همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Technical Analyst Attention Network (TAAN): An Interpretable Deep Learning Model for Algorithmic Trading in Crypto and Forex Markets

نویسندگان :

Ali Tavassolian¹ Fatemeh Yousefloei² Mojtaba Vahidi Asl³ Monireh Abdoos⁴

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی 3- دانشگاه شهید بهشتی 4- دانشگاه شهید بهشتی

کلمات کلیدی :

Algorithmic Trading،Interpretable Deep Learning،Attention Mechanism،Human-Inspired AI،Automated Feature Engineering،Technical Analysis،Forex،Cryptocurrency Trading

چکیده :

Designing profitable algorithmic trading systems for highly volatile and complex markets, such as Forex and cryptocurrency, remains a significant challenge. While powerful predictors, many contemporary deep learning models lack the intrinsic ability to replicate the analytical logic of an expert trader. A human analyst operates hierarchically: they first evaluate signals across different time periods within a single indicator, and then synthesize insights from the most important indicators for a final decision a process for which common deep learning architectures are not designed. To address this, we introduce the Technical Analyst Attention Network (TAAN), a novel architecture specifically engineered to emulate this analytical workflow. At its core, TAAN features a custom-designed hierarchical attention mechanism. The first attention layer identifies the most salient signals by assessing multiple time periods within each indicator. The second layer then finalizes the analysis by weighing the importance of the indicators themselves (e.g., market structure vs. momentum). We evaluated this architecture on high-volume Forex (EUR/USD, GBP/USD) and cryptocurrency (BTC/USD, ETH/USD) data. TAAN significantly outperformed benchmark models, achieving an accuracy of 95.43\% and a Matthews Correlation Coefficient (MCC) of 0.81. Furthermore, a backtest on EUR/USD yielded a 192.16\% cumulative return and a 3.6 Sharpe ratio. This superior performance, achieved with a remarkably smaller parameter count, validates our design approach.