Can Generative AI help Investors Avoid the Disposition Effect?

Last registered on September 12, 2025
View Trial History

Pre-Trial

Trial Information

General Information

Title
Can Generative AI help Investors Avoid the Disposition Effect?
RCT ID
AEARCTR-0016753
Initial registration date
September 09, 2025

Initial registration date is when the trial was registered.

It corresponds to when the registration was submitted to the Registry to be reviewed for publication.

First published
September 12, 2025, 10:27 AM EDT

First published corresponds to when the trial was first made public on the Registry after being reviewed.

Locations

Country
Canada
Region
Country
United Kingdom of Great Britain and Northern Ireland
Region
Country
United States of America
Region
national

Primary Investigator

Name
Jeffrey Livingston
Affiliation
Bentley University
Contact Primary Investigator

Other Primary Investigator(s)

PI Name
Kobe Rankich
PI Affiliation
University of Washington
PI Name
Maria Adelaida Martinez
PI Affiliation
Bentley University
Contact Investigator
PI Name
Adrian Aamodt
PI Affiliation
Brown University
PI Name
Ja-Nae Duane
PI Affiliation
Brown University
PI Name
David Lauer
PI Affiliation
Urvin Finance

Additional Trial Information

Status
On going
Start date
2025-09-05
End date
2025-09-30
Keywords
Finance & Microfinance
Additional Keywords
disposition effect, generative AI
JEL code(s)
D9, G4
Secondary IDs
Prior work
This trial does not extend or rely on any prior RCTs.
Abstract
Beginning with Odean (1998), many studies have shown that investors, professional and otherwise, are often vulnerable to the disposition effect: a tendency to hold on to assets that have declined in value for too long, and to sell assets that have risen in value too early. Weber and Camerer (1998) developed an experimental task to measure the disposition effect where subjects buy and sell assets that have constant but unknown probabilities of going up or down in price from round to round. Previous studies have used this task and have shown that this effect can be mitigated (e.g. Fischbacher et al. 2017), and recent work shows that robo-advisors that have been conditioned to give good advice can be particularly effective (Back et al. 2023).

We conduct an artefactual field experiment with investor subjects recruited from the community of Urvin Finance. Urvin Finance has the following mission statement: "Urvin will truly democratize access to financial data and tools, and provide a unique platform for enabling communities to organize, analyze and publish compelling research and analysis. Urvin will, at all times, seek to serve the underserved, focusing on retail investors around the world, and always putting their interests first. Urvin will also provide comprehensive educational resources, with access to experts, unique content, and always in the most open and accessible way possible." it was co-founded by David Lauer, a co-author on this project.

We follow a similar strategy to that of Back et al. (2023). We study whether ChatGPT can help investors avoid the disposition effect in a version of the Weber and Camerer 91998) task. Subjects are randomized into one of three treatment groups: 1) Control, which simply does the task as normal; 2) Treatment No Prompt, in which subjects have access to a ChatGPT window and can ask it anything they want; or 3) Treatment Prompt, which also gives access to ChatGPT and pre-enters prompts that explain the details of the game and update it with the new situation after each round.
External Link(s)

Registration Citation

Citation
Aamodt, Adrian et al. 2025. "Can Generative AI help Investors Avoid the Disposition Effect?." AEA RCT Registry. September 12. https://doi.org/10.1257/rct.16753-1.0
Sponsors & Partners
Experimental Details

Interventions

Intervention(s)
we use a version of the Weber and Camerer (1998) task to measure the disposition effect where subjects buy and sell six assets that have constant but unknown probabilities of going up in price by 6% or down in price by 5% in each round over 14 rounds. The probabilities of increasing in price range from 40% to 60%. Subjects have $2000 in experimental currency to invest in shares of the six assets. The price of a share of each asset starts at $100 in round 1, then evolves according to pre-generated price paths that are simulated according to the assigned probabilities.

We follow a similar strategy to that of Back et al. (2023). We study whether ChatGPT can help investors avoid the disposition effect in a version of the Weber and Camerer (1998) task. Subjects are randomized into one of three treatment groups: 1) Control, which simply does the task as normal; 2) Treatment No Prompt, in which subjects have access to a ChatGPT window and can ask it anything they want; or 3) Treatment Prompt, which also gives access to ChatGPT and pre-enters prompts that explain the details of the game and update it with the new situation after each round.
Intervention (Hidden)
Intervention Start Date
2025-09-05
Intervention End Date
2025-09-30

Primary Outcomes

Primary Outcomes (end points)
The disposition effect measure, which is the proportion of gains realized - the proportion of losses realized. See our pre-analysis plan and/or Back et al. (2023) for more details.
Primary Outcomes (explanation)

Secondary Outcomes

Secondary Outcomes (end points)
o Proportion of Losses Realized
o Proportion of Gains Realized
o Profit/Loss
o Shares invested in risky assets
o Shares invested in the highest priced asset
o Shares invested in the lowest priced asset
Secondary Outcomes (explanation)

Experimental Design

Experimental Design
Subjects are randomized into one of three treatment groups: 1) Control, which simply does the task as normal; 2) Treatment No Prompt, in which subjects have access to a ChatGPT window and can ask it anything they want; or 3) Treatment Prompt, which also gives access to ChatGPT and pre-enters prompts that explain the details of the game and update it with the new situation after each round.
Experimental Design Details
Randomization Method
We use a randomization procedure and design that is similar to that of Back, Morana, and Spann (2023). We generate 98 sets of price paths for the randomization by first randomizing the type of each asset, then randomizing a price path for each asset (according to the asset’s assigned probability of increasing in price) across the 14 rounds.

The subjects are recruited from the community of Urvin Finance, members of which are day traders and/or individuals interested in investing in the stock market. They are recruited by email announcements and social media postings, so they show up to participate in the experiment not all at once but rather at different times of their choosing. We first randomly assign each show up position (the first to begin the experiment is position 1, the next person to start is position 2, etc.) to a group of 3, then within each group, randomize each of the three subjects to one of the three treatment groups. Then, each group is randomly assigned to one of the 98 price paths.
Randomization Unit
individual
Was the treatment clustered?
No

Experiment Characteristics

Sample size: planned number of clusters
n/a, though there are 98 possible price paths that the subjects might face and 3 subjects are randomly assigned to each one, then the three subjects are randomized within each price path group to the treatment conditions.
Sample size: planned number of observations
294
Sample size (or number of clusters) by treatment arms
98
Minimum detectable effect size for main outcomes (accounting for sample design and clustering)
We generated 98 price paths and assign three subjects to each one. Within these price path groups, the subjects are randomly assigned to one of the three treatment groups. This will yield a total of 294 observations. Calculated by simulation and b sampling from the data of Back, Morana and Spann (2023), this sample size allows us to detect a minimum detectable difference in DE size between two treatments of 0.137 with 80.2% power. The mean sample DE is just below 0 with a standard deviation of about 0.3, so this effect size represents about 0.42 sample SDs.
Supporting Documents and Materials
IRB

Institutional Review Boards (IRBs)

IRB Name
Bentley University
IRB Approval Date
2024-05-24
IRB Approval Number
240524095
Analysis Plan

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Post-Trial

Post Trial Information

Study Withdrawal

There is information in this trial unavailable to the public. Use the button below to request access.

Request Information

Intervention

Is the intervention completed?
No
Data Collection Complete
Data Publication

Data Publication

Is public data available?
No

Program Files

Program Files
Reports, Papers & Other Materials

Relevant Paper(s)

Reports & Other Materials