Transfer‐entropy‐based dynamic feature selection for evaluating Bitcoin price drivers
提出一种三步法，先用转移熵分析比特币收益与金融时间序列的因果关系，再筛选变量进行多步预测，最后用可解释AI揭示各特征的贡献。
Abstract Despite the growing literature in cryptocurrency forecasting and their price drivers, the relationship between their price and other financial time series is an ongoing matter of debate. This study proposes a three‐step methodology to cover these arguments. First, we conduct an ad hoc analysis using transfer entropy (TE) to study the causal relationship between Bitcoin (BTC) returns and a vast array of financial time series. Then, we utilize variables with a significant amount of information flow toward BTC returns to forecast multi‐step‐ahead BTC returns. Finally, we use explainable artificial intelligence post hoc analysis methods to discover the contribution of each input feature to the overall forecasting. The results indicate a significant change in the information flow pattern in the first days of the COVID‐19 pandemic outbreak. Additionally, our proposed TE‐based feature‐selection method outperforms both benchmarks, a nonfeature‐selection model, and backward stepwise regression.