTS and higher-order Song FTS) for predicting daily Bitcoin prices from January 2021 to September 2025. Model performance was evaluated using MAE, RMSE, MAPE, residual diagnostics, and long horizon scenario analysis, while 95% confidence intervals and a Diebold-Mariano test quantified forecast uncertainty and validated model superiority. Results indicate that the MLP model achieved the lowest short-term forecast errors (MAE = 2,396.18; RMSE = 3,022.10; MAPE = 2.16%), effectively capturing nonlinear price dynamics, whereas ETS provided robust and interpretable forecasts. Higher-order fuzzy models outperformed first-order variants, highlighting the importance of memory effects. Tree-based models were conservative, and SVR underperformed under non-stationarity. These findings demonstrate that multi-model evaluation with uncertainty quantification is essential for accurate and reliable cryptocurrency price forecasting, offering a benchmark framework for risk-sensitive investment and strategic decision-making.