本研究旨在應用機器學習技術於加密貨幣短線交易預測，針對市場特性迥異的比特幣（Bitcoin, BTC）與狗狗幣（Dogecoin, DOGE）進行價格變動與交易訊號之分析與預測。研究整合技術指標、期貨市場微結構資料與情緒因子（恐懼與貪婪指數），建構小時級高頻資料集，並採用隨機森林（Random Forest）與極端梯度提升樹（XGBoost）兩種監督式學習模型，搭配滑動視窗與特徵排序策略，強化模型辨識短期市場變動的能力。 實驗結果顯示，XGBoost 在高波動環境下具備較佳穩定性，並且在 DOGE 模型中納入情緒變數後，可顯著提升其預測精度與 Recall 表現；相對而言，BTC 模型則主要依賴技術面與市場結構特徵，顯示其價格行為較不受情緒影響。研究成果驗證了將情緒指標納入模型可有效提升對迷因幣等情緒驅動型資產的掌握力。 本研究不僅強調模型預測準確性之提升，更著重於其實務應用價值，包含即時交易訊號產生、風險控管與量化交易策略建構之可行性。期望本研究能提供投資人與交易系統開發者具體且可行的資料驅動預測工具，促進短線操作決策之科學化與自動化。 ;This study applies machine learning techniques to short-term cryptocur-rency trading prediction, focusing on two assets with distinct market character-istics: Bitcoin (BTC) and Dogecoin (DOGE). By integrating technical indicators, futures market microstructure data, and sentiment metrics (e.g., Fear & Greed Index), a high-frequency hourly dataset was constructed. Two supervised learning models—Random Forest and XGBoost—were implemented with a sliding window mechanism and feature ranking to enhance responsiveness to rapid market shifts. Experimental results demonstrate that XGBoost outperforms in volatile scenarios and achieves notable improvements in precision and recall for DOGE when sentiment indicators are included. Conversely, BTC prediction perfor-mance is primarily driven by technical and structural features, indicating its rela-tive independence from emotional factors. The findings validate the effective-ness of incorporating sentiment variables for assets with strong psychological dynamics. Beyond predictive accuracy, this study highlights the practical applicability of these models in real-world settings—supporting real-time signal generation, short-term trade execution, and risk-aware quantitative strategies. The pro-posed framework offers actionable insights and tools for investors and devel-opers aiming to systematize and automate decision-making in highly dynamic cryptocurrency markets.
