كنفرانس بين المللی مهندسی صنايع و سيستم ها

صفحه اصلی / دوازدهمین كنفرانس بين المللی مهندسی صنايع و سيستم ها

Adaptive Barrier Labeling and Deep Sequential Learning for Regime-Aware Forex Trading Signal Prediction

نویسندگان :

Zahra Karim¹ Erfan Hassannayebi² Majid Khedmati³

1- دانشگاه صنعتی شریف 2- دانشگاه صنعتی شریف 3- دانشگاه صنعتی شریف

کلمات کلیدی :

Algorithmic Trading،Deep Learning،Forex Market،Target Labeling،ATR_Based Labeling،Regime_Aware Labeling،Average Directional Index

چکیده :

This paper investigates the impact of barrier-based labeling on the predictive and financial performance of deep learning-driven intraday Forex trading systems. We propose a hybrid CNN-GRU-Attention framework designed to classify trading signals by capturing both sequential dependencies and non-stationary market dynamics. We compare three distinct labeling methodologies-static Fixed-pip, volatility-adaptive ATR-based, and regime-aware ADX-ATR-to determine their effectiveness in enhancing risk-adjusted returns. The experimental analysis utilizes 15-minute EUR/USD data from January 2026 to July 2026. Strategy outcomes are evaluated across various holding horizons using a comprehensive suite of metrics, including cumulative returns, annual return, Sharpe ratio, maximum drawdown, and value at risk. Empirical results confirm that label design critically drives trading performance, with the proposed ADX-ATR scheme at an extended horizon achieving superior cumulative and annual returns along with optimal risk-adjusted metrics compared to ATR and Fixed-pip baselines. The underperformance of the Fixed-pip approach highlights the inherent limitations of static thresholds in the heteroskedastic nature of currency markets. Consequently, adaptive label engineering should be treated as a core component of financial model design rather than a simple preprocessing step.