Post-Earnings-Announcement Drift Prediction: Leveraging Postevent Investor Responses with Multitask Learning

Post-earnings-announcement drift (PEAD) refers to the phenomenon in which a company’s stock price tends to drift persistently in response to the information released during the earnings announcement event. Predicting PEAD is of great interest to both investors and researchers because the magnitude of PEAD is economically significant. Whereas decades of studies have explored various approaches to forecasting PEAD, prior research has largely overlooked postevent investor responses—a critical intermediary in the PEAD mechanism—because of the limitations of single-task learning (STL). This study addresses this gap by introducing a multitask learning (MTL) framework that explicitly incorporates investor responses as auxiliary tasks while mitigating look-ahead bias. To further enhance model performance, we propose GradPerp, an adaptive task weighting method that assigns greater weight to auxiliary tasks that contribute diverse and informative training signals. Our model employs a multilevel, multiquery transformer architecture to facilitate cross-task learning, effectively integrating structured financial features with lengthy earnings call transcripts. Evaluation of the model from 2010 to 2022 demonstrates that the proposed design innovations not only outperform benchmark models in terms of prediction accuracy but also generate a daily risk-adjusted return (alpha) two to three times larger than the traditional earnings surprise–based models. This study contributes to the stream of information systems (IS) literature at the intersection of artificial intelligence (AI), finance, and design science research. Our work provides valuable decision support modules and managerial implications for investment and other financial decision makers.