A Machine Learning–Based Long–Short Decision Framework for Hourly EUR/USD Forecasting under Strict Temporal Alignment

Abstract

This study develops and evaluates a machine learning framework for hourly EUR/USD directional forecasting that emphasizes temporal alignment, economic interpretability, and out-of-sample validation. Despite extensive research on algorithmic trading strategies, a critical disconnect persists between reported classification accuracy and actual economic profitability, often arising from methodological issues related to temporal misalignment between predictions, positions, and realized returns. Employing hourly EUR/USD data spanning 2005 to 2020, this research implements a logistic regression classifier with simple price-based features evaluated through walk-forward validation. The model achieves approximately 58.5 percent mean out-of-sample directional accuracy across multiple validation folds, demonstrating statistically stable predictive performance. Translation of probabilistic forecasts into trading positions occurs through a confidence-based long-short strategy that exploits bidirectional price movements while remaining inactive during periods of low prediction certainty. Under strict temporal alignment ensuring causal consistency between information availability and return realization, the machine learning strategy generates positive cumulative returns and superior risk-adjusted performance compared to passive buy-and-hold benchmarks. The buy-and-hold strategy experiences severe drawdowns and terminal cumulative returns of approximately negative 19 percent, while the machine learning approach maintains positive terminal returns of approximately 12 percent with substantially improved downside protection across heterogeneous market regimes including the 2008 financial crisis and European sovereign debt crisis.