Machine Learning techniques for predictive modeling of Bitcoin price dynamics

This thesis compares econometric and machine-learning approaches to forecasting the next-day log return of Bitcoin, and asks whether conditioning on the volatility regime adds economic value once a volatility-targeting sizer is already in place. The data are eleven years of daily observations, from 1 January 2014 to 31 December 2024, a total of 4,018 days, transformed into a 24-feature panel that spans technical, onchain, macroeconomic, sentiment, volume, and lagged-return signals; the two machine-learning models also receive a one-step GARCH ? forecast as a further input. Six strategies are evaluated under 5-fold anchored walk-forward cross-validation, with the test years 2020–2024 pooled into one out-of-sample record: a Naive benchmark, ARIMA, and ARMA-GARCH as univariate baselines, a Random Forest and a BiLSTM on the full panel, and a Regime-Reactive Ensemble that selects among them by volatility regime. Forecast accuracy is measured by the root mean squared error, the mean absolute error, and the coefficient of determination; trading performance is measured by the annualised Sharpe ratio net of a 10 bps round-trip cost, the directional hit rate, and the maximum drawdown. The Diebold-Mariano test, the Hansen Superior Predictive Ability test, the Pesaran-Timmermann test, and a stationary block bootstrap supply the statistical backing, and the whole pipeline is rerun across six random seeds. Forecast accuracy is low and very similar across all six strategies, while their trading results differ widely, and that difference comes almost entirely from the volatility-targeting sizer; the forecasts themselves add little. The Regime-Reactive Ensemble records the highest net Sharpe at the canonical seed, 1.254, but no active strategy significantly improves on a volatility-targeted Buy & Hold benchmark, and the strongest active model changes with the seed. The central finding is that, on daily Bitcoin over this period, economic performance came from managing volatility, and the return forecasts contributed little of their own.