Mining Intraday Risk Factor Collections via Hierarchical Reinforcement Learning based on Transferred Options

Traditional manually designed risk factors, such as beta, size/value, and momentum, often lag behind market dynamics when measuring and predicting volatility in stock returns. Furthermore, statistical models, such as principal component analysis (PCA) and factor analysis frequently fail to capture hidden nonlinear relationships. While genetic programming (GP) has advanced in identifying nonlinear factors automatically, it often lacks an internal mechanism for evaluating factor quality, and the resulting formulas are typically too complex. To address these challenges, we propose a Hierarchical Proximal Policy Optimization (HPPO) framework for automated factor generation and evaluation. The framework leverages two PPO models: a high-level policy and a low-level policy. The high-level policy learns and assigns weights to stock features, while the low-level policy identifies latent nonlinear relationships by combining operators such as $\mathit{sin}()$, $\mathit{+}$, $\mathit{**}$, and $\mathit{/}$. The Pearson correlation coefficient between the generated risk factors and realized return volatility serves as the reward signal, quantifying factor efficacy. Additionally, we incorporate transfer learning into HPPO by pre-training the high-level policy on large-scale historical data from the same High-Frequency Trading (HFT) market. The policy is then fine-tuned with the latest data to account for newly emerging features and distribution shifts. This Transferred Option (TO) enables the high-level policy to leverage previously learned feature correlations across different market environments, resulting in faster convergence and higher-quality factor generation. Experimental results demonstrate that, compared to baselines, the HPPO-TO algorithm achieves a 25\% excess return in HFT markets across China (CSI 300 Index/CSI 800 Index), India (Nifty 100 Index), and the United States (S\&P 500 Index).