Introduction
In quantitative active investing, the primary objective is to outperform market benchmarks by selecting high-performing stocks. Traditionally, this process is fragmented into discrete, siloed steps: manual factor creation, rule-based factor selection, stock scoring, and finally, portfolio construction. This "piecewise" approach often fails to capture non-linear relationships between factors and often optimizes individual modules in isolation, leading to sub-optimal global performance. Furthermore, while deep learning models offer powerful predictive capabilities, their "black box" nature often hinders adoption by institutional investors who require transparency.
Researchers have introduced E2EAI, the first holistic, end-to-end deep learning framework designed to unify these disparate steps into a single trainable pipeline. By integrating factor selection and portfolio optimization, E2EAI not only outperforms traditional methods but also introduces a "Directional Buffer" to make its complex "deep factors" interpretable to human fund managers.
The E2EAI Framework
The E2EAI framework replaces the manual, multi-step pipeline with a unified neural network architecture that processes raw market and fundamental data into a final portfolio.
Key components of the architecture include:
- Dynamic Factor Selection: Instead of static rules, the model employs a gated attention block to dynamically mask and select the most relevant input factors based on current market conditions.
- Relational Neutralization: Utilizing Graph Attention Networks (GAT), the model accounts for inter-stock influences, such as industry-specific movements. This allows the system to "neutralize" industry noise and isolate the true alpha signal for individual stocks.
- The Directional Buffer: To address the interpretability gap, the authors developed a mechanism that tracks whether an original factor (e.g., Value or Growth) contributes positively or negatively to the final "deep factor." This allows managers to verify the economic logic behind the model's bets.
- Dual Optimization: The pipeline is trained using a global optimizer for the entire system alongside a specialized cross-sectional optimizer to refine factor return estimations.
Experimental Setup and Datasets
The model was rigorously tested on real-world data from the Chinese stock market spanning from 2010 to mid-2022. The evaluation targeted three major indices representing different market segments:
- CSI 300: Large-cap, highly liquid stocks.
- CSI 500: Mid-cap stocks.
- CSI 1000: Small-cap stocks.
Input data comprised both fundamental factors (from financial statements) and technical factors (market trading data), categorized into groups such as Value, Growth, Momentum, Quality, Size, and Liquidity.
Key Insights and Results
E2EAI demonstrated significant performance gains over standard linear models and state-of-the-art deep learning baselines like DMFM.
- Superior Alpha Generation: In the CSI 1000 (small-cap) universe, E2EAI achieved an Excess Return (α) of 20.12%, markedly higher than the 16.36% achieved by the next best baseline.
- Risk-Adjusted Performance: The Information Ratio (IR)—measuring return relative to risk—was consistently higher for E2EAI. In the CSI 1000 index, it reached 4.35, compared to just 1.35 for traditional linear models.
- Dynamic Factor Importance: The model revealed intuitive market shifts, such as "Liquidity" being a negligible factor for large-caps but a dominant driver for small-cap stocks.
- Interpretability through Linear Approximation: The authors demonstrated that their complex deep factors can be linearly approximated using original factors without significant loss in performance, providing a "bridge" for human trust.
Conclusion and Future Work
E2EAI represents a significant step forward in applying deep learning to active management by proving that end-to-end optimization is superior to piecewise construction. While the current study focused on the Chinese market, the underlying principles of relational neutralization and dynamic factor selection are globally applicable. Future research will focus on adapting the framework to more complex institutional constraints, such as risk-parity mandates and diverse sector-exposure requirements, to better serve client-oriented investment objectives.
Summary updated