Can machine improve the wisdom of crowd from social media? Evidence from a field experiment

How to cite

Liu, Zhiyu; Zhang, Tianyu; Shu, Susan; Cheng, Zihao. 2026. "Can machine improve the wisdom of crowd from social media? Evidence from a field experiment." AEA RCT Registry. AEARCTR-0018180. https://doi.org/10.1257/rct.18180-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Zhiyu Liu* , Tianyu Zhang , Susan Shu , Zihao Cheng

* Primary investigator

Registry ID
AEARCTR-0018180
DOI
https://doi.org/10.1257/rct.18180-1.0
First registered
Last updated
Status
In development
Trial dates
2026-01-01 – 2026-12-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/18180

Abstract

In this study, we examine whether and how introducing greater diversity of viewpoints through machine can improve the wisdom of crowds on an investor-oriented social media. Prior research shows that aggregate opinions extracted from social media posts can predict firm-level earnings surprises and announcement returns. However, user behavior on social media is often shaped by platform-specific biases, such as echo chambers and the spread of misinformation. This project addresses these concerns by using AI-generated, standardized supporting, dissenting, and neutral comments to introduce structured opinion diversity into investor discussions. We expect this intervention to improve how investors process information on social media, strengthen the predictive power of social media signals for earnings outcomes and announcement returns, and shed light on how these effects vary with intervention intensity. This study contributes to the literature by examining how opinion diversity affects the function of the wisdom of crowds on social media and how AI can be used to enhance information quality in capital markets.

Experimental design

Firms are randomly assigned to a control group or to one of three treatment groups: dissenting, supporting, or neutral. For firms in treatment, we post standardized AI-generated replies to qualifying posts shortly after publication. Depending on treatment assignment, replies either challenge, support, or provide non-directional information related to the original post. Control firms receive no intervention. To study treatment intensity, firms within each treatment arm are further randomized so that replies are posted to approximately one-quarter, one-half, three-quarters, or all qualifying posts during the intervention window. The intervention takes place over the nine trading days from day −10 to day −2 relative to the firm’s annual report disclosure date.

Design details

Randomization method
randomization done in office by a computer
Randomization unit
firm
Clusters
719 firms
Observations
The study is expected to include approximately 50,000 to 100,000 posts. The exact number of observations cannot be predetermined.
Arms
Dissenting group 179 firms Supporting group 181 firms Neutral group 181 firms Control group 178 firms
Primary outcomes
(1) how users interact with our comments; (2) whether these comments enhance the predictive power of posts with respect to firm-level earnings and stock returns; and (3) where the effective boundaries of these interventions.

Cited by

No citation data available for this registration yet.

Keywords