Intraday trading strategy selection under transaction costs with machine learning

Abstract

Machine learning has shown promising results for financial forecasting; however, many studies focus primarily on predictive accuracy while overlooking the economic implications of trading decisions under realistic market conditions. This study systematically evaluates the predictive and economic performance of Decision Tree, Random Forest, and XGBoost models for intraday price direction forecasting across 40 financial assets, including stocks, foreign exchange pairs, and cryptocurrencies. Forecasting horizons of 1, 3, 5, 10, 15, and 30 min were analyzed using a walk-forward validation framework with Bayesian hyperparameter optimization. Unlike many previous studies, the evaluation explicitly incorporates transaction costs, benchmark comparisons, risk-adjusted performance metrics, and execution robustness. The results show that the forecasting horizon is one of the most influential factors affecting both predictive and economic performance. Across most assets, the 30-min horizon emerged as the best-performing option within the evaluated candidate set, indicating a tendency toward improved performance at longer forecasting horizons. After incorporating realistic transaction costs, only a limited subset of assets remained profitable, highlighting the importance of evaluating trading strategies beyond predictive metrics alone. Benchmark comparisons and risk-adjusted performance measures further illustrated the economic performance of the profitable machine learning strategies relative to passive, persistence-based, and random trading approaches. Overall, the proposed framework provides a comprehensive approach for jointly assessing the predictive and economic performance of machine learning-based intraday trading strategies under the transaction-cost and execution assumptions considered in this study.