自注意力模型於比特幣短期價格走勢預測之應用

Abstract

隨著區塊鏈技術的發展，加密貨幣市場在近年來迅速普及與成長。比特幣作為這個領域最具代表性的加密貨幣，其價格的高度波動性不僅吸引了大量投資者和學術研究者的關注，也使得即時掌握短期價格走勢更為重要。然而，大多數現有的研究往往聚焦於使用日交易數據進行預測，而這對於快速波動的市場可能無法及時提供有價值的信息。 本論文著重於使用自注意力機制預測比特幣的短期價格走勢。通過自注意力機制，模型能夠捕捉時間序列中的長期依賴關係，結合更短的時間週期，使用更精細的K線和交易行為數據，使我們能夠捕捉到比特幣價格的細微變化。此外，通過位置編碼、滾動窗口驗證和早停法等機器學習技巧來強化模型的學習能力。結果表明，本模型在預測比特幣短期價格走勢上表現出色，達到了55.29%的準確率和0.561的F1分數，表現出比迴歸模型更優異的性能，也優於基於訓練資料漲跌比例作為預測機率的隨機選擇模型。兩者都證明了模型架構的有效性。

With the advancement of blockchain technology, the cryptocurrency market has experienced rapid proliferation and growth in recent years. Bitcoin, as the most emblematic cryptocurrency in this domain, possesses highly volatile prices, not only captivates a large number of investors and academic scholars but also accentuates the necessity for real-time monitoring of short-term price trends. Nevertheless, the majority of the existing research primarily concentrates on utilizing daily trading data for predictions, which may not adequately furnish timely insights in such a fluctuating market. This study employs the self-attention mechanism to forecast the short-term price trend of Bitcoin. Through the self-attention mechanism, the model is adept at capturing long-term dependencies in time series data. By integrating this with shorter time frames and employing more granular candlestick and trading behavior data, we are able to discern the slight fluctuations in Bitcoin prices. Additionally, the study incorporates machine learning techniques such as positional encoding, rolling window validation, and early stopping to augment the model’s learning capabilities. The results illustrate that this model excels in predicting short-term price trends of Bitcoin, achieving an accuracy of 55.29% and an F1 score of 0.561. This exhibits superior performance in comparison with regression models. It also improves upon random selection models, which make predictions based on the proportions of upward and downward movements in the training data as prediction probabilities. Both corroborate the efficacy of the model architecture.