The evolution of financial data availability and computational power has fundamentally transformed asset management practices, shifting the industry toward systematic, model-driven portfolio construction. In China, the rapid maturation of the A-share market—reflected in the growing depth and representativeness of benchmarks such as the CSI 300 Index—presents both opportunities and challenges for modern portfolio managers. While factor-based strategies offer a disciplined framework for capturing systematic risk premia, their widespread implementation often remains static, applying fixed weightings to pre-selected factors. This rigidity poses a significant limitation in a market characterized by inherent volatility, regulatory sensitivity, and pronounced sector rotations, necessitating a more adaptive portfolio management approach capable of real-time factor exposure adjustment.

Quantitative asset management in China has matured alongside core benchmarks like the CSI 300 Index. As illustrated in Fig. 1.1, the index's market capitalization and trading volume have expanded substantially since its inception, providing the necessary liquidity and breadth for institutional-scale portfolio strategies.

While foundational models such as the Capital Asset Pricing Model (Sharpe, 1964) and the Fama-French multi-factor (Fama & French, 1993) frameworks are widely employed, their static formulations frequently struggle to adapt to the unique dynamics of China's market. This study addresses this gap by developing a reinforcement learning–graph neural network framework designed to empower portfolio managers with dynamic factor integration capabilities, enabling real-time optimization of both security allocation and underlying factor sensitivities.

Among AI techniques, Reinforcement Learning (Sutton & Barto, 2018) is particularly suited for portfolio management, framing it as a sequential decision-making problem where an agent learns to maximize a risk-adjusted reward (see conceptual diagram, Fig. 1.3). Concurrently, Graph Neural Networks (Kipf & Welling, 2017) excel at modeling the intricate relational structures within markets—such as sectoral dependencies and supply-chain links—offering a richer context for prediction. By integrating these advanced models with traditional factors and real-time event analysis, we aim to create a strategy that is not only predictive but also adaptive and resilient to the unique challenges of the A-share market.

This paper makes three primary contributions to asset management practice:

Methodological: We propose a novel, three-component AI architecture (GBDT + RL + GNN) for dynamic, multi-factor portfolio optimization, explicitly incorporating event-study signals.

Empirical: We provide extensive backtesting and event-study evidence from the Chinese A-share market (2021–2025), demonstrating significant outperformance and effective risk management.

Practical: We detail a fully implementable pipeline using publicly available data and models, offering a reproducible blueprint for asset managers and researchers.

Existing literature has extensively documented factor efficacy and explored AI's predictive power. However, a salient gap exists in the deployment of AI for integrated portfolio governance—the continuous, dynamic management of factor exposures within a live portfolio. This study positions itself at this intersection, proposing an RL-GNN framework not merely as a superior alpha model, but as the core of an adaptive portfolio management system that makes holistic allocation decisions.

Machine learning models have become increasingly adopted in professional asset management for analyzing non-linear patterns in financial data. Ensemble methods like Gradient Boosting (Chen & Guestrin, 2016) and deep learning architectures like LSTMs (Hochreiter & Schmidhuber, 1997) are widely used for return prediction and market state classification within quantitative investment teams. In the context of China's A-share market, these models must capture the high-momentum behavior of growth stocks, as exemplified in Fig. 2.1, which shows the strong, trend-driven price action of a stock like Cambricon (688256.SH)—a characteristic highly relevant for active managers seeking momentum alpha.

Predictive Accuracy: The integrated RL-GNN strategy established a clear "Alpha Frontier," outperforming baseline models like standalone GBDT and linear regression (Fig. 4.4). This superior predictive capability translates directly to improved portfolio construction for quantitative asset managers.

Dynamic Allocation: The RL agent successfully performed "intra-tech" rotation, shifting capital between technology sub-sectors (e.g., from AI Hardware to FinTech) based on real-time momentum and sentiment signals, as visualized in the sector allocation chart (Fig. 3.2). Such dynamic sector rotation represents a systematic enhancement over static allocation approaches commonly used in traditional asset management.

Factor Importance: Feature importance analysis (Fig. 4.5) revealed that Momentum (38%) and Sentiment/Liquidity (30%) were the primary alpha drivers (Chen & Guestrin, 2016; Hochreiter & Schmidhuber, 1997), followed by GNN-captured sectoral co-movement (20%). Traditional valuation factors played a smaller, stabilizing role (12%). These findings provide asset managers with clear guidance on factor prioritization when constructing AI-enhanced portfolios.

The results validate the efficacy of the proposed integrated AI framework for asset management applications. The outperformance stems from several interconnected strengths relevant to portfolio managers: Adaptive Factor Timing: The RL agent moves beyond static factor weighting, dynamically tilting exposures based on the prevailing market regime signaled by GBDT forecasts and alternative data—a capability particularly valuable for active asset managers seeking to navigate regime changes. Structural Alpha from GNNs: By modeling sectoral relationships, the GNN enables the strategy to understand and exploit momentum clusters and systemic movements that are invisible to single-asset models, offering institutional investors insights into cross-asset dependencies. Event-Driven Edge: The systematic incorporation of event-study findings allows the strategy to proactively position for and profit from predictable market reactions to corporate news, providing a systematic approach to event-driven investing within a disciplined framework. The "black-box" nature of the ensemble and its dependence on high-quality, timely data are acknowledged limitations for practical deployment in asset management firms. However, the rigorous backtesting and explicit risk controls mitigate concerns about overfitting and support implementation in professional investment processes.

This study demonstrates that a carefully architected integration of AI techniques—specifically GBDT (Chen & Guestrin, 2016), RL (Sutton & Barto, 2018), and GNN (Kipf & Welling, 2017)—coupled with empirical event-study validation, provides a powerful and robust framework for quantitative asset management in China's A-share market. The framework successfully converts complex market data, inter-asset relationships, and temporary informational advantages into sustained, risk-adjusted alpha, offering portfolio managers a systematic approach to enhance returns while managing downside risk. Future work could focus on enhancing model interpretability for investment committees, incorporating a broader universe of assets relevant to multi-asset portfolios, and testing the framework's resilience across different emerging markets to support global asset allocation decisions.