<![CDATA[The rapid growth of cryptocurrency, particularly Bitcoin, has introduced high-return investment opportunities accompanied by extreme price volatility, posing challenges for accurate forecasting. Previous studies have applied various machine learning models for Bitcoin price prediction; however, limited attention has been given to how different training data horizons affect model performance and generalization. This study addresses this gap by comparing three machine learning algorithms: Linear Regression (LR), XGBoost, and Long Short-Term Memory (LSTM). The analysis examines different training periods, with a primary focus on a 3-year training scenario. Historical Bitcoin data (1-minute intervals) from Kaggle was aggregated into daily observations and processed using strict chronological splitting (80:20) without data leakage. Feature engineering was applied using lag-based variables, moving averages, and volatility indicators, while LSTM utilized sequence windowing with 30–60 time steps. Empirical results from the 3-year training scenario show that LR and XGBoost achieve strong predictive performance (R² = 0.9757 and 0.9667), whilst LSTM performs moderately (R² = 0.72) with higher prediction errors. Additional exploratory experiments on shorter training horizons (e.g., 6 months) indicate a decline in performance across models, reflected in unstable generalization and negative R² values on test data, suggesting overfitting. However, directional accuracy remains above 55% in the primary scenario. These findings suggest that model performance is sensitive to the length and stability of historical data. While simpler models such as linear regression and tree-based methods demonstrate consistent performance in the evaluated setting, conclusions regarding model superiority should be interpreted within the scope of the experiment.]]>
