Scholarlink

<p>This paper examines whether Machine Learning models can extract decision relevant signals from traditional and attention based markets to support Bitcoin trading. Using daily data from 2018 to 2024, we consider Bitcoin returns at several horizons and combine information from global equity indices, commodities, exchange rates, volatility measures, and Google Trends indicators. In the first stage, we compare several models to explain Bitcoin returns and use SHAP values to identify nonlinear dependencies and horizon specific drivers. In the second stage, Machine Learning based signals are translated into weekly buy/sell decisions through tree based classifiers. We then evaluate an all in/all out trading strategy against Buy&amp;Hold and a purely random benchmark, accounting for realistic transaction costs. The results show that ensemble methods, particularly XGBoost, deliver superior explanatory power at lower frequencies. Random Forest achieves competitive or superior performance relative to Buy&amp;Hold in two of the three test years and consistently outperforms the random benchmark, especially during episodes of extreme returns. The findings indicate that Machine Learning based signals can complement traditional approaches to timing and risk management in cryptocurrency markets.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Machine Learning perspective to explore cryptocurrencies’ dynamics and derive buy/sell trading signals

  • 期刊论文

  • 发表日期：

    2026年4月20日

查看全文
  • Rosella Castellano,
  • Federico Cini,
  • Annalisa Ferrari

摘要

This paper examines whether Machine Learning models can extract decision relevant signals from traditional and attention based markets to support Bitcoin trading. Using daily data from 2018 to 2024, we consider Bitcoin returns at several horizons and combine information from global equity indices, commodities, exchange rates, volatility measures, and Google Trends indicators. In the first stage, we compare several models to explain Bitcoin returns and use SHAP values to identify nonlinear dependencies and horizon specific drivers. In the second stage, Machine Learning based signals are translated into weekly buy/sell decisions through tree based classifiers. We then evaluate an all in/all out trading strategy against Buy&Hold and a purely random benchmark, accounting for realistic transaction costs. The results show that ensemble methods, particularly XGBoost, deliver superior explanatory power at lower frequencies. Random Forest achieves competitive or superior performance relative to Buy&Hold in two of the three test years and consistently outperforms the random benchmark, especially during episodes of extreme returns. The findings indicate that Machine Learning based signals can complement traditional approaches to timing and risk management in cryptocurrency markets.