Immaterial Apart, Material Together: paper, data, model scores and code

Abstract

Companion to the paper Immaterial Apart, Material Together: Statement-Level Disclosure Regulation under Cheap Machine Aggregation by Lisha Ai and Xiaoyu Chen. Chinese listed firms answer thousands of small investor questions on the exchanges' interaction platforms; each answer passes review on its own. A language model that reads a firm's answers together, with the firm's name removed, predicts the direction of the earnings preannouncement the firm has yet to publish, and keeps that accuracy on events that follow its documented training cutoff, where supervised text classifiers fall to chance. The archive holds the paper and its online appendix, one row per event with every model score and label for the main sample of 4,053 preannouncements (2020–2022), the FY2023 and FY2024 post-cutoff holdouts, the identity probes and the name-only, entity-masked and same-batch arms; the scripts that turn those scores into every number in the paper, with an offline entry that recomputes and checks them without any model access; the four figures; and the prompt used to read the replies. Firms and events carry pseudonymous six-digit codes that preserve the exchange; announcement and reply dates are real. Firm names, real security codes, the reply text, the daily price files and the model transcripts are withheld; the replies are public on the Shenzhen and Shanghai platforms, and the reply and preannouncement records come from CNRDS and CSMAR under licence. Version 4, September 2026, with the September 2026 text of the paper and its online appendix.