A Generalized Portfolio Network Architecture to Improve Financial Prediction Model Leveraging Differentiable Layer

In quantitative finance, the standard approach involves predicting stock returns to optimize asset allocation, aiming to maximize returns and minimize risks. This predict-then-optimize method traditionally focuses on minimizing prediction errors and optimizing a risk-adjusted objective function. Such a dichotomy often leads to suboptimal outcomes due to the misalignment between prediction and optimization goals [1]. In response to these challenges, while end-to-end deep portfolio architectures [2] – [4] have been proposed, their "black box" nature often fails to adapt to rapidly changing market conditions, resulting in inefficiencies and non-explainable outcomes [5].This paper introduces a novel network architecture that aligns the goals of minimizing prediction error with optimizing allocation by integrating a differentiable optimization layer [6]. We applied this architecture to Qlib, an AI-oriented quantitative investment platform, to simulate market environments and perform rigorous backtesting. Our results show significant enhancements in key financial metrics such as cumulative returns and average risk-adjusted objective values, surpassing traditional two-stage models.Additionally, we conducted portfolio experiments during the COVID-19 period to emphasize the importance of dynamic constraint adjustment in portfolio optimization and demonstrate our model’s effectiveness compared to end-to-end models.