Surpassing stock market's noise-nonstationarity tradeoff by causal-based domain discovery
提出一个模型无关的领域发现框架，利用因果关系从历史市场环境中筛选与目标预测期相似的训练样本，结合分位数正则化技术，提升神经网络在非平稳股票数据上的预测性能。
While neural networks have been extensively applied in stock prediction, forecasting stock returns remains a significant challenge due to the nonstationary nature of financial markets. Existing models struggle to balance two competing demands: learning from large historical datasets to reduce noise, and maintaining temporal relevance in ever-changing market regimes. To address these challenges, we propose a novel Domain Discovery framework that leverages causal relationships to identify training samples from historical market environments that closely resemble the target prediction period. This selection mechanism enables models to learn more effectively from nonstationary data by avoiding outdated or mismatched information, while significantly increasing the number of useful labeled samples. Further, to avoid non-implementable alphas (i.e. profits unachievable under real-world market constraints), embedded within the Domain Discovery framework is a quantile-focused regularization technique that prioritizes the top-performing predictions, aligning training with practical investment goals. The entire approach is model-agnostic and integrates with a wide range of neural network architectures. Extensive experiments across six state-of-the-art architectures demonstrate consistent improvements in out-of-sample performance, underscoring the robustness and versatility of the proposed framework.