Post-Earnings-Announcement Drift Prediction: Leveraging Postevent Investor Responses with Multitask Learning
提出多任务学习框架，通过将公告后投资者反应（如分析师预测修正、机构交易）作为辅助任务而非直接输入，避免前瞻偏差，提升盈余公告后漂移预测的准确性，为量化金融从业者提供嵌入领域知识的新方法。
Post-earnings-announcement drift (PEAD) remains one of the most persistent market anomalies, yet traditional models struggle to predict it effectively. Prior research has relied on single-task learning (STL), which treats PEAD prediction as an isolated task, overlooking key postevent investor responses—such as analyst forecast revisions and institutional trading—that drive stock price movements. However, incorporating these signals directly as model inputs introduces look-ahead bias, making real-world predictions impractical. Our study proposes a multitask learning (MTL) framework that circumvents this issue by treating postevent investor responses as auxiliary tasks rather than direct inputs. This enables the model to learn from these critical signals without “cheating.” Additionally, we introduce GradPerp, a novel adaptive task weighting method that prioritizes diverse, meaningful training signals, further improving predictive performance. A key insight from our research is that leveraging MTL in real-world contexts requires deep domain knowledge and novel designs. More importantly, our MTL framework opens new opportunities for practitioners to enhance deep learning models by incorporating their financial expertise through carefully chosen auxiliary tasks. Unlike traditional AI models that rely solely on automated feature selection, our approach provides a structured way for investment professionals to embed domain-driven signals into predictive modeling, unlocking new potential in quantitative finance.