Reproduce “Finance-Grounded Optimization For Algorithmic Trading” on OpenResearch

About this paper

Kasymkhan Khubiev, Mikhail Semenov, Irina Podlipnova, Dinara Khubieva

Researchers at Sirius University of Science and Technology and Kazan Federal University developed a framework for deep learning models in algorithmic trading that utilizes finance-grounded loss functions and a turnover regularization mechanism. This approach enables models to directly optimize for risk-adjusted returns and controlled trading activity, consistently outperforming traditional statistical objectives and achieving superior economic performance in both single-alpha and multi-alpha portfolio construction.

Key insights

• Optimizing deep learning models directly with finance-grounded loss functions, particularly those focused on drawdown control like LogMDDLoss, yields substantially superior economic performance and risk-adjusted returns compared to traditional statistical loss functions (e.g., MSELoss).
• The standard Mean Squared Error (MSELoss) is consistently ineffective for financial decision-making, leading to negative Sharpe ratios and poor economic outcomes.
• Turnover regularization, especially the proposed band-based approach, is critical for enhancing strategy stability, controlling transaction costs, and ensuring practical deployability by guiding trading activity within economically meaningful bounds.

Summary by alphaXiv.