Predicting Internal Control Deficiencies from Investor Online Interaction Texts

Abstract

This study adopts a text-based perspective to transform Q&A on exchange-hosted investor interaction platforms into forward-looking signals and fuses them with financial ratios within a unified deep learning framework to predict internal control deficiencies. Empirical results show that, relative to a no-text baseline, the early-fusion model that incorporates textual information improves Average Precision (AP) and Area Under the ROC Curve (ROC-AUC) by approximately 14.6% and 4.1%, respectively; further adopting hierarchical text fusion yields cumulative improvements of 18.3% and 5.1% over the baseline. The findings document the incremental predictive value of text data for internal control deficiencies and provide effective tools for regulatory early-warning, audit sampling, and corporate self-assessment, with important theoretical and practical implications.