Scholarlink

<p>This paper evaluates machine learning (ML) algorithms for forex trading based on directional forecasting over the 2018–2023 period. We conduct a rigorous comparative analysis of seven ML techniques—Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, AdaBoost, XGBoost, and Neural Networks—across eight currency pairs against the US dollar, including major (EUR, JPY, CHF, AUD), managed float (CNY), and emerging market (MXN, ZAR, TRY) currencies. Our methodology implements comprehensive hyperparameter optimization through time series cross-validation, realistic dynamic transaction costs based on institutional spreads, and extends evaluation to future validation periods. Uniquely, we compare traditional accuracy-based optimization with the profit-aware Mean Absolute Directional Loss (MADL) function for model selection for the Logistic Regression model. Results demonstrate that simpler, interpretable models achieve superior risk-adjusted returns, with Logistic Regression optimized using MADL obtaining the highest Risk-Adjusted Performance Index (RAPI) scores of 1.45–1.58. All ML strategies show statistical significance after multiple testing corrections (Hansen’s SPA test, t-statistics: 2.96–2.89), with medium to large economic effect sizes (Cohen’s d: 0.47–0.60). However, performance varies dramatically with transaction costs—while major pairs remain profitable with round-trip costs below 0.4%, emerging market currencies become unviable at institutional spreads exceeding 1.0%. Value-at-Risk backtesting reveals that ML models calibrated for directional prediction fail catastrophically for risk estimation (29.1% average violations versus 5% expected), mandating separate risk management frameworks. Our findings challenge the complexity bias in financial ML, demonstrating that interpretable models with proper profit-aware optimization outperform black-box approaches in real-world trading conditions. Clinical trial number: not applicable. </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

Directional forecasting for eight forex pairs against the US dollar using machine learning techniques

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月27日

查看全文
  • Francisco López-Herrera,
  • Jaime González Maiz Jiménez,
  • Adán Reyes Santiago

摘要

This paper evaluates machine learning (ML) algorithms for forex trading based on directional forecasting over the 2018–2023 period. We conduct a rigorous comparative analysis of seven ML techniques—Logistic Regression, Decision Trees, Random Forest, Gradient Boosting, AdaBoost, XGBoost, and Neural Networks—across eight currency pairs against the US dollar, including major (EUR, JPY, CHF, AUD), managed float (CNY), and emerging market (MXN, ZAR, TRY) currencies. Our methodology implements comprehensive hyperparameter optimization through time series cross-validation, realistic dynamic transaction costs based on institutional spreads, and extends evaluation to future validation periods. Uniquely, we compare traditional accuracy-based optimization with the profit-aware Mean Absolute Directional Loss (MADL) function for model selection for the Logistic Regression model. Results demonstrate that simpler, interpretable models achieve superior risk-adjusted returns, with Logistic Regression optimized using MADL obtaining the highest Risk-Adjusted Performance Index (RAPI) scores of 1.45–1.58. All ML strategies show statistical significance after multiple testing corrections (Hansen’s SPA test, t-statistics: 2.96–2.89), with medium to large economic effect sizes (Cohen’s d: 0.47–0.60). However, performance varies dramatically with transaction costs—while major pairs remain profitable with round-trip costs below 0.4%, emerging market currencies become unviable at institutional spreads exceeding 1.0%. Value-at-Risk backtesting reveals that ML models calibrated for directional prediction fail catastrophically for risk estimation (29.1% average violations versus 5% expected), mandating separate risk management frameworks. Our findings challenge the complexity bias in financial ML, demonstrating that interpretable models with proper profit-aware optimization outperform black-box approaches in real-world trading conditions. Clinical trial number: not applicable.