基於 Transformer 模型在預測加密貨幣價格的訓練管道增強：一項實證研究

Abstract

在本研究中，我們使用訓練管道增強 (TPE) 方法以增強基於 Transformer 的加密貨幣價格預測模型。 我們提出了三種訓練管道增強機制：零偏移、時間序列加權和多符號迭代訓練。 並且，我們在加密市場數據集（包括比特幣、以太坊和幣安幣）上對這些機制進行定量評估，以證明它們在提高性能方面的有效性。此外，我們將討論每種機制結果的潛在原因。

In this study, we enhance transformer-based models for crypto price forecasting using Training Pipeline Enhancement (TPE) methods. We propose three Training Pipeline Enhancement mechanisms: Zero Shift, Time Series Weighting, and Multi-symbol Iteration Training. We also quantitatively evaluate these mechanisms on crypto market datasets, including Bitcoin, Ethereum, and Binance Coin, to demonstrate their effectiveness in enhancing performance. Furthermore, we will discuss the potential reasons for the results obtained from each mechanism.